Contents

Conventions xi
Preface xiii
Note to Students xvii
Acknowledgments xxi

Introduction 1

Chapter 1
On the Notion of a Generalized Conversational Implicature 11
1.0 The Argument 11
1.1 Grice's Program 12
1.2 Three Layers versus Two in the Theory of Communication 21
1.3 The Argument from Design: The Maxims as Heuristics 27
1.4 A Typology of GCIs 35
1.4.1 The First (Q) Heuristic 35
1.4.2 The Second (I) Heuristic 37
1.4.3 The Third (M) Heuristic 38
1.4.4 Interactions Between Implicatures 39
1.5 Non-monotonicity and Default Reasoning 42
1.5.1 Typology of Nonmonotonic Reasoning Systems 42
1.5.2 Nonmonotonic Inference and Implicature 45
1.5.3 Investigating the Defeasibility of Scalar Implicatures 49
1.6 Against Reduction of GCIs to Nonce Speaker-Meaning 54
1.6.1 Sperber-Wilson Relevance 55
1.6.2 Implicature as Accommodation 60
1.7 Generalized Implicature and Stable Patterns of Lexicalization 64
1.8 Conclusions 71
viii Contents

Chapter 2
The Phenomena 73
2.1 Introduction 73
2.2 The Q Principle 75
2.2.1 Q Inferences 75
2.2.2 Entailment Scales 79
2.2.3 Q-Contrasts Based on Other Kinds of Lexical Opposition 98
2.2.4 Residual Problems: Scalar Implicature, GCIs, and PCIs 104
2.2.5 Clausal Implicatures 108
2.3 Exploring I-Inferences 112
2.3.1 Formulating the Maxim or Heuristic 112
2.3.2 Some Prominent I-Implicatures 122
2.4 M-Implicatures and Horn's Division of Labor 135
2.4.1 Horn's (1984) Division of Pragmatic Labor 137
2.5 The Joint E¨ect of Q-, I- and M-Implicatures 153
2.5.1 The Projection Problem 155

Chapter 3
Generalized Conversational Implicature and the Semantics/Pragmatics
Interface 165
3.1 Background 165
3.2 The Received View: Semantics as Input to Pragmatics 170
3.2.1 Grice's Circle: Implicatural Contributions to ``What Is Said'' 172
3.2.2 Disambiguation 174
3.2.3 Indexical Resolution 177
3.2.4 Reference Identi®cation 180
3.2.5 Ellipsis Unpacking 183
3.2.6 Generality Narrowing 184
3.2.7 Some Interim Conclusions: Responses to Grice's Circle 186
3.3 Intrusive Constructions 198
3.3.1 Comparatives 199
3.3.2 The Conditional 205
3.3.3 Metalinguistic Negation and Other Negatives 210
3.3.4 Conclusions Regarding Intrusive Constructions 213
3.4 The Argument from Reference 217
3.4.1 How Implicatures Can Determine De®nite Reference 217
3.4.2 Implicaturally Determined Reference and Donnellan's Referential/
Attributive Distinction 225
3.4.3 The Obstinate Theorist's Final Retort on Reference 230
3.4.4 Presemantic Pragmatics versus Postpragmatic Semantics 232
ix Contents

3.5 Some Implications 236

3.5.1 Disposing of the Existing Responses 236
3.5.2 Modularity and Control 243
3.5.3 Sag's Proposal and Possible Ampli®cations 245
3.5.4 Kadmon's DRT Proposal and Possible Extensions 248
3.5.5 Some Future Directions: DRT and Intrusive Constructions 251
3.5.6 A Residual Problem: How to Get from Semantic Representations to
Propositions 256
3.6 Conclusions 259

Chapter 4
Grammar and Implicature: Sentential Anaphora Reexamined 261
4.1 Grammar and Implicature 261
4.2 Implicature and Coreference 267
4.2.1 The Pragmatics of Local Anaphora 267
4.2.2 Inferring Coreference 273
4.2.3 Inferring Disjoint Reference 277
4.3 Binding Theory and Pragmatics 280
4.3.1 Introduction 280
4.3.2 The A-First Account: Pragmatic Reduction to Binding Conditions B
and C 285
4.3.3 The B-First Account, with a Pragmatic Reduction of Binding Conditions A
and C 327
4.4 The B-then-A Account: Synthesis of the A-First and B-First Accounts 345
4.5 Conclusions 359
4.5.1 Summary 359
4.5.2 Pragmatics versus Parameters in Language Learning and Language
Change 361
4.5.3 Pragmatics and the Generative Program 362

Chapter 5
Epilogue 367
5.1 Predictive Power of the Theory of GCIs 368
5.2 Presumptive Inference and General Reasoning 371
5.3 Role of GCIs in Linguistic Theory 374

Notes 379
References 425
Name Index 451
Subject Index 457
Conventions

1. TYPOGRAPHICAL

Italic is used for emphasis and mention of linguistic expressions.

Double quotes indicate utterances (i.e., uses of expressions and sentences).
Single quotes are used to gloss foreign-language expressions and prag-
matic inferences.

2. SYMBOLS AND ABBREVIATIONS

& logical conjunction

4 logical disjunction
! material conditional (logical ``if '')
@; s negation
1 logical equivalence
p, q (occasionally sentential and propositional variables
f, j, P, Q)
F, G predicate variables
K (as in Kp) epistemic modi®er, to be read as `the speaker knows p'
P (as in Pp) epistemic modi®er (the dual of K: Kp 1@P@p), to
be read as `for all the speaker knows, p'
` entails, as in p ` q
‡> implicates, as in ``p'' ‡>`q' (uttering ``p'' impli-
cates `q')
‡‡> communicates (the sum of what is said and what is
implicated)
xii Conventions

Q‡> implicates under the Q-principle

I‡> implicates under the I-principle
M‡> implicates under the M-principle
i¨ if and only if
ha; bi ordered pair
fa; bg unordered set
hS; W i Horn scale, of linguistic expressions, such that
linguistic expression S is an informationally stronger
element than W.
GCI generalized conversational implicature
PCI particularized conversational implicature
* sentence ungrammatical sentence
?? ``sentence'' pragmatically odd or unacceptable utterance
Further symbols put to limited use are introduced in passing.
Preface

In this book I attempt to defend the notion of a generalized conversational

implicature as a species of preferred interpretation. I do not pretend that
this notion is straightforward, and I grant that it is not easy to defend. Yet
I believe that the idea of a preferred (or default) interpretation is much too
important to be ignored just because it is a di½cult concept. Moreover,
the existence of preferred intepretations has (for me at least) something of
that brute self-evidence that Dr. Johnson invoked when he kicked a stone
to dispatch Berkeley's idealism. We grope to grab and harness the beast
without the slightest doubt about its existence.
I attempt in the book to justify the notion of a generalized conversa-
tional implicature, ®rst, by trying to show that some general theoretical
sense can be made of the notion of a preferred interpretation; second, by
defending the notion against its detractors and exposing the de®ciencies of
the pro¨ered alternatives; and third, by showing that a rather simple ®rst
approximation to a theory of such things can give us coverage of a broad
range of important interpretive phenomena. This is the thrust of the ®rst
two chapters.
Then I try to show just how useful and important such a theory of pre-
ferred interpretation is for our understanding of how language works. To
make the point with maximum dramatic e¨ect, I have gone for the lin-
guistic jugular, as it were. On the one hand, I try to show that the phe-
nomena of preferred intepretations force us into a radical reconstruction
of the entire theory of meaning. On the other, I attempt to reduce much of
the alleged grammar of anaphoraÐcentral to many recent developments
in syntaxÐto matters of preferred interpretation. I feel that these attacks
are successful enough to lend powerful credence to the theory I provide.
My ambitions, in this book at least, do not extend beyond this. There
are no general claims being made about the underlying mechanisms
xiv Preface

constituting a fundamental aspect of human cognition or the essence of

human communication. On the contrary, I believe that a theory of pre-
ferred interpretations is just another piece in the jigsaw puzzle of the
theory of meaning, that complex multifaceted landscape which the ex-
plorers of twentieth-century philosophy and linguistics have proved to be
rugged terrain indeed. Thus my modest ambitions contrast strikingly with
another recent attempt to recast Grice's theory of implicature, viz. Sperber
and Wilson's (1986) Relevance theory, which is o¨ered to us as a wide-
ranging cognitive principle from which all pragmatic facts follow just as
do facts about human attention, memory, and so on. Nevertheless, I do
not believe that my narrow, technical slant is devoid of general con-
clusions; on the contrary, I believe that the theory of preferred interpre-
tation that I sketch, although crude, is su½cient to show that central
problems in the theory of meaning and the theory of grammar have been
completely misconstrued. A scalpel cuts deep just because it is thin.
Finally, those who know other strands of my work may be puzzled that
I should have written this book: the rationalist, universalist tinge may
seem inconsistent with my current attempt to divert cognitive science into
a proper consideration of social and cultural factors, or my previous ef-
forts to understand some of the bases of social interaction. But I do not
believe there is any inconsistency. Current perspectives on the relation
between universal human nature and cultural factors often seem to me to
be inverted: for example, language is held to be essentially universal,
whereas language use is thought to be more open to cultural in¯uences.
But the reverse may in fact be far more plausible: there is obvious cultural
codi®cation of many aspects of language from phoneme to syntactic
construction, whereas the uncodi®ed, unnoticed, low-level background of
usage principles or strategies may be fundamentally culture-independent.
This is not to suggest that culture is a mere veneer that sticks just to the
more rule-bound aspects of human behavior. Indeed, I think I have dem-
onstrated that culture-speci®c semantic concepts can run deep in human
thinking (Levinson 1996, 1997). But underlying presumptions, heuristics,
and principles of usage may be more immune to cultural in¯uence simply
because they are the prerequisites for the system to work at all, precon-
ditions even for learning language. Whether the principles explored in this
book really have universal application is, given the primitive state of our
knowledge in pragmatics, merely a working hypothesis, which at least
seems to ®t what we currently know. (Incidentally, the idea that usage
principles may be much more uniform and simpler than the conventions
xv Preface

of language is interestingly at variance with Chomsky's (1975: 25) oft-

repeated pessimism about ``our very limited progress in developing a
scienti®c theory of any depth to account for the normal use of lan-
guage,'' perhaps because ``human science-forming capacities simply do
not extend to this domain.'')
If this picture inverting the normal presumptions is correct, then prog-
ress in this area of pragmatics, the theory of preferred interpretations,
may be a prerequisite to progress in semantics, for we may eventually be
able to use well-established pragmatic principles to help unravel what
content is actually coded in lexemes and constructions in other languages,
where we are far too prone to assume universal patterning. It may also
help in the other direction, so to speak, by connecting to other, more
complex, areas of pragmatics like the study of verbal interaction. Here,
our patterns of preferred interpretation should provide an additional
semiotic out of which the moves in social interaction must willy-nilly
be constructed (see Levinson 1987a). And here, there are perhaps con-
nections to broader strands of thinking in the social sciences, such as
Bourdieu's (1977) notion of habitus, a structure of dispositions that gen-
erates tendencies to act and interpret in certain ways.
Note to Students

The question for students naturally is: ``Do I really have to read it all?''
The answer depends on your interests. Students wishing simply to get a
feel for the phenomena that modern Gricean pragmaticists wrestle with
may ®nd chapter 2 all they really need. Those with theoretical interests in
the theory of meaning will ®nd chapter 3 central, and they may want to
work back to check the reasoning in chapters 1 and 2. Those students
whose primary interests are in the relation between form and meaning will
want to concentrate on chapters 2 and 4. But the reader who attempts
to shortcut the development of the book in this way will need to use the
index to check back on concepts developed earlier.
This book presupposes some familiarity with the standard neo-Gricean
or Radical Pragmatics approach to conversational implicature but treats
only a portion of that subject. Levinson (1983, chap. 3) provides a text-
book treatment that summarizes the Gricean approach and the work of
Horn (1972, 1973), Gazdar (1979), Atlas and Levinson (1981), and others.
Updates have been provided by Horn (1984, 1989: chap. 4 and 5 [which is
also very useful bibliographically], 1992a, 1996b) and Huang (1994). The
collections of Cole (1978, 1981) contain a number of relevant papers, as
does the 1990 16 th Proceedings of the Berkeley Linguistics Society. The
readers edited by Davis (1991) and, more extensively, by Kasher (1998)
reprint in handy form critical papers by Grice and the important early
paper by Harnish (1976), with much other relevant material. Grice's
(1967) William James lectures are (largely) published, together with other
papers, in Grice 1989 with an epilogue, and there is useful discussion of
Grice's general philosophy (together with Grice's reply) in Grandy and
Warner 1986. A word of warning here: reprints of Grice's work (includ-
ing his own compilation of 1989) are rarely completeÐGrice 1981 for
instance has substantial remarks on GCIs that are missing, without
xviii Note to Students

warning, from his 1989 chapter 17 (they were clearly felt to be redundant
with earlier chapters of that book). This explains why I sometimes cite
unpublished or earlier versions of what is apparently the same paper
(quite often, for example, typographical emphasis has been removed in
the reprints).1
Hirschberg (1985) attempts to recast many of the problems in a theory
of particularized implicature but also o¨ers useful discussion of the stan-
dard approach. Atlas (in press) presents discussion of many outstanding
issues. Although Horn 1989 is a di½cult, if not always unhumorous, work
dedicated to negation, it contains so many important observations on
generalized implicature that it is a ``must'' for the dedicated pragmaticist:
it is a goldmine, but be prepared to use your pick and shovel (and the
excellent index).
The approach developed here is sketched in Levinson 1995a, which
may be a useful introduction to this book (I have drawn on parts of that
paper for chap. 1). Further applications of this particular system of inter-
acting principles have been o¨ered by Levinson (1987b, 1991) (which I
have in part drawn on for chap. 4) and especially Hawkins (1991) and
Huang (1991, 1994). I have tried to lay out a wide range of data because I
feel the range of application of pragmatic principles has been under-
appreciatedÐindeed the generality of the neo-Gricean account is pre-
cisely what recommends it. (Much further grist for the pragmatic mill can
be found in work on linguistic typology, especially Haiman's (1985a,
1985b) pioneering studies of iconicity in language.) The danger, on the
other hand, is that many of the examples will be found to be under-
analyzed, but I felt it was better to point (possibly with a wobbly ®nger)
in the direction of many future dissertations than to withhold such direc-
tions in the interest of more scholarly reticence. My inadequacies, dear
student, are your opportunities!
Any intellectual program is often best understood by glancing at the
rival accounts. An alternative approach with much, if not more, currency
is given by Sperber and Wilson (1986), perhaps best approached though
the Behavioral and Brain Sciences article with attendant commentary
(Sperber and Wilson 1987; see also Blakemore 1992). Sperber and Wilson
attempt to capture all implicatural phenomena with a single dualistic
principle but are especially interested in particularized conversational im-
plicatures. The second edition of Relevance (Sperber and Wilson 1995)
contains a very useful bibliographical update keyed to the new ``post-
face.'' On the whole, the two lines of theorizingÐRelevance Theory and
xix Note to Students

the neo-Gricean line exempli®ed by this bookÐare focused on quite dif-

ferent phenomena. However, work by Kempson (1986), Carston (1988),
and others attempts to reduce generalized conversational implicatures
to particularized ones using this principle of Relevance. A direct attack
from this angle on my approach (of an earlier vintage) may be found in
Carston 1995; my own criticisms of Relevance Theory can be found in
Levinson 1989, which I have not repeated here. Another direct attack,
partly from a Relevance Theory perspective, is directed at my treatment
of anaphora (here enlarged in chap. 4) by Ariel (1994), although I believe
she's mistaken the target. Another approach, never fully developed but
still much presupposed in formal semantics, involves Lewis's concept of
accommodation (see Thomason 1990). Situation semantics (Barwise and
Perry 1983) at one time looked set to swallow within semantics many of
the phenomena discussed here, but again, no explicit account has been
published. A more recent kind of reductionist approach, based on ex-
tended versions of discourse semantics, has been proposed by van Kup-
pevelt (1996) and Scharten (1997). There is a fresh interest in implicature
in computational and arti®cial intelligence circles; a good idea of what is
going on can be got from glancing at the proceedings of the 1996 AAAI
symposium on implicature (see also Hobbs, Stickel, Appelt, and Martin
1990).
Finally, there are a number of interesting recent dissertations on im-
plicature to which the interested reader should refer. Those in Sperber and
Wilson's framework are listed in the second edition of Relevance (1985,
p. 296, n. 9), but especially relevant to this book are those by Hirschberg
(1985), McCa¨erty (1987), Wainer (1991), and Welker (1994). Further
bibliographical leads have been given throughout on the various topics
covered.
Acknowledgments

My debts in this book reach far, far back. As an undergraduate I was

trained by an ardent structuralist anthropologist, Edmund Leach, and it
was at Berkeley as a graduate student that, in reaction to structuralism, I
found Paul Grice's ideas about the derivative nature of conventional
meaning quite revolutionary. His central idea, that ``every arti®cial or
non-iconic system is founded upon an antecedent iconic system'' of rep-
resentation and communication (Grice 1989: 358), is still too radical for
most current thinking in linguistics and philosophy. This book explores
one intriguing margin between the noniconic and the iconic systems, and
it tries to establish in detail one way in which the noniconic systems trade
on the iconic background. Those who share Grice's vision think that lan-
guage will never be understood solely from the inside: linguistic com-
munication is not explained by a direct form-meaning mapping but only
by taking into account the intentional and inferential and indeed inter-
actional umbrella (see, e.g., Sperber and Wilson 1986; Clark 1996).
At Berkeley in the early 1970s, Grice's ideas were being explored all
around me: in sociolinguistics by John Gumperz, in grammar and mean-
ing by George and Robin Lako¨, in philosophy by John Searle and
Stephen Schi¨er, and of course, there was Grice himself. Subsequently,
Larry Horn, Jerrold Sadock and others in the States, and Gerald Gazdar,
Ruth Kempson, and Deirdre Wilson in the U.K. rescued the crucial ideas
from the collapse of Generative Semantics. Gerald Gazdar's ideas of that
vintage (later published as Gazdar 1979) play a rather special role in this
work, and I learned a great deal from him, re¯ected in my textbook of
pragmatics (1983). After diversions in anthropological linguistics, a paper
by Larry Horn (1984) brought the whole subject back to life for me and
made me rethink the typology of implicatures, and this led to the paper
(1987a, given at Viareggio, 1985) where I ®rst explored the tripartite
xxii Acknowledgments

scheme employed here, encouraged by John Lyons, Peter Matthews, Ni-

gel Vincent, and Yan Huang, then all at Cambridge. Meanwhile, work by
Dan Sperber, Deirdre Wilson, Ruth Kempson, and associates has spurred
and goaded me into some response; while at about the same time the op-
portunity to teach with Larry Horn at the Linguistics Institute at Stanford
in 1987 greatly increased both my con®dence and competence in this area,
and it was there that I ®rst adumbrated all the themes to be found in this
book. That opportunity was arranged by Ivan Sag, with whom I have
discussed at length on the issues in chapter 3 (we had indeed hoped to
pursue some of these issues together). A year at Stanford followed, where
the Linguistics Department and CSLI provided ideal conditions for dis-
cussion of many pertinent themes, and I taught a graduate course on the
material in this book, where I received much helpful feedback. Herb
Clark, Ray Perrault, Jerry Hobbs, and other participants in the CSLI
implicature group forced me to clarify my ideas by describing their own
alternative ones. Also at CSLI, conversations with Ivan Sag, Stan Peters,
Craige Roberts, Nirit Kadmon, and David Perlmutter were important;
and a seminar on abduction organized by Jerry Hobbs and Doug Ed-
wards at SRI helped me understand work on the theory of inference in
AI. Back in Cambridge, Yan Huang once again made me rethink my
pragmatic reduction of the Binding Conditions (I had already been un-
settled by discussions with Ivan Sag and K. Mohanan at Stanford, and
thus recalled an earlier conversation with Ann Farmer and Robert
Harnsh on the Viareggio beach). An invitation to give the Nijmegen
Lectures in December 1988 gave me the opportunity to condense a me-
andering manuscript and gave me some valuable audience response; a
workshop led by Rob van der Sandt was also most helpful. At that point,
the book still lacked the survey in chapter 2, but I was given the un-
refusable opportunity by the Max Planck Society to pursue a quite dif-
ferent line of work (see, e.g., Levinson 1996), which has proved all-
consuming for many years. In the meantime, I had expected others to ®ll
the gap. This didn't happen. The need for a book of this kind remained.
Encouraged by Max Planck colleagues and fellow pragmaticists, I deter-
mined at last to recast the manuscript in line with current developments. I
hope that I have succeeded in acknowledging most of the recent impor-
tant contributions, but the canvas is large.
Throughout all these wanderings through the thickets of meaning, Jay
Atlas has been a constant guide over 15 years; chapter 3 ®nds me follow-
ing him in a direction I managed to resist for 10 yearsÐsomehow I had to
xxiii Acknowledgments

come up with my own arguments before I could believe his conclusions. I

notice a number of others making the same acknowledgment, and I look
with renewed interest at all the places I still believe him to be wrong.
A number of scholars have commented on the manuscript (some, alas,
in print). Most recently, I have received most helpful comments from
Felix Ameka, Bob Arundale, Penny Brown, Eve Clark, Yan Huang, Eric
Pederson, and David Wilkins on parts of this book, for which I am most
grateful. But special mention must be made of the rich annotations I was
lucky enough to receive on the whole manuscript from Jay Atlas, Kent
Bach, and Larry Horn (the last two in thin disguise as referees for MIT
Press). Such generous help has rescued me from numerous errorsÐfaults
and infelicities that remain are of course my own. I am also most grateful
to Edith Sjoerdsma for many kinds of assistance with the preparation of
the manuscript, to Elizabeth LaurencËot for copyediting and Amy Brand
for seeing the book through the Press.
Finally, I would like to thank the KroÈller-MuÈller Museum, Otterlo,
The Netherlands, for permission to use the reproduction of ``Still life with
oil lamp'' by Juan Gris on the dustcover, Dover publications for use of
the Rembrandt sketch in ®gure 0.1 (from Toney 1963), and Cambridge
University Press for permission to reprint some paragraphs from two
earlier articles of mine that appeared in the Journal of Linguistics.
Introduction

This book is about meaning, but about a rather special subpart of gen-
eral utterance meaning,1 the pragmatic penumbra closely surrounding
sentence-meaning: it is about utterance-type meaning, not the utterance-
token meaning that has been a major focus of recent work in pragmatics.
Utterance-type meanings are matters of preferred interpretationsÐthe
presumptive meanings of the title of this bookÐwhich are carried by the
structure of utterances, given the structure of the language, and not by
virtue of the particular contexts of utterance. In certain ways, the ideas in
this book are conservative, and the majority of them are not new (the debt
to early work by Atlas, Harnish, Horn, and Gazdar will be particularly
evident). But these ideas are in danger of being eclipsed and forgotten
before their promise has been properly appreciated, so in the repetition of
old information I remain unapologetic. Part of the motive of the book is
therefore to collect these previous ideas and data together in one place, so
that the systematicity of the observations can be directly judged and their
implications properly appreciated. In other respects, however, the ideas
presented here extend the old ideas just far enough for us to see that they
are potentially quite explosive: they threaten many of the orthodoxies
about meaning and form in the language sciences.
I have tried to avoid technicalities. In part, this is because half a dozen
years interposed between the ®rst and second drafts of this manuscript,
and this time revealed to me the passing nature of many of the formalisms
and formal theories to which the material in this book could be related.2
The observations in this book seem to have a genuine independence from
such passing clouds, and so it seems worthwhile to detach them from the
technicalities. Nevertheless, I suppose it is inevitable that there will be
passages that are rather hard going. I have therefore thought it worth-
while to sketch here in the introduction in a quite informal way the kind
2 Introduction

of picture of linguistic communication that will emerge from this study. I

will do so partly through analogy.
Consider the Rembrandt sketch in ®gure 0.1 (from Toney 1963, plate
94). We interpret this sketch instantly and e¨ortlessly as a gathering of
people before a structure, probably a gateway; the people are listening to
a single declaiming ®gure in the center. You and I may see slightly dif-
ferent details adumbrated: for me, the declaiming character has a beard,
and the man in the foreground to the left a hat, and the sitting ®gures
seem female, but you may have your own interpretations. Knowing
something about Rembrandt's times and the subjects he was fond of, we
can be fairly sure this is Christ telling parables to a crowd before a city
gate. But all this is a miracle, for there is little detailed information in
the lines or the shading (such as there is). Every line is a mere sugges-
tion; some lines seem to make no sense except as ¯ourishes of the pen,
but yet we struggle to ®nd some sense within them (see for example,
the three little ¯ourishes by Christ's right ``hand,'' which suggest perhaps
the heads of further listeners). So here's the miracle: from a merest, sketch-
iest squiggle of lines, you and I converge to ®nd the adumbration of a
coherent scene (that we may diverge, according to fantasy and our knowl-
edge of art history, is of course less miraculous).
How is it possible? Although the theory of vision is considerably
more advanced than the theory of the language capacity, current knowl-
edge cannot really explain our understanding of visual information as
``degraded'' or physically impoverished as this (but see Changeux 1994).
What we do know is that there is tremendous visual sensitivity to outlines
and shading and that we can use these (although normally reinforced by
multiple further sources of information like parallax, stereopsis, color,
texture, and re¯ection) to extract three-dimensional models out of two-
dimensional retinal arrays. Somehow, then, Rembrandt is exploiting the
fact that we are ``built'' to see three-dimensional scenes in squiggles on a
plane surface. He is also relying of course on cultural conventions about
how to represent scenes in drawings, using the post-Renaissance prefer-
ence for temporal instantaneousness and correct perspective, as opposed
to, for example, the Central Australian cultural conventions in drawing
a bird's-eye view from above, which represents entire episodes over time.
Another important ingredient is that the sketch is presented as a repre-
sentation, not, say, as scratch marks on linoleum, and as a result, there is
some kind of compact between sketcher and viewer that warrants the
3 Introduction

Figure 0.1
Rembrandt sketch
4 Introduction

pursuit of represented detailÐthere is an assumption that with proper

attention we should be able to recover the artist's intentions.
The problem of utterance understanding is not dissimilar to this little
visual miracle. An utterance is not, as it were, a veridical model or
``snapshot'' of the scene it describes, although much talk of truth con-
ditions might lead one to suppose the contrary (the work of the early
Wittgenstein explored this initially attractive idea, which has never been
quite washed out of our thinking since). Rather, an utterance is just as
sketchy as the Rembrandt drawing. Given the way we are built and the
cultural conventions that specify a particular language and its appropriate
deployment, we are inexorably led, at least in most cases, to a common
understanding (and, as in the Rembrandt sketch, we even have a clear
sense of where we may diverge). It is the apparently deterministic nature
of this process that underlines the miracle. A sentence like ``It will be
ready soon'' is indeed sketchy like the Rembrandt drawing: it may refer to
anything under the sun as long as it is not a human adult, will be speci®es
an in®nite series of future temporal spans, and ready might be interpreted
as `cooked' (as when the utterance is a response to the question ``When is
supper?'') or `surfaced' (as in response to ``When will the motorway be
open?'') or `refuelled' (in response to ``When does the ¯ight leave?''). And
clearly, soon can take on quite di¨erent values in the context of serving
a meal or ®nishing a doctoral dissertation. But in each of these di¨er-
ent contexts, ``It will be ready soon'' has its appropriate and apparently
determinate interpretation e¨ortlessly delivered to us.
How? That is the question that a theory of utterance comprehension
has to answer. As with the visual processes behind the interpretation of
the Rembrandt sketch, let us confess that we don't really have the faintest
idea how it works. First principles even seem to suggest that the recovery
of a speaker's intentions on the basis of what he has said is in principle
impossible (Levinson 1995b).3 Books like those by Sperber and Wilson
(1986) or Atlas (1989) or Horn (1989), or the present e¨ort, which attempt
to spell out some of the pragmatic processes involved, are pretty much
stabs in the dark. What I have tried to do in this book is detach a little
part of this problem and argue that there must be powerful heuristics that
give us preferred interpretations without too much calculation of such
matters as speakers' intentions, encyclopedic knowledge of the domain
being talked about, or calculations of others' mental processes. Such
preferred interpretations may be overridden by, and are certainly supple-
mented with, calculations of just this complex kind. But it seems implau-
5 Introduction

sible that the phenomenology of instantaneous, determinate interpretation

could be delivered solely by reasoning about such matters as the poten-
tially in®nite regress of what the speaker is thinking the hearer will think
that the speaker is thinking, and so on, ad in®nitum.4
Issues of cognitive processing have played an explicit role in only one
pragmatic theory so far: in Sperber and Wilson's (1986) Relevance theory,
the addressee infers as much as he or she can for the processing buck,
following a mini-max strategy. Somehow this must have some general
correlation with the speaker's communicative intention. I do not think
that the Sperber and Wilson framework can o¨er an account of the phe-
nomena discussed in this book, precisely because the inferences in ques-
tion are relatively invariant over changes in context and background
assumptions, despite being defeasible (i.e., they do not go through in the
presence of contrary assumptions). It is this relative invariance that gives
these inferences their linguistic importance, as systematic feeders to se-
mantic processes and language change, and I think the rich observations
that have been collected by Horn (1972, 1989) and others need to be pre-
served. Explicit processing considerations do not enter the framework
o¨ered here, but they do form part of the background, for the character
of the inferences in question as default inferences can, I think, be under-
stood best against the background of cognitive processing. The evidence,
so far as it goes, from the psycholinguistic literature is that hypotheses
about meaning are entertained by the hearer incrementallyÐas the words
come in, as it were. It seems unlikely that this could be achieved without
rich heuristics, because there will be early moments in utterance process-
ing where propositional or clausal information for inferencing is simply
not yet available. For this reason, the Sperber and Wilson framework is
not psycholinguistically plausible, or at least it is not plausible for early
moments in utterance processing. In contrast, some of the heuristics to be
o¨ered here can proceed on a word-by-word basisÐfor example, a scalar
quanti®er like some will, as I will show, already invoke default enrich-
ments before the predicate is available.
Close observers of this line of thinking have noticed that it constitutes a
renaissance of information-theoretic ideas:
These notions went out of fashion in theoretical linguisticsÐand I think this is the
right way to put itÐwhen Chomsky (1956, 1959, etc.) criticized (correctly) their
association with radical, tabula rasa, behaviourism and ®nite-state models of lan-
guage acquisition and of grammaticality. Their rehabilitation within the frame-
work of more satisfactory models of the structure and use of language is very
much to be welcomed. (Lyons 1995: 239±240)
6 Introduction

Let me return now to the kernel idea behind the present work, which
is disarmingly simple. The central background fact, an information-
theoretic observation, is that human speech encoding is relatively very
slow: the actual process of phonetic articulation is a bottleneck in a sys-
tem that can otherwise run about four times faster (a point taken up in
chap. 1). The pressures that this exerts on language are easy to seeÐfor
example, the pressure for frequent words to be reduced, as has been long
documented (e.g., by Zipf 1949). Zipf (and, more recently, Horn 1984,
1989) saw the resulting pattern as a balance between two forces: a
speaker's desire for economy and an auditor's need for su½cient infor-
mation. This tension, real enough, is not the focus of the current work.
Instead, I assume a kind of coincidence of interests, treating linguistic
communication as a ``game of pure coordination'' in the game-theoretic
sense of Schelling (1960), a picture I think presupposed by Grice (1957):
the speaker is trying to ®nd an economical means of invoking speci®c
ideas in the hearer, knowing that the hearer has exactly this expectation.
Now, the solution to the encoding bottleneck, I suggest, is just this: let not
only the content but also the metalinguistic properties of the utterance
(e.g., its form) carry the message. Or, ®nd a way to piggyback meaning on
top of the meaning.
How can this be piggybacking be achieved? Only by utilizing the form,
the structure, and the pattern of choices within the utterance to signal the
extra information beyond the meanings of its constituents. So, here I
propose just three simple heuristics, which will serve to amplify utterance
content:
1. If the utterance is constructed using simple, brief, unmarked forms,
this signals business as usual, that the described situation has all the
expected, stereotypical properties;
2. If, in contrast, the utterance is constructed using marked, prolix, or
unusual forms, this signals that the described situation is itself unusual
or unexpected or has special properties;
3. Where an utterance contains an expression drawn from a set of con-
trasting expressions, assume that the chosen expressions describe a
world that itself contrasts with those rival worlds that would have been
described by the contrasting expressions.
This description is imprecise of course. However, one can imagine for-
mal treatments of this meaning ampli®cation along the following lines.
Let the metalinguistic information control the model in which the utter-
7 Introduction

ance is interpreted: let it pick out of the set of possible worlds a proper
subset that has the kinds of properties signaled by the metalinguistic
properties of the utterance; let it determine the domain of discourse; let it
restrict the model to one in which certain extra properties obtain. In the
body of the book, especially in chapter 2, many examples will be given
that indicate the more precise directions in which these constraints on
interpretation must be formulated. The essential point here is that no
special, exotic, hermeneutic principles are being proposed: metalinguistic
scrutiny by the addressee of utterance form, including the consideration of
salient alternates, will be su½cient to amplify the content of utterances
and ease the pressure on the encoding bottleneck; knowing which, the
speaker is bound to conform to the same heuristics.
Approaches to meaning are so diverse and fractionated that it may help
the reader to know where the writer stands on a wider range of issues than
are directly addressed in this book. So, let me declare my cards:
1. The distinction between semantics and pragmatics is one of a
number of essential distinctions in the study of meaning; it may be that in
the long run the distinction will dissolve into a larger set of distinctions,
but nothing is gained by lumping in our attempt to understand human
communication.
2. Semantics is not to be confused with ``conceptual structure'' or the
``language of thought''; the semantics of a language is a language-speci®c
phenomenon, and the mapping into a ``language of thought,'' a nontrivial
relation between nonisomorphic structures (Levinson 1997).
3. Aspects of semantic content (usually only when enriched by prag-
matic speci®cations) can be speci®ed by the apparatus of a recursive truth
de®nition, but this is unlikely to have a direct cognitive counterpart.
Consider the analogy to vision: we can objectively specify the relation
between the world that is viewed and the signals reaching the visual cortex,
and compare the subjective visual experience. But this psychophysical
mapping is not always truth-preserving: there are numerous conditions
under which we see things that do not exist, fail to see things that do, and
so onÐthe stock of illusions that psychophysicists take as a central part
of their job to explain. Similarly, truth-conditional semantics viewed in
the realist wayÐas a direct veridical mapping of semantic structures onto
states of a¨airs (bypassing the head as it were)Ðis useful as a yardstick of
human performance. Something like this is what the cognitive processes
must do, but as in the case of visual illusions, they may fail to do so,
and how they generally do so will be unrelated to the machinery of truth-
8 Introduction

conditional semantics. So we can have our cake and eat it, too: we can use
the insights of truth-conditional semantics without buying Realism, and
without caring that it obviously fails to meet any criteria for adequacy as
a cognitive model. It is a bit like buying a pair of shoes: we are glad that
the salesperson can measure our feet and get the size roughly right, but in
the end, it's a subjective question whether the shoes ®t.
4. There is no algorithm that, given a syntactic string in a language,
cranks out its unique logical form or semantic structure. The view that
there is such an algorithm forms the basis of much linguistic theorizing,
from Montague to Chomsky's most recent views. But it is patently absurd
to hold such a view. First, there is the enormous range of ambiguities in
natural language (requiring at least a one-to-many correspondence). Sec-
ond, syntactic structures may actually be indeterminate in certain respects
(Matthews 1981: 17±21). Third, pragmatic resolution is crucial before
semantic interpretation or the assignment of semantic structure: nobody
disputes the role of deixis here, but there are many other aspects of prag-
matic resolution from the determination of anaphoric reference to the
assignment of scope. The overall picture sketched by Sadock (1991) or
Jackendo¨ (1997) is here at least correct: phonology, syntax, and seman-
tics are areas each with their own generative capacities, and there are
signi®cant mismatches between the structural strings in each representa-
tion that, in the end, come to be associated with one another. What puts
them into association with one another are correspondence rules or pro-
cesses, which are not deterministic in character. Jackendo¨ suggests that
the picture with lexical items is not in principle di¨erent: they are corre-
spondences across phonological, syntactic, and semantic representations,
but the correspondences may be partial and not one to one (compare idi-
oms, or the mismatch between phonological and syntactic words).
5. Insofar as we can get from syntactic structures and lexical material
directly to a semantic representation (which in the most part, I argue, we
cannot), then such semantic representations are only partially speci®ed;
templates of partial information far too unspeci®ed to determine truth
conditions.
6. The overall role that pragmatics plays in such a picture is funda-
mentally di¨erent from that sketched in the textbooks (including mine of
1983). There is no scheme of the kind that syntactic structures are mapped
onto semantic structures which themselves represent full-¯edged proposi-
tions, these semantic structures being the input into pragmatics, which
yields additional inferences or restrictions on meaning. Rather, pragmatic
9 Introduction

processes play a crucial role in the correspondence rules mapping syntac-

tic structures onto semantic representations, and again mapping semantic
representations onto communicated thoughts or utterance meanings.
The picture of the overall theory of meaning that will emerge from
this book is thus radically di¨erent from the starting point from which
Gricean pragmatics began. The distinction between semantics and prag-
matics was construed in the standard theory as the distinction between
sentence-meaning and utterance-meaning, with the output of the seman-
tics being the input to the pragmatics. For reasons detailed at length in
chapter 3, this cannot be right. Instead, I argue, we should stop thinking
of the distinction in terms of levels of representation. Instead, we should
think about both semantics and pragmatics as being component processes
that o¨er their own distinctive contributions to a single level of represen-
tation. The processes remain distinct in kind, and thus the distinction
between semantics and pragmatics must be retained.
This new view of the architecture of a theory of meaning is one of the
explosive consequences, mentioned at the beginning, of careful observa-
tions about the apparently minor phenomena of preferred interpretations.
Another far-reaching set of consequences has to do with constraints on
syntactic form, explored in chapter 4 through consideration of anaphors,
where it is argued that patterns of preferred interpretation explain the
distribution and typology of anaphoric expressions.
As background to all this, chapter 1 lays out the reasons to suppose
that preferred interpretations do indeed exist and have a life of their own.
Chapter 2 o¨ers a catalog of these phenomena in terms of three simple
neo-Gricean principles, and sketches their interaction. The reader will ®nd
many familiar facts here assembled in perhaps unfamiliar ways. A catalog
may make tedious reading, but I think there is no other place where the
student can ®nd all these phenomena laid out, and it is important for the
thesis of this book that the theory of preferred interpretations is seen to
have perfectly general application.
This then is the structure of the book: a chapter that introduces the idea
of presumptive meanings, a chapter that lays out many diverse examples
and a scheme that organizes them, a chapter that explores the conse-
quences for theories of semantics, and a fourth that explores the impli-
cations for syntax.
Chapter 1
On the Notion of a Generalized Conversational Implicature

1.0 THE ARGUMENT

In this chapter I introduce and attempt to defend the notion of a gener-

alized conversational implicature (abbreviated to GCI), where what is at
stake is the generalized nature of such inferences. A generalized implica-
ture is, in e¨ect, a default inference, one that captures our intuitions about
a preferred or normal interpretation. The notion of a preferred interpre-
tation is not one that has any special currency in the theory of language
at the present time. Indeed, it is a notion that is currently either ignored
or subject to virulent attack by some pragmaticists, psycholinguists, and
workers in arti®cial intelligence. In defense of this notion, I develop a
number of arguments both positive and negative.
First, looking back at Grice's program, we see that such a notion
(contrary to some recent commentary) was central to his concerns; and
those concerns, whatever the di¨erences in modern implementation, seem
perfectly current today. Second, I diagnose that one source of the current
theoretical resistance to GCIs is a simplistic overall scheme for a theory of
communication, within which there is no place for a theory of preferred
interpretation. I argue that there is overwhelming evidence for the need
for such a theory, evidence much broader than that for GCIs themselves,
and consequently that our overall scheme must provide a niche of an
appropriate sort in any case. When this is appreciated, much of the resis-
tance to GCIs should wither away.
Next, I produce an argument from design, to the e¨ect that a system of
preferred interpretation would be much too e¨ective a device in commu-
nication for any naturally evolved system of communication to fail to
have developed such a thing. At this point, I digress to introduce the
typology of GCIs that will be central to the book.
12 Chapter 1

I then attempt to deal frontally with another source of resistance to the

concept of a GCIÐnamely, the question of whether any theoretical
sense can be made of the notion ``default interpretation.'' I argue that we
need to address this question in the context of the di¨erent kinds of
nonmonotonic reasoning systems that have recently been developed.
Although signi®cant puzzles remain, these recent developments do o¨er
some reassurance that formal models of default reasoning are not beyond
our reach and thus that some light can be thrown from that direction
on the notion of a preferred or default interpretation. (Whether these
systems are essential to a proper account of default intepretations in lan-
guage use is another matter, because default inferences in language may
arise because of the very nature of communication.)
I turn next to some recent accounts of implicature that leave no room
for a notion of a generalized implicature: Sperber and Wilson's (1986,
1995) Relevance theory, and the theories of implicature based on Lewis's
(1979) notion of ``accommodation'' that have been developed by Hobbs
(1987) and Thomason (1990). According to these theories, all implicatures
are a matter of nonce (i.e., once-o¨ ) inference to the best interpretation
(however that is de®ned), given the full richness of background presump-
tions. These theories, I argue, simply cannot handle the phenomena that
are focal to a theory of GCIs.
As a ®nal argument in favor of the concept of the GCI, I repeat Horn's
observations that certain kinds of GCI systematically block lexicalization
of certain concepts. Because these observations have an immense regu-
larity and crosslinguistic generality, it is really quite inconceivable that
anything less than a generalized, preferred species of interpretation could
be responsible.

1.1 GRICE'S PROGRAM

Much of the argument in this book is developed under the assumption of

what I shall call the Gricean umbrella, a general approach to the study of
meaning and communication. Let us recall the essential features of these
umbrella assumptions (no full explication will be made here; see, e.g.,
Levinson 1983: chap. 3). First, a theory of communication has as its
target the full scope of Grice's (1957) nonnatural meaning (meaningnn ),
where this constitutes communicational e¨ects that fall under a special
kind of complex re¯exive intention. He suggested that a communicator S
meantnn something by x if and only if S ``intended the utterance of x to
13 On the Notion of a Generalized Conversational Implicature

produce some e¨ect in an audience by means of the recognition of this

intention'' (1957: 385). Lest such a re¯exive intention, including reference
to itself, prove problematic, Grice (1969, 1989: chap. 5) also phrased
meaningnn in another more explicit way, as sketched in example (1).1
(1) Grice's theory of utterer's meaning
S meansnn p by ``uttering'' U to A i¨ S intends:
a. A to think p
b. A to recognize that S intends (a)
c. A's recognition of S's intending (a) to be the prime reason for A
thinking p
We will call this special kind of intention an m-intention (for ``meaning
intention''); one of its interesting properties is that it is an intention that is
achieved or satis®ed just by being recognized. There have been many
attempts at revision of this formula (see Avramides 1989 for discussion),
but because nothing much will depend on this in this book we may leave
the matter there. The point here is just that meaningnn (or something of
the sort) draws an outer boundary on the communicational e¨ects that a
theory of communication is responsible for.
The second aspect of the Gricean umbrella that is important here is the
claim that meaning is not homogeneous. Instead, within the circle of
delimited m-intentions we may distinguish numerous di¨erent genera and
species of meaning. Grice himself suggests that the semiotic pie might be
cut up as in (2).
(2) Genera and species of meaningnn

This picture is partial. Grice allows that there may be other subtypes of
signi®cation, tooÐfor example, presupposition, nonconversational non-
conventional implicatures, and so forth. None of these distinctions are
straightforward (see Harnish 1976: 331¨ ). Additionally, the modern
pragmaticist will want to add many other kinds of inference, such as those
14 Chapter 1

based on details of conversational organization (see, e.g., Levinson 1983:

chap. 6 for review), on speci®c assumptions about interactional politeness,
and so on (see, e.g., Brown and Levinson 1987, Leech 1983). But it is this
corner of the semiotic pie that concerns us here.
Chapter 3 is quite largely concerned with the di½culties with the dis-
tinction between what is said and what is implicated, although I have little
to say here about conventional implicatures. It should be noted, however,
that in this scheme, what is coded by the linguistic system is the sum of
what is said (roughly the truth-conditional content) and what is conven-
tionally implicated. In contrast, what is conversationally implicated is not
coded but rather inferred on the basis of some basic assumptions about
the rational nature of conversational activity, as stated in the Cooperative
Principle and its constituent maxims of conversations. For reference,
let me brie¯y recapitulate Grice's maxims, with abbreviations I will use
below.
The cooperative principle
``Make your contribution such as is required, at the stage at which it
occurs, by the accepted purpose or direction of the talk exchange in which
you are engaged.'' (Grice 1989: 26)
The maxims of Quality
``Try to make your contribution one that is true,'' speci®cally:
``Do not say what you believe to be false''
``Do not say that for which you lack adequate evidence'' (Grice 1989: 27)
The maxim of Relation (Relevance)
``Be relevant'' (Grice 1989: 27)
The maxims of Quantity
Q1: ``Make your contribution as informative as is required (for the pur-
poses of the exchange)''
Q2: ``Do not make your contribution more informative than is required''
(Grice 1989: 26)
The maxims of Manner
``Be perspicuous,'' speci®cally:
M1: ``Avoid obscurity of expression''
M2: ``Avoid ambiguity''
M3: ``Be brief (avoid unnecessary prolixity)''
M4: ``Be orderly'' (Grice 1989: 27)
15 On the Notion of a Generalized Conversational Implicature

For the purposes of this bookÐexploring default inferencesÐonly the

maxims of Quantity and Manner are the focus of attention, and a further
simpli®ed scheme will be introduced shortly. But here, the point is that
Grice argued that conversational implicatures come about on the basis of
the assumption that the maxims are being preserved, at least at some
fundamental level. He assumed that we can distinguish between what a
speaker has said by virtue of the conventional meaning of the words or
sentence (Grice 1989: 25, 87±88, 120±121) and what a speaker conversa-
tionally implicates by saying those words. Grice de®nes conversational
implicature as follows (my abbreviation from Grice 1989: 30±31): By
saying p, utterer U conversationally implicates q i¨:
1. U is presumed to be following the maxims
2. the supposition q is required to maintain (1)
3. U thinks the recipient will realize (2)
(Grice's distinction between saying and implicating was complicated by
his notion of conventional implicature. But it is also problematic for other
reasons, explored in detail in chapter 3, and I defer discussion till then.)
As a result of this inferential derivation, conversational implicatures are
held to display various distinctive properties:
. Cancellability (i.e., defeasibility)Ðthe property of being an inference
defeatable by the addition of premises (see section 1.5)
. NondetachabilityÐany expression with the same coded content will
tend to carry the same implicatures (a principled exception has to be
made for Manner implicatures)2
. CalculabilityÐthe more or less transparent derivation of the inference
from premises that include the assumption of rational conversational
activity
. NonconventionalityÐthe noncoded nature of the inferences and their
parasitic dependence on what is coded.
To Grice's list, we may add:
. ReinforcabilityÐit is often possible to add explicitly what is anyway
implicated with less sense of redundancy than would be the case if one
repeated the coded content (Sadock 1978, Horn 1991b)
. UniversalityÐbecause the inferences are derived ultimately from fun-
damental considerations of rationality, we expect a strong tendency to
universality (unlike with coded meanings, of course); conversational
implicatures are motivated, not arbitrary
16 Chapter 1

Apart from these reminders, I presume familiarity with all this and
proceed at once to the ®nal distinction here, that between particularized
and generalized conversational implicatures. Some abbreviations are now
in order: I shall often use the term implicature as an abbreviation for
conversational implicature where no misunderstanding is likely to arise;
and the heroes of this book, the generalized conversational implicatures
will be dubbed GCIs, and their nongeneralized cousins, the particularized
conversational implicatures are abbreviated as PCIs. I also use the sym-
bol ‡> to signify ``conversationally implicates''; more precisely, I use the
convention: ``S'' ‡> `p' to signify `the utterance of the sentence S would
(normally, or in the context sketched) conversationally implicate that the
speaker knows (or believes or thinks likely) the proposition p.'
Grice (1975: 56f ) provides only the following characterization of the
distinction between particularized and generalized conversational impli-
catures (or PCIs vs. GCIs): a PCI from saying p would ``leave no room
for the idea that an implicature of this sort is NORMALLY carried by
saying that p''; whereas in the case of a GCI ``one can say that the use of a
certain form of words in an utterance would normally (in the ABSENCE
of special circumstances) carry such & such an implicature'' (emphasis in
the original).3 Grice's intention here is clear enough although his own
discussion lacks detail; the following informal formulation captures, I
think, the intention:
(3) The distinction between PCIs and GCIs
a. An implicature i from utterance U is particularized i¨ U
implicates i only by virtue of speci®c contextual assumptions that
would not invariably or even normally obtain
b. An implicature i is generalized i¨ U implicates i unless there are
unusual speci®c contextual assumptions that defeat it
The intuition behind the contrast between PCIs and GCIs may be high-
lighted by a simple example. Consider the sentence Some of the guests are
already leaving, and consider that it might be uttered in two rather dif-
ferent contexts:
(4) Context 1
A: ``What time is it?''
B: ``Some of the guests are already leaving.''
PCI: `It must be late.'
GCI: `Not all of the guests are already leaving.'
17 On the Notion of a Generalized Conversational Implicature

(5) Context 2
A: ``Where's John?''
B: ``Some of the guests are already leaving.''
PCI: `Perhaps John has already left.'
GCI: `Not all of the guests are already leaving.'
Naturally enough, in the two contexts the utterance-form ``Some of the
guests are already leaving'' carries di¨erent implicaturesÐit is used as a
partial answer to di¨erent questions and implicates di¨erent propositions
accordingly. These PCIs may be attributed to the maxim of Relevance (or
Relation), construed as a principle of attending to interlocutors' goals or
plans (Holdcroft 1987, Levinson 1987a, Thomason 1990).4 But in both
contexts there is also a shared inferenceÐnamely, that not all of the
guests are in the process of leaving. This inference has an entirely general
currency: any statement of the form ``Some x are G'' will, other things
being equal, have the default interpretation `Not all x are G'. One might
be tempted to think about this as part of the meaning of some, but the
semantic compatibility of Some, but not all, x are G seems to rule out such
an explanation. It is this sort of inference that, by virtue of its prevalence
and its lack of salience, has the closest connection to linguistic analysis
and misanalysis and that is the focus of this book.
Grice's own treatment of the distinction between generalized and par-
ticularized implicatures is unfortunately not extended. In the immediate
context of a discussion of the distinction, Grice provides only one, none
too clear, example (1975: 56, 1989: 37), viz. the inference from the inde®-
nite article to the assumption that the speaker is not in a position to be
speci®c. Thus the assertion of (6) might normally carry the GCI inference:
(6) ``I saw a woman in my o½ce.''
GCI: `I saw someone other than my wife/girlfriend/mother/etc.'
because ``the speaker has failed to be speci®c in a way in which he might
have been expected to be speci®c, with the consequence that it is likely to
be assumed that he is not in a position to be speci®c'' (Grice 1989: 38).
Grice notes than an expression of the form an X may sometimes carry the
reverse implicature in certain cases: for example, ``I cut a ®nger'' impli-
cates `I cut my ®nger' because it would require rather special assumptions
to make plausible an interpretation where S cut someone else's ®nger
but the identity of the victim was unknown or irrelevant (perhaps two
manicurists might so talk!). Incidentally, I show later that the essence of
18 Chapter 1

Grice's claim here can be handled in a rather di¨erent way by a system-

atic scalar inference.
From the brief discussion in the published versions of the William
James Lectures (Grice 1989) it has sometimes been supposed that Grice
attributed no great importance to the distinction between generalized
versus particularized implicatures.5 However, the lectures show just the
contrary. Grice was interested in the whole phenomenon of implicature
largely because it promised an account of the generally associated but
defeasible inferences that rise from the use of the logical connectives.6
And he was particularly interested in the generalized implicatures, just
because they are hard to distinguish from semantical or conventional
content: ``noncontroversial examples are perhaps hard to ®nd, since it is
all too easy to treat a generalized conversational implicature as if it were
a conventional implicature'' (1989: 37). Elsewhere, he explained:
I also distinguished . . . particular conversational implicatures that depended on
particular contextual features . . . and ones that I thought of as relatively general
which I called GENERALIZED IMPLICATURES. These are the ones that seem
to me to be more controversial and at the same time more valuable for philosoph-
ical purposes, because they will be the implicatures that would be carried (other
things being equal) by any utterance of a certain form, though, as with all impli-
catures, they are not to be represented as part of the conventional meaning of the
words or forms in question. And I thought that this notion of a GENERALIZED
conversational implicature might be used to deal with a variety of problems, par-
ticularly in philosophical logic, but also in other areas. In these areas there seemed
to me to be quite good grounds for suspecting that some people have made the
mistake of taking as part of the conventional meaning of some form of expression
what was really not part of its conventional meaning, but was rather a non-
conventional implication which would normally be carried, except in special cir-
cumstances. It is di½cult to ®nd noncontroversial cases just because, if this
mistake has been committed, it has been committed on such a wide scale. (Grice
1981: 185, capitals in the original).

With respect to philosophical logic, he had especially the connectives

in mind. Here he wished to maintain a univocal or monosemous sense,
explaining frequently associated but nevertheless defeasible additional
meanings as pragmatic overlays due to generalized conversational impli-
catures. Thus, for instance, he assumed that good arguments could be
made for treating the following inferences as GCIs:
(7) a. ``There is a spade either in the attic or in the basement.''
‡> `The spade is not in both places and S doesn't know which.'
19 On the Notion of a Generalized Conversational Implicature

b. ``He brushed his teeth and went to bed.''

‡> `He did so in that order.'
c. ``If the President signs the treaty, then the de®cit will be repaid.''
‡> `There is some non-truth-functional relation (e.g., of cause-
e¨ect) between the two events.'
Grice's own arguments are somewhat di¨use and, in the case of the con-
ditional, convoluted, but it is worth repeating some of the ways in which
he thought the inferences might arise:7
(8) The disjunction
a. ``p or q''
‡> `S doesn't know that p, or that q.'
Rationale: If S did know that one of the disjuncts was true, that
would be more informative (and generally more relevant);
therefore by the ®rst maxim of Quantity, the fact that S hasn't
said ``p'' or ``q'' implicates he's not in a position to make the
stronger statement.
b. ``p or q''
‡> `there is some non-truth-functional connection between p
and q.'
Rationale: Because the second maxim of Quality requires that S
have adequate evidence for what he says, there must be some
argument that concludes with `p 4 q.'
(9) The conjunction
``He took o¨ his trousers and went to bed''
‡> `in that order.'
Rationale: By the fourth Maxim of Manner (M4, `be orderly'), the
speaker should describe events in the order in which they occurred.
(10) The conditional
a. ``if p then q''
‡> `S doesn't know that p, nor that q.'
Rationale: No one would be interested in the truth value of the
whole if they weren't interested in the truth value of the parts;
therefore in saying the conditional when one knew, for example,
the consequent was true, would be an infringement of the ®rst
maxim of Quantity; therefore S must be assumed to have
evidence only for the less informative conditional.
20 Chapter 1

b. ``if p then q''

‡> `there are non-truth-functional grounds for the connection
between p and q.'
Rationale: Given the second maxim of Quality (`have adequate
evidence for what you say'), there must be an argument (almost
certainly non-truth-functional) that supports `p ! q.'
(One should note that although Grice's defense of the equation of ``if ''
with the material conditional no longer meets with general approval (see,
e.g., Gazdar 1979: 83¨; Strawson 1986), the pragmatic arguments are
largely una¨ected. Despite that, modern accounts of the implicatures of
the conditional utilize the maxim of Quantity.)
Grice's interest in developing these arguments was to suggest that
although at ®rst sight these inferences may appear to be part of the con-
ventional meaning of these connectives, that may not be the case. Thus it
may be possible to maintain a relatively simple logical analysis of the
meaning of the connectives, explaining the systematic divergence in ordi-
nary usage as due to generalized conversational implicatures. It is the fact
that the inferences normally accompany the connectives that gives rise to
the temptation to view them as part of the coded content, with all the
analytical mistakes that would follow. It is this normal, general, default
tendency of interpretation that makes it nonobvious that what we are
dealing with here is mere pragmatic inference. Thus the whole line of
argument only has force if sense can be made of the notion of a general-
ized conversational implicature. Grice saw this kind of analysis as having
general application to lexical meaning: he proposed (1989: 47±48) a
modi®ed Occam's razor to the e¨ect that ``senses are not to be muliplied
beyond necessity'' and suggested instead that we should always prefer an
analysis where a word has ``a less restrictive rather than a more restrictive
meaning,'' where the more restrictive interpretations can be achieved by
``a superimposed implicature.''
In modern GCI theory, the connectives are treated rather di¨erently
(I sketch this later); in particular, the maxims of Quantity are utilized
rather than the second maxim of Quality. But the basic intellectual moves
still seem well motivated. Wherever we are faced with a linguistic expres-
sion that is apparently systematically ambiguous, we should entertain the
possibility that the correct analysis is in fact a single univocal, semanti-
cally broad sense with a defeasible set of generalized pragmatic restric-
tions.8 If the restrictions were ad hoc, varying from context to context,
21 On the Notion of a Generalized Conversational Implicature

there would be no temptation towards a (multiple) ambiguity claim. But,

if the pragmatic restrictions can constitute a preferred interpretationÐ
that is, a systematic tendency to read an expression in one particular
wayÐthen the temptation arises and it is here that GCI theory has most
to contribute to linguistic analysis.

1.2 THREE LAYERS VERSUS TWO IN THE THEORY OF

COMMUNICATION

As already mentioned, in recent work there is in general both a neglect of,

and an hostility to, the notion of a preferred or default interpretation.9
The root of this resistance to the idea would seem to lie in some largely
implicit, and I believe inadequate, assumptions about the nature of an
overall theory of communication. In trying to understand the place of
GCIs in a theory of communication, it is essential to make these issues
explicit.
I believe that Grice was essentially correct in thinking of meaning as
a composite notion; that is to say, he correctly considered that the full
import of an utterance could only be captured by distinguishing many
di¨erent kinds of contentÐeven the coded content was divided between
``the said'' and ``the conventionally implicated'' (and later he added ``the
presupposed''), whereas the inferred content was divisible between par-
ticularized and generalized implicatures and perhaps other kinds of infer-
ence altogether. I believe that this recognition of the heterogeneous nature
of meaning marked a fundamental advance in the theory of meaningÐ
but others of a holistic temperament (e.g., as in so-called cognitive
linguistics) consider the very variety of proposed mechanisms a sign of
premature failure. We can happily concede the favoring of parsimonious
explanatory principles where they seem potentially su½cient, but meaning
(conceived broadly as the full scope of meaningnn ) is quite clearly not a
unitary ®eld but rather one involving cognitive principles, knowledge
factors, and interactional principles, and no doubt much besides. At a
time when we are happy to consider that some apparently unitary ®eld
like the heart of grammar is best treated as the outcome of distinct
modules of grammatical principles, we ought surely to be willing to con-
sider that pragmatic inference is the outcome of a set of rather di¨erent
kinds of pragmatic principles.
In the composite theory of meaning, the theory of GCIs plays just a
small role in a general theory of communication. In this regard, GCI
22 Chapter 1

theory is not in direct competition with holistic theories like Sperber and
Wilson's theory of Relevance, which attempts to reduce all kinds of
pragmatic inference to one mega-principleÐGCI theory is simply not a
general theory of human pragmatic competence. Instead it attempts to
account for one relatively small area of pragmatic inference. The theo-
retical importance of that area derives from the generalized, stable nature
of such inferences that consequently have much closer interactions with
the stable body of rules of formation and construal that form a gram-
matical system. The point being emphasized is that a theory of GCIs has
to be supplemented with a theory of PCIs that will have at least as much,
and possibly considerably more, importance to a general theory of com-
munication. It is just to linguistic theory that GCIs have an unparalleled
import.
The overall picture of a general theory of communication that emerges
is rather di¨erent from the standard picture. Although at the present time
there could hardly be said to be a general theory of communication, there
is nevertheless a prevailing assumption. According to the standard line
(more often presupposed than justi®ed), there are just two levels to a
theory of communication: a level of sentence-meaning (to be explicated by
the theory of grammar in the large sense) and a level of speaker-meaning
(to be explicated by a theory of pragmatics, perhaps centrally employing
Grice's notion of meaningnn ).10 (I use the term ``layer'' or ``level'' only to
indicate distinct categories of phenomena; whether these indicate distinct
levels of representation or successive stages of processing is a theory-
dependent matter, which as I show in chap. 3 is by no means simple.)
Speaker-meaning, or utterance-token-meaning, will be a matter of the
actual nonce or once-o¨ inferences made in actual contexts by actual
recipients with all of their rich particularities. This view, although par-
simonious, is surely inadequate, indeed potentially pernicious, because
it underestimates the regularity, recurrence, and systematicity of many
kinds of pragmatic inferences.
What it omits is a third layer, what we may call the level of statement-
or utterance-meaning (see Atlas 1989: 3±4), or, as I will prefer below,
utterance-type-meaning. This third layer is a level of systematic pragmatic
inference based not on direct computations about speaker-intentions but
rather on general expectations about how language is normally used.
These expectations give rise to presumptions, default inferences, about
both content and force; and it is at this level (if at all) that we can sensibly
23 On the Notion of a Generalized Conversational Implicature

talk about speech acts, presuppositions, conventional implicatures, felicity

conditions, conversational presequences, preference organization, and so
on and of special concern to us, generalized conversational implicatures.
It is also at this level, naturally, that we can expect the systematicity of
inference that might be deeply interconnected to linguistic structure and
meaning, to the extent that it can become problematical to decide which
phenomena should be rendered unto semantic theory and which unto
pragmatics (q.v. the long-standing disputes about the semantic or prag-
matic status of illocutionary force and presupposition).
The supposition of this third, intermediate layer in a theory of com-
munication is nothing new. Austin (1962: 100¨ ), for example, clearly
had something of this kind in mind when he proposed the three-way
distinction between locutionary, illocutionary, and perlocutionary acts.
The locutionary level corresponds to the level of sentence-meaning, the
illocutionary to our intermediate layer formed of conventions or habits of
use, and the perlocutionary (partly) to the level of speaker-meaning (i.e.,
the speaker's intentions to get the addressee to believe or do something as
consequences of the utterance).11 Other theorists have energetically tried
to defend the notion of a convention of use to be distinguished from a
convention of language; for example, such a distinction seems essential if
we are to retain the idea that indirect speech acts are both partially con-
ventional and inferentially motivated (Searle 1975; see also Bach and
Harnish 1979: 192±202 on standardization as opposed to conventionaliza-
tion of a nonliteral use). Without admitting the existence of such an in-
termediate layer, how are we to explain the use of routine formulae (like
Good luck, Bless you, See you later) which although meaning what they
literally mean simultaneously perform habitual everyday rituals (Morgan
1978)? Why is it that I can introduce myself with My name is Steve, but
not I was given the name Steve; that I can express sympathy with you with
I am sorry but not conventionally with That saddens me; that I express
outrage with Really! but not with In truth!; that I can say I am delighted to
meet you but not idiomatically I am grati®ed to meet you; that I can
choose a pastry by saying I'd like that one but not I'd admire that one, and
so on. And to every speci®cation of proper usage there tends to be a cor-
responding restriction on interpretation (Levinson 1979). There is a great
body of language lore here, beyond knowledge of grammar and seman-
tics, extensively studied of course by traditional rhetoric, the ethnography
of speaking, and students of translation and second language learning.
24 Chapter 1

That two ways of ``saying the same thing'' might be unequal in their
conversational import, or that one way of saying something might pre-
empt another, these are surely not radical doctrines.
The theory of GCIs is not of course a theory of conventional idioms,
clicheÂs, and formulae, but it is a generative theory of idiomaticityÐthat is,
a set of principles guiding the choice of the right expression to suggest a
speci®c interpretation and, as a corollary, a theory accounting for pre-
ferred interpretations. GCI theory o¨ers a systematic account of why, for
example, saying ``See you on Tuesday'' when tomorrow is Tuesday would
suggest not seeing you tomorrow, or why saying ``Some of my colleagues
are competent'' would suggest that not all of them are, and so on,
matching a way of putting things with a favored interpretation in each
case. The theory thus belongs to the intermediate level of a theory of
communication, the layer of utterance(-type)-meaning.
One question that arises is whether GCIs are just the same things as
other kinds of pragmatic inference that might belong to this level. First,
there are perhaps more general kinds of inference that might subsume the
GCIs and that also lie between the conventions of language on the one
hand and the full-blown nonce inference to speaker-meaning on the other.
For example, Bach has proposed the idea of a standardized inference
which ``shortcircuits the steps of the inferences, both as intended by the
speaker and as carried through by the listener'' (Bach 1975: 235, 1995).12
Such inferences go through by default, without full calculation. But
Bach's notion relies on compression by precedent (1995: 685), and GCIs,
on my conception at least, are generative, driven by general heuristics and
and not dependent on routinization. Second, there are more speci®c
notions which at ®rst blush one might want to equate with GCIs but
shouldn't. One of these, the notion of short-circuited implicature (Brown
and Levinson 1987 [1978]: 290, n. 35; Morgan 1978; Horn and Bayer
1984), has been developed to handle a very speci®c phenomenon, also
arising from routinization, whereby utterances acquire morphosyntactic
markers of their implicated content (examples are indirect speech acts and
Neg-raising constructions; see chap. 2). Finally, various theorists have
argued that not every kind of inference driven by Gricean reasoning is an
implicature proper: it is variously an explicature (Sperber and Wilson
1986) or an impliciture (Bach 1994), and it may be thought that GCIs
belong to one of these categories. These are essentially conceived of as a
¯eshing out of what was expressed (thus utterances only have one of
them, but they may have many GCIs). I argue against this equation in
25 On the Notion of a Generalized Conversational Implicature

chapter 3, but meanwhile it should be clear that my conception of a GCI

is on the one hand rather catholic in the sense that it includes phenomena
that others may want to consider nonimplicatures (e.g., explicatures or
implicitures), and on the other hand it is orthogonal to the very general
principles of short-circuiting of inference by precedence.
There are thus a number of distinct phenomena that may belong to
this intermediate layer in a theory of communication. Nevertheless, that
layer is constantly under attack by reductionists seeking to assimilate it
either to the level of sentence-meaning or to the level of speaker-meaning;
thus, for example, in the case of GCIs, Kamp, Peters, Kempson, van
Kuppevelt, and others have all suggested that they should be in e¨ect
semanticized, whereas Sperber and Wilson and arti®cial intelligence local-
pragmatics theorists have presumed that on the contrary they should be
assimilated to matters of nonce inference at the level of speaker-intention.
But GCIs are not going to reduce so easily in either direction, for they sit
midway, systematically in¯uencing grammar and semantics on the one
hand and speaker-meaning on the other.
There is another way of course to make these distinctions, by invoking
the type/token distinction.13 On the two-leveled theory of communication
we have a theory of sentence-meaning, which is a theory of the meaning
of sentence-types rather than sentence-tokens,14 on the one hand, and a
theory of speaker- or utterance-meaning, which is a theory of utterance-
token rather than utterance-type meaning, on the other. The theory of
utterance-meaning is a theory of utterance-token meaning because theo-
rists of this ilk conceive of speaker-meaning as a matter of nonce inference
to the best interpretation taking all the particularities of the context into
account. Now, from the point of view of a three-leveled theory of mean-
ing, we need an additional distinction between utterance-token meaning
and utterance-type meaning: the intermediate level is the level at which
generalizations about whole classes of utterances (utterance-types) must
reside. Thus GCI theory is one kind of theory about utterance-type
meaning.
Reductionists, if they are willing to countenance the idea of a preferred
interpretation or an utterance-type meaning at all, might propose that the
phenomena can in fact be explained in terms of the structure of contexts.
Faced for example with the tendency to read John has three children as
`John has exactly three children', an explanation might be o¨ered in terms
of nonce speaker-meaning together with a general preponderance of con-
texts in which absolute cardinality is relevant (Kempson (1986) perhaps
26 Chapter 1

holds this view). Thus we could hold onto the idea that all inferences
are constructed de novo without reference to a default rule or a preferred
interpretation, by hypothesizing that the background premises remain
su½ciently constant to yield the observable bias in interpretations. It may
not be easy to distinguish between this view and a theory of default
interpretations, especially because such a contextual bias might itself be
su½cient to engender a mental counterpart, a default rule of inference.
The material di¨erence would be that instead of having two di¨erent
kinds of inference and one kind of context (as in GCI theory, where we
must distinguish default GCIs from nonce PCIs), we will have one kind of
inference and two kinds of context, the default context and the fully par-
ticular nonce context (amounting to a reduction of GCIs to PCIs).
My own view is that a theory of preferred interpretations is at least as
plausible and much more workable; I am simply not sanguine about the
possibilities of a theory of contexts able to predict the bias in inter-
pretations in the same straightforward way. Moreover, on the theory that
is advanced in this book, it is the choice of utterance-form and content, in
line with perfectly general principles, that triggers the inferences in ques-
tion. Such general patterns, in which, for example, marked forms suggest
marked extensions, can't come out of a theory of contextual biases alone.
A theory of favored contexts would have to be married to a theory of the
correlation of utterance-form to kinds of contexts (some suggestions can
be found in Zipf 1949). The present theory presumes such a rough corre-
lation but goes beyond it, in that it proposes that speakers exploit such
correlations for semiotic purposes.
Now it may be thought that for quite di¨erent reasons the notion of
utterance-type meaning would force us into a theory of context-types. The
argument would be that if an utterance is construed, as it is classically, as
a sentence-context pair, the only way to abstract a type out of such a
token would be to generalize over context-types.15 But we may need to
change the classical analysis of the utterance anyway. The need for such a
change emerges fairly clearly from recent work in semantics; for example,
in both DRT and Situation Semantics, sentence-meaning is already a
matter of the relation between context and content, so we can no longer
make the distinction between sentence-meaning and utterance-meaning
by dragging in the context. We need to reconstruct all these concepts. One
way may be to split the notion of context into di¨erent constitutive com-
ponents and abstract over each of these. Thus to obtain the semantic
content of ``We are here'' we may abstract over speakers and places,
27 On the Notion of a Generalized Conversational Implicature

whereas to obtain the proposition expressed we must instantiate individ-

ual speakers and places of speaking; that alone will not be anything like
su½cient to give us a speaker-meaning. For that, we'll need to know who
are the speaker's companions and how the place including the speaker is
to be e¨ectively delimited, and so on. If we are playing hide-and-seek,
then the speaker-meaning may be intentionally less speci®c. Another way
to make the relevant distinctions might be to make speaker-meaning a
three-way relation between speaker, context, and the proposition intended,
whereas utterance-type meaning is just a relation between context and
classes of proposition (those that the context circumscribes as possible
speaker-meanings). But these are di½cult issues and the precise moves we
make will depend on the nature of the semantic theory we are utilizing. In
any case, it is clear that we need more distinctions than the simple oppo-
sition between sentence and utterance provides, and it is likely that in
talking about sentence-meaning we will already be generalizing over
classes of contexts. And it is quite clear (as, e.g., demonstrated by Situa-
tion Semantics) that generalizing over classes of contexts does not amount
to having a theory of default or preferred or statistically normal con-
textsÐthat should be a rather more empirical matter.
To return to the central issue, it would seem incontrovertible that any
theory of utterance interpretation would have to admit the contribution
of a level at which sentences are systematically paired with preferred
interpretations. I will therefore presume that we need such a three-tiered
theory of communication. This presumption does not presuppose that the
distinctions between the middle layer of utterance-meaning and the upper
and lower levels is in any way cut and dried. Indeed, there is every reason
to suppose that matters of utterance-meaning will shade into speaker-
meaning at the one end and sentence-meaning at the other. This is in
part because there is plenty of evidence that language use is the source
for grammaticalized patterns and that there is a diachronic path from
speaker-meanings to utterance-meanings to sentence-meanings. Thus gray
areas at the boundaries do not constitute evidence against the tripartite
view, and evidence for it is the existence of preferred interpretations,
default presumptions of the kind I illustrate in detail in section 1.4.

1.3 THE ARGUMENT FROM DESIGN: THE MAXIMS AS HEURISTICS

The apparently e¨ortless speed of daily communication through language

is built on much specialized and highly evolved machinery. Some of the
28 Chapter 1

machinery is anatomical, like the structure and positioning of pharynx

and larynx; some is neurophysiological like the language specialized areas
of the brain, and the special pathways for the analysis of speech signals;
and some is perceptual like the precise match between the window of
human auditory perception and the frequencies of human speech.16 All
this betokens a deep evolutionary history and a long honing by the pro-
cesses of selection toward an optimal system of linguistic communication
characterized (all things considered) by an almost miraculous speed,
accuracy, and automaticity.17
Against this background, it is nevertheless possible to identify a signi®-
cant bottleneck in the speed of human communicationÐa design ¯aw, as
it were, in an otherwise optimal system. The bottleneck is constituted by
the remarkably slow transmission rate of human speech (conceived of as
the rate at which phonetic representations can be encoded as discrim-
inable acoustic signals), with a limit in the range of seven syllables or 18
segments per second.18 It is instructive (if in various ways misleading) to
compare the transmission rate of human speech to the rates of informa-
tion transmission that we are familiar with between relatively simple
electronic devices like personal computers. As users of such devices will
know, this rate is measured in bits of information transferred per second
(bps), and transmission rates of 300 bps used to be the absolute and frus-
trating minimum at which information was transferred (the information
will re¯ect on a screen, for example, at a much slower rate than one can
read it); 2400 bps or higher is now the standard minimum and directly
linked devices may communicate at 115,200 bps or higher. If we now
compare human speech encoding rates we ®nd that, making various
rather optimistic assumptions, the maximal sort of data-transfer rate for
normal human speech is under 100 bps.19 That is brutally slow.
In contrast, the psycholinguistic evidence suggests that all other aspects
of speech production and comprehension can run at a much higher rate.
For example, prearticulation processes in speech production seem to run
at a speed three to four times faster than the actual articulation rate
(Wheeldon and Levelt 1995), and rates at which speech can be parsed and
comprehended show the same kind of advantage over articulation rates
(Calvert 1986: 178; Mehler et al. 1993).20 It is this mismatch between
articulation rates on the one hand, and the rates of mental preparation for
speech production or the speed of speech comprehension on the other
hand, which points to a single fundamental bottleneck in the e½ciency of
human communication, occasioned no doubt by absolute physiological
constraints on the articulators.
29 On the Notion of a Generalized Conversational Implicature

Grice routinely adopted a design perspective in order to tackle funda-

mental problems in philosophy. He would imagine arti®cial life-forms,
which he dubbed pirots, and ask how they would have to be minimally
endowed in order to display the particular aspect of human behavior he
was interested in.21 Now it is quite clear that pirots, conceived of as in-
telligent agents with the asymmetrical abilities in thinking and speaking I
have just elucidated, would ®nd a way around the articulatory bottleneck
( just as, as a matter of fact, evolution has). The essential asymmetry is:
inference is cheap, articulation expensive, and thus the design require-
ments are for a system that maximizes inference.22 (Hence the perspective
outlined in the introductionÐlinguistic coding is to be thought of less like
de®nitive content and more like interpretive clue.)
From a Gricean perspective, communication involves the inferential
recovery of speakers' intentions: it is the recognition by the addressee of
the speaker's intention to get the addressee to think such-and-such that
essentially constitutes communication. The question has been just how
this recognition of others' intentions is possible. Skeptics have assumed
that all accounts will simply smuggle in the notion of a conventional
signal (Zi¨ 1975; Chomsky 1975: 62±77). Griceans have put forward
various accounts that trade for the most part on a notion of mutual
salience. Mutual salience was shown to have quite dramatic e¨ects on the
possibility of tacit coordination, or joint action in the absence of com-
munication, by Schelling (1960: 55), a student of strategy in economics
and politics. He showed that if, for example, two subjects are told they
will win a prize only if they both choose either heads or tails without
communicating, nearly 90% will coordinate on heads. The trick rests on
our ability to double-guess what the other party is likely to think the ob-
vious solution, which in turn rests on what he thinks we think he will
think prominentÐthat is, on some notion of what is mutually salient.
This rather astonishing ability (utilized further below) was seized on by
Griceans (Lewis 1969; Schi¨er 1972), because it suggests how it may be
possible to produce a communicative action that suggests the m-intention
that underlies it, even in the absence of any communicative convention or
signal. For example, a communicator may produce an utterance or action
(say, waving an umbrella) knowing that his behavior will suggest one
salient idea (say, that it may rain), and that the addressee will be looking
for just such a salient suggestion. (Note that a coordination problem can
involve each party doing di¨erent things, e.g., one ®nding the correct
signal, and the other its intended interpretation; see Lewis 1969, Clark
1996.) How tacit coordination really works is admittedly a mystery,
30 Chapter 1

although some have suggested it is based on the fact that humans are
endowed with a mental re¯ex to process information in particular ways
(Sperber and Wilson 1986). But it is clear that any mutual expectations
can serve to drastically limit the search-space for coordination solutions
(see Clark 1996: chap. 3).
There is a further, related problem that has received less attention, what
might be called the logical problem of reconstructing speaker's intentions
(Levinson 1995b). Let us (following Aristotle) assume that we reason
from goals to actions utilizing a logic of action, or a practical reasoning.
Now, some theorists have assumed that intention-recognition is simply a
matter of running that reasoning backwards (``turned upside down'' as
von Wright (1971: 96) puts it): we observe the behavior and ®gure out the
underlying intention by the same rules that we convert intentions into the
actions that will e¨ectuate them. The logical problem is that this cannot
work, for the simple reason that for all inference systems one cannot
work backwards from a conclusion to the premises from which it was
deducedÐthere is always an in®nite set of premises which might yield the
same conclusion (consider, e.g., the conclusion q from the sets fp; p ! qg,
fp & qg, fp4q; @pg, etc.). Given that we do indeed recognize others'
intentions (not only communicative of course), there have to be further
constraints that limit the search space of sets of premises. Let us call such
constraints heuristics. They may be various in kind: I may take your
intentions to be circumscribed because you are a shop attendant and I am
a customer, or because I just asked you a question, or because I happen to
know your predilections, or because we have tacitly agreed to maximize
the chances of successful communication by following some speci®c heu-
ristics (e.g., on this radio channel we'll only talk about air-tra½c control).
Now we have two particular design requirements for our communicat-
ing pirots. First, they must ®nd a way around the articulatory bottleneck
and use inferential enrichment to bypass this physiological limitation,
given that the rest of the linguistic apparatus seems to be able to outstrip
the articulation rate many times over. Second, the inferential enrichment
must be circumscribed in such a way that communicators can indeed
zero-in or coordinate (in the Schelling sense) on the correct commu-
nicative forms and intentions. What our pirots clearly need is a set of
heuristics, mutually assumed by sender and receiver, that can serve to
multiply the coded information by a factor of, say, three, by licensing
inferential enrichment of what is actually encoded by choice of a speci®c
signal. These heuristics must at the same time constrain that enrichment
in such a way that the overall message can still be correctly recovered, by
31 On the Notion of a Generalized Conversational Implicature

guiding (or coordinating) the match between the chosen signal and the
recoverable, augmented message.
How could such heuristics work? It may be useful to sketch a little
world about which our pirots may talk to one another. But ®rst, recollect
the idea, independently developed in slightly di¨erent ways by Popper
(1959) and Bar-Hillel and Carnap (1952), that the semantic information
content of a message can be measured in proportion to the number of
states of a¨airs that the message e¨ectively rules out, given a domain of
discourse. For example, in a domain consisting of Bill and Ben, and con-
sidering just the states of a¨airs characterized by two predicates, say
`married' and `middle-aged', the assertion ``Bill is married'' will rule out
just 8 out of 16 possible states of a¨airs (e.g., a state where Ben is married
and middle-aged and where Bill is not middle-aged and not married). But
the assertion ``Bill is married and middle-aged'' will rule out twice as
many states of a¨airs and, by this measure, will be twice as informative.
Clearly, there are limits to the utility of such a characterization of infor-
mativeness (e.g., rather a lot depends on what properties we are actually
interested in).23 But, it is useful as a ®rst approximationÐand besides, it
is just about the only measure of semantic information available (for a
critique, see, e.g., Cohen 1971: 135¨ ).
Thus, to increase the informativeness of a coded message which itself
rules out some number n of states of a¨airs, and thus to overcome the
limit on the rate of encoding, all we need are some heuristics that will
serve to rule out some further possible states, yielding an incremented
content n ‡ m. To illustrate this graphically, imagine a simple ``blocks
world'' of the sort made familiar by Winograd's (1972) program for
understanding natural language. Let us assume that in this blocks world
there are a set of cubes, cones, and pyramids of di¨erent colors. Now we
can invent three simple heuristics and demonstrate their e½cacy.
Heuristic 1
What isn't said, isn't
As an example, take the assertion in (11).
(11) ``There's a blue pyramid on the red cube,''
Licensed inferences: `There is not a cone on the red cube'; `There is
not a red pyramid on the red cube'
The heuristic licences the inferences (and others of their ilk, depending
on the number of objects and the number of colors in the blocks world).
Because these inferences rule out a number of possible states of a¨airs,
32 Chapter 1

they multiply the informational load of what has been said by a signi®-
cant factor. The heuristic is wildly underspeci®ed, of course. It must be
understood to be restricted to a set of salient contrastsÐfor example, the
ruling out of the presence of a cone on the red cube depends on the salient
opposition in the blocks world of fcones, pyramids, cubesg, whereas the
ruling out of the red pyramid depends on the salient opposition {red,
blue}. If the heuristic is unrestricted, then of course whatever one did
not specify would not be the case, and that would be such a powerful
heuristic it would inhibit one from saying anything (for fear of having to
list everything that is the case)! Thus this heuristic depends crucially on
clearly established salient contrasts. But suitably restricted, such a heu-
ristic o¨ers tremendous increases in communicational load.24
Consider now the e½cacy of a di¨erent kind of heuristic, one that
trades on assumptions about what is normal. Such a heuristic might be
phrased, to emphasize the contrast with the previous one, as ``what isn't
said, is as usual.'' Or it might be phrased as ``don't bother to say what can
be taken for granted.'' But for reasons that will become clear, I prefer the
following sort of formulation:25
Heuristic 2
What is simply described is stereotypically exempli®ed
Returning to our blocks world, consider the application of such a heu-
ristic to:
(12) `The blue pyramid is on the red cube.''
Licensed inferences:
`The pyramid is a stereotypical one, on a square, rather than, e.g.,
a hexagonal base.'
`The pyramid is directly supported by the cube (e.g., there is no
intervening slab)';
`The pyramid is centrally placed on, or properly supported by, the
cube (it is not teetering on the edge, etc.)';
`The pyramid is in canonical position, resting on its base, and not
balanced, e.g., on its apex.'
First, we are likely to assume that the pyramid in question has a four-
sided base, although the de®nition of pyramid allows any number of sides
from three up. Similarly, we expect it to be a regular ®gure, not leaning to
one side, or anything of that kind. If these conditions were not met, we'd
expect to be warned by a phrase like a pyramid with polygonal base, an
33 On the Notion of a Generalized Conversational Implicature

irregular pyramid, a sort of pyramid, and so on. The heuristic licenses

these assumptions.
But let us concentrate on the speci®cation of the relation between the
two blocks, given by the preposition on. As a ®rst approximation, we may
take Miller and Johnson-Laird's (1976: 386f ) gloss of the relation
ON(x; y) as specifying that y supports x, and x is `in the region of inter-
action' of a surface of y's. Thus the de®nition of on will allow things to
intervene between x and y (as when we say ``the mug is on the table'' even
though there's a book underneath it) and it will certainly allow x and y to
be in a momentary or precarious support relation. But the heuristic will
license more restrictive assumptions, speci®cally for (11) that the pyramid
is directly on the red block, and that it is not teetering on the edge of the
red block, or in the process of falling o¨. Instead, we assume, because of
the lack of warning to the contrary, that the pyramid is ®rmly, directly,
and indeed centrally placed on the block. Similarly, the heuristic licenses
the inference that the pyramid is resting on its base, the most stable way in
which it could be supported.
Such a heuristic is extremely powerfulÐit allows an interpreter to bring
all sorts of background knowledge about a domain to bear on a rich
interpretation of a minimal description.26
This heuristic, because of its very power, carries with it the need for
another complementary one that would allow the communicator to cancel
the assumptions that would otherwise follow from a heuristic specifying
``normal description betokens normal situation.'' If what is simply
described can be presumed to be stereotypically exempli®ed, then what is
described in a marked or unusual way should be presumed to contrast
with that stereotypical or normal exempli®cation. I formulate the com-
plementary heuristic as follows:
Heuristic 3
What's said in an abnormal way, isn't normal; or Marked message
indicates marked situation
Consider the application of this heuristic to:
(13) ``The blue cuboid block is supported by the red cube.''
Licensed inferences: `The blue block is not, strictly, a cube'; `The blue
block is not directly or centrally or stably supported by the red cube.'
Thus, given the third heuristic, an interpreter of (13) is likely to presume
that the situation described is not one in which a canonical cube is sitting
34 Chapter 1

directly on the middle of another. Speci®cally, the presumption is likely to

be that there are reasons for avoiding the use of the simple expressions
cube and on, the reason being that the second heuristic would license
assumptions that the speaker doesn't want to encourage.
How do the suggestions come about? Note that any cube is also a
cuboid (i.e., a solid with rectangular faces), but the use of the less familiar
word suggests that the familiar extension is not what the speaker intends.
The procedure is general: the use of the phrase the squarish block would
also suggest `not a cube', even though a cube is surely (at least) squarish.
Similarly, opposing modes of description of the ON-relation, carry con-
trastive suggestions:
(14) a. The blue pyramid is on the red cube.
b. The blue pyramid is supported by the red cube.
The use of the simple preposition in (14a) serves by the second heuristic to
rule out various marked states of a¨airs (like indirect support), whereas
the use of the circumlocution or partial paraphrase in (14b) induces the
complementary interpretation by the third heuristicÐthe marked expres-
sion suggesting the description of a marked state of a¨airs. Thus the third
heuristic again excludes certain interpretations, now the ones that would
have risen had the stereotypical assumptions been encouraged by the use
of an unmarked expression.
Here then are three simple heuristics that serve to get around the bottle-
neck created by the articulation rate of human speech. Each ampli®es
the communicative content of what is said by the measure suggested by
Bar-Hillel and Carnap (1952): for each heuristic allows the recipient of a
message to rule out a great many states of a¨airs that are in fact com-
patible with the content of the coded message. Taken together, the set of
three heuristics serves to multiply the informational content of any mes-
sage by a factor of perhaps a score, transforming the slow coding rate
of human speech into something approximating the perceived speed of
human communication. All that is required for such a system to work is a
tacit agreement between communicators that such heuristics can be
assumed to be operative unless there are indications otherwise. At the
same time, because the heuristics e¨ectively narrow the range of intended
extensions of expressions, they will signi®cantly constrain the search space
for speakers' intentions and thus help to resolve the logical problem of
intention-recovery.
35 On the Notion of a Generalized Conversational Implicature

The question that can now be raised is: could any rational and intelli-
gent communicators, faced with a transmission bottleneck, fail to derive
and utilize such a simple solution to the problem of narrow bandwidth? It
seems highly unlikely. And if that is so, one of the central preoccupations
of pragmatic theory should be to discover and elucidate the particular
heuristics that are actually employed. And although the force of this
argument does not depend on any particular proposed heuristics, the
three I have sketched here seem to have both intuitive and empirical sup-
port, which I now turn to show.

1.4 A TYPOLOGY OF GCIs

There is in fact a fairly transparent relationship between the three heu-

ristics introduced in the previous section and three of Grice's maxims of
conversation. All that is necessary to see the connection is to introduce a
di¨erent way of thinking about the maxims. Instead of thinking about
them as rules (or rules of thumb) or behavioral norms, it is useful to think
of them as primarily inferential heuristics which then motivate the behav-
ioral norms. That some of the maxims and not others have a special status
as inducers of GCIs more or less follows from the nature of generalized
inference, as will become clear. By assimilating the new perspective to the
old, we can bene®t from the many observations made under more classi-
cal Gricean assumptions.
But the purpose here is to show that the three heuristics sketched in the
prior section do indeed seem to be good candidates for such background
inferential principles, and this can be done by taking them one by one,
showing how they relate to Grice's scheme and how they collect prior
observations into a tripartite typology of GCIs. At this stage, I simply
wish to enlist intuitive support for the three proposed heuristics, and I
shall make various rather drastic simpli®cations that will be modi®ed later
(see especially chap. 2).

1.4.1 The First (Q-) Heuristic

The ®rst heuristic (``What isn't said, isn't'') is more or less transparently
related to Grice's ®rst Maxim of Quantity, QI: Make your contribution as
informative as is required. Grice's Q1 maxim is the one normally held to
be responsible for the classic scalar implicatures, as well as Gazdar's
(1979) clausal implicatures. The essential concept behind the scalar and
36 Chapter 1

clausal implicatures is the notion of a contrast set, of linguistic expres-

sions in salient contrast, which di¨er in informativeness (see chap. 2).27
And, as I have just shown, the ®rst heuristic depends crucially on a
restriction to a set of salient alternates. For example, there is a scalar
contrast set hall, somei, such that saying (15a) implicates the rationale
being that the speaker would have chosen the stronger alternate if he was
in a position to do so. Similarly, for the clausal alternates h(since p, q),
(if p, q)i, the use of the weaker conditional stands in opposition to the use
of constructions that would entail the embedded sentences (e.g., since p, q).
(15) a. ``Some of the boys came.''
‡> (scalar implicates) `Not all of the boys came.'
b. ``If there is life on Mars, the NASA budget will be spared.''
‡> (clausally implicates) `There may or may not be life on Mars.'
Scalar implicature is thus just a special case of a whole family of impli-
catures based on salient alternates (mostly, but not necessarily, ranked as
informationally weaker or stronger), as indicated schematically below:
Q Contrast Sets
The Q heuristic (``What you do not say is not the case'') has to be
restricted: for sets of alternates, use of one (especially a weaker) implicates
inapplicability of another (especially an otherwise compatible stronger
alternate).
scalar: hall, somei, ``some'' ‡> `not all'
negative scales: hnone, not alli, ``not all'' ‡> `not none, i.e., some'
clausal: hsince-p-q, if-p-qi, ``if p then q'' ‡> `p is uncertain'
nonentailment scales: hsucceed, tryi ``try'' ‡> `not succeed'
nonentailment sets: fyellow, red, blue, . . .g ``yellow'' ‡> `not red, etc.'
Given the heuristic (``For the relevant salient alternates, what isn't said is
not the case''), such sets of salient alternates then provide the basis of the
following sorts of inference:
(16) a. ``Some of the boys came''
‡> `not all'
b. ``Three boys came in''
‡> `not four'
c. ``Possibly, there's life on Mars''
‡> `not certainly'
d. ``Not all of the boys came''
‡> `some did'
37 On the Notion of a Generalized Conversational Implicature

e. ``If John comes, I'll go''

‡> `maybe he will, maybe he won't'
f. ``John tried to reach the summit''
‡> `he didn't succeed'
g. ``Her dress was red''
‡> `not red and blue'

1.4.2 The Second (I) Heuristic

The second heuristic (``What is expressed simply is stereotypically exem-
pli®ed'') may be related directly to Grice's second Maxim of Quantity,
Q2: Do not make your contribution more informative than is required.
The underlying idea is, of course, that one need not say what can be taken
for granted.
I develop this heuristic in a particular way below, generalizing some-
what to gather in inferences to a rich interpretation that may not strictly
be inferences to the stereotype. As a rough gloss, I o¨er ``minimal
speci®cations get maximally informative or stereotypical interpretations.''
Brief and simple expressions thus encourage, by this heuristic, a tendency
to select the best interpretation to the most stereotypical, most explana-
tory exempli®cation. This rubric allows us to subsume under this heuristic
a great number of well-known interpretive tendencies. Hence from the use
of semantically general expressions we obtain the generality-narrowing in
(17a±c), the strengthening of conditionals to biconditionals in (17d) (i.e.,
conditional perfection) and the strengthening of negations from contra-
dictories to contraries in (17e), the enrichment of conjunctions by the
assumption of temporal sequence and causality as in (17f ) (i.e., conjunc-
tion buttressing), the assumption that conjoined subjects acted together as
in (17g) (i.e., together-implications), a preference for local coreference
as in (17h), and for the ®nding of local antecedents as in the bridging
inference in (17i).
(17) a. ``John's book is good''
‡> the one he read, wrote, borrowed, as appropriate.
b. ``bread knife''/``kitchen knife''/``steel knife''
c. ``a secretary''
‡> female one
``a road''
‡> hard-surfaced one
d. ``If you mow the lawn, I'll give you $5.''
‡> `I¨ you mow the lawn, will I give you $5.'
38 Chapter 1

e. ``I don't like garlic''

‡> contrary `I dislike garlic.'
``I don't believe p''
‡> `I believe not-p.'
f. ``John turned the switch and the motor started.''
‡> p and then q, p caused q, John intended p to cause q, etc.
g. ``John and Jenny bought a piano''
‡> together
(cf.: ``The Americans and the Russians launched a satellite in
1962.'')
h. ``John came in and he laughed''
Preference for local coreference: John ˆ he
i. ``The picnic was awful. The beer was warm.''
Bridging: ‡> `the beer is part of the picnic.'
Once again, a great deal of support can be found for the operation of
some heuristic of this kind. A version of such a principle is given by Atlas
and Levinson (1981), who dubbed it the Informativeness Principle, and
following this, I call the heuristic the I-principle.

1.4.3 The Third (M) Heuristic

The third heuristic (``What's said in an abnormal way isn't normal'') can
be related directly to Grice's maxim of Manner (``Be perspicuous''), spe-
ci®cally to his ®rst submaxim ``avoid obscurity of expression'' and his
fourth ``avoid prolixity'' (Grice's M1 and M4). The underlying idea here
is that there is an implicit opposition or parasitic relationship between our
second and third heuristics: what is said simply, brie¯y, in an unmarked
way picks up the stereotypical interpretation; if in contrast a marked
expression is used, it is suggested that the stereotypical interpretation
should be avoided. Thus we have complementary interpretations: marked
expressions pick up the complement of the stereotypical extensions that
would have been suggested by the use of the corresponding unmarked
forms, had they been used (Horn 1984; all this developed in chap. 2 of
this book).
Again, there seems plenty of intuitive support for the existence of such
a heuristic. Prototype examples are the pragmatic e¨ects of double neg-
atives; thus the simple positive in (18a) suggests that the plane may be late
as planes often are, but the double negative in (18b) suggests that there's a
rather more remote possibility.
39 On the Notion of a Generalized Conversational Implicature

(18) a. ``It's possible the plane will be late.''

‡> (by I) `likely to stereotypical probability n'
b. ``It's not impossible that the plane will be late.''
‡> (by M) `rather less likely than n'
In the same way, use of a periphrastic alternative to a simple causative
verb suggests some deviation from the expected chain of events:
(19) a. ``Bill stopped the car''
‡> (by I) `in the stereotypical manner with the foot pedal'
b. ``Bill caused the car to stop''
‡> (by M) `indirectly, not in the normal way, e.g., by use of the
emergency brake'
Similarly for more elaborate paraphrases:
(20) ``The corners of Sue's lips turned slightly upward.''
M ‡> `Sue didn't exactly smile'
And a particularly important instance of the opposition between
unmarked and marked expressions may prove to be that between pro-
nouns (inducing assumptions of local coreference) and full lexical NPs
(resisting local coreference) as in:
(21) ``John came in and the man laughed.''
‡> `the man denotes someone other than he would have'

1.4.4 Interactions Between Implicatures

We now have three heuristics, each supporting a distinct family or genus
of implicature. The implicatures will be generalized, that is will have the
status of preferred interpretations, because the heuristics will be under-
stood to be generally in forceÐit is that which gives them their commu-
nicational e½cacy.
But di¨erent modes of GCI generation may yield inconsistent implica-
tures. This gives rise to an interesting and complex projection problem,
which will be the subject of later attention (see chap. 2).28 But it is im-
portant to know that although the problem is complex, it is not as in-
tractable as it might appear. For it seems that, where inconsistent GCIs
arise, they are systematically resolved by an ordered set of priorities:
Level of genus: Q > M > I (read > as `defeats inconsistent')
Level of species: (e.g.) Q-clausal > Q-scalar
40 Chapter 1

Such ordered priorities seem to account for the observable preferred

interpretations where, in principle, two or more inconsistent inferences
might arise (as will be described in chap. 2 in detail).
Why should the priorities be like this? There is no complete explanation
at present, but here is perhaps part of the answer. First, there is a dis-
tinction between the Q- and M-inferences based primarily on linguistic
alternates on the one hand, and the I-inferences based primarily on stereo-
typical presumptions about the world on the other. It would be essen-
tial to be able to utilize the resources of the language in order to indicate
that the normal, stereotypical, rich presumptions about the world do
not hold; and that is what ordering Q and M before I achieves. Both Q-
and M-heuristics essentially induce a metalinguistic mode of inference,
in contrast to the I-heuristic. They di¨er though in the kind of meta-
linguistic contrast that they rely onÐQ relies on sets of alternates of
essentially similar form with contrastive semantic content, whereas M
relies on sets of alternates that contrast in form but not in inherent
semantic content.
These properties provide a set of simple diagnostics for the three main
genera of GCI, which I represent in a simpli®ed way in table 1.1.29 A few
remarks on the table are in order. First, although there is some termino-
logical confusion and some di¨erences between lumpers and splitters, on
the whole neo-Griceans are agreed about how the phenomena divide. For
example, the reasons that Horn subsumes both (what are called in this
book) Q and M inferences under one label should already be clear from
the tableÐthey share many properties, even though they are distinct at a
®ner level of granularity. In this book, I make no direct argument for the
tripartite scheme here employed (for which see Levinson 1987a) but
rather exemplify with how such a scheme handles the data, and so let the
scheme argue for itself. Now a word on the listed properties of each type
of implicature. First, by negative inference I have in mind that Q and M
inferences are essentially negative, not incidentally so, because what is
implicated is that the speaker is avoiding some stronger (Q) or some
simpler (M) expression, and thus indicating that he or she is not in a posi-
tion to use those other expressions. Making further psychological assump-
tions, the recipient may thus be led to believe that the speaker does not
know whether the relevant sentence with those alternate expressions
would be true, or even knows that those expressions if employed would
express a falsehood. For this reason, both Q and M implicatures are
41 On the Notion of a Generalized Conversational Implicature

Table 1.1
Diagnostics for the three types of GCI
Corresponding terminologies
Present book1 Q M I
Gricean maxims Q1 M1 & M4 Q2
Horn's (1984) terms Q Q R
Levinson's (1985/7) terms2 Q Q/M I
Properties of each type
Negative inference yes yes no
Metalinguistic basis yes yes no
Contrast between
semantically strong/weak yes no N/A
synonymous surface forms no yes N/A
Within the scope of metalinguistic yes yes no
negation
Inference to stereotype no no yes
Overriding GCIs none Q Q, M
1 I am thus reverting e¨ectively to the terminology in Atlas and Levinson 1981.
2 The Q/M formulation was intended to record that in Horn's view these infer-
ences are Quantity inferences, while in mine they are Manner ones. I am guilty of
using yet another terminology in Levinson 1995a, where for purposes of perspic-
uous reference to Grice's maxims, I used Q1, M, and Q2 respectively.

metalinguistic in the sense that they can only be recovered by reference to

what else might have been said but was not. In particular, there is meta-
linguistic reference to salient sets of linguistic alternates but of di¨erent
kinds in the two cases: informationally ordered sets in the case of Q; sets
of synonyms di¨ering in markedness in the case of M. This last feature is
what distinguishes between Q and M implicatures. Their shared meta-
linguistic character, however, makes it possible to deny the relevant
implicature using metalinguistic negation with special stress, as in You
didn't eat SOME of the cookies, you ate ALL of them. All these properties
contrast with I implicatures, which do not have a metalinguistic basis in
speaker avoidance of other expressions but are direct inferences from
unmarked expressions without stronger contrastive alternates to infor-
mationally rich, often stereotypical interpretations. Finally, the three
kinds of inference di¨er in strength, and consequently they interact in a
speci®c way: Q inferences take priority over inconsistent inferences of
other kinds, and M inferences over I inferences.
42 Chapter 1

We now have in place our cast of characters, the kinds of GCI that will
play an essential role in the argument throughout this book. In the next
chapter, I describe them and the heuristics that give rise to them in much
more detail. But now I return to the work of this chapterÐthe attempt to
establish the very existence of a mode of inference of a kind exempli®ed
by generalized conversational implicatures.

1.5 NONMONOTONICITY AND DEFAULT REASONING

1.5.1 Typology of Nonmonotonic Reasoning Systems

GCIs are inferences that appear to go through in the absence of infor-
mation to the contrary; but additional information to the contrary may be
quite su½cient to cause them to evaporate. Thus the mode of inference
appears to have two important properties: it is a default mode of reason-
ing, and it is defeasible. Here, I try to set these properties in the context
of various di¨erent kinds of nondeductive reasoning that have been
explored. The underlying purpose is to show that there are a number of
proposed kinds of reasoning to which the inference of GCIs might be
usefully assimilated.
A reasoning system is said to be defeasible (or when instantiated in an
argument nonmonotonic) if an inference or argument in that system may
be defeated by the addition of further premises.30 Deductive systems, of
course, are monotonic or nondefeasible. It seems most unlikely that
implicatures are derived as deductive inferences (contrary to Sperber and
Wilson 1986) because implicatures are clearly defeasible:31
(22) a. Assertion: ``John ate some of the cookies.''
b. Default implicature: `John did not eat all of the cookies.'
c. Cancellation of b: ``John ate some of the cookies. In fact he ate
all of them.''
Indeed, it is quite possible, as Johnson-Laird (1983) has argued, that true
deduction plays little part in informal human reasoning.
Our understanding of the range of alternative inference methods is still
too limited to permit a proper typology. Nevertheless, a tripartite dis-
tinction between deduction, induction, and abduction is a useful starting
place (Josephson and Josephson 1994: 27±29). A crude but e¨ective way
to approach these three modes of reasoning is through a comparison of
their central syllogisms (Pople 1973):
43 On the Notion of a Generalized Conversational Implicature

(23) Deduction
A(x)(P(x) ! Q(x)) (major premise)
P(a) (minor premise)
D
Q(a) (conclusion)
(24) Induction
P(a) (observed fact)
Q(a) (observed fact)
I
A(x) (P(x) ! Q(x)) (induced law)
(25) Abduction
A(x)(P(x) ! Q(x)) (known law)
Q(a) (observed fact)
A
P(a) (hypothesized explanation)
The idea behind this formulation is, of course, that whereas deduction
gives one conclusions by instantiation in a general law, induction enables
one to conclude a general law from multiple observations of singular
facts, and abduction is the hypothesizing of a fact that would, in con-
junction with a general law, explain an observed fact. Both induction and
abduction are clearly nonmonotonic; in the former case, the discovery of
a fact (e.g., `Q(b) & @P(b)') not in accord with the hypothesized law is
su½cient to annul the law; in the case of the latter, the conclusion P(a)
will be defeated if we already know that @P(a).
To these three contrasting syllogisms we may add the following two,
which characterize two further kinds of nondeductive reasoning that are
currently being explored:
(26) Practical reasoning (Atlas and Levinson 1973 after Kenny 1966)
A(x) (P(x) ! Q(x)) (known law)
Goal(Q(a)) (desire)
KL
Goal(P(a)) (consequent desire)
(27) Default ``logics'' (see Ginsberg 1987)
A(x) ([P(x) & M(Q(x))] ! Q(x)) (conditionally assume Q(x))
P(a) (known fact)
DL
Q(a) (assumed consistent fact)
Of these four kinds of defeasible reasoning, inductive reasoning has
been the focus of by far the most thought; indeed there has been a lively
debate since at least the time of Francis Bacon, and I make no attempt
44 Chapter 1

here to survey the various logical systems that have been proposed to
capture notions like degree of inductive support (see, e.g., Cohen 1971;
Holland, Holyok, Nisbett, and Thagard 1989).32 However, the other
kinds of system are less familiar, and I should brie¯y sketch the motiva-
tions behind their development.
Abduction, like practical reasoning, has a striking property illustrated
in syllogism (25) aboveÐit is an application of what in deduction would
be the fallacy of a½rming the consequent. This captures the essence of the
informed guess, the presumption that if a fact ®ts with a known law, it is
an instantiation of it. It may be objected here that this is a poor charac-
terization of abduction, at least as originally conceived by PeirceÐfor
even though he provides the simple syllogistic account in (25) above,33 he
seems to have intended that the term ``abduction'' should refer to the
creative hypothesis of the major premise or scienti®c law itself and not the
hypothesis of a mere fact that would make applicable a preexisting law.
However, we can trivially manipulate the abductive syllogism to make the
major premise the conclusion:34
(28) (p & (p ! q)) ! q (logical truth)
q (observation)
A
p & (p ! q) (abduced law)
Clearly, what is fallacious in deductive reasoning (a½rming the con-
sequent) may be legitimate in hypothesis formation only under strict
constraints on the antecedents one may hypothesize; otherwise any ante-
cedent whatsoever may be posited. The strict constraints must capture, in
some way, our sense of the best available explanation for the factuality of
the consequent. At the same time, these constraints must suggest some
heuristics that would allow the selection of reasonable hypotheses, given
that, in general, there will be inde®nitely many antecedents that may be
hypothesized. The best explanation or hypothesis should be that one, it
has been suggested (Thagard 1978), that gives us maximal predictions for
minimal assumptions. In short, hypothesis formation is conservative as
regards the modi®cation of existing theory or assumptions but expan-
sionist as regards the data that may be covered.
Theories of ``inference to the best explanation'' in philosophy of science
may seem far a®eld from theories of utterance interpretation; utterance
interpretation is not leisurely, ratiocinative, and globally ambitious like
scienti®c theory. But some of the same processes may well be involved
nevertheless (it would be odd if our scienti®c ability had no origins in
45 On the Notion of a Generalized Conversational Implicature

everyday thinking), and transitional ®elds like medical diagnosis or diag-

nosis of mechanical failure make this plausible. AI work in diagnostic
systems requires some kind of abductive inference and there has been
much experimentation here.35 But more directly pertinent is the abductive
approach to implicature being developed by Hobbs and co-workers,
which I review in the next section.
Practical-reasoning systems, closely allied to AI planning systems, aim
to capture the reasoning from ends to means that will achieve those ends,
or, if one likes, from goal to plan. They are also clearly nonmonotonic;
for example, instantiating in (26), given that if one takes the 10:00 train
from Cambridge one arrives in London at 11:15, and given that one
wants to be in London at 11:15, then one should want to catch the 10:00
train. But suppose that one hates traveling by train; well, then, the plan is
not adopted, at least if there's some alternative. Developed ®rst in philo-
sophical theories of action (see, e.g., Kenny 1966; von Wright 1971) and
applied speci®cally to implicature by Atlas and Levinson (1973; see also
the sketch in Brown and Levinson 1987: 64±65, 87±90), such modes of
reasoning have since been explored in arti®cial intelligence (see, e.g.,
Allen 1983; Charniak and McDermott 1987: chap. 9).
Default logics aim to capture a rather di¨erent mode of reasoningÐ
namely, the notion of a reasonable presumption, a ceteris paribus assump-
tion. There are a great many di¨erent formalisms under investigation (see
Ginsberg 1987), but the kind of inference that these are designed to
capture can again best be indicated by instantiation in the syllogism in
(28). Suppose we are told ``The noise of the gun frightened o¨ the birds.''
It is a natural presumption that the birds ¯ew away; but of course the
presumption is unwarrantedÐthere are birds like penguins that do not ¯y
at all, and others like quail that may run o¨ along the ground or swans
that may swim away. The inference from `x is a bird' to `x can/will ¯y' is
an inference we seem to make unless there is some inconsistent assump-
tion already made (e.g., the birds in question are penguins). So we can set
up a syllogism that says, in e¨ect, if x is a bird and x's ¯ying is consistent
with what is known, presume x ¯ies. The inference will be defeasible given
any fact that is inconsistent with it.

1.5.2 Nonmonotonic Inference and Implicature

Reviewing these forms of nonmonotonic inference, it is clear that each of
them gives, by design, an account of defeasibility, the essential property
for modeling inferences like implicatures. Induction alone, however, does
46 Chapter 1

not recommend itself as the basis for a theory of implicature just because
implicatures in general, but GCIs in particular, seem to be presumed ab-
solutely, not with such-and-such degree of inductive support; nor in gen-
eral do we seem to think that a speaker meant p by U with probability n.
Abduction is more attractive because it o¨ers a circumscribed presump-
tionÐbut attempts to construct an abductive system for implicature pre-
sumption require an elaborate weighting of inference rules, and even a
weighting of individual facts, in order to constrain guesswork by making
certain presumptions more costly than others (see Hobbs and Martin
1987, Hobbs 1987, Hobbs et al. 1993).
Practical reasoning systems, on the other hand, have a certain attrac-
tion for the modeling of implicature on ®rst principles: if speaker-meaning
is (as on the classic Gricean account) a matter of having a complex plan
or intention, then understanding (and thus implicature-recovery) must
be largely a matter of plan reconstruction (see, e.g., Cohen, Morgan, and
Pollack 1990). A number of attempts have thus been made to model
implicature in these terms (Atlas and Levinson 1973; Brown and Levinson
1978; McCa¨erty 1987; Thomason 1987, 1990; Welker 1994). But there is
a striking problem that has never been properly addressedÐnamely, the
fact that planning systems (like all logical systems) are asymmetrical, in
the sense that plan recognition is not simply plan construction run back-
wards (Levinson 1995b). Just as a conclusion p might be reached from an
in®nite set of distinct premises (e.g., p & q; or q & (q ! p); or ( p4q) &
@q) so inde®nitely many plans might converge on a single utterance.
Thus even if we had an account of how a speaker plans an utterance
complete with its implicatures (and see Levelt 1989 for a glimpse of the
complexities), this would not directly give us an account of how the
implicatures were recovered in comprehension.
Finally, only default logics clearly have the two properties that gener-
alized implicatures (GCIs) clearly exhibit: defeasibility and default or
preferred presumption. On the face of it, though, they may seem to have
little to o¨er to a general theory of implicature because they provide a
restricted set of limited inference rules, incapable of modeling the open-
ended, creative, and inde®nite set of inferences that come under Grice's
theory. However, as an account of the inference of GCIs, just those
implicatures that seem to have the default property, they may indeed
have something to o¨er. There are a great many rather di¨erent models of
default reasoning (Ginsberg 1987), but I will sketch (simplistically) how
47 On the Notion of a Generalized Conversational Implicature

such an application might work by taking the simplest of such non-

monotonic systems, the default-rule systems originating with Reiter (1980).
We can conceive of default rules, on this interpretation, as extra-logical
rules of inference that can be used to expand the data base (the set of
available premises). Such rules have the canonical form (where a and b
are sentences, and M b means `b is consistent with what is known'):
(29) a: Mb
b
which is read `if a is true, and b is consistent with the data-base (what
is known), then assume b'. For example, we might represent our earlier
default assumption as:
(30) bird(a): M ¯ies(a)
¯ies(a)
Thus one might assume that if a is a bird, then a ¯ies, unless one has
speci®c knowledge to the contrary. Turning now to GCIs, we might treat
the scalar implicature from the use of the weaker existential quanti®er in a
precisely similar way (where a is a simplex arbitrary sentence frame):
(31) a (some): M (a(not all))
a (not all)
Then an utterance like ``Some of them came'' will implicate `Not all of
them came' wherever this further assumption is consistent with the data
base. If we interpret the data base as the common ground (i.e., what is
publicly taken for granted in the conversation so far), then a rule of this
sort serves to augment the common ground in just the sort of way gener-
ally envisaged in pragmatic theory. Of course, the rule as stated loses all
the generality that is the signal virtue of the pragmatic analysisÐwe want
the inference to follow just from the fact that some is a weaker element in
a Horn scale of the schematic kind hSTRONG, WEAKi. We can capture
that generality, perhaps, by setting up a meta-rule of the sort in (32),
which will apply to any simplex sentence frame a that contains a weak
item paired with a strong item in a predetermined set of scales.
(32) a (WEAK): M (a(not STRONG))
a (not STRONG)
48 Chapter 1

A more complex alternative is to treat the default rule as part of a modal

logical system, rather than as an extralogical way of augmenting the
common ground. Thus in Moore's (1985) autoepistemic logic, we can
recast our paradigm scalar inference as a modal inference:
(33) [B(a(some)) & @B@ (a(not all))] ! a(not all)
which is read `if it is believed that a (some), and it is not believed that it is
not the case that a (not all), then a (not all)'. The operator here is just the
strong L operator in a standard modal logic (weak S5), so the properties
of such systems are well known. The intuition behind these systems is
that of a rational agent inspecting his own beliefs and making the strong
assumption that what he does not know is not the case (Levesque 1990).
Wainer (1991) explores in detail the application of these ideas directly to
the modeling of GCIs. He notes that the inference from `John is in the
attic or in the garden' to `the speaker does not know which' (i.e., Gazdar's
(1979) clausal implicatures) falls out as a consequence in such a system.
But unfortunately the system also over-generates, predicting many pseudo-
implicatures too.
Wainer (1991) also considers, for the purposes of modeling GCIs, an-
other class of nonmonotonic reasoning systems based on circumscription
(McCarthy 1980). Circumscription involves restricting the entities that
satisfy predicates, or restricting entities in a domain to those known to
exist. McCarthy (1986) notes that communication conventions rely on such
circumscription, on the assumption that things are normal unless other-
wise stated. This suggests that inferences due to the I heuristic (Grice's
Q2) may be quite generally modeled in such systems. Wainer (1991: chap.
4) also points out that scalar GCIs might be directly generated in such
systems: if one says ``Two boys came,'' then it can be assumed that no
more than two came. He notes, however, a number of problems, and in
the end he outlines a model in which GCIs are generated with the seman-
tic structure, and then such nonmonotonic reasoning systems are used to
model defeasibility rather precisely under complex conditions (Wainer
1991: chap. 5).
There are a number of di½culties in the direct application of such sys-
tems to the modeling of GCIs: apart from the need to capture the gen-
eralizations inherent in GCI theory, we also need to handle many such
rules that may potentially yield inconsistent inferences (although again see
Wainer 1991: 175). Responding to earlier suggestions of mine, Carston
(1995) draws attention to further di½culties: I implicatures, for instance,
49 On the Notion of a Generalized Conversational Implicature

are often interpretations to stereotypical extensions and thus have a

piecemeal character not easily captured by default rules (although better
handled by circumscription). As well, there are patterns of defeasibility
that do not seem to match this format, which I detail immediately below.
Although anyone interested in formalizing GCIs should clearly investi-
gate the potentials of this whole family of nonmonotonic logics, there are
reasons to think that perhaps we should invert the argument: rather than
seek understanding of implicature in theories of default inference, perhaps
we should instead think of generalized implicatures as the prototype of
default inferences. Meanwhile, we can look carefully at the core phe-
nomenon itselfÐthe defeasibility of implicaturesÐto discern the exact
properties that such a formal theory would have to cope with.

1.5.3 Investigating the Defeasibility of Scalar Implicatures

The mechanism I call Gazdar's bucket gives a good ®rst approximation to
the conditions under which GCIs evaporate or are defeated. Gazdar
(1979), following Hamblin (1971), Stalnaker (1972), and others, proposed
that the communicative content of an utterance should be considered an
ordered n-tuple of entailments, implicatures of various kinds, presuppo-
sitions, and so forth. The process of updating the common ground with
the communicative content of a new utterance can then be treated as the
ordered incrementation of that background with the entailments, impli-
catures, and presuppositions of the utterance.36 Such an incremental
model of dialogue has a number of virtues. For example, it allows us to
capture the way in which utterances can presume background assump-
tions and the way in which their communicative force consists in changes
to the context. Thus a permission both presumes a background prohibi-
tion and changes that background by removing the prohibition. Presup-
position, illocutionary force, and anaphora all have natural accounts in
this kind of model, now widely presumed. But an additional interest is
how the incremental model might be used to handle the defeasibility of
pragmatic inferences.
Informally and metaphorically, we can think of the common ground as
a bucket, holding all the facts mutually assumed, either because they're
common knowledge or because they have been asserted and accepted.
Then, assuming that implicature generation is accounted for by other
means, we can capture many aspects of the defeasibility of implicatures
(GCIs in particular) in the following way: a new assertion will have its
50 Chapter 1

content chucked in the bucket (i.e., context) strictly in the following order
and only if each incrementation is consistent with the contents of the bucket:
(34) Order of incremented information
a. Entailments
b. Q GCIs
i. Clausal
ii. Scalar
c. M GCIs
d. I GCIs
This accurately captures the following kinds of defeasibility (where ‡>*
indicates canceled or nonarising implicature). First, there are cases where
GCIs are e¨ectively canceled by inconsistent background assumptions:
(35) A: ``A Saudi Prince has just bought Harrods.''
B: ``Some Saudi princes must be pretty wealthy.''
‡>* `not all Saudi Princes are pretty wealthy.'
Thus B's assertion would by default give rise to the inference that not all
Saudi Princes are wealthy, but in a situation where the assumption is that
all Saudi Princes are wealthy, the inference will simple disappear. This
behavior is captured because one can only add implicatures to the bucket
if they are consistent with what is already assumed.
Second, there are cases where an utterance has an entailment incon-
sistent with an implicature, in which case the implicature is invariably
canceled:
(36) ``Some Saudi Princes, and in fact all of them, are pretty wealthy.''
Entailment: `All Saudi Princes are pretty wealthy.'
Scalar implicature: `Not all Saudi Princes are pretty wealthy.'
This e¨ect is captured by ensuring that entailments are added to the
bucket before implicatures, so that when one comes to add an inconsistent
implicature the consistency requirement prevents the addition.
Third, there are cases where there are inconsistent implicatures:
(37) ``Some, if not all, Saudi Princes are wealthy.''
Scalar implicature: `Not all Saudi Princes are wealthy.'
Clausal implicature (due to conditional): `Possibly all Saudi Princes
are wealthy.'
Here, an implicature of the consequent of a conditional is explicitly men-
tioned in the antecedent, and under such circumstances ( just like pre-
51 On the Notion of a Generalized Conversational Implicature

suppositions, see Levinson 1983: 224), the implicature is regularly ®ltered

or suspended. The context-increment model explains this in terms of
the priority of di¨erent kinds of implicatures: a clausal GCI of epistemic
uncertainty arising from conditionals has a higher priority than a scalar
implicature from the consequent. Because clausal implicatures are thrown
in the bucket before scalar ones, when we come to add the scalar infer-
ence, we will ®nd the addition prevented by the consistency rule, correctly
predicting the suspension of the scalar implicature only.
Thus the simple device of an ordered incrementation together with a
consistency check su½ces to account for many aspects of the defeasibility
of implicatures. And such a consistency ®lter is consonant with the pos-
sibility of modeling implicature generation with default rules, as sketched
in section 1.5.2. However, there are di½culties. The least of these, per-
haps, is that implicatures can in fact be canceled at more or less arbitrary
distance as in:
(38) ``Some Saudi Princes are pretty wealthy. . . . Indeed, there is
indubitable evidence that they are all billionaires.''
Implicature of ®rst sentence: `Not all Saudi Princes are pretty
wealthy.'
Entailment of second sentence: `All Saudi Princes are wealthy
(billionaires).'
This would require either inde®nitely postponing the incrementation pro-
cess, or tagging the implicatures and modifying the consistency ®lter so that
later entailments could cancel earlier implicatures inconsistent with them.37
Much more interesting though are cases of implicature cancellation
that are not triggered by an inconsistency at all. Here is a simple example:
(39) A: ``Is there any evidence against them?''
B: ``Some of their identity documents are forgeries.''
Predicted (but nonoccurring) scalar implicature: `Not all of their
identity documents are forgeries.'
Assuming no special knowledge on the part of the interlocutors, there
would be no inconsistency with the common ground or the entailments of
B's utterance if not all of the documents are forgeries. It seems, though,
that we do not let the inference through. That's because, intuitively, A is
only interested in whether there is at least some evidence against the
criminals; given A's question, all that is relevant is the possession of at
least some evidence.
52 Chapter 1

Perhaps the following example is clearer. I argue in chapter 2 that the

numerals form a Horn scale in the normal way, so that saying ``I have $9''
will by default implicate `I do not have $10'. Then in the following
exchange (which we may imagine occurring at the entrance to a cinema),
the default inference from B's utterance seems to evaporate just as in the
previous example:
(40) A: ``It's going to cost us ten dollars to get in and I didn't bring a
cent.''
B: ``Don't worry, I've got ten dollars.''
Predicted (but evaporated) scalar implicature: `I have no more
than ten dollars.'
Again, the inference would be (let us suppose) consistent with the com-
mon ground (pockets being private) and with all higher ordered infer-
ences from the utterance itself, yet the implicature is canceled. And again,
the explanation would seem to be that satisfying the very speci®c goal
here (i.e., knowing that we can get into the cinema) can be presumed to
exhaust the communicational load of the utterance.
In these cases, we can think of the situation as setting up a restricted
goal: the question investigates a potential obstacle to the goal, and the
reply indicates that the obstacle is not actual (see Allen 1983). I attribute
this business of plan recognition and the inferences that follow from it to
a maxim of Relevance, construed much as Grice intended and not at all
as Sperber and Wilson (1986) reconstruct (for similar presumptions, see
Holdcroft 1987, McCa¨erty 1987, and Thomason 1990). It seems then
that Relevance implicatures, or inferences about speaker's goals, can limit
the amount of further inference that is warranted. Thus even where these
further inferences are entirely consistent with all that is known, they do
not go through.
These sorts of examples su½ce to show that consistency with an incre-
mental common ground is not the only constraint on the addition of
default implicatures to what is taken for granted. This may cast some
doubt on whether default logics are the right formalism for modeling
GCIs, or it might show that they simply need to be bounded by a yet to be
described maxim of Relevance. But because the inferences are defeated by
recognition of a goal, and because that kind of defeasibility is precisely
what practical reasoning systems are set up to capture, perhaps these
examples are indications that the plan-generation and plan-recognition
types of nonmonotonic inference are what are really involved in implica-
53 On the Notion of a Generalized Conversational Implicature

ture generation. Plan generation and plan recognition are, after all, what
must be involved if meaning in general has the character of Grice's
meaningnn . Clearly, though, much closer work is required on the exact
nature of conditions under which inferences are abandoned. For these are
the facts that are likely to adjudicate between the di¨erent potential
models of defeasible reasoning.
Two ®nal speculations. First, one may wonder whether the intuitions
that lie behind the attempts to create default logics are not in fact direct
intuitions about how humans reason, but rather are primarily intuitions
about communicational inferencesÐin fact, intuitions about the very
default pragmatic inferences we are concerned with. GCIs would indeed
seem to be the prototype default inferences; the root intuition is not, one
may argue, that thinking of birds one assumes ceteris paribus that they ¯y.
It is rather that, anyone saying ``I scared away the birds'' would have
misled his audience if he did not intend to suggest (by I-inference to the
stereotype) that the birds ¯ew away. This suggestion is due of course in
part to the way the world is (most kinds of birds ¯y); but that would only
warrant a probabilistic inference. The default nature of the inference, to
the surety that the birds ¯ew away, relies on the presumption that other-
wise we would have been told! That is, given a mutual presumption about
how we use stereotypes to augment the communicational load of utter-
ances (recollect the I-heuristic: `what is stereotypical may be presumed'),
the e¨ect will be to convert Bayesian probabilities or hazardous assump-
tions into the certainties of communicational presumptions.
There is some empirical support for the rather extraordinary e¨ect that
the mutual assumption of a common heuristic can have on the proba-
bilities of co-ordinated action. As mentioned above, Schelling (1960)
showed, in some well known informal experiments, that if, for example,
two people are asked to guess the same number without communicating
to one another, they can do so with much greater reliability than
chanceÐthey search for a salient number that each thinks the other will
also ®nd the most salient (e.g., 1 or 100). Or, if two persons are lost in a
department store they can guess where each would guess the other would
go, so ®xing on a single meeting place. If we think of communication as a
coordination problem of the same sort (the speaker ®nding the signal that
the recipient will interpret just as the speaker intended), a heuristic of the
sort I have been describing would play the same sort of roleÐnarrowing
down the solutions to the coordination problem. Thus a mere probability
that the speaker meantnn such and such can become converted, against
54 Chapter 1

tremendous odds, into a near certainty. And arguably this is just what
default logics model: presumptions about interpretation based on the fact
that an interpreter may assume a certain heuristic (e.g., the inference to
the stereotype) and thus a particular interpretation (e.g., restricting the
extension of birds to `¯ying birds') on the basis that the speaker will have
coded the message assuming the operation of the very same heuristic. Or
to put it another way: when I think ``the dog chased away the birds,'' I
will remember the particularities of the event (whether the quail ran or
¯ew away); it is only when I say it that default presumptions come into
action (on the mismatch between thought and speech, see Levinson
1997).38
And this brings us to the second and ®nal speculation: perhaps this
account of pragmatic inference as based on stable heuristics is all we need
to give an account of this species of defeasible inference. There is no need
for a nonmonotonic logic to model GCIs, it might be claimed, because
GCIs are simply instantiations of communicational heuristics and indeed
the prototype default inferences. Just as Johnson-Laird (1983) argues that
we do not need formal deduction to model deductive properties of human
reasoning, relying instead on the reasoning based on instantiated mental
models, so perhaps we do not need formal models of nonmonotonic
argument to capture the processes of default interpretation. Some phi-
losophers think that much of the e¨ort expended on defeasible logics or
inference systems is futile, stemming from a confusion between reasoning
and rule-based inferenceÐreasoning is always subject to revision, but
deduction as least is strictly cumulative (see Harman 1986). If this cate-
gory mistake has been made, it is because in the context of strong com-
municational assumptionsÐof tacit coordination and m-intentionsÐ
default inferences come to have the normative certainty that we associate
with rule-based inference systems.39

1.6 AGAINST REDUCTION OF GCIs TO NONCE SPEAKER-MEANING

We have mentioned the current tendency to reduce utterance-meaning to

speaker-meaning, or, in other words, utterance-type meaning to utterance-
token meaning.40 Thus a theory of communication is left, on these
accounts, with only the level of sentence-meaning on the one hand and
speaker-meaning on the other, with no provision being made for the
notions of conventions of use, idiomatic use, partially conventionalized
meaning, preferred interpretation, default interpretation, and so on. To
55 On the Notion of a Generalized Conversational Implicature

theorists of this persuasion, as mentioned, the theory of GCIs is a sort of

category mistake, treating pragmatic principles of inference quite inap-
propriately as ``code-like rules'' able to account for only ``quite untypical
examples of implicature'' (Sperber and Wilson 1986: 37).41 Instead, such
theorists attempt to reduce all kinds of implicature to what we may call
nonce inferences, one-o¨ inferences to a speci®c speaker-meaning given
the full rich body of assumptions held by the participants on that partic-
ular occasion.
The main argument against such reductionism in the case of GCIs is
simply that there is no way to preserve the many simple but powerful
generalizations that hold at the level of utterance-types. Thus the body of
this book attempts to justify the notion of a GCI by ostension as it were in
chapter 2, and then by showing how generalizations only available to a
theory of utterance-types can contribute signi®cantly to other areas of
linguistic theory (chap. 3 and 4). However, here we pause to argue directly
but brie¯y against two rival accounts that would reduce GCIs to nonce
inferences.

1.6.1 Sperber-Wilson Relevance

The most prominent of the reductionist views is the theory of Relevance
promoted by Sperber and Wilson (1986). This is not the place to review
that theory (see Levinson 1989 for my own response and the general peer
commentary in Behavioral & Brain Sciences 10:4, 1987), but it su½ces to
say that the theory holds that implicature is a side e¨ect, as it were, of a
mental automatism, a tendency to extract the maximal inferences for the
minimal psychic e¨ort.42 There are no maxims, no heuristics, no special
modes of reasoning involved in implicature derivation. Thus the fact
that the GCI phenomena do not seem properly attributed to a maxim of
Relevance is, on this view, beside the point; Sperber-Wilson Relevance
(henceforth SW-R) is intended to replace the whole Gricean apparatus.
Nor is it pertinent, apparently, that SW-R appears to be a very inade-
quate characterization of what pretheoretically would generally be con-
sidered the nature of relevance (as already mentioned, this seems best
captured in terms of the local goals of participants at a particular locus
in a conversation). SW-R is in e¨ect a proportional measure balancing
informational richness (corresponding, perhaps, approximately to my I
principle or Grice's Q2) with processing e¨ort (capturing, arguably, some
aspects of my M). And the central claim is that low-cost inferences
(deductive in character) will be invariably preferred.
56 Chapter 1

SW-R seems to have overwhelming disadvantages as an account of

implicature in general, but it faces even worse problems as an account of
GCIs in particular. I raise two sample problems here.

1.6.1.1 SW-R Employs Deduction, but GCIs (and PCIs) are Non-
monotonic One problem, salient in the context of the last section, is
that according to SW-R all inference involved in implicature derivation
is deductive, hence the inferences must be monotonic. But I have shown
that they are nonmonotonic in character. Sperber and Wilson have not
seriously addressed this problem, but presumably they would claim that
the nonmonotonicity is illusory along the following lines:
1. where implicatures are generated, the deductive premises are part of
the context;
2. where implicatures appear implicitly ``canceled,'' the necessary prem-
ises are missing, so no inference in fact arises;
3. where implicatures are explicitly ``canceled,'' there is an implicit
adjustment of enthymematic (unexpressed) premises.
It is this last point that seems especially insuperable for such an account.
Consider, for example, the following:43
(41) a. John has two children, if not more.
b. John has two children and perhaps more.
c. John has two children and in fact, now I come to think of it,
maybe more.
d. John has two children; he had two more by his ®rst wife but they
are grown up now.
e. John has two children. Actually Sue says he has three, but I've
never seen the third so I doubt he or she exists.
Intuitively, in these cases, the speaker asserts that John has two chil-
dren, thereby suggesting he has no more; the additional material then
suspends or cancels that suggestion. Without the suggestion, the addi-
tional material would be unmotivatedÐit is added precisely to suspend or
cancel the suggestion. So the implicature `at most two' must have arisen
and thus on this account must have been deduced in the context (using,
e.g., some premise of the sort; `the speaker can be expected to know the
exact cardinality of John's children', `the exact cardinality is of interest to
the recipient'). But if it is deduced, it can only be ``canceled'' by the dele-
tion of a premise and the construction of a new deduction based on new
57 On the Notion of a Generalized Conversational Implicature

premises (a context lacking, e.g., the premise `the speaker can be expected
to know the exact cardinality of John's children'). Thus the speaker has
produced an utterance yielding inconsistent deductionsÐa contradiction
at the level of utterance-meaning; from which every proposition under the
sun will follow. Given that cancellation phrases may arise at arbitrary
distance from the phrases that gave rise to the original implicature (as,
e.g., in (41e)), this account would suppose that we are forced constantly to
make sense of self-contradictory speakers.
Further, it won't account for the important (and long-noted)44 asym-
metry between the nonmonotonic behaviour of implicatures (allowing
cancellation) and the monotonic nature of entailments (which do not
allow cancellation):45
(42) a. John has two children and, in fact, a total of three.
b. *John has two children and, in fact, a total of one.
(43) a. John has two children. Actually, Sue says he has three, but if
so I've never seen the third.
b. ??John has two children. Actually, Sue says he has only one, and
I've never seen the second.
Although all implicatures (arguably by de®nition) exhibit such non-
monotonic properties, GCIs pose a special problem for any would-be
deductive account because they can so plainly arise as a preferred inter-
pretation only to be canceled or annulled later at some distance without
any sense of self-contradiction by the speaker.

1.6.1.2 SW-R, if it Makes any Clear Predictions, Probably Makes the

Wrong Ones The second problem with an SW-R account of GCIs that I
address is its inability to yield empirical predictions, at least of the right
sort. Elsewhere (Levinson 1987c, 1989) I have suggested that SW-R is
incapable of making clear predictions, partly because the theory is not
clearly articulated but partly because the factor of cognitive e¨ort, an
essential ingredient in the proportional measurement of Relevance, is not
empirically measurable (or at least not empirically measured). So if the
theory is to be of any use, we must have some rough and ready concept
of a low-cost inference such that it would be just the sort of inference
favored by SW-R. Following the examples used in Sperber and Wilson
(1986), I was led to think that the schema in (44) would constitute just the
right kind of example.
58 Chapter 1

(44) A: ``If p then q''

B: ``p''
Implicature: q
where q constitutes the low-cost deduction typical of an interpretation
favored by SW-R. I therefore proposed (1987c) that (45) might be a
counterexample to SW-R and an example that favored GCI theory.
(45) A: ``If the spy had possibly more than two passports, then he may
yet escape.''
B: ``He had two passports.''
SW-R should favor a reading of B's utterance that will make available the
low-cost deduction of the consequent of A's utterance. B has said some-
thing that is consistent with the antecedent of A's utterance, and indeed
on a rough gloss of the semantic content of the word two as `at least two,
not excluding more', B has said something almost equivalent. Thus, A's
antecedent should be considered met (any minor inferences required being
made), and the consequent should be deduced as:
(46) SW-R predicted interpretation (my guess)
A: `If the spy had possibly more than two passports, he may yet
escape.'
B: `He had at least two and possibly more passports.'
Deductive implication (alias R-implicature): `He may yet escape.'
Now contrast the account according to GCI theory, wherein the
utterance-types are su½cient to engender preferred interpretations. Two
pertinent GCIs here are the I inference from the use of the conditional in
A's utterance to the biconditional interpretation (i.e., conditional perfec-
tion), and of course the scalar Q inference from two to `at most two' in
B's utterance. This gives the following interpretations and a di¨erent
deductive implication from the implicaturally-strengthened premises (in
GCI theory deductive implications are just that, deductive implications, not
implicatures):
(47) GCI predicted interpretations
A: `The spy may yet escape if and only if he may have had more
than two passports.'
B: `He had only two passports.'
Deductive implication: `It's not the case that the spy may yet
escape.'
59 On the Notion of a Generalized Conversational Implicature

SW-R would not favor such an interpretation, because in order to get the
same yield (one deductive implication) we have had to make two gratu-
itous further inferences (the GCIs).
Intuitively, it seems to me that GCI theory makes the right predic-
tionÐthat is, that we tend to read the exchange as suggesting that the spy
will not escape. But, SW-R (on my construction of it) predicts the con-
trary reading, that the spy may yet escape. Thus I concluded (Levinson
1987c) that GCI theory makes the right predictions just where SW-R
makes the wrong ones and, more generally, that a theory of preferred
interpretations can never be superseded by a theory of nonce inference.
However, in their response, Sperber and Wilson (1987: 748) contend
that their theory makes no such predictions and speci®cally it ``does not
predict that the most relevant interpretation conceivable is the right one.''
Rather, in this case, everything will depend on the details of the contex-
tual assumptions and speci®cally on whether the interlocutors take B to
be an authority on the cardinality of the spy's passports. Thus, without
such further details, the authors claim that SW-R simply makes no pre-
dictions whatever. But this response would hardly seem to help the theory
wriggle out of examples of this kind, presented here as interlocutors often
experience them, without mutually assumed background assumptions of
the sort SW-R calls for. There is, intuitively, a clear preference for the
reading predicted by GCI theory, and it is a preference that GCI theory
holds would be expected in the absence of any further speci®cities about
the context.
Thus, either SW-R seems to make the wrong predictions about the
phenomena generally considered GCIs (my claim), or it makes no general
predictions about them (Sperber and Wilson's claim), and both alter-
natives seem equally damaging. No doubt there is plenty of room for
maneuver in the application of a theory as general as SW-R to speci®c
examples; and perhaps I continue (as alleged) to misunderstand SW-R.
But certainly the two theories ought to make rather di¨erent predictions
in a whole range of cases, and as far as I can see, GCI theory is bound to
make the right sort and SW-R the wrong sort as far as the inferences
central to GCI theory are concerned. Regardless of the success or failure
of present versions of GCI theory, these rather complex examples only
serve to reinforce a simple enough, indeed almost self-evident, point:
expression-types do tend to have preferred interpretations (or more
exactly, there are default restrictions on their extensions), interpretations
that tend to go through in the absence of information to the contrary.
60 Chapter 1

1.6.2 Implicature as Accommodation

There is another class of attempts to reduce implicature wholesale
(including GCIs) to matters of nonce inference. These theories have in
common the use of Lewis' (1979) concept of accommodation, which had
earlier been used to account for certain properties of presupposition
(Heim 1982: 370±384). Let us consider two rather di¨erent theories
(Thomason 1987, Hobbs 1987) that claim that the concept of accommo-
dation coupled with a nonmonotonic inference system may be all we need
for a theory of implicature. Thomason (1987) suggests we need accom-
modation plus a plan-recognition inference system,46 whereas Hobbs
(1987) suggests accommodation plus his abduction system (a weighted
backward-chaining, based on a constrained use of the logical fallacy of
a½rming the consequent). The basic idea behind accommodation is that
there are reasonably well-de®ned felicity conditions or usage constraints
on the use of particular expressionsÐfor example, the X is felicitously
used (it might be claimed) only where the common ground provides a
uniquely salient X. Against this background of expected usage, deviant
uses can then be accommodated by a cooperative interlocutor. Suppose I
say ``My car broke down. The diesel froze,'' in a context in which a
unique mass of diesel has not been previously introduced into the dis-
course; well, it will take no great imagination to construct the assumption
that the speaker's car runs on diesel and that the mass in question is the
diesel in the fuel tank of that car.
There are thus two essential ingredients in the proposed mechanism
whereby utterance U implicates proposition p:
1. U must contain a trigger T, some expression whose felicity conditions
require that a proposition of the class P must be an element of the
common ground if T is to be felicitously used;
2. if U is uttered and there is no proposition of the class P presumed in
the context, then the recipient should accommodate T by inferring p
such that p A P, and by adding p to the common ground.
The essential di¨erence between Hobbs's account and Thomason's comes
down, then, as mentioned, to the mode of inference envisaged (abduction
or practical reasoning).47
Clearly, these accounts work only where there are appropriate trigger
expressions, expressions that require certain contextual assumptions that
have not already been met. Thus they give reasonably good accounts of
61 On the Notion of a Generalized Conversational Implicature

presuppositional inferences; Thomason concentrates on factive predicates

(e.g., know) and Hobbs on the bridging inferences often required by de®-
nite referring expressions.48 Both theorists claim that their accounts gen-
eralize to implicatures and indeed generalize right across the di¨erent
kinds of implicature. But Thomason provides little reason to believe this,
and Hobbs's extensions are interesting but remain within the class of
presupposition-like triggers (e.g., he extends the account to noun-noun
compounds like oil alarm, which require the inference of some unique
relation between the extensions of the two nouns).
Without conceding that the account is persuasive even in these cases,
we can nevertheless easily show that it o¨ers little hope of reducing GCIs.
The essential reason for this is that GCIs are not inferences that have
to be made in order to maintain some particular felicity condition; the
expressions that give rise to GCIs do not require any inferenceÐotherwise
cancellation would not be possible. Rather, GCIs are inferences that are,
from the point of view of accommodation theory, essentially gratuitous.
Another way of putting it is that accommodation theory is an account of
how rule-breaking behavior is brought back into line with expectations.
For example, I use a presuppositional trigger where no such presupposi-
tion is actually presumed; my interlocutor accommodates the usage by
accepting the presumption. But in the case of the use of a GCI-inducing
expression like scalar some, no rule has been broken; yet the inference to
`not all' is gratuitously made. And the very nature of GCI defeasibilityÐ
namely, cancellation in the face of inconsistent or unsupportive assump-
tionsÐmakes it impossible for a GCI-inducing expression to force a pre-
sumption into the common ground.
Just to make this concrete, compare the kinds of case favorable to an
accommodation account, like (48a), to the classic GCIs like scalar impli-
catures in (48b).
(48) a. ``Sue walked into the room. The chandelier was magni®cent.''
‡> `The room had a chandelier.'
b. ``Sue has two children.''
Q‡> `Sue has at most two children.'
c. ?``Sue walked into the room. The chandelier was magni®cent. It
was in the cupboard.''
d. ``Sue has two children. She has a third by her former husband.''
Note too the ease with which the scalar implicature can be cancelled in
(48d), compared to the bridging inference in (48c). The problem in (48c) is
62 Chapter 1

simply that we have to ®nd another way (other than the bridging infer-
ence) of satisfying the felicity condition of uniqueness required by the
de®nite article in the chandelier. It is the gratuitous nature of the scalar
inference, compared to the forced nature of the felicity-condition require-
ment, that makes cancellation so straightforward in the scalar case. Thus
the accommodation account signally fails to generalize to the class of
GCIs as a whole.
However, even in the presuppositional-type cases, an accommodation
account will not be su½cientÐwe still need GCI theory. The point can be
made quite well by considering further facts about de®nite and inde®nite
reference. Compare the sentences in (49).
(49) a. ``Jerry Rich came in. He walked over to the window.''
b. ``Jerry Rich came in. The man walked over to the window.''
(our M ‡> nonlocal coreference)
Here both the expressions he and the man are de®nite referring expres-
sions, with a uniqueness condition attached. In an abductive system of
inference like Hobbs's, an attempt would be made to unify the discourse
referents by linking both expressions to the antecedent Jerry Rich, thereby
satisfying the de®niteness conditions. Indeed, both the pronoun and the
de®nite description would be predicted to be interpreted in the same way.
But that, of course, is contrary to intuition: the pronoun invites such
linkage to Jerry Rich, but the description the man resists it. GCI theory
gives a concise account of this: pronouns are minimal expressions that
invite maximal interpretations by the I principle discussed above, whereas
the use of a marked de®nite description where a pronoun could have been
used suggests by the M-principle that the speaker did not intend such
local coreference (see chap. 4).
We can press the critique further. Consider the possible analysis of the
inde®nite article in accommodation theory. If the de®nite article requires
a previously existing discourse referent, the inde®nite may be claimed to
require the setting up of a new discourse referent (Heim 1982; Kamp
1981). On this account, the sequence of sentences in (50a) would seem to
presume distinct discourse referents; this seems correct. But now examine
(50b). Here we seem happy to co-identify the ship and a ®ne galleon,
contrary to the theory.
(50) a. ``The man entered. A man coughed.''
disjoint reference
63 On the Notion of a Generalized Conversational Implicature

b. ``The ship broke up. A ®ne galleon of the Spanish ¯eet had been
irretrievably lost.''
‡> possible conjoint reference
c. Scalar explanation: hDEFIN, INDEFi; use of inde®nite Q-
implicates that the use of the de®nite would be misleading
The best explanation of these facts seems once again to be given by
GCI theory (sketched indeed originally by Grice). We should reject the
analysis of the inde®nite article as requiring the setting up of a new dis-
course referent aÁ la Heim (1982), while accepting that de®niteness does
involve conditions on usage (like uniqueness).49 We can then can set up a
scalar opposition (more details in chap. 2) as in (50c), where an inde®nite
expression is weaker because it simply lacks the uniqueness and ``given-
ness'' conditions associated with the more speci®c and informative de®-
nite expression. Given a scalar opposition of this kind, a speaker using an
inde®nite will Q-implicate that the use of the de®nite would be misleading
in at least one way. Thus in (50b), a speaker who said The ®ne galleon of
the Spanish ¯eet might be thought to be communicating that there was
only one ®ne galleon in the Spanish ¯eet and, wishing to avoid this infer-
ence, would use the inde®nite even though it is to be coidenti®ed with the
immediately preceding discourse referent. So again, accommodation-plus-
abduction gives the wrong account here, at least without rescue by GCI
theory.
In sum then, whatever the success of accommodation theory in the
presuppositional domain, it will not supersede GCI theory as Thomason
and Hobbs have both independently suggested. On the one hand, accom-
modation theory simply fails to extend to the gratuitous inferences that
are the focus of GCI theory, and on the other hand, without GCI theory
it make the wrong predictions, over-unifying discourse referents in some
cases (where M-implicatures warn otherwise) and failing to unify them in
others (where, for example, a Q-inference from the use of an inde®nite
explains the use of the inde®nite as a coreferential expression).
We may conclude this discussion of SW-R and Accommodation Theory
by noting the inadequacy of such reductionist approaches in the face of
the clear evidence in favor of a level of utterance-type meaning. These are
not the kinds of problems that might be addressed by tinkering with the
reductionist theories in limited ways. It is just because these theories are
theories of speaker-meaning, or nonce inference, that they are unable in
principle to capture aspects of meaning associated with the general, nor-
mal use of expressions.
64 Chapter 1

1.7 GENERALIZED IMPLICATURE AND STABLE PATTERNS OF

LEXICALIZATION

We may give one ®nal positive argument in favor of a theory of general-

ized implicature before passing on to consider in depth and by illustration
the richness of generalizations that are made possible by a theory of pre-
ferred interpretations of this sort. This argument concerns one such rich
generalization which has much empirical support, and which de®es ex-
planation in any other terms. The argument is based on work by Larry
Horn over 15 years, summarized in Horn 1989.
Recollect that we have described Q-implicatures as involving contras-
tive expression-alternatives, often arranged in scales of semantic strength,
like the prototype hall, somei scale, where the use of the semantically
weaker some will ceteris paribus Q-implicate the inapplicability of the
stronger expression all. Recollect, too, that for each positive scale of the
form hS, Wi, there will be a corresponding negative scale of the form
h@W, @Si, such that asserting (in a suitable sentence frame) @S will
Q-implicate @@W (i.e., W). So, instantiating, given a positive scale
hall, somei, we have the negative scale hnot-some (none), not alli, such
that asserting ``Not all the boys came'' will implicate `not-none of them
came', (i.e., `some of them came').
Now, the relationship between positive and negative scales is important
for a number of reasons. First, the relationship has not been understood
by most commentators (despite the explanation in Atlas and Levinson
1981). Thus Gazdar (1979), Hirschberg (1985: 73), and Kadmon (1987)
all claim that negation blocks implicature projection instead of seeing that
negation in the case of scales simply inverts the scale, so that negative
scales induce implicatures that just happen to be di¨erent from the
implicatures induced by positive scales. But a second reason why the
relationship between positive and negative scales is important is that it
yields real insight into the nature of logical operators and in particular
into the traditional square of opposition.
The square of opposition, that device conceived by Boethius out of
Aristotle to plague monks and schoolchildren for two millennia, has the
prototype instantiation shown in ®gure 1.1.50
I remind the reader of the following de®nitions:
p and q are contraries if p and q can't both be True,
but they can both be False (``John is happy/sad'')
65 On the Notion of a Generalized Conversational Implicature

Figure 1.1
The traditional square of opposition

p and q are contradictories if p and q can't both be True,

and nor can p and q both be FalseÐi.e., the negation of the one
implies the other, so there's no middle ground:
(e.g., ``The number of students is odd/even'')
Then we can see that the essential geometry of the square is given by
entailment relations, represented in ®gure 1.1 by the vertical lines (with the
entailing concepts at the top), contrary relations represented by the top
horizontal line, and contradictory relations represented by the diagonal
lines. The subcontrary relationship is the focus of interest here and its
proper interpretation has been controversial over the ages. Note that I
employ the traditional labeling of the vertices, A, I, E, and O, following
the mnemonics a½rmo for the positive side of the square and nego for the
negative side.
Note that in essence the traditional square is nothing but a pair of
scales, here hall, somei and hnone, not alli arranged vertically, so that the
strong elements all and none fall at the A and E vertices respectively, in a
66 Chapter 1

Figure 1.2
Square of quanti®ers

Figure 1.3
Square of logical (alethic) modals

contrary relationship to one another. It then follows by the geometry of

the square that all and not all (likewise none and some) are contra-
dictories, which the reader may verify by referring to the de®nition above.
Onto the geometry of the square one can map not only the quanti®ers
shown in ®gure 1.2 but also all the sets of logical operators shown in
®gures 1.3±1.5.
The reason that these distinct sets of logical operators map onto the
same square is that there are traditional ways of paraphrasing each set in
terms of the quanti®ers, and the quanti®ers in terms of the connectives.
For example, the alethic modal necessary that p can be paraphrased
(assuming a possible-world semantics) as `in all worlds, p', while possibly
p can be paraphrased as `there is some world in which p'; while the
quanti®cational all Fs are G can (assuming a ®nite number of entities, x1 ±
xn that are F ) be paraphrased as x1 is G and x2 is G and . . . xn is G, and so
67 On the Notion of a Generalized Conversational Implicature

Figure 1.4
Square of deontic modals (and deontic causatives)

Figure 1.5
Logical connectives

on, enumerating all x, while some F is G can be paraphrased as either x1 is

G or x2 is G or . . . xn is G, enumerating all x.
And one may state the following well-known logical equivalencies
directly on the relationship between the vertices of the square, thus gen-
eralizing across all the di¨erent squares (I illustrate with the quanti®ers):
(51) The duals
A ˆ @I . . . @ (All . . . ˆ Not some not . . .)
I ˆ @A . . . @ (Some . . . ˆ Not all not . . .)
@A ˆ I@ . . . (Not all . . . ˆ Some not . . .)
@I ˆ A . . . @ (Not some . . . ˆ All not . . .)
So we come now to the pertinence of all this to GCI theory. There has
in fact been two and a half millennia of confusion over the subcontrary
68 Chapter 1

relation, the relation between the I (e.g., some) and O (e.g., not all )
vertices of the square. The question has been whether the relation should
be thought of as a logical one, with the I vertex implying the O one, or
whether on the contrary it should be understood as a nonlogical sugges-
tion. Horn (1972, 1989) documents the many inconsistent positions that
theorists have taken over the ages, showing how, for example, Aristotle
held that in the case of the modals the I/O relation was logical but in
the case of the quanti®ers it was a nonlogical suggestion. Hamilton and
Jespersen held the relation is logical for all the squares, De Morgan and
J. S. Mill that it is nonlogical for all the squares, and so on.
The theory of GCIs helps to explain the confusion. The I (some) corner
of the square carries a generalized scalar implicature to the e¨ect that the
O (not all ) corner also holds. It is the generalized nature of the inference
that explains the confusion even among these eminent scholars thinking
deep and hard about the problem. ``Some'' strongly suggests `not all', and
``Not all'' strongly suggests `some', but we can say without contradiction
``Some in fact all'' or ``Not all, indeed none,'' indicating that the sugges-
tion cannot be a logical relationship. The correct answer to this perennial
puzzle is given by the nature of scalar implicatures and the way in which
negation reverses scales: the vertices always form scales of the form hA, Ii
and hE, Oi; so the assertion of ``I'' will implicate `@A', which is of course
the contradictory of A, and therefore equivalent to O:
(52) a. hAll, somei; hNone, not alli
b. ``Some of the boys came'' (I)
‡> `Not all of the boys came' (@A ˆ O)
c. ``Not all of the boys came'' (O)
‡> `Not-none (i.e., some) of them came' (@E ˆ I)
Thus, given the nature of scales, it follows that there will be a regular
predictable inference from the assertion of a weak item on a scale to the
contradictory of the strong item on the scale, which will itself correspond
to the weak item on the scale of opposite polarity. Thus there will always
be a pragmatic inference from the assertion of the I (some) corner to the
inference of the O (not all ) corner, or from the assertion of the O (not all )
corner to the inference of the I (some) corner. Despite the fact that this
relation is only a defeasible pragmatic suggestion, its occurrence is so
regular and predictable that theorists were misled into believing it was a
semantic relation.
69 On the Notion of a Generalized Conversational Implicature

Here then is the ®rst argument one can make from consideration of
the I/O relation in favor of a theory of GCIs: it is evidently not easy to
separate GCIs from semantic content, witness two millennia of analysts'
perplexity. But if these inferences were nonce inferences, sometimes going
one way, sometimes another, the pragmatic nature of the I/O relation
would be self-evident. It is not self-evident precisely because there is a
strong default inference, a strongly preferred interpretation that arises
from the assertion of either corner to the conclusion that the other corner
holds. Teasing apart preferential interpretations from semantically stipu-
lated ones is a di½cult analytical task.
A second argument for GCI theory that can be made arises from an
important observation by Horn (1989: 252±267): as a matter of empirical
fact, there is a strong crosslinguistic tendency for the O corner of the
square to not be instantiated by a single (or at least a simple, nonderived)
lexical item. In other words, the O corner of the square hardly ever lexi-
calizes; it is nearly always expressed by complex phrases of the form
exempli®ed by English not all, possibly not, not both, and so on. This is in
contrast to the strong negative apex E, where the forms routinely lexi-
calize, as in English none, impossible, neither, and so forth. In this context,
the absence of *nall (meaning `not all'), *nalways (meaning `not always')
or *noth (meaning `not both') is certainly noteworthy; and the fact that
this generalization (the O corner of the square hardly ever lexicalizes)
appears to hold widely across languages, indeed universally as far as we
now know, is certainly remarkable.
(53) English exempli®cations of the universal lexicalization constraint on
the O corner
not all has no lexicalization *nall (cf. none)
permit not has no form *permitn't (cf. forbid )
not always has no form *nalways (cf. never)
not . . . and . . . has no form *nand (cf. nor)
not both . . . and . . . has no form *noth (cf. neither)
The patterns in English in (52) are thus typical of other languages as well.
The details are telling, too: thus Priests can not marry is ambiguous in
English between the E-reading `are forbidden to' and the O-reading `are
permitted not to', but as soon as it is contracted and thus lexicalized
(Priests can't marry) only the E-reading is available (Horn 1989: 259).
The uncontracted form already has a preferred E-reading, and this is also
a crosslinguistic, diachronic tendency, which Horn (1989: 262±263) dubs
70 Chapter 1

``O to E drift.'' Note that mustn't lexicalizes `must not', but mayn't (which
might be expected to lexicalize the O corner `may not') can only mean
`must not', too. The generalization is that we expect no lexicalization of a
necessity modal that incorporates a negation outside its scope (e.g., `not
must'), or a lexicalization of a possibility modal with a negation inside its
scope (e.g., `may not')Ðboth kinds of lexicalization would be at the O
corner.51
GCI theory provides an account of this lexicalization constraint along
the following lines. First, as we have just established, I and O are related
by implicature: asserting I will carry a generalized Quantity scalar impli-
cature that O. Second, it seems that what is conveyed by a GCI is gener-
ally not lexicalized; that is to say, there is a redundancy constraint on
lexicalization, roughly as follows:
(54) Redundancy constraint
For any lexical item W that carries a generalized Quantity
implicature P, there will be no fully lexicalized counterpart W 0 that
lexicalizes the content of P. Because I items on a Square will always
implicate O items, there will be no lexicalized O items.
Another way of putting this is: if the I item entails p and implicates q,
then the corresponding O item will entail q and implicate p. Thus in an
appropriate sentence frame the assertion of ``some'' will entail `at least
one', and implicate `not all', whereas the assertion of ``not all'' will entail
`not all' and implicate `some'. Thus the I and O items have the same
overall communicational load, simply reversing the distribution of the
load between what is entailed and what is implicated.52
We need one more ingredient in the explanation. Given that the asser-
tion of the I (some) or the O (not all ) corner will carry the same commu-
nicational load, why should I (some) be lexicalized but not O (not all )?
The answer is provided by the relatively complex nature of negation (as
shown by much psycholinguistic work). Given a choice between a positive
term and a negative term carrying the same load, the positive term will
be chosen as the basis from which to apply the redundancy constraint
against lexicalization of the negative expressions. One should note that
there is no general ban against lexicalization of negatives; that is shown
by the tendency for the E terms to incorporate negation.53 But given a
choice between choosing the I (some) expressions or the O (not all )
expressions as the basic ones and excluding the other corner from lexi-
71 On the Notion of a Generalized Conversational Implicature

calization by application of the redundancy constraint, the processing

complexity of negatives favors the I (some) corner.
Anyone who rejects the notion of a GCI or a preferred interpretation
will have to ®nd another explanation for this lexicalization pattern. The
nonce-implicature theorist cannot explain itÐaccording to any such
theory, there are no general tendencies to be found, or if there are they have
the status of mere behavioral tendencies, playing no role in the systematic
generation of implicatures. On such a theory, there is no more reason for
not all to resist lexicalization than for not some to undergo it (none). Only
a theory of preferred or default interpretations can explain how some-
thing as defeasible and protean as a pragmatic suggestion can impose
such fundamental constraints on the lexicon.
In conclusion, then, these patterns statable across the traditional square
of oppositions appear to be strong evidence in favor of a theory of gen-
eralized implicature. The nonce-implicature theorist cannot in principle, it
seems, explain how there could be such a strong, reliable preference for
interpretation in one direction that (a) generations of theorists themselves
could be confused about whether inferences from one logical operator to
another were to be explained by semantic or by pragmatic content and
that (b) the language itself is adapted (by the provision of lexical gaps) to
the prevailing pragmatic winds, as it were.

1.8 CONCLUSIONS

In this chapter, my prime objective has been to try and defend the notion
of a generalized conversational implicature. Throughout, I have argued
against the dominant current line that a theory of communication
should allow only a level of speaker-meaning (or utterance-token mean-
ing). Instead, I have argued for a third level or layer in a theory of com-
munication, the level of utterance-type meaning interposed between the
level of speaker-meaning and the level of expression-meaning. This level
is to capture the suggestions that the use of an expression of a certain
type generally or normally carries, by default. We may concede that the
concept of a preferred, default interpretation is not an easy one to expli-
cate and it lacks the kind of theoretical underpinning we may perhaps
have for other concepts in the theory of meaning, without conceding that
there is any doubt at all about the need for such a notion in a theory of
communication.54
72 Chapter 1

My arguments for the need for such a level of utterance-type mean-

ing in general, and for the theory of GCIs in particular, have been (I
recapitulate):
1. There are all sorts of other phenomena, like illocutionary force (both
direct and indirect), conversational routines, presuppositions, felicity
conditions, and so forth, that would themselves justify the setting up of
such a level.
2. The theory of implicature as originally developed by Grice (e.g., in his
application of the theory to the natural-language connectives) presup-
poses such a level.
3. There is an argument from design to the e¨ect that a stable set of
heuristics inducing generalized implicatures would be far too powerful
and e¨ective an instrument of communication to neglect.
4. There are a range of formal models of nonmonotonic default reason-
ing that o¨er some hope for formalization in this area. Such models,
whatever their ultimate relevance to the theory of GCIs, show that it is
perfectly possible to formulate relatively precise theories of preferred
interpretation; thus there is nothing to inhibit the development of
clearly articulated theories of utterance-type meaning.
5. I have argued that existing theories of nonce implicature when applied
to the phenomena that are central to GCI theory will either make the
wrong predictions or make no clear predictions at all, just where in fact
there are clear predictions to be made. In contrast, GCI theory seems
to make the necessary generalizations in a straightforward way.
6. Finally, I have argued that if there were no such things as GCIs or
preferred default interpretations, it would be impossible to account for
the strength and generality of the pragmatic inferences associated with
the natural language logical operators. In particular, we would have
no account at all of the reliable crosslinguistic tendency to ban lexi-
calization of the O corner of the traditional square of oppositions.
These arguments will be at least su½cient, I hope, for the reader to
entertain the idea of a generalized conversational implicature without
prejudice. We can then proceed to show how the possession of a theory of
generalized implicature may contribute to the description and analysis of
a great variety of linguistic facts, as detailed in the following chapters.
In making these arguments, I have introduced in a cursory way our cast
of characters, the three main genera of GCI. In the next chapter, I explore
this typology and its consequences in detail.
Chapter 2
The Phenomena

2.1 INTRODUCTION

In chapter 1, I argued that Grice's notion of a generalized conversational

implicature (GCI) is an essential explanatory notion. It amounts to the
claim that there is a special species of pragmatic mechanism that yields
inferences that are both defeasible and default in character. At least some
of the properties of such inferences can be captured within a number of
currently available default, nonmonotonic reasoning systems. But these
inferences, generated under the mutual assumption of tacit coordination
through speci®c heuristics, have the force of strong presumptions. They
belong to a broad third category or layer of meaning, midway between
sentence-meaning and speaker-meaning (or utterance-token-meaning),
namely utterance-type meaning, in which types of linguistic form come to
have preferred, idiomatic readings. Despite many attempts to reduce this
third layer away, either to aspects of sentence-meaning or aspects of
speaker-meaning, it remains a robust, irreducible category of meaning, as
shown by Horn's generalization of universal patterns of constraints on
lexicalization that can be directly attributed to the omnipresent operation
of inference patterns of this default sort.
The three main species of GCI (according to my typology) were intro-
duced in the last chapter in a cursory way. The purpose of this chapter is
to lay out the kind of phenomena that motivate the theory of GCIs. These
data are the crucial observations that reductionist theories of utterance
meaning have never come to terms with, and indeed cannot, as far as I
can see. The phenomena are of course much more important and robust
than the theory I am advocating; indeed I will not hide the fact that they
are not all adequately accounted for. Nor are the observations below
exhaustive; the reader should be able to ®nd many more, similar kinds of
74 Chapter 2

default inference associated with speci®c kinds of linguistic expression.

But there will be su½cient examples, I hope, to demonstrate that there is a
large class of preferred interpretations, which have a systematicity both
within and across languages and whose regularity makes them a frequent
source of grist for the grammaticalization mill.
Although the observations have a real theory independence (having
been observed as far back as Aristotle; see Horn 1973, 1989: chap. 1),
I present them in terms of the tripartite classi®cation already introduced
in chapter 1Ðthat is, in terms of Q-, I-, and M-principles or heuristics,
which I feel yields more or less natural classes. So at the same time I will
attempt to clarify these principles and show how the relevant examples
fall under their respective predictive power. Various critiques and alter-
native treatments will be dealt with en passant, with the consequence that
my assessments of them will be more cursory than some of these pro-
posals deserve (for which I apologize in advance). Finally, at the end of
the chapter, I describe the interactions between these principles, the pro-
jection problem this interaction gives rise to, and the solution that I have
to o¨er.
Why three and just these three principles? Other estimates vary from
one (Sperber and Wilson 1986) to ten or more (Leech 1983), and there is
little point in arguing from ®rst principles (but see Levinson 1987a, where
I take on the lower estimates, and Atlas and Levinson 1981). As explan-
atory principles, we want as few as will do the job, while recognizing
essential distinct genera and respecting evidence from conversation that
speci®c principles are being oriented to (again, see Levinson 1987a).
Some discussion will occur throughout this book, but a few pointers may
assuage irritation at this point. Readers will note that the three principles
are derived partly by elimination from Grice's original four maxims (with
a total of nine submaxims). The maxim of Quality (``try to make your
contribution one that is true'') plays only a background role in the gen-
eration of GCIs. (There has been a tendency to assume that the Quality
maxim generates implicatures of belief or commitment, despite Grice's
explicit warnings against this.)1 The maxim of Relation or Relevance
(which Grice (1989: 26±27) stated simply as ``Be relevant'' within the
overall Cooperative Principle, ``Make your contribution such as is
required, at the stage at which it occurs, by the accepted purpose or
direction of the talk exchange in which you are engaged'') has pertinence
only to the immediate, ever variable, conversational goals: it generates
PCIs, not GCIs. This leaves two important subprinciples of Quantity,
75 The Phenomena

canonized here as the Q- versus I-principles, and four subprinciples of

Manner, of which two are captured in my M-principle (I suspect there is
scope also for the one Grice labelled ``be orderly,'' perhaps along the lines
of Haiman's 1985a principle of iconicity). Still, it may be objected that
this shows much too much reverence for Grice's ramshackle scheme, a
sort of joke calqued on Kant.2 Reverence is indeed beside the point
(especially for such a delightfully irreverent intellect as H. P. Grice), but
the fact is each of these three principles seem to capture a large range of
phenomena (as has been noted by a number of di¨erent scholars under
various rubrics). The three principles are in various ways incommensu-
rable and even antagonistic, and therefore successful reduction looks
unlikely. But for the purposes of this chapter, I ask the reader to suspend
these questions and just see how the recommended scheme seems to ®t the
facts.

2.2 THE Q PRINCIPLE

2.1.1 Q Inferences
In chapter 1, I introduced the Q principle as a heuristic based on Grice's
®rst maxim of Quantity, and sketched the way in which it induces infer-
ences from the use of one expression to the assumption that the speaker
did not intend a contrasting, usually informationally stronger, one. Con-
sequently, implicatures from this principle have two essential character-
istics: they are metalinguistic (and paradigmatic) in the sense that what is
implicated makes essential reference to what might have been said but
wasn't (i.e., to a set of linguistic alternates in paradigmatic opposition),
and they are negative propositions, what is implicated being a presump-
tion that such and such is not the case.
Now this species of inference has been more extensively explored in the
literature than any other kind of implicature, and I will not repeat here
the whole history of observations and commentaries (see, e.g., Levinson
1983: chap. 3). Su½ce it to say that Grice (1967, 1978) employed the ®rst
maxim of Quantity to account for the implicatures from the use of the
disjunction and conditional to the unknown truth value of the constituent
sentences, but that it was Horn's (1972) treatment, and the subsequent
formalization by Gazdar (1979), that established what we may call the
standard doctrine (as reviewed in Levinson 1983: chap. 3). Since then, a
number of critiques and elaborations have appeared, which I will refer
to later. The version reproduced here (based on Levinson 1987b) is
76 Chapter 2

essentially the standard one (it is thus equivalent to ``Quantity'' in Atlas

and Levinson 1981, to Horn's 1984 ``Q-principle'' minus its Manner ele-
ments, and so on). A number of problems have since been raised with this
formulation, which I consider below. Additionally, I go on to assimilate a
range of further inferences not normally considered to fall under its
rubric.
(1) Q-principle3
Speaker's maxim: Do not provide a statement that is
informationally weaker than your knowledge of the world allows,
unless providing an informationally stronger statement would
contravene the I-principle. Speci®cally, select the informationally
strongest paradigmatic alternate that is consistent with the facts.
Recipient's corollary: Take it that the speaker made the strongest
statement consistent with what he knows, and therefore that:
a. if the speaker asserted A(W ), where A is a sentence frame and W
an informationally weaker expression than S, and the contrastive
expressions hS; W i form a Horn scale (in the prototype case,
such that A(S) entails A(W )),4 then one can infer that the speaker
knows that the stronger statement A(S) (with S substituted for
W ) would be false (or K @ (A(S)) in Gazdar's (1979) notation,
read as `the speaker knows that it is not the case that (A(S))')
b. if the speaker asserted A(W ) and A(W ) fails to entail an
embedded sentence Q, which a stronger statement A(S) would
entail, and fS; W g form a contrast set, then one can infer that the
speaker does not know whether Q obtains or not (i.e., @K(Q) or
equally, fP(Q); P @ (Q)g read as `it is epistemically possible that
Q and epistemically possible that not-Q')
The recipient's corollary will induce the well-known scalar and clausal
Q-implicatures (Gazdar 1979; Levinson 1983: chap. 3) as illustrated here:
(2) a. Scalar Q-implicatures
``Some of my best friends are linguists.''
‡> K @ (all of my best friends are linguists)5
b. Clausal Q-implicatures
``John believes there is life on Mars.''
‡> fP(there is life on Mars), P @ (there is life on Mars)g
Note that the scalar implicatures are induced from ranked sets of
alternates (e.g., quanti®ers or scalar adjectives), whereas the clausal ones
77 The Phenomena

derive from contrasts between one expression that entails its embedded
sentence(s) and another that does not. Further kinds of Q inferences are
introduced below. I shall have relatively little to say about clausal Q-
implicatures (see section 2.2.5), merely expounding Gazdar's account,
whereas scalar implicatures have received considerable recent attention
that will require review.
I have introduced the inferences derived from the principle with the
epistemic modi®cations suggested by Gazdar (1979).6 Now, these have in
fact been the subject of much dispute. A ®rst issue concerns the di¨eren-
tial strength of implicatures from the same scale. Horn (1972: 90) origi-
nally suggested that not all scalar implicatures have the same epistemic
force: in general, for scales with more than two members, for example, a
scale of the form ha,b,c,di, and an arbitrary sentence frame S, assertion
of ``S(d)'' implicates that the recipient must infer K @ S(a) but only that
the recipient may infer K @ S(b), K @ S(c). Thus, for example, saying
``Some came'' implicates that the speaker knows that not all came, and
possibly that not many or not most came (given the scale hall, most,
many, somei). Gazdar's formulation has equally strong epistemic com-
mitment by the speaker for all items on the scale above the asserted one,
and one may complain that this neglects Horn's correct intuitionÐ
namely, that the most salient implicature is the denial of the head of the
scale. Gazdar's formulation was in fact designed to handle the projection
problem; he wanted to capture the observation (see section 2.5.1) that
clausal implicatures would cancel scalar ones where they were inconsis-
tent, and he achieved this by assuring that the epistemic modi®cation did
not mask inconsistency and by stipulating that clausal implicatures take
precedence. Because clausal inferences are of the form fPp, P @ pg, they
are inconsistent with a scalar implicature of the form K @ p. Atlas (1993a:
209¨ ) ®nds the clausal inference Pp implausible but will then need to
reconstruct a projection solution (see section 2.5.1 and Kay 1992 for a
defense of Gazdar's approach to the projection problem).
A second issue is that many authors feel that the epistemic commit-
ments on the part of the speaker proposed by Gazdar are just much too
strong (see, e.g., van der Sandt 1988: 126¨ ). Soames (1982, 1998: 115±
120) suggests that the epistemic force of scalar implicatures is the external
negation of the Hintikka operator K, so that asserting, ``Some came'' will
carry the generalized implicature that the speaker doesn't know whether
all came or @K(all came) or P(all came), which may be strengthened by
the contextual assumption that the speaker would know whether or not
78 Chapter 2

all came, to a particularized implicature that the speaker de®nitively

knows that not all came, or K @ (all came). He provides both telling
examples and a Gricean argument to that e¨ect. Horn (1989: 214, 233±
234, 543 n. 5) essentially concurs and suggests that scalar implications are
of the form @Kp (`the speaker doesn't know that p', rather than Gazdar's
K @ p, `the speaker knows that not p'), unless they are strengthened by an
additional assumption that S is an expert on the subject p.7 Soames (1998)
goes on to show that if this tack is taken, scalar implicatures are not after
all (at least in the classic cases) inconsistent with clausal implicatures, and
Gazdar's special rule that scalars cancel clausal implicatures where they
are mutually inconsistent can be dispensed with.8
Both more elaborate and simpler solutions are possible. Hirschberg
(1985: 75¨ ) argues that it would be desirable to distinguish between the
strong epistemic commitment associated (by the maxim of Quality) with
the asserted content of utterances, and the weaker commitment, perhaps
in terms of belief rather than knowledge, associated with implicatures.
She therefore claims that scalars actually implicate a disjunction itself:
asserting ``Some came'' implicates that S believes that not all came or that
S doesn't know whether all came.9 Additionally, one might wish that the
relative defeasibility of di¨erent species of pragmatic implication (e.g., of
scalar vs. clausal Q-implicatures) was also re¯ected in variable strengths
of speaker commitment. On the other hand, Atlas (1993a) holds that the
whole issue betrays a deep confusion (promoted by Grice's reductionist
program) between utterance-meaning on the one hand and psychological
states of speakers on the other: when one utters U implicating p, p is
conveyed whether one believes it or not. Even if a speaker's mental states
are involved in the computation that U implicated p, they are not part of
the content of U.10 On this account, the entire debate is beside the point:
implicatures come naked, without epistemic commitments.
One central point needs to be preserved in this debate, and that is the
intuitive di¨erence between scalar implicatures, which do come with a
strong presumption, arguably something close to epistemic certainty, and
clausal implicatures where what is conveyed is precisely epistemic uncer-
tainty. It is for this reason that, in what follows, I simply presume, fol-
lowing the Gricean tradition, that implicatures do come with some form
of epistemic modi®cation that will retain this distinction. None of the
pro¨ered solutions seem quite right,11 and although a correct solution
would be required for any formalization, I will o¨er none here and will
remain agnostic about the directions in which such a solution may be
79 The Phenomena

found. Henceforth, I shall not in general mark the epistemic modi®cation

on implicatures: thus when I write ``Some came'' Q ‡> `Not all came',
this may be understood as a shorthand for `the uttering of ``Some came''
will, ceteris paribus, implicate that the speaker judges that not all came'
(where ``judge'' can be cashed out in various ways).12 Where the distinc-
tion between the the scalars and the clausal implicatures needs to be
highlighted, I revert to Gazdar's (1979) notation, using K and P, but with
a weaker ``belief '' rather than ``know'' interpretation (where the impli-
cated suggestion seems especially weak, I indicate this in the gloss), so
that scalars will generate implicatures of the Kp form and clausal impli-
catures will have the form Pp, P @ p, as I will explain.13

2.2.2 Entailment Scales

We may loosely de®ne a Horn scale (or entailment scale, to di¨erentiate
such from non-entailment scales to be mentioned below) as an ordered
n-tuple of expression alternates hx1 ; x2 ; . . . ; xn i such that, where S is an
arbitrary simplex sentence-frame and xi > xj , S…xi † unilaterally entails
S(xj ).14 Classic scales of this kind come from many grammatical classes
and include the quanti®ers hall, most, many, somei, connectives like hand,
ori, modals hnecessarily, possiblyi, hmust, should, mayi, adverbs like
halways, often, sometimesi, degree adjectives like hhot, warmi, and even
verbs like hknow, believei, hlove, likei. They all induce Q-implicatures in
the way that should already be familiar: assertion of a lower ranking
(rightwards) alternate implicates that the speaker is not in a position to
assert a higher ranking oneÐfor if the speaker was in that position, he or
she should, by the Q-principle, have asserted the stronger sentence. The
one-way semantic entailment that orders these sets guarantees of course
that leftmost items are informationally richer than rightwards ones.
Further conditions have to be put on Horn scales, lest they over-
generate (Gazdar 1979: 57f; Atlas and Levinson 1981: 44). Clearly we do
not implicate the negation of every proposition that entails what we say.
First, items in the same scale must be in salient opposition: of the same
form class, in the same dialect or register, and lexicalized to the same
degree. These considerations exclude for example a Horn scale hi¨, if i,
for i¨ (if and only if ) belongs to a specialized register and is not mono-
lexemic like if. This lexicalization constraint holds especially for the
stronger items in scales, and for good Gricean reasons: if the hearer sus-
pects that the speaker may have avoided a stronger statement simply be-
cause it would have been clumsy and prolix (and thus contrary to Manner
80 Chapter 2

maxims), the hearer can clearly draw no conclusions about the speaker
not being in a position to make the stronger statement (Matsumoto
1995).15 More precisely then, the lexicalization constraint requires that
stronger items on a scale be lexicalized to an equal or greater degree than
weaker items.
Second, scalar items must be from the same semantic ®eld, ``about'' the
same semantic relations, and thus in conceptually salient opposition.
Hence there is no scale hregret, knowi because regret introduces semantic
parameters additional to those involved in know.16 Similarly, in conjunc-
tion with the requirements for sameness of word class, semantic issues
rule out a scale h p because q, p and qi. The same multiple constraints
rule out the putative scale hnecessarily-p, pi, thought by Burton-Roberts
(1984) to be a serious anomaly in the Gricean account (because the plain
assertion of, e.g., ``Two and two is four'' would then imply that it is not
necessarily so).
There is one apparent systematic exception to these constraints, which
has to do with negative scalar items. Gazdar (1979: 56±57) blocks impli-
catures arising in negative contexts: otherwise, he claims, the assertion of
``It is not the case that Paul ate some of the eggs'' would implicate `It is
not the case that Paul ate not all of the eggs' (i.e., that Paul ate all of the
eggs). The same blocking constraint is imposed by Hirschberg (1985: 73),
who correctly notes, however, that Gazdar's blocking of other logical
functors is unnecessary. Even Horn (1989: 234) subscribes to a blocking
function that includes both overt and covert negatives (and indeed all
downward-entailing operators). Now, in fact, this move arises from a
misunderstanding of negative scales. The correct generalization, as
pointed out by Atlas and Levinson (1981), is that negation reverses scales:
for example, if hall, somei form a Horn scale, so do hnone (i.e., not some),
not alli.17 Hence ``It is not the case that Paul ate some of the eggs'' (or
more simply ``Paul didn't eat some of the eggs'') asserts an item at the
strongest end of a scale (not some ˆ none), from which as always no
implicature follows. The process is entirely general, and no special pro-
jection apparatus is required. However, given the generalization noted by
Horn (1973, 1989) described in chapter 1, that the (not all ) O corner of
the traditional square of oppositions regularly fails to lexicalize, whereas
the (none) E corner does, this arguably constitutes a regular exception to a
constraint that items in Horn scales must be of equal lexicalized status.
When the constraint is formulated, as above, so that it requires strong
scalars to be lexicalized to a same or greater degree than weak items on
81 The Phenomena

the same scale, negative scales are in fact no longer exceptional (note, e.g.,
that not some lexicalizes to none on the hnone, not alli scale). Still, there is
a special pattern here to the nature of negative expressions, given the
systematic O-corner resistance to lexicalization. This special pattern may
be recognized by introducing negative scales by meta-rule: for any scale of
the form hx1 ; x2 ; . . . ; xn i, there will be a corresponding scale of the form
h@xn ; . . . ; @x2 ; @x1 i. The psychological counterpart to this move is the
claim that we process such negative scale items as none by reference to the
equivalent expressions containing the relevant positive scalar items, like
`not some', a decompositional claim that is consistent with the known
processing di½culties associated with negatives.18
There are some well-known diagnostics for scalar implicature. If we
entertain the hypothesis that two expressions, call them S and W for
``strong'' and ``weak,'' form a scale hS; Wi, then they should permit the
following embeddings:
(3) a. Canceling phrases
``W and even S'' (e.g., ``Some and even all of them
came.'')
``Not only W, S'' (e.g., ``Not only some of them, all of
them came.'')
``W in fact/indeed S'' (e.g., ``Some, in fact all of them,
came.'')
b. Suspending phrases
``W or possibly/even S'' (e.g., ``Some or possibly all of them
came.'')
``W if not S'' (e.g., ``Some, if not all of them,
came.'')
The rationale, of course, is that because implicatures unlike entailments
are defeasible, it is possible to assert the contrary, or explicitly raise its
possibility, without any sense of contradiction. The last type, the suspen-
sion of the implicature by mention in an if-clause, or by mention of the
strong alternate in an or-clause, was taken as an especially important
diagnostic (called ``suspenders'') by Horn (1972, 1989: 234±234). For non-
entailment scales where we cannot order expressions on simple entailment
grounds, Horn suggested we could use ``suspenders'' to establish that
items belong to a single scale and their ordering within it. Gazdar (1979)
assumed that the mechanism in these suspending if-clauses was simply
implicature cancellation by a general projection mechanism (whereby
82 Chapter 2

clausal implicatures cancel scalar ones inconsistent with them); for more
recent commentary, see Hirschberg 1985 (118±120) and Atlas 1993a
(211±217), who suggests that these concessive if-clauses do not have the
semantics of full conditionals. In addition to these diagnostics, one may
also note that the head of a scale (the A or E or topmost items in the
square) can collocate with modi®ers like absolutely, barely, not quite (as in
absolutely all vs. ??absolutely many, not quite all vs. ??not quite some, etc.).
One should note that these diagnostics establish scalar relations between
expressions in a context, but they do not establish the GCI status of the
scale (see the discussion of Hirschberg's scales in section 2.2.4.)
A summary of the points made so far is given in (58).
(4) a. Positive Horn scales
A set of linguistic alternates hx1 ; x2 ; . . . ; xn i such that S(xi )
unilaterally entails S(xj ), where S is an arbitrary simplex sentence-
frame, and xi > xj , and where x1 ; x2 ; . . . ; xn are
i. equally lexicalized items, of the same word class, from the
same register; and
ii. ``about'' the same semantic relations, or from the same
semantic ®eld.
b. Negative Horn scales
For each well-formed positive Horn scale of the form
hx1 ; x2 ; . . . ; xn i, there will be a corresponding negative scale of
the form h@xn ; . . . ; @x2 ; @x1 i, regardless of the relative
lexicalization of the negation.
c. Diagnostics for GCI scales
For any item xj weaker than xi on a scale, ``xj , in fact/if not/or
even/or possibly xi '' should be a felicitous phrase in an
appropriate sentence-frame.
Before attending to the issue of other kinds of scales, let us reap the har-
vest so far sownÐfor there are highly systematic generalizations to be
made here.

2.2.2.1 Quanti®cational and Modal Operators: Horn's Arithmetic Square

of Oppositions I introduced in chapter 1 the traditional square of oppo-
sitions and reviewed Horn's (1973) argument that the subcontrary rela-
tion (between I and O corners of the square) is implicatural.19 Further,
the regularity that strong negative scalar items incorporate negation
(none) whereas weak ones tend not to (not all ) was attributed to the
83 The Phenomena

regular (i.e., generalized) implicature from some to `not all'. Indeed I

argued that this kind of regularity cannot be captured without a theory of
default or generalized inference. This area is relatively well understood in
English, and I provide just su½cient detail for those studying other lan-
guages to see how to proceed (much more will be found in Horn 1989).
A problem that has arisen in understanding the quanti®ers is the exact
relationship between midscale negative and positive quanti®ers. For ex-
ample, is Few came the negative counterpart of Many came, and are they
contraries or subcontraries? And what is the contradictory of Many came?
There are a great many quanti®ers in English whose position in the square
is not immediately self-evident (e.g., several, a few, many, very many, most
or very few, a minority, not many). Clearly, if we are to understand
quanti®cation, we must understand the relationships between quanti®ers
that are midway on, or not at either end of, Horn scales. Diagrammati-
cally, the problem is shown in ®gure 2.1.
What we seek is a general solution in terms of the square of oppositions
to the problem of whether the negative counterpart of a given scalar item
is a contrary or a subcontrary, and what its contradictory is. Then, the
solution will carry over from the quanti®ers to the modals, logical con-

Figure 2.1
The problem of the location of midscale expressions
84 Chapter 2

Table 2.1
Counterparts across the quanti®ers, modals, and connectives within the square of
opposition
A-Corner E-Corner I-Corner O-Corner
Quanti®ers all none some not all
both neither one not both
everything nothing something not everything
everybody nobody somebody not everybody
always never sometimes not always
Modals
alethic necessary impossible possible possible not
epistemic certain impossible possible/ possible not/
may may not
deontic must/need to must not may/can may not
obligated forbidden permitted permitted not to
Connectives (both) . . . and neither . . . nor (either) . . . or not both

nectives, and other domains. Following Jespersen (1917), Sapir (1930),

and generations of philosophical and logical thought (all nicely summa-
rized in Horn 1989), we may put into correspondence on the square the
sets of notions given in table 2.1. The placement of scalar items on the
square requires in the ®rst instance a distinction between the positive (A-I
dimension) scales and the negative (E-O dimension) scales. This may be
established partly by reference to the well-known duals (repeated here
from chapter 1):
(5) The duals
A . . . ˆ @I . . . @ (e.g., all . . . implies not some . . . not . . .)
I . . . ˆ @A . . . @ (e.g., some . . . implies not all . . . not . . .)
@A . . . ˆ I @ . . . (e.g., not all . . . implies some . . . not . . .)
@I ˆ A . . . @ (e.g., not some . . . implies all . . . not . . .)
It may partly also be established by reference to Horn's ``suspender'' tests
(1989: 235±236):
(6) Using suspending clauses to
a. Establish ordering in positive scales
i. ``Some or even all came.''
ii. ??``All or even some came.''
85 The Phenomena

Figure 2.2
Arithmetic scale for English quantifying phrases

b. Establish ordering in negative scales

i. ``Not all or even none came.''
ii. ??``None or even not all came.''
c. Disallow mixed positive and negative scalar items20
i. ??``Some or even none came.''
ii. ??``None or even some came.''
To establish a correspondence between items on the positive and cor-
responding negative scales, Horn suggests that, as a heuristic device, we
arithmetize the square by assigning values in the range of 0 to 1.0 on the
positive and 0 to ±1.0 on the negative scale (in each case, in the direction
weak W to strong S ends of the scale). Using tests to be discussed, Horn
(1989: 237) gives ®gure 2.2 as the square for English quanti®ers. The
halfway 0.5 position marks the line above which positive items have neg-
ative items as contraries (e.g., ``all came'' entails `not none came'); below
that line, positive items have negative items as subcontraries (i.e., the
relationship is implicatural; e.g., ``some came'' implicates but does not
entail `not all came'). How do we establish that, for example, most is
above the line and many below it? First, we can show that most is higher
on the positive scale than many by Horn's ``suspender'' tests: one can say
``Many if not most of them came'', but not ??``Most if not many of them
86 Chapter 2

came.'' Second, we can use LoÈbner's (1987) test for tolerant versus intol-
erant logical operators: logical contraries are intolerant of conjunction
with their internal negations (i.e., an intolerant quanti®er Q does not com-
bine in an acceptable sentence of the form ``Q(x) P(x) and Q(x) @ P(x)'') ,
whereas implicated subcontraries are tolerant of such a conjunction.21
(7) LoÈbner's tolerance test
a. Intolerant quanti®ers
i. *``All of the boys came but all of them didn't.''
ii. *``Most of the boys came but most didn't.''
iii. *``A majority of the boys came but a majority didn't.''
b. Tolerant quanti®ers
i. ``Many of the boys came, but many didn't.''
ii. ``Quite a few of the boys came, but quite a few didn't.''
iii. ``Several of the boys came, but several didn't.''
Putting the two tests together, Horn's ``suspender'' tests for order and
LoÈbner's test for tolerance, we can order all the items on both the positive
and negative scales, and place them correctly with regard to the crucial
midway line establishing contrary versus subcontrary (implicatural) rela-
tions. Items aligned horizontally on the positive and negative scales are
taken to be directly opposed as contraries (if above the midline) or sub-
contraries (if below it).22 Using these methods, we can obtain the kind of
model illustrated in ®gure 2.2, with the two scales (positive and negative)
systematically related. Then we can predict not only the intrascalar GCIs
(e.g., Several xj'd will implicate `Not many j'd'), but the crossscalar
inferences between the expressions near the I and O corners (e.g., A few
j'd will implicate `Many didn't j' and vice versa). The same procedures
of course can be employed over various sets of logical operators. Readers
may like to play with quanti®cational frequency adverbs of the type
exempli®ed by seldom, often, and rarely and see if they can construct the
positive and negative scales, the contrary versus subcontrary and contra-
dictory relationships.23

2.2.2.2 Entailment Scales over Nonlogical Predicates

2.2.2.2.1 General Vocabulary There are many entailment scales that

cannot be arranged in a square of oppositions even though they come as
paired positive and negative scales. Consider for example gradable adjec-
tives hhot, warmi and hcold, cooli; these form two scales linked by a
87 The Phenomena

(sub)contrary relationship rather than one uni®ed scale, as shown by the

usual tests. For example, one can say It's warm, in fact quite hot, but not
?It's cool, in fact quite warm, because warm and cool belong to di¨erent
scales. Moreover, the comparative warmer means `further towards hot'
(not, on analogy to, e.g., redder, `more towards warm'), whereas cooler
means `further towards cold'Ðthat is, the scales are headed in di¨erent
directions (Horn 1989: 239±240). But the two scales form no square
because such predicates allow no inner negations, and thus hot and cold
are not related like all and none as A to E corners of the square (Horn
1989: p. 245, 548±549). The full scales in this case seem to be hboiling,
hot, warmi and h freezing, cold, cool, lukewarmi. Similar paired scales
abound; for example, hexcellent, good, all righti and hawful, bad,
mediocrei. Further examples are given in (62).
(8) a. i. ``The handwriting is similar.''
ii. ‡> `not identical'
iii. hidentical, similari
iv. ``The handwriting is similar, if not identical.''
b. i. ``This is a partial solution.''
ii. ‡> `not complete'
iii. hcomplete, partiali
iv. ``This is a partial solution, and indeed perhaps a complete
one.''
and similarly for hlong-term, short-termi (or h permanent, durablei),
hhuge, largei, hancient, oldi, and so on.
In addition to gradable adjectives, verbs form scales like hadore, love,
likei and its negative counterpart hloathe, hate, dislikei. Some scalar
verbs like those in the scale hknow, believei, which permit internal nega-
tions, have been of considerable philosophical interest: knowing p entails
believing p, and indeed not believing that not p, but stating ``John believes
p'' implicates `John doesn't know p'. It is likely that our analysis of the
lexical meaning of verbs has systematically failed to take scalar implica-
tures into account.

2.2.2.2.2 The Number Words One kind of scale that has exercised
analysts is that formed by the in®nite series of numbers words or numer-
ical expressions in English, h . . . , ®ve, four, three, two, onei. Horn (1972)
assumed that the numerical expressions should fall within his scalar
approach, carrying the normal implicatures: thus an assertion ``He has
88 Chapter 2

three children'' should implicate but not entail that he has no more. In
general, the reasons for assuming that there is such a scale with cor-
responding GCIs still seem good: the assertion of (9a) would have the
semantic content paraphrased in (9b) (the lower bound `at least three' is
entailed)24 and the GCI in (9c) (the upper bound is implicated, as in
scales generally). The normal suspension tests apply as in (9d), showing
that the content of (9c) cannot be entailed. Finally, in the right context,
where an exact speci®cation is irrelevant, the implicature is implicitly
cancelled as in (9e)
(9) a. ``John has three children.''
b. ``John has at least three children.''
c. ``John has at most three children.''
d. ``John has three children, if not four/and perhaps more/or
possibly ®ve.''
e. A: ``Does John qualify for the Large-Family Bene®t?''
B: ``Sure, he has three children all right.''
Despite these attractions, there are a number of di½culties with a scalar
approach to number words. First, mathematics would be in danger: when
we assert that the square root of 9 is 3 we do not mean that the square
root of at least 9 is at least three, otherwise we could as well say the square
root of 9 is 2 (Sadock 1984)Ðspecial conventions are in force of course,
and we may want to distinguish sharply between natural language nu-
merical expressions and cardinal numbers (Atlas 1993: chap. 2). Second,
Horn (1972) himself drew attention to the fact that in certain speci®c cir-
cumstances the scales could be inverted, for example in tallying golf
scores where the higher score is the lesser number one might felicitously
say ``He can do the round in seventy two, if not seventy.'' This possibility
of scale-reversal does not appear to arise with other quanti®ers (Sadock
1984: 143) and indeed otherwise seems restricted to pragmatically de®ned
scales; this raises the general issue of the extent to which all scales are in
the last resort pragmatic in nature, an issue addressed in section 2.2.4.
But this possibility of scale reversal, coupled with the fact that the
implicature can evaporate in contexts like (9e), has led some analysts to
abandon the scalar approach to numerical expressions. One position is to
claim that the number words are just ambiguous between an `exactly' and
an `at least' interpretation, perhaps disambiguated in speci®c linguistic
contextsÐfor example, as `exactly n' when acting as determiners (see
Kempson 1975; this is Kamp's position as reported in Kadmon 1984).
89 The Phenomena

But the prospect of in®nite ambiguities not only here but then elsewhere
too is very unattractive (especially compared to the scalar approach,
where the two uses of numeral expressions follow from general properties
of any scalar series whatsoever). Another approach is to assume that the
meaning of, say, three is simply indeterminate at the semantic level
between `at least three' and `exactly three' and is then further speci®ed by
a particularized conversational implicature in context as appropriate, or
perhaps restricted by semantic constraints in particular constructions
(thus Atlas (1983) suggests that collective NPs have the `exactly' inter-
pretation). Kempson (1986) and Atlas (1993b) take this approach but for
di¨erent reasons: Kempson because she wishes to reduce GCIs to PCIs
compatible with Relevance theory,25 and Atlas because he wishes to as-
similate the unspeci®ed nature of cardinal semantics to his account of not,
which has a similar property. Carston (1985) took a similar line, arguing
that three is semantically unspeci®ed between three readings, `at least
three', `at most three' (as in the reverse-scale cases), and `exactly three',
particular contexts being responsible for our understanding in each case.
The trouble with much of this argumentation is that it confounds gen-
eral problems of the very existence of GCIs as a species of inference,
pragmatic intrusion into truth conditions, the general nature of semantic
representations before pragmatic enrichment, so-called metalinguistic
negation, and the projection problem for implicaturesÐwith the rather
speci®c di½culties associated with the English number words. The speci®c
di½culties boil down to the following facts.26 First, in math-literate cul-
tures, there will be possible confusion between English three and the
numeral `3', with a consequent bias towards an `exactly three' interpreta-
tion. Second, this may be associated with the fact that in some cases the
number words are associated with upper-bounding speci®cations that are
not easily defeasible (Kadmon 1984: 30¨; Horn 1989: 251). For example,
compare ``He has one child'', which seems to stipulate only one, with ``He
has a child'', where the upper-bounding implicature is more easily lifted.
This may well be due however precisely to an M-implicature associated
with that opposition. Similarly, ``I have two hundred and three dollars''
more strongly suggests an upper bound than ``I have two hundred dol-
lars'', which can again perhaps be attributed to a Manner implicature: to
bother to say ``two hundred and three'' where the shorter ``two hundred''
alone might be su½cient suggests that there is reason to be precise, and
the speaker is acting in accordance with that reason. Third, as Horn
(1989: 251±252) points out (see also Brown and Levinson 1987: 258¨ ), all
90 Chapter 2

implicatures are potentially subject to a process of conventionalization,

and the number words may be under pressure to lexicalize the `exactly'
reading: where `three' is lexically incorporated, as in three-sided, triple,
thrice, three-ply, or triumvirate, no `at least' reading is possible (Horn
1978, Atlas 1983), a fact already clear in the ordinals like fourth prize.
Fourth, there are perhaps genuine di¨erences in interpretive freedom
between number words in di¨erent syntactic and thematic positions, as
Atlas (1983), Kadmon (1987), Fretheim (1992), Van Kuppevelt (1996),
and Scharten (1997) argue, which might be expected if the partial con-
ventionalization account is correct.
These complicating factors and confusions are su½cient to make the
number words not the correct testbed for the whole theory of scalar
implicature (as, e.g., Kempson (1986: 86) seems to suppose). Horn
(1992b: 172±175) indeed abandons the classic scalar approach to just the
number words, pointing out a number of additional special properties.
Still, when due allowance is made for the special role of number words in
math-literate cultures, and consequent possible conventionalization of the
`exactly' readings, there are a number of reasons to hang on to a scalar
interpretation of ordinary language numeral expressions in general. One
central piece of evidence is provided by those languages that have a ®nite
series of numerals. Many Australian languages, for example, have just
three number words, which are glossed as `one', `two' and often `three'.
The scalar prediction is clear in these cases: we have a ®nite scale h`three',
`two', `one'i, where `one' or `two' will implicate ceteris paribus an upper
bound; but because there is no stronger item `four', the cardinal `three'
should lack this clear upper bounding by GCI. And this is clearly the case
in, for example, Guugu Yimithirr: nubuun can be glossed `one', gudhirra
`two', but guunduu must be glossed `three or more, a few'.27 And this is
the general report (see Dixon 1980: 108). None of the other theories make
this correct prediction. Finally, Cruse (1986: 69±70) points out that by
certain tests for the better establishment of one sense over another, the
`exactly' interpretations of the English number words are psychologically
more salient than the `at least' interpretations, and by the same tests it can
be seen that this corresponds with the favored interpretations of the
quanti®ers (those with doubts about the very existence of GCIs may ®nd
this bias interesting).28

2.2.2.2.3 Closed-class Morphemes, Function Words Entailment scales

can be found among any restricted set of lexemes with semantic asym-
91 The Phenomena

metries of the right kind. One happy hunting ground, barely exploited to
date, is the area of closed-class morphemes and function words, although
not all of these are organized as entailment scales. Treating the quanti®ers
in this way has yielded, as we have seen, rich insights. The connectives
were of course the original focus of Grice's (1967) theory, and GCI theory
explains many of their propertiesÐfor example, the scalar relations of
hand, ori explain the preferred exclusive reading of the disjunction, and
the clausal implicatures of if and or rely on the availability if stronger
locutions that entail their constituent sentences like since and and.
GCI theory makes a number of interesting predictions with respect to
sets of closed-class morphemes or function words covering single semantic
domains:
. On the whole, we expect sets of contrasting function words, because the
pragmatic load can then be increased through pragmatic oppositions
without further increasing the set of expressions;
. We can expect scalar oppositions due to informational asymmetries, in
which case the usual scalar predictions should apply (e.g., for a scale
hS; W i, we do not expect lexicalization of `not S');
. There are likely to be Manner contrasts due to the M-principle intro-
duced in chapter 1 and further described below (where a set of marked/
unmarked expressions covering the same domain will tend to pick up
complementary oppositions, the unmarked ones picking up the stereo-
typical state of a¨airs); for example, a contrast between simplex on and
the compound preposition on top of.
Here, I concentrate on the scalar implicatures, returning at the end of the
section to the prospect of handling closed classes in terms of multiple
pragmatic principles.
An important example is provided by the articles, although Grice's
(1975: 56) remarks here seem to have de¯ected proper analysis (including
Atlas and Levinson 1981: 49; Carston 1988: 214). Grice noted the co-
nundrum that ``I went into a house'' suggests that the speaker went into
someone else's house, whereas ``I cut a ®nger'' suggests I cut my own
®nger. There is in fact no scale hmy, ai because possession and indef-
initeness are not ``about'' the same semantic relations; consequently, there
is no GCI from ``a NP'' to `not my NP', and nonscalar inferences of var-
ious kinds would seem to be involved. However, there is an inference
from inde®nite reference to the inappropriateness of de®nite reference, for
the articles form a Horn scale hthe, ai by the usual criteria:
92 Chapter 2

(10) a. ``I saw the man with a red hat.''

Entailed: `I saw a man with a red hat.'
b. ``I saw a man with a red hat.''
Implicated: `I saw someone who I would hesitate to describe as
the man with a red hat.'
c. ``I saw a man with a red hat, in fact the very man with a red hat
you saw yesterday.''
Grice's examples are thus indirectly explained: when one says ``I went into
a house'' one Q-implicates `I didn't go into the house', where the de®nite
suggests (I-implicates) my very own house. On the other hand, when I say
``I cut a ®nger'' I merely implicate that it was no unique, otherwise salient
®nger (say, the one I cut before) that su¨ered, and that interpretation is
compatible with the assumption that it was my own ®nger (which in turn
is a more stereotypical reading than one that involves the chopping of
other peoples' ®ngers).
Hawkins (1991) has argued that if this analysis is adopted, many puz-
zles in the analysis of de®niteness seem more tractable. He holds on to an
essentially Russellian analysis of the de®nite article, thereby explaining
the entailment relationship between the X and an X (the representation of
the contains the representation of a as a conjunct), supplemented with a
conventional implicature associated with the that speci®es the kind of
mutually accessible set in which the referent must be unique. The seman-
tic representation of a is simply the existential quanti®er, and the ba¿ing
range of uses that are associated with it follow in large part from the
Q-implicatures that derive from the contrast with the. In general, an
inde®nite phrase an X seems to be associated with the presumption of
nonuniqueness (as in A senator resigned ), but there are plenty of circum-
stances in which this presumption is defeated by background knowledge
or a constructional meaning, as in existential statements like Nine has a
positive square root, or France has a capital. The explanation would be
that an X picks up its nonuniqueness presumption by contrast to the
stronger de®nite article, in the normal scalar way, except where this is
overridden by background knowledge or semantic conditions.
In those circumstances where the inde®nite is read as referring to a
unique entity, a de®nite article is also grammatical, but it tends to force
the assumption of a di¨erent contextual set in which the referent is
unique. Consider (11), where one might expect only a de®nite article to be
possible, given the unique nature of the o½ces in question, but in fact
both the de®nite and inde®nite article can be used.
93 The Phenomena

(11) a. England has a prime minister, and America has a president.

b. England has the prime minister and America has the president.
Hawkins (1991: 421±422) notes that we read the existential construction
in (11a) as assigning types of constitutional leadership, o½ces with unique
incumbents, to countries. In contrast, we read (11b) in a di¨erent way: the
de®nite forces us to ®nd some further contextual set in which the referent
is unique (e.g., England is hosting the prime minister of France, whereas
America is hosting France's president). The explanation seems to be that
existential statements tend to take an inde®nite object (whose uniqueness
may be given by background knowledge), so that a de®nite NP encour-
ages us to ®nd another construal. These kind of complexities seem con-
sistent with the pragmatic account but not with some more restrictive
semantics for the inde®nite article (e.g., specifying lack of uniqueness).
Note that on this account, the anaphoric potential of the de®nite, as
opposed to the inde®nite, follows from the meaning (contrast discourse
semantic accounts like Heim 1982). The range of data that can thus be
explained by this approach far exceeds other existing accounts of the
nature of the de®nite/inde®nite contrast. Before we can fully appreciate
this, I need to introduce some further types of implicature, and further
observations about the articles are given later in this chapter (see, e.g.,
remarks on bridging inferences in section 2.3 and on the opposition
between article and zero-article in section 2.4.1).
Lyons (1977) advances the view, well supported by diachronic evidence,
that the de®nite article may in some respects be treated as a deictic
determiner unmarked for the proximal/distal distinction expressed by this
and that. Hawkins (1991: 415) suggests that the contrast is essentially that
the demonstratives presume accessibility to perception or via discourse,
whereas the de®nite article is neutral (although both presume uniqueness
in some set). If this were correct, one might expect a scale h{this, that},
thei, such that by saying ``Could you lend me the book?'' one would
suggest that the book is not perceptually available in the situation of
speech (for then this book or that book would be more informative).
Notice that there is indeed a contrast between (12a) and (12b), the ®rst
being sayable with my hand in my pocket, or on the telephone, whereas
the second requires a demonstration or exhibit of the ®nger (ignoring for a
moment the anaphoric uses of the determiners).29
(12) a. The index ®nger on my left hand is hurting.
b. This index ®nger on my left hand is hurting.
94 Chapter 2

More tentatively, we may note that the marked deictic determiners also
come in closed classes and may be susceptible to a similar analysis. For
example, there may well be a Q-contrast between hthis, thati, where this
has more precise criteria of application than that: the proximal deictic is
marked for proximity, the so-called distal deictic presumes accessibility
(in space or discourse) but is unmarked for proximity, and thus picks up
the complementary interpretation. This predicts that that has a wide dis-
tribution, potentially overlapping with this, as indeed seems to be the
case. Thus for the same spatial arrangement of a book on a table I can say
both I wrote this book and I wrote that book, but if there are two books,
that book will suggest the one I would not describe as this book. This kind
of opposition between deictically speci®ed and unspeci®ed has already
been suggested by Wilkins and Hill (1995) for `come' and `go' verbs in
various languages, and may carry over to other sets like hhere, therei.
I turn now to anaphoric uses of these determiners. Prince (1981) argues
that there is a hierarchy of anaphoric expressions in English, according to
the degree to which their referents are mentally activated, as it were. She
argues that ``the use of an NP representing a certain point on the scale
implicates that the speaker could not have felicitously referred to the same
entity by another NP higher on the scale'' (1981: 245). Developing these
ideas further, Gundel, Hedberg, and Zacharski (1989, 1993) outline a
Givenness Hierarchy (see ®gure 2.3) along the following lines, where
the stronger requirements are leftwards and the weaker ones rightwards,
because any referent that is ``in focus'' (in central attention in the dis-
course) is also ``activated'' (previously referred to, etc.), ``familiar,'' and
so on. On the basis of quantitative textual evidence, they claim that the
English anaphoric expressions pattern as shown in ®gure 2.3. But the use
of a rightwards expression only implicates that a leftwards one does not

Figure 2.3
Givenness Hierarchy (after Gundel, Hedberg, and Zacharski 1993)
95 The Phenomena

apply ( just as noted above for the use of the inde®nite vs. the de®nite
article), but this is su½cient to suggest a change of referent. For example,
use of that signals a change of referent from it (Gundel, Hedberg, and
Zacharski 1989: 94).
(13) He had to show it to the agency before he sent it to the newspaper,
and when the editor asked for a change, he had to show that to the
censors too.
Whatever the success of this particular analysis, implicature does seem
to have a systematic role in signaling changes of referent, and this is
addressed directly in chapter 4.
Grammatical morphemes, with their contrastive oppositions, may also
frequently exhibit Q-contrasts parasitic on, and additional to, semantical
contrasts. Consider tense in English.30 Harnish (1976, 1991: 354) notes
``the use of the past tense often carries the implication that the activity or
state indicated by the verb no longer is present''; thus the pattern in (14a±
b) (see also Comrie 1985: 24). Harnish might have gone on to suggest a
scale as in (14c), where it is more informative to know if the event still
holds rather than just that the event held at some previous time (because
knowing the current situation gives us a unique known instance in time
where the event is known to hold, whereas a past tense only assures us
that there was some instant at which the event occurred).
(14) a. ``John used to live in Rome.''
Implicature: `John no longer lives in Rome.'
b. ``John used to live in Rome, and in fact he still does.''
c. hPRESENT, PASTi
d. ``John lives in Rome.''
Most formal treatments of tense re¯ect the asymmetry in informational
speci®city by stating in e¨ect that the present tense of an event holds if the
time of an event coincides with the designated utterance time, whereas the
past tense holds if there is at least some time preceding utterance time at
which the event occurred (e.g., Cann 1993: 242±243). Given the linear
nature of time, it is also the case that any present event becomes a past
event: the past is more event-populated than the present! These consid-
erations, together with the tendency for present relevance of present
events, su½ce to ensure an informational asymmetry between the tenses
even where no entailment relationship holds (strictly, of course, it does
not follow from the uttering of John is sleeping at tj that John was sleeping
96 Chapter 2

at ti , the moment before, although it is likely, and indeed it is certain at tk ,

the moment after the utterance, that he was sleeping before).
There are many other implicatures from tenses that need a proper
pragmatic account (see Comrie 1985: 25, 71). Consider, for example, the
perfect tenses, where the extra complexity of the perfect over the simple
tenses suggests that a simple tense speci®cation would not have been
adequate (see Harder 1994: 73):
(15) a. ``John had been living in London.''
Implicature: `He was no longer living in London.'
b. ``John had been living in London, and in fact it turned out he
still was.''
(16) a. ``John will have ®nished by tomorrow.''
Implicature: `He has not yet ®nished.'
b. ``John will have ®nished by tomorrow, in fact he has already
done so.''
Languages that make do without tenses and temporal conjunctions like
after or before often use oppositions between aspects to implicate sequence
and anteriority. Bohnemeyer (1998) shows for example that speakers of
Yucatec, a language without tense or temporal conjunctions, can system-
atically communicate temporal information through two kinds of gener-
alized implicature: from aspectual unboundedness one can infer temporal
overlap (as in ``John came in. Mary was sleeping''), and from aspectual
boundedness one can infer temporal nonoverlap (as in ``John came in.
Mary got up'', with order then indicated iconically).
Another area that would be worth reexamining from an implicatural
point of view are spatial expressions in closed sets, whether local cases,
prepositions or post-positions. Here we can expect all our three pragmatic
principles to be involved. For example, hat, neari form an entailment
scale, such that X is at the station entails X is near the station; if one then
asserts the weaker item, as in ``John is near the station'' one implicates
`not at the station' (Levinson 1995b: 192±194). To understand the full
range of pragmatic oppositions between di¨erent spatial prepositions, one
will need to bring in our other pragmatic principles. Thus, for example,
fon, on top of g form an M-opposition, such that ``on top of the desk''
suggests a noncanonical placement (e.g., the object is not in contact with
the desk, resting instead on other objects which are themselves directly on
the desk) or a noncanonical viewing angle (e.g., the speaker is under the
desk).
97 The Phenomena

I know of no treatment of local cases along these kind of lines,

although it is notable that the locative case, being normally unopposed to
more speci®c local cases, tends to pick up stereotypical associations (in,
e.g., Tamil) by the I-principle, so that `Mango is table-LOC' suggests `The
mango is on the table', `Mango is bowl-LOC' suggests `The mango is in
the bowl', which could have been more exactly speci®ed by more prolix
postpositional phrases. These latter do not constrain the interpretation of
the locative because they are not in Q-contrast but rather in M-contrast
relation with the simple case marker.
Other case contrasts, or contrasts between case and other constructions,
should be susceptible to pragmatic treatment. For example, Levine (1990)
examines the contrast in Russian between possessive forms of NP and
dative NPs with possessive meanings, where the latter carry implicatures
of noncanonical or special relationship. HeinaÈmaÈki (1994: 213±214) notes
that Finnish has the following contrast between partitive and accusative
cases:
(17) a. Terttu luki kirjaa.
Terttu read book-PART
`Terttu was reading a (part of the) book.'
b. Terttu luki kirjan.
Terttu read book-ACC
`Terttu read (all) the book.'
c. Terttu luki kirjan vain puolivaÈliin.
Terttu read book-ACC only halfway-through
`Terttu read the book only halfway through'
HeinaÈmaÈki notes that the inference in (17a) to the e¨ect that only part of
the book was read is cancellable; and so is the inference in (17b) to the
e¨ect that all of the book was read. The partitive in conjunction with a
durative verb like `read' is speci®cally noncommittal about whether the
event is bounded or not, whereas the accusative is speci®c that the event
is indeed bounded. We then have in e¨ect a meta-scale of the from
hBOUNDED; GBOUNDEDi, instantiated by the accusative and parti-
tive cases respectively, such that the use of the unspeci®ed partitive form
suggests that a bounded event is not being referred to. The speci®c inter-
pretations from boundedness to `read all the book', and from unbound-
edness to `read part of the book', can be attributed to the I-principle,
which induces stereotypical interpretations; these can also be canceled
as indicated in (17c), where the `read all the book' inference is canceled
98 Chapter 2

by explicit indication that the bound of the reading was only half-way
through the book.
These illustrations suggest that an implicature analysis of closed classes
of morphemes may yield many important insights and resolve many puz-
zles that arise from the apparently elusive and variable content of these
expressions.

2.2.3 Q-Contrasts Based on Other Kinds of Lexical Opposition

If Q-implicatures are based on the opposition between an informationally
stronger and an informationally weaker expression, it is conceivable that
there are other kinds of Q-implicature than those based on Horn scales or
Gazdar's clausal oppositions. There may for example be lexical sets or
semantic ®elds giving rise to similar inferences where however the criteria
for Horn scales, for example, are not met. To the extent that such infer-
ences are based directly on the structure of the lexicon, rather than ancil-
lary knowledge about the speci®c context, one might expect them to have
a generalized character. In this section, a number of putative kinds of
such inference are discussed, but it should be borne in mind here that the
phenomena are hardly explored and de®nitely not fully understood. This
section is included only in the hope that it will spur further work.
It has long been clear that we can get scalar-like inferences from sets of
linguistic expressions that fail the entailment criterion for Horn scales
(see, e.g., Harnish 1976). Thus the pair of verbs succeed, try behave as if
they belonged to a Horn scale, but as (18c) shows, there is no entailment
from x succeeded to x tried.
(18) a. Putative scale: hsucceed, tryi
b. ``John tried to reach the peak.''
Q-implicature: `he didn't succeed.'
Suspension: ``John tried to reach the peak, and indeed/or
perhaps he succeeded.''
c. ``John succeeded without even trying.''
The correct theoretical moves here are perhaps not clear. A Gricean
argument based on Q is still applicable, because there seems to be an
informational asymmetry between the expressions: trying involves trying
to succeed, hence the listener may be expected to have an interest in the
success of the attempt; the failure to supply information about success
thus implicating that it was not achieved. Although succeeding does not
entail trying, a succeeded in jing contrasts with the simpler a j'd and thus
99 The Phenomena

perhaps M-implicates that `a j'd with directed e¨ort'.31 The implicature

from try to `not succeed' is clearly generalized, so we cannot assign it to
Hirschberg's (1985) solution for particularized scales discussed below. It is
not presuppositional, as it is nondetachableÐthat is to say, the inference
is attached to the semantical content not the lexical form, and thus the
same inference appears from the more or less synonymous pair hmanage,
attempti. Tentatively, then we shall adopt the view that there are non-
entailment scales, which nevertheless carry generalized conversational
implicatures.32
We may be tempted by other putative scales, where the phenomena are
actually rather di¨erent:
(19) Pseudo-scale: hmountain, hilli
``John can climb hills.''
Implicature: `John can not climb mountains.'
Here there is not only no entailment relation between the items on
the pseudo-scale (the one term excludes the other), there is no intrinsic
informational asymmetry. Rather, we are in the domain of the assertion
of one alternate implicating the inapplicability of the other, which is lower
ranked in some way. This kind of implicature is taken up later, but it is at
best marginal as far as generalized conversational implicature is concerned.
A more interesting marginal case is provided by deictic preemption: as
Fillmore (1975) notes, we cannot, if today is Tuesday, say ``Tuesday''
meaning `today' without danger of being mistaken to mean next Tuesday.
In a similar way, you preempts the name of the addressee, here preempts
the name of the place, and so on. One analysis is that whereas, in our way
of reckoning, Tuesdays are multiple (this week's and next week's, for
example), during this diurnal span there is only one today. Thus today
is informationally richerÐit uniquely refers, whereas Tuesday does not:
there is a scale of referential determinateness with a genuine informational
asymmetry, where asserting the weaker item implicates that the stronger
does not apply:
(20) Scale: htoday ˆ Tuesday, Tuesdayi
``Let's meet on Tuesday.''
Implicature: `Let's not meet today, this Tuesday, but rather next
Tuesday.'
I believe that this kind of referential asymmetry is crucial to the under-
standing of anaphora, as will be spelled out in chapter 4: determinate
100 Chapter 2

reference is informationally richer than potentially indeterminate refer-

ence, and scalar-like inferences are thus invoked by the use of the infor-
mationally weaker expression.
Still, it is clear that there are implicatures of contrast that have nothing
to do with informational asymmetry. Harnish points out the following
kind of inference based on lexical alternates, which he felt could be
attributed to the ®rst maxim of Quantity (Harnish 1976, 1991: 336):
(21) Alternates: hwhite, red, blue, . . .i
``The ¯ag is white.''
Q ‡> `The ¯ag is not white and red.'
``The ¯ag is white and red'' entails ``The ¯ag is white.''
It is clear that, whatever our initial intuitions, the statement ``The ¯ag is
white and red'' must entail that ``the ¯ag is white'', because if ``the ¯ag
is white'' is false, then the conjunction must also be false. One way of
looking at this is to claim that nonranked alternates, in addition to the
ranked scalar alternates, give rise to systematic implicatures: assertion of
the one alternate will ceteris paribus imply the inapplicability of others
unmentioned, otherwise by the Q-principle the speaker should have men-
tioned them.
Harnish (1976, 1991: 320) notes that, unlike the case where ``the ¯ag is
white'' implicates `the ¯ag is white all over', there is no corresponding `all
over' inference associated with such other adjectives as spotted, stained,
torn, dented, and so on. On the other hand, it can be noted that there is
indeed a corresponding inference with smooth, rough, warm, cold, shiny,
wooden, iron, and so forth. One stain is su½cient to warrant the descrip-
tion stained, whereas one spot of white is not su½cient to warrant the
description white (without further speci®cation). It is notable that the
adjectives without the `all over' inference seem to lack underived alter-
nates or contrastive terms: the best we can do are the contradictories like
unspotted, unstained, untorn (whole?), undented. And that perhaps is the
root of the di¨erence: white belongs to a salient lexical set of alternates
that are at most contraries33 and may simultaneously apply to a single
referentÐthe use of the one term then implicates that the use of the
others, which could equally apply, are not applicable in this case.
These di¨erent kinds of example raise the possibility that there is a
whole set of distinct but related types of Q-implicature, each bearing a
family resemblance to one another, and each deriving from di¨erent kinds
of contrast set provided by the structure of the lexicon. They would be
101 The Phenomena

related under the Q umbrella by the fact that in each case the inferences
are (a) negative, (b) metalinguistic, in the sense that they depend on a
contrast set of expressions, and (c) they rely on the fact that the unmen-
tioned alternates are either more informative alone (the scalar case) or a
conjunction of two or more of them would have been so (the alternates
case).
It is worth looking back at the work of the structuralist semanticists
(like Lyons, Lehrer and Lehrer, Haas, Cruse), those who have been
especially interested in the way in which the lexicon can be seen to be
organized into semantic ®elds, and the ways in which the semantic values
of expressions are (at least partially) de®ned by the sense relations that
hold between lexemes. In what follows, I make special use of the work by
Cruse (1986) and con®ne myself to di¨erent kinds of contrast set (e.g.,
taxonomies, metonymies, chains, helices). Such an exercise does not
appear to reveal simple generalizations, and indeed the remarks below are
extremely tentative. The generalized character of these inferences is in
general dubious. Nevertheless, looking through these types one can see
some general tendency for implicature production of two types (similar,
independent observations can be found in Hirschberg 1985: 60±61, 65,
108±113):
(22) a. Type I: use of a superordinate (more general) expression,
(especially above the ``basic level'' of Rosch 1977), suggests that
the speaker is not in a position to use a (more speci®c, more
informative, ``basic level'') subordinate expression, or deems it
irrelevant:
``I just saw a horrid animal in the larder.''
‡> `speaker is not sure whether it was a rat, a shrew, a mouse,
or what.'
b. Type II: use of an alternate from a contrast set suggests the
inapplicability of another alternate, even when both could apply:
``He lectures on Wednesdays.''
‡> `As far the speaker knows, he does not lecture on
Thursdays.'
c. Type IIb: the denial of an alternate in a contrast set suggests
that the other alternates may apply:
``He doesn't lecture on Wednesdays.''
‡> `For all the speaker knows, he lectures on another day of
the week.'
102 Chapter 2

The form of the Type I implicature is a particularly weak, disjunctive

implicature. Note, however, that it is based on an informational asym-
metry, and this suggests a parallel to Q-inferences. The Type II implica-
tures, like those associated with the color terms, are more theoretically
puzzling and indeed less systematic but are nevertheless regular in at least
some cases. They seem to arise especially in connection with two kinds of
lexical sets: sets of compatibles, like the color terms, where an expectation
of exhaustivity is raised by the use of one term; and sets that exhibit some
kind of intrinsic order (like the days of the week). The inferences in other
cases may require special contexts (e.g., following a question: ``Does he
have pets?'', ``He has cats'' implicates clearly that he has no dogs; see
Hirschberg 1985), or special intonation (Ladd 1980; Horn 1989: 230f,
410; Ward and Hirschberg 1985), and thus lie beyond a theory of GCIs.
They seem to have a secondary type (labeled Type IIb above), as sug-
gested by Hirschberg (1985), where the denial of an alternate implicates
the possible applicability of another. Type II inferences still seem to fall
under the rubric of the Q-principle, or Grice's ®rst maxim of Quantity,
because they are based on the absence of speci®ed information that one
would expect to be provided if applicable.
The generalized nature of most of these inferences must be in doubt.
Type I implicatures may lack a generalized status in part because of the
potential depth of taxonomic structure in the vocabulary and thus an
unclarity of opposition. Matsumoto (1995) suggests that it is choice of a
term above the ``basic level'' terms that gives rise to the inference that
the speaker cannot be more speci®c. He illustrates with kinship terms in
Japanese, where, for example, the basic-level terms are more speci®c than
English, so that use of the `brother' term implicates that the speaker
cannot specify `younger-' or `older-brother', which are the basic-level
terms in Japanese. It is noteworthy that such inferences seem on occasion
to have been su½ciently persistent to be a major source for diachronic
shifts in lexical meaning. Thus where there are pairs of terms where one
(the hyponym) is semantically marked for some feature, say ‡F, and the
other (the superordinate) is unmarked, then the use of the unmarked term
can come to implicate the alternative value for the feature (i.e., ±F), the
superordinate thus coming to function as its own hyponym, an ``auto-
hyponym'' (Horn 1984: 32¨; Lyons 1977: 308±311). A thumb is a kind of
®nger, hence we have ®ve ®ngers. But normally ®nger will denote `digit
other than thumb'. Similarly, dog can come to denote `male dog' in con-
103 The Phenomena

trast to bitch because the speaker by withholding the sex-marked term

implicates that the other sex must be intended. Note that we can think of
these inferences in a similar way to scalar implicatures: hthumb, ®ngeri
and hbitch, dogi form ordered pairs where the ®rst member entails the
second in a suitable sentence frame, and the use of the second member
thus implicates that the speaker is not in a position to assert that the
stronger expression holds (hence ``I cut my ®nger'' suggests the speaker
knows he did not cut a thumb). Diachronically, implicated autohypon-
ymy leads to systematic polysemy (Horn 1984; Kempson 1980).
Type II inferences seem, as mentioned, to have a generalized status in
only a few cases, usually over sets of compatibles (i.e., sets of alternate,
contrasting, terms that may be predicated of the same individual without
inconsistency). The color terms are the prototype case here, as described
above. In just the same way, if one asserts ``The marble is smooth'', one
implicates that it is not also rough in some places (i.e., that it is smooth all
over), although there is clearly no incompatibility between being smooth
in some places and rough in others. The negative Type IIb inferences
seem to arise rather reliably from the negation of one alternate to the
suggestion that another applies. Here of course one needs to distinguish
between the cases where such an inference is more or less necessary (as
from ``The ¯ag is not white'' to the conclusion that it must be another
color), as opposed to cases where it is freely invited (as from ``He doesn't
lecture on Wednesdays'' to the inference that he does lecture on other
days).
It is clear that considerably more work is required before such infer-
ences will be fully understood. They depend in part on the structure of the
particular semantic ®elds. Thus ``It will be ready on Wednesday'' suggests
`it will not be ready on the prior Tuesday', due to the helical structure of
the set of names for days of the week. Research is thus required into the
distinct kinds of semantic ®elds (e.g., taxonomies, meronomies; see Cruse
1986) and the way in which they may set up a basis for speci®c kinds of
inference. Many of the inferences that arise in connection with semantic
®elds are weak suggestions, which are only clearly attributable to the
speaker as implicatures in particular contexts; for example, in the context
of an answer to a question, as illustrated in (23) below. They thus lie at
the borderline between GCIs and PCIs, because on the one hand they rely
on the structure of the vocabulary, and on the other hand, they depend on
speci®c contexts to emerge. This prompts a review of this borderline.
104 Chapter 2

(23) a. A: ``Did John climb a mountain?''

B: ``He climbed a hill.''
‡> `As far as the speaker knows, he did not climb a
mountain.'
b. A: ``What did you see in the larder?''
B: ``I saw a horrid animal.''
‡> `speaker is not sure whether it was a rat, a shrew, a
mouse or what.'

2.2.4 Residual Problems: Scalar Implicature, GCIs, and PCIs

The classic scalar implicatures considered so far derive from the structure
of the vocabulary, speci®cally from salient, contrastive alternates ordered
in informational strength. For example, the quanti®cational scalar impli-
catures have the systematic properties they have just because the set of,
for example, English quanti®ers have the logical relations between them-
selves that are captured in the traditional square of oppositions. The
implicatures are parasitic on that structure. It is for this reason that they
have the default character they have: they are generalized conversational
implicatures because they depend on invariant salient properties of lan-
guage structure rather than variable contexts.
But, from the outset of the development of the theory of scalar impli-
cature, it was recognized that in various ways scales may be said to be
pragmatically de®ned. It has already been noted that the cardinal scale
can be inverted in special pragmatic contexts, although it appears to be
unique for vocabulary-de®ned scales in this respect. Fauconnier (1975)
pointed out that many scalar implicatures seem to be derived from scales
that depend on a structure of assumptions beyond the lexicon. To provide
a fresh example, the pattern in (24) suggests a scale of driving di¨erent-
sized trucks, but clearly this is not an entailment scale, nor one based on
lexical patterning of another sort, but on general assumptions about the
world (permits, or abilities, to drive large trucks normally permit driving
small ones, but not vice versa; but clearly the world could be otherwise).
(24) ``He can drive small trucks.''
Implicature: `He can't drive big ones.'
hDriving big trucks, driving small trucksi
``He can drive small trucks, and indeed/if not big ones too.''
Hirschberg (1985) o¨ers a systematic generalization of these pragmati-
cally given scales and the inferences they allow, but in so doing she
105 The Phenomena

presumes that she can subsume, and thus reduce, all scalar GCIs to PCIs,
an issue I will return to after introducing the theory. Hirschberg denies
that there is any principled distinction between GCIs and PCIs (1985: 42)
and o¨ers instead a theory ranging both over scales given by the lexicon
and those given by real-world or pragmatic factors. Consider for example
Fred's marathon attempt to cycle from the West Coast to New York City:
if we are told he has made it to Utah, we infer that he has not got further
east.
(25) A: ``How is Fred doing?''
B: ``He's got to Salt Lake City.''
‡> `not Chicago, New York' etc.
Given the scale: hNew York, Chicago, Salt Lake City, Renoi
The inference depends on a nonce scale, a contextually given ad hoc scale,
which I dub a Hirschberg scale. Similarly with the following example:
(26) A: ``Did you get Paul Newman's autograph?''
B: ``I got Joanne Woodward's.''
‡> `not Paul Newman's.'
Given scale of autograph prestige: hNewman, Woodwardi
She also independently notes that lexical sets of incompatibles can give
interesting implicaturesÐa½rmation of one implicates denial of another,
whereas denial of one alternate implicates a½rmation of another:
(27) A: ``Do you speak Spanish?''
B: ``I speak Portuguese.''
‡> `I don't speak Spanish.'
She o¨ers the following generalization. A set of expressions may act as a
Hirschberg scale, providing there is any partial ordering relation de®ned
over them in a contextually salient way. A partially ordered set, or poset, is
any set of values V (items ranked)34 associated with such a partial order-
ing O; a particular poset is de®ned by the set of values V and the speci®c
ordering relationship O.35 Because the ordering relation O might consist
of an is-a-kind-of relation, we could get a partial ordering hfoak, mapleg
treei or if O is the ordering relation has-parts, then hbook, fchapter 1,
chapter 2, . . .gi. We can accommodate the classical Horn scales as just
the special case where O is the entailment relation.36 Alternates (like
fblack, white, red, . . .g are therefore distinctive by virtue of not being
ordered by the ordering relation. The rules for implicature derivation are
then:
106 Chapter 2

(28) For ordered items

a. Stating a lower ordered value L implicates the speaker doesn't
believe a higher value H or doesn't know if H obtains
(schematically ``L'' ‡> not-H)
i. ``Some came''
‡> `not all' (O ˆ entailment, V ˆ hall, somei)
ii. ``It's a tree''
‡> `not sure which kind' (O ˆ is-a-kind-of, V ˆ hfoak,
maple, . . .g, treei)
iii. ``I've read chapter 1''
‡> `not the whole book' (O ˆ has-parts, V ˆ hwhole book,
fchapter 1, chapter 2, . . .gi)
b. Denying a higher ordered value H implicates that the speaker
believes a lower one obtains or may obtain (schematically ``not
H'' ‡> L)
i. ``Not all came''
‡> `some did'
ii. ``It's not an oak''
‡> `it is a tree'
iii. ``I haven't read the whole dissertation''
‡> `I have read some parts'
(29) For unordered items
a. A½rming an alternate L1 implicates other alternates L2 or Ln
are false or unknown (``L1 '' ‡> not-L2 †
i. ``I've read chapter 1''
‡> `not chapter 2'
ii. ``You need a typhus shot''
‡> `not cholera too'
b. Denying an alternate L1 implicates that other alternates L2 or Ln
may be true or unknown (``not L1 '' ‡> L2 †
i. ``It's not an oak''
‡> `it maybe a maple or . . .'
ii. ``She doesn't have a boy''
‡> `she has a girl'
Some further examples of the kind of pragmatically given partial order-
ings that one can ®nd underlying particularized implicatures of these sorts
follow (where O is the ordering relation):
107 The Phenomena

(30) a. A: ``Are you Greek?

B: ``I speak some Greek''
‡> `I am not Greek.'
b. O ˆ has-attribute
Poset ˆ hGreek, fGreek-speaking, Greek relatives, Greek
residency, Greek ancestrygi
(31) a. A: ``So is she married?
B: ``She's engaged''
‡> `she's not yet married.'
b. O ˆ has-prior-stage
Poset ˆ hmarriage, engagement, going-steady, datingi
(32) a. ``Smoking is a misdemeanor.''
‡> `Smoking is not a felony.'
b. O ˆ is-a-lesser-o¨ense-than
Poset ˆ htort, misdemeanor, felony, capital crimei
It must be conceded this is a good account of some elusive PCIs, like
those in the immediately preceding examples. But there are a number of
problems.37 The most serious is that the theory will overgenerate impli-
catures because it o¨ers no constraints on scalehood. One might try to
limit the ordering-relations structuring posets in some way, but that is not
in the spirit of the theory and it will not be su½cient any way. As we have
seen, even the entailment relation will overgenerate if not restricted by our
constraints on Horn scales: we do not implicate the negation of every-
thing that entails what we said (if we did, then by saying ``two plus two
equals four'' I would implicate `It is not necessarily the case that two plus
two equals four').38 Hirschberg (1985: chap. 6) suggests that the relevant
poset must be salient to both speaker and addressee, but the question then
is how this comes about. Although she points to discourse context, pros-
ody, and the like, it is clear that there is no general solution here.39 All
implicatures are made dependent on the contextually salient ordering
relation, so we have no account of implicature generation without an
account of how this is arrived at.
Hirschberg thus sketches a theory under which scalar GCIs and scalar
PCIs can be seen to be ultimately related: there are scalar implicatures
based on the lexicon (GCIs) and ones based on salient-contrast-in-the-
world (PCIs). I see no real con¯ict between GCI theory and her broader
theory, and I do not think it establishes, as she claims (1985: 42), that ``the
108 Chapter 2

traditional distinction between generalized and particularized implicature

is a false one.'' The GCI theorist is simply claiming that speakers carry
their lexicons on their backs as it were, from context to context, and it is
mutual knowledge of this fact that elevates the Q-heuristic to a default
mode of inference. And indeed Hirschberg (1985: 141) more or less con-
cedes this:
all other things being equal, a domain independent ordering will be more likely to
be assumed salient for the purpose of supporting scalar implicature than will a
domain dependent ordering.

2.2.5 Clausal Implicatures

As we saw in chapter 1, Grice was much concerned with the pragmatic
inferences associated with the logical connectives. Among those he wished
to account for are inferences of epistemic uncertainty that arise from both
disjunctions and conditionals (see, e.g., Grice 1989: 8±10). For example, if
I say ``Sue is either in the attic or in the garden'', I implicate that I do not
know which. In such a case, the inference cannot be due to the scalar
opposition between hand, ori, because Sue can clearly not be in both
places at once. In any case, the scalar inference would be to `not both p &
q', and the inference we are interested in here is to `the speaker does not
know whether p, or whether q'. The inference is general, but defeasible
(as shown by ``I hid the prize either in the attic or the garden, but I am
not telling you which''), and thus has all the hallmarks of a GCI. Grice's
explanation was that, if one knew p or knew that q, it would be un-
cooperative to utter ``p or q'', and therefore that the speaker must have
some chain of reasoning in mind that concludes with p4q and yet which
does not involve as premises plain p or plain q.
Gazdar (1979: 59±62) has in¯uentially advanced a di¨erent way of
obtaining the inferences of epistemic uncertainty. He suggests that just as
with the scalar implicatures, the inference rests on opposing alternative
expressions but, in this case, of a constructional kind. Wherever there is a
construction X-f, such that X-f entails f, and an alternative of roughly
equal brevity Y-f which is otherwise semantically identical to X-f but fails
to entail f, then asserting the weaker Y-f will Q-implicate that the speaker
does not know whether f obtains or not. The Gricean reasoning is, of
course, that if the speaker was in the position to utter the stronger X-f he
or she should do so by the ®rst maxim of Quantity (my Q principle), and
if the speaker utters instead the weaker Y-f, he or she can be presumed to
109 The Phenomena

be in ignorance of the truth of f. Gazdar (1979) proposed that the impli-

cature from the weaker expression Y-f should have the form fPf; P @ fg
(i.e., `for all the speaker knows f, for all the speaker knows @f'), but I
gloss the implicatures informally with perhaps or epistemic may.
In the case of the connectives, neither If p then q nor p or q entail their
constituent sentences. Moreover, there are contrasting constructions of
equal brevity that entail their constituent sentencesÐfor example, Since p
then q is the entailing counterpart to If p then q, and p and q the more
informative counterpart to p or q. Hence we obtain the clausal implica-
tures illustrated below where (33a) and (34a) carry the implicatures that
immediately follow, and (33b) and (34b) are alternative phrasings that
would avoid the implicatures (because the constituent sentences are
entailed).
(33) a. ``If eating eggs is bad for one, we should give up omelets.''
Q ‡> `(For all the speaker knows) eating eggs may be bad for
one, or it may not be bad for one; we perhaps should give up
eating omelets, or we perhaps should not.'
b. ``Since eating eggs is bad for one, we should give up omelets.''
(34) a. ``Sue is a linguist or an anthropologist.''
Q ‡> `For all the speaker knows, Sue is perhaps a linguist, or
perhaps not a linguist, perhaps an anthropologist, or perhaps
not an anthropologist.'
b. ``Sue is a linguist and an anthropologist.''
These implicatures capture the inference of epistemic noncommittedness
associated with the conditional and the disjunction. Notice that there is a
Horn scale hand, ori giving rise to an additional but distinct scalar
implicature from the use of the disjunction, viz. that the speaker is not in a
position to use the conjunction (an implicature which would be satis®ed if
the speaker knew one of the disjuncts to be false).
If this account is accepted (and not all ®nd it plausible; see Atlas
1993a),40 then it generalizes to all cases of embedding constructions
where strong and weak constructions are both available. In particular,
many verbs come in doublets of the appropriate kind. Consider, for
example, the contrastive pair know, believe, where know-f entails f, but
believe-f does not entail f or @f. There are sentences of roughly equal
brevity like (35a±b), which di¨er only in that one (35a) has a linguistic
expression that entails an embedded sentence S, whereas the other (35b)
110 Chapter 2

has in the same slot a linguistic expression that fails to entail S. Then if
one knew that S was true, one would be in breach of the ®rst maxim of
Quantity (my Q-principle) if one used (35b); hence the sentence in (35b)
implicates that the speaker is not in a position to use the stronger sentence
in (35a).
(35) a. ``The doctor knows that the patient will not recover.''
b. ``The doctor believes that the patient will not recover.''
Q ‡> `The doctor may or may not know that the patient will
not recover.'
There are many such pairsÐfor example, the verbs realize, think or
reveal, claim, as illustrated in (36).
(36) a. ``John realized that the accountant had made a mistake.''
Entails: `The accountant had made a mistake.'
b. ``John thought that the accountant had made a mistake.''
Q ‡> `The accountant may have made a mistake, or he may
not have.'
c. ``The government revealed that the ®les had been destroyed.''
Entails: `The ®les had been destroyed.'
d. ``The government claimed that the ®les had been destroyed.''
Q ‡> `The ®les may have been destroyed, or they may not have
been.'
In these cases, such paired expressions also form Horn scales and give
rise to the corresponding scalar implicatures. But they are distinct. For
example, from ``John believes Sue came'' we have, via the scale hknow,
believei, the scalar inference `(The speaker knows that) it is not the case
that John knows that Sue came'. That implicature is compatible with the
speaker knowing that Sue did not come. But the clausal implicature
from ``John believes Sue came'' is `For all the speaker knows, Sue came,
or Sue did not come'. Thus the clausal implicatures indicate epistemic
uncertainty about the truth of the embedded sentences, not the speaker's
knowledge of (or belief about) the negation of the matrix clause as in
scalar implicatures from the same construction.
The class of relevant verbs giving rise to clausal implicatures includes
most of the verbs of propositional attitude (unless like know or realize
they entail the truth or falsity of their complements) and most of the verbs
of saying (again, unless like disclose, divulge, admit they entail the truth or
falsity of their complements) (Gazdar 1979: 61). Table 2.2 gives some
111 The Phenomena

Table 2.2
Verbal doublets giving rise to clausal implicatures
Weak verbs not entailing Strong verbs entailing
their complements their complements
say, claim, hold disclose, divulge, reveal
deny, reject disprove, refute
predict foresee
believe know

examples of relevant verbs, with the column on the left displaying verbs
that do not entail their complements, and which will therefore in a frame
of the kind V that f give rise to clausal implicatures, whereas the column
to the right contains more informative expressions (which do entail their
complements) that might have been employed to avoid the implicature.
Note that there are not the same strict conditions here on expression alter-
natives that there are in the case of Horn scalesÐthe requirement simply
is that there should be available some alternative expression of roughly
equal brevity that due to its semantics would not carry the implicature.
It should be clear from these brief remarks that clausal implicatures
have not been heavily researched since Gazdar (1979) proposed this dis-
tinction. Most researchers have acknowledged that these are a distinct
subspecies of inference and that, by the Gricean argument sketched, they
should fall under the ®rst maxim of Quantity or, as here, the Q-principle
(but see Atlas 1993b for doubts). Some though consider there is no reason
at all to group them together with their scalar cousins (see Frederking
1996).
One issue I have left aside here is the relation of these implicatures
to presuppositions. Most analysts hold that presuppositions cannot be
reduced to matters of implicature and that presuppositions are attached to
their lexical or syntactic triggers (and are thus not detachable in Grice's
sense; see Levinson 1983: chap. 3 for review). On Gazdar's (1979: 59±60)
theory whereby clausal implicatures cancel inconsistent presuppositions,
it was therefore necessary to exclude the possibility that some expression
of the form X-f both presupposes and clausally implicates f, otherwise f
would be inevitably canceled. This was achieved by stipulation in that
theory (by blocking the troublesome clausal implicatures), but the rela-
tionship clearly requires further research.
112 Chapter 2

2.3 EXPLORING I-INFERENCES

2.3.1 Formulating the Maxim or Heuristic

Recollect that Grice (1975) proposed two maxims of Quantity:
Q1: ``Make your contribution as informative as is required (for the
current purposes of the exchange)''
Q2: ``Do not make your contribution more informative than is
required.'' (Grice 1975)
Is the second maxim (which I abbreviate Q2) otiose? Grice (1975: 46)
was unsure:
It might be said that to be overinformative is not a transgression of the CP
[Cooperative Principle] but merely a waste of time. However it might be answered
that such overinformativeness may be confusing in that it is liable to raise side
issues; and there may also be an indirect e¨ect, in that hearers may be misled as
a result of thinking that there is some particular POINT in the provision of the
excess information.
Perhaps the maxims of Relevance or Manner would handle these
``indirect e¨ects.'' And Grice (1989: 372) later repeated his doubts.
But the doubts seem misplaced, even from a strictly Gricean point of
view. Thus he found no trouble imagining the correlate of such a principle
in a nonverbal activity such as passing too many screws when one was
required (1975: 47). Zipf (1949) has provided statistical data from many
kinds of human activity, including language, that seem to support it. Two
observations seem to have worried Grice. One was that Relevance might
secure the same e¨ects because too much information would surely be
irrelevant (indeed, an infringement of the second maxim of Quantity may
also constitute an infringement of the maxim of Manner). The other was
that the principle did not, it seemed to him, yield the rich set of inferences
one would expect in the category of ¯outed or exploited maxims (only one
clear example being provided in Grice 1975: 52). But that, as it turns out,
is the wrong place to look.41 It is in the observance of the maxim that its
e¨ects are most apparent, for there is indeed strong evidence in favor of
its interpretive corollary, which as a ®rst approximation might be phrased
``Don't provide unnecessary information, speci®cally don't say what
would be obvious anyway'' (for much crosslinguistic evidence in favor of
such a heuristic, see Haiman 1985a). Such a principle might account for
our tendency to interpret utterances in line with our knowledge about
113 The Phenomena

what is normal or typical, as in (37a). It is clear that providing too much

information defeats this kind of interpretation, as in (37b), where the
excess information does indeed lead the recipient to think ``that there is
some particular POINT in the provision of the excess information'' as
Grice imagined.
(37) a. ``He opened the door.''
‡> in the normal way by turning the handle.
b. ``He opened the door by turning the handle quickly
anti-clockwise.''
‡> not in the normal way.
The evidence for such a tendency towards economy is overwhelming. As
Haiman (1985a: 150) puts it, ``there is a powerful tendency in languages
. . . to give reduced expression to the familiar and the predictable,'' a cor-
relation that is ``one of the most well-attested in human language.'' But a
tendency of this Zip®an sort could have sources other than the speaker's
orientation to some maxim or heuristic.
There is in fact a great deal of evidence for the existence of a speaker's
preference for minimal speci®cations, and its interpretive corollary, to be
found in the extensive literature on the empirical properties of conversa-
tion (reviewed in Levinson 1987a: 79±98). I will cite just one general ob-
servation here. Researchers have noted that there is a decided preference
for reference to persons to be achieved by the shortest expression, with the
least descriptive content, that will do the job. That this has the status of a
maxim, or expected procedure, is shown by the fact that when a speaker is
in doubt about whether a brief form will do, the form is often produced
with a distinctive rising intonation, and only when lack of response indi-
cates that the recipient has not understood, is a fuller speci®cation given,
also with rising intonation; and so on recursively, until recognition is
achieved. For example (from Sacks and Scheglo¨ 1979: 19, where the
question mark indicates the rising intonation):
(38) A: . . . well I was the only one other than that that the uhm tch
Fords?
uh Mrs. Holmes Ford?
You know uh the the cellist.
[
B: Oh yes. She's she's the cellist.
A: Yes.
114 Chapter 2

Unless there is some kind of maxim in operation, it would be hard to

understand why a speaker would not initially choose, or immediately
escalate to, a descriptive phrase that he knew would be more than su½-
cient to achieve successful reference. Note that a parallel pattern can be
observed in self-identi®cations on the telephone (Scheglo¨ 1979), where
(a) intimates expect their identity to be conveyed just by the sample of
voice quality in their ®rst ``Hello'', only escalating step by step, when
overt recognition is withheld, with further examples of voice quality, then
nicknames or ®rstnames: ``Hello (:) It's me (:) Steve (:) Levinson'';42 (b)
persons less intimate may escalate to self-descriptive phrases: ``Hello. It's
Les (:) Garston (:) the man who called about the plumbing.'' Those who
doubt the existence of some maxim of minimization (e.g., Carston 1995:
223) must therefore ®nd not only an account of the general tendencies for
minimization but also an account of this apparent orientation of speakers
to just such a principle.
I o¨er the following (after Levinson 1987b) as a ®rst approximation to
a characterization of the maxim involved (more elaborate formulations
and discussions can be found in Atlas and Levinson 1981, Levinson
1987a,b). Following the practice initiated by Atlas and Levinson (1981), I
call this rendition of Grice's Q2 maxim the Principle of Informativeness,
or the I-Principle for short.
(39) I-Principle
Speaker's maxim: the maxim of Minimization. ``Say as little as
necessary''; that is, produce the minimal linguistic information
su½cient to achieve your communicational ends (bearing Q in
mind).
Recipient's corollary: the Enrichment Rule. Amplify the
informational content of the speaker's utterance, by ®nding the
most speci®c interpretation, up to what you judge to be the
speaker's m-intended point,43 unless the speaker has broken the
maxim of Minimization by using a marked or prolix expression.
Speci®cally:
a. Assume the richest temporal, causal and referential connections
between described situations or events, consistent with what is
taken for granted.
b. Assume that stereotypical relations obtain between referents or
events, unless this is inconsistent with (a).
115 The Phenomena

c. Avoid interpretations that multiply entities referred to (assume

referential parsimony); speci®cally, prefer coreferential readings
of reduced NPs (pronouns or zeros).
d. Assume the existence or actuality of what a sentence is about if
that is consistent with what is taken for granted.
This characterization contains a number of unde®ned terms. The direc-
tion in which one would hope for clari®cation is mostly clear enough, but
none of these are trivial notions. For now, I can o¨er only the following
glosses:
1. Speci®city: p is more speci®c than q if (a) p is more informative than
q (e.g., p entails q); and (b) p is isomorphic with q (i.e., each term or
relation in p has a denotation that is a subset of the denotations of the
corresponding expressions in q). (For further remarks, see below.)
2. Stereotype: As Putnam (1975: 249¨ ) notes, a stereotype needn't
have a close relation to reality or statistical tendency (e.g., ``as ®erce as a
gorilla'' or ``absent-minded professor''). On Putnam's view, meanings are
stereotypes that may fail to determine correct extensions; on my view,
stereotypes are connotations associated with meanings, but not part of
them, which nevertheless play a role in interpretation.44
3. Informativeness: p is more informative than q if the set of states of
a¨airs that q rules out is a proper subset of the set that p rules out (Bar-
Hillel and Carnap 1952; Popper 1959). This account makes a number of
idealizations and thus can only be considered a ®rst approximation (see
discussion in Levinson 1987b: 404±406).
4. Minimal: One needs to distinguish semantic generality (crucial here;
e.g., building vs. hall ) from expression brevity (e.g., hall vs. auditorium).
Zipf (1949) argues that these tend to con¯ate (e.g., a pronoun is both
semantically general and brief ). This follows from the conjunction of his
Law of Abbreviation (`the more use, the shorter the expression') and his
Principle of Economic Versatility (`the more semantically general, the
more use'). Equally in the Prague School theory of markedness, the
unmarked expression in an opposition is both formally less complex and
semantically general, whereas the marked expression is formally more
complex and semantically more speci®c (Jakobson 1939; see also discus-
sion in Horn 1989: chap. 3).
This characterization is compatible with the original proposal in Atlas
and Levinson (1981). We sketched there a two-level account: ®rst, all the
competing interpretations are generated; then the most informative is
116 Chapter 2

selected. The principles generating the set of competing interpretations

were left open in that account, but clearly they should delimit the set of
possible interpretations. In the sketch just given, the relevant principles
are given a bit more ¯esh: they include the search for maximal cohesion,
or temporal, spatial, causal, and referential connectedness, and the pre-
sumption of stereotypical relations and the actuality of referents. Again,
in the earlier account, it was claimed that I-inferences:
enrich `what is said' by reshaping the range of possible states of a¨airs associated
with `what is said' to a narrower range of possible states of a¨airs associated with
`what is communicated.' `What is communicated' is MORE PRECISE than `what
is said'.'' In contrast, Q-inferences were held ``to shrink the range of possible states
of a¨airs . . . `What is communicated is MORE DEFINITE than `what is said',
(Atlas and Levinson 1981: 35±36, italics added)
The contrast intended was the idea that I-inferences are not just more
informative in the sense that they entail what is said (that is equally true
of most Q-implicatures): they introduce semantic relations absent from
what is said, and in that sense can be said to reshape the proposition
expressed (whereas Q-implicatures of the scalar type only introduce a
negative bound from within the same semantic ®eld). I have tried to cap-
ture this notion by building into the concept of more speci®city (replacing
the earlier concept of more precision) the requirement that although the
interpretation is (partially) isomorphic with the content of what is said,
each constituent may have a more restricted sense and denotation. Thus
the phrase a spoon in the cup may have the interpretation `a metal spoon
partially inside the ceramic teacup', went and bought may have the inter-
pretation `went and then bought', took a drink the interpretation `imbibed
an alcoholic drink', she bit her nails the interpretation `she bit her own
®nger nails', and so on. The isomorphism holds of course not at the level
of English phrases and glosses but at the abstract level of propositional
representation: what is implicated is a specialization of one or more of
each of the intensions of what is said.45 It corresponds to the fact, noted
below, that it is usually di½cult to express an I-implicature as an un-
related proposition conjoined to `what is said'.46
Now, under the I-Principle rubric, I want to collect a whole range
of inferences that appear to go in just the reverse direction to that in
which Q-implicatures tend. To emphasize the tension, we may say that
I-implicatures are inferences from the lack of further speci®cation to the
lack of need for it, whereas Q-implicatures are inferences from the lack of
informational richness to the speaker's inability to provide it.47 Under the
117 The Phenomena

rubric of the I-Principle, we can gather a range of well-known phenomena,

as follows (‡> stands for `implicates'; ‡‡> stands for `communicates',
i.e., the sum of what is `said' and the I-implicature in question):
(40) Conditional perfection (Geis and Zwicky 1971)
``If you mow the lawn, I'll give you ®ve dollars.''
I ‡> `If you don't mow the lawn, I will not give ®ve dollars' or
perhaps `If I give you $5, you will have mown the lawn.'
‡ ‡> `(If and) only if you mow the lawn, will I give you ®ve
dollars.'
(41) Conjunction buttressing (e.g., Atlas and Levinson 1981)
An utterance of the form ``p and q'' where p and q describe events,
such as ``John turned the key and the engine started.''
‡ ‡> `p and then q' (temporal sequence)
`p therefore q' (causal connectedness)
`A did X in order to cause q' (teleology, intentionality)
(42) Bridging (Clark and Haviland 1977)
``John unpacked the picnic. The beer was warm.''
‡> `The beer was part of the picnic.'
(43) Inference to stereotype (Atlas and Levinson 1981)
``John said `Hello' to the secretary and then he smiled.''
‡ ‡> `John said `Hello' to the female secretary and then John
smiled.'
(44) Negative strengthening (Horn 1989: chap. 5)
``I don't like Alice.''
‡> `I positively dislike Alice.'
(45) Preferred local coreference
``John came in and he sat down.''
‡ ‡> `John1 came in and he1 sat down.'
(46) Mirror maxim (Harnish 1976: 359)
``Harry and Sue bought a piano.''
‡> `They bought it together, not one each.'
(47) Noun-noun compounds (NN-relations) (Hobbs et al. 1993)
``The oil compressor gauge.''
‡> `The gauge that measures the state of the compressor that
compresses the oil.'
118 Chapter 2

(48) Specializations of spatial terms (Herskovits 1986)

a. ``The nail is in the wood.''
‡> `The nail is buried in the wood.'
b. ``The spoon is in the cup.''
‡> `The spoon has its bowl-part in the cup.'
(49) Possessive interpretations (Sperber and Wilson 1986)
a. ``Wendy's children''
‡> `those to whom she is parent.'
b. ``Wendy's house''
‡> `the one she lives in.'
c. ``Wendy's responsibility''
‡> `the one falling on her.'
d. ``Wendy's theory''
‡> `the one she originated.'
It is immediately clear that by comparison to, say, scalar Q-implicatures,
these inferences are heterogeneous, and the kinds of procedures involved
in actually deriving the inferences in question may di¨er signi®cantly. The
claim therefore is not that the same mental algorithms compute this range
of inferences but rather that there is a single overarching principle that
licenses them, legitimating the application of any of the subprocesses
involved, and that this by virtue of its overall relation to the Q- and M-
Principles regulates the conditions under which the particular inferences
are blocked or canceled. To belabor the point: the I-principle here oper-
ates as an instruction to ®nd an interpretation that meets certain require-
ments (positive, stereotypical, highly speci®c interpretations).
Now a question that immediately arises is: in what sense are these
I-inferences generalized ? Most of these inferences interact with shared
background presumptions, which might in principle vary, and thus the
inferences might have none of the cross-context, even crosslinguistic,
invariance that are the hallmarks of GCIs. But at a su½cient level of
abstraction, it is quite clear that the kinds of inferences here collectedÐ
for example, conjunction-buttressing, negative-strengthening, preferred
patterns of coreferenceÐdo hold as preferred interpretations across
contexts and indeed across languages. And at a slightly higher level of
abstraction, the di¨erent types collected can be seen to share the property
of maximizing the informational load by narrowing the interpretation to
a speci®c subcase of what has been said.
119 The Phenomena

Perhaps the sense of commonality between the di¨erent subtypes of

I-inferences can be reinforced by contrasting their properties with those of
the Q-inferences:
(50) Q- vs. I-inferences
a. Common properties of I-inferences
i. They are inferences to more speci®c interpretations: what is
implicated is a specialization of what is said.
ii. The inference is positive in character: the extension of what
is implicated is a proper subset of the extension of what is
said, the extension being restricted positively (not as with
scalar implicatures, by complementarity).
iii. The inference is typically guided by stereotypical
assumptions; to spell such inferences out more fully would
be both redundant and onerous, and perhaps (in the case of
euphemism, etc.) socially undesirable.
iv. The inference makes no essential reference to something else
that might have been said but wasn't (call this the absence of
a metalinguistic element).48
b. Contrasting properties of the Q scalar implicatures
i. Q-inferences are inferences to more precise interpretations:
e.g., the scalar inference `Some and not all x are G' is a
subcase of what was said: ``Some x are G'.
ii. The inference is negative in character: the extension of what is
implicated is a proper subset of the extension of what is said
because the complement of that proper subset is ruled out by
the negative scalar implicature (e.g., ``some'' ‡> `not all').
iii. There is no reference to stereotypical or indeed any
background (nonlinguistic) knowledge in the derivation of a
scalar GCI (although such knowledge may cancel a scalar
implicature).
iv. The inference is metalinguistic: it makes essential reference
to something else (the stronger scalar item) that might have
been said but wasn't.
Before proceeding, I should pause to deal with some problems that
have arisen with these observations. First, there are rival accounts of some,
if not all, of these phenomena. Let us turn ®rst to conditional perfection.
Van der Auwera (1995, 1997) claims that much older accounts in the heri-
tage of Ducrot o¨er a better explanation than the standard I-implicature
120 Chapter 2

one: conditional perfection is actually just a scalar Q-implicature; the

scale is of the form h((if p, then q) & (if r, then q)), if p then qi, and by
asserting the conditional, one implicates that there is no other conditional
with consequent q. As he acknowledges, this runs afoul of my constraint
on Horn scales requiring equal lexicalization,49 but he jettisons that con-
straint and in general reanalyzes the nature of scales in favor of ranked
propositions of arbitrary complexity. This is just a di¨erent theory of
Q-type inferences, and I shall leave it to the reader to adjudicate (the
theory seems interestingly allied to Hirschberg's PCI account of scalar
implicature, which might subsume it). Carston (1995: 221±222) notes that
the conditional perfection implicature is actually generalized to the point
that it may run against stereotypical assumptions: ``If you don't mow the
lawn, I'll give you ®ve dollars'' now suggests that not mowing the lawn
is a necessary condition for the speaker's gift. In fact, it is noteworthy that
any form of words that suggests a su½cient condition is liable to enrichment
to a necessary condition: ``Mow the lawn and I'll give you ®ve dollars'', ``Go
on: earn ®ve dollars and mow the lawn'', and so on.50 Thus the nature of
the inference licensed by the I-principle must be to the necessary condition,
provided that is consistent with what is taken for granted (in ``If the door
is locked, I have a key'', the inference does not go through).
Second, some commentators have pointed out a failure of precision, at
least in earlier formulations. Carston (1995) for example points out that
whereas I have represented Q-implicatures as just the content implicated,
not the conjunction of implicature with what is said, I have represented
I-inferences in terms of the total signi®cation of what is said plus what is
implicated (Levinson 1987a). Correcting this, the conditional-perfection
implicature from ``If you mow the lawn, I'll give you ®ve dollars'' should
read `If you don't mow the lawn, I won't give you ®ve dollars', with the
total communicated content `I¨ you mow the lawn, will I give you ®ve
dollars'. Carston goes on to note that the implicature alone does not
entail the content of what is said, as she thought I claimed, and moreover
that as a result my distinction between Q- and I-inferences collapses.51
The failure of precision was regrettable, and occasionally where it matters
I use the symbol ‡> (and the term implicate) for the pure implicated
content, and ‡‡> (and the barbaric *implicate) for the total communi-
cated or conveyed content (or at least for the sum of what is said and the
relevant implicatures being discussed). But there is in fact a reason for the
lapse: it is di½cult, in probably most cases of I-implicature, to represent
the implicature alone without that implicature entailing the content of
121 The Phenomena

what is said; for example, in the case of conjunction buttressing where

utterances of the form p and q implicate something like `p and then q' (or
`after p, q'), the latter entails the content of what is said. This is not an
accident: it follows from the nature of I-implicatures that they are enrich-
ments of what is said, whereas Q-implicatures merely exclude something
else that might have been said but wasn't. Thus we can often represent the
Q-implicature as a conjunct which does not itself include the content of
what is said, whereas I-implicatures are not so satisfactorily represented
in this way. For this reason, I will normally cite I-*implicatures in their
inclusive form without special typographical conventions.
However, the lapse has none of the dire theoretical consequences that
Carston (1995: 220) hopes for. As stated (apparently not clearly enough)
in Levinson 1987a and three paragraphs above, in the case of both Q- and
I-inferences (and indeed in all implicatures other than ¯outs, exploitations,
or tropes), the communicated content will entail what is said (otherwise,
what would be the value-added content of implicatures?). The di¨erence
lies in the character of the value added to what is said: in the case of
I-implicatures the inferences are positive strengthenings towards the
stereotypical, which make no metalinguistic reference; in the case of Q-
inferences they are inferences to the negation of statements employing
stronger alternative linguistic expressions. I-inferences import rich assump-
tions about the world; Q-implicatures circumscribe what has been said by
what else might have been said but wasn't. One would of course wish to
be more precise, but the characterization seems to me intuitively clear.
The di¨erence in character between Q- and I-implicatures can perhaps
best be appreciated by considering how they might yield inconsistent
implicatures. As mentioned, they appear to generate implicatures in op-
posite directions: informally the Q-principle speci®es ``What you haven't
said you do not mean (otherwise you should have said so)'', whereas the
I-principle can be glossed ``What you haven't bothered to say, you expect
me, the recipient, to supply.'' Thus, if we permit a scale hsince, andi,
which in fact is barred by the constraint that scalar items must be ``about''
the same semantic relations, then we would obtain implicatures that are
not only inconsistent but actually contradictories:
(51) ``John turned the key and the engine started.''
I ‡‡> `Since John turned the key, the engine started.'
Pseudo-scale: hsince, andi
no Q ‡> `It is not the case that: Since John turned the key, the
engine started.'
122 Chapter 2

In this case the clash is averted by the constraints on Horn scales (see
section 2.2.2). But there are cases where genuine clashes occurÐthat is,
where the potential implicatures from Q- and I-principles are inconsistent.
In these cases, it appears, as a matter of observation, that Q-inferences
always take precedence (or cancel) I-implicatures. It will transpire that
single sentences will often engender multiple implicatures. For example,
in (52) the disjunction gives rise to a clausal Q-implicature that defeats the
I-inference to conditional perfection.
(52) ``If you come, he'll forgive you, or even if you don't come he may
forgive you anyway.''
Q-implicature: `It's possible that even if you don't come he may
forgive you anyway.' [clausal implicature from disjunction]
Potential I-implicature: `If you don't come he will not forgive you.'
Q-implicature defeats I-implicature.

2.3.2 Some Prominent I-Implicatures

2.3.2.1 Conjunction-buttressing and Parataxis Let us now review in a

bit more detail some of these kinds of I-implicature. I have noted that
conjunction is ripe for I-enrichment: when events are conjoined, they tend
to be read as temporally successive and, if at all plausible, as causally
linked.52 There are many rival accounts of this phenomenon (see Lang
1984 for an attempt to embed this problem in the larger range of possible
conjunction interpretations). Partee (1984) has suggested that the tempo-
ral succession reading works a bit like anaphora: a past-tense event de-
scription establishes a reference interval R (intuitively just after the event)
that is picked up by the next past tense; if the succeeding clause describes
an event, R includes the event; if it describes a state, the duration of the
state includes R.53 Similarly, Kamp and Reyle (1993: 521¨ ) note a pre-
sumption of temporal vicinity, which in the case of descriptions of states
will be read as temporal inclusion and in the case of events as temporal
succession. They observe that this seems not to depend solely on world
knowledge and appear to conclude that the matter is semantic or that it
at least must be captured in the semantic representation or DRS. The
pragmaticist has in fact no quarrel with the representational ideas here
but insists on the pragmatic character of the inference, with important
implications for the role of pragmatics in the construction of semantic
representations, an issue taken up in chapter 3. The formal-semantics
tradition, incidentally, says nothing about the causal and teleological
123 The Phenomena

implicatures from p and q, so in any case it needs supplementing with

a pragmatic account. But what the pragmatist will insist on is that the
succession inference is pragmatic, hence defeasible under conditions where
the contrary is assumed, or where it is explicitly or implicitly canceled:
(53) a. ``Hans wrote a novel and he sold the rights to Macmillan.''
I ‡‡> `Hans wrote a novel and then sold the rights
to Macmillan.'
b. ``Hans wrote a novel and he sold the rights to Macmillan in
advance.'' (implicature canceled by the interpretation of in
advance)
c. ``Hans wrote a novel and sold the rights to Macmillan but not
necessarily in that order.'' (implicature canceled by explicit denial)
d. ``Hans wrote a novel and made up the plot as he commuted to
work.'' (implicature defeated by our knowledge that one must
make up the plot while or before writing a novel).
Equally, of course, our knowledge of the world may defeat default inter-
pretations. Thus sequentiality is overruled by assumptions of simultaneity
in (54a±b) (Schmerling 1975: 214).
(54) a. The lights went o¨ and I couldn't see.
b. Joan sung a ballad and accompanied herself on the guitar.
Occasionally the very search for some rich connectedness may require the
rejection of the `and then' interpretation in favor of a reversed sequence
(prosodic clues being important here):
(55) a. He got a Ph.D. and he only did a month's research!
b. But she won an Olympic medal and she was born in a slum and
undernourished as a child.
Most authors agree that although the inference to temporal sequence is
defeasible, and in some conjunctions hardly seems to even arise, it is
nevertheless crosslinguistically the default interpretation of two conjoined
past-tensed event descriptions without special prosodic or other marking
(see, e.g., Haiman 1985c; Lascarides and Asher 1993). Nonce-implicature
theorists may want to insist that it all depends on content and context and
that there is no such default interpretation, but Carston (1995: 232) points
out that even when our knowledge of the world would suggest `q before
p', still we attempt to read ``p and q'' as `p before q':
(56) He opened the door and she handed him the key.
124 Chapter 2

Clearly there are language-speci®c semantic constraints on this sequen-

tial interpretation, arising from tense, aspect, and Aktionsart, gapping,
and reduction, which are highly complex. Although He researched the
subject and gave some lectures carries the normal implicature, there are no
sequential implicatures associated with He is researching the subject and
he is giving some lectures, He was researching the subject and giving some
lectures, He researches the subject and he gives some lectures, or He used to
research the subject and he used to give some lectures, which encode over-
lapping, ongoing activities. If the Aktionsarten of the verbs are mis-
matched, one suggesting a state the other a dynamic event, we tend to get
an inclusion reading as in He knew he was wrong and apologized, or He
slept and fell o¨ the chair, or we may force an habitual simultaneous
reading out of the dynamic verb as in He lived in Cambridge and traveled
to London. As well, some sentences with and simply do not have an
underlying conjunction semantics: She went and spoiled her dress or I've
got to try and ®nd it involve complex idiomatic expressions of the form go/
try and VP (Schmerling 1975). In fact, conjunction reduction in English
seems often to grammaticalize speci®c interpretations (as Schmerling
documents). All of these factors are well known, if not well understood,
and I mention them simply to emphasize that the I-implicature can only
go through when not blocked by this web of semantic factors.54
Now these same interpretive tendencies are observable in parataxis or
the unmarked adjunction of clauses, which is as predicted: there is no
possible Horn scale between nothing (i.e., sheer adjunction) and other
linguistic expressions that might give rise to Q-implicatures that would
block the I-inference to maximal connectedness (however, I will show that
the possibility of a Manner contrast between explicit conjunction and
parataxis does arise). Hence (57a) and (57b) suggest their respective
implicatures, pretty much as the corresponding sentences with and would.
(57) a. ``John turned the switch. The motor started.''
I ‡‡> `John turned the switch and then as an intended result
the motor started.'
b. ``John lost his grip on the cli¨. He fell and broke his leg.''
I ‡‡> `John lost his grip on the cli¨ and then as an
(unintended) result he fell and broke his leg.'
c. ``John fell and broke his leg. He lost his grip on the cli¨.''
I ‡‡> `John fell and broke his leg. The reason was he (had just
before) lost his grip on the cli¨.'
125 The Phenomena

d. ``John fell and broke his leg and he lost his grip on the cli¨.''
No implicature: `John fell and broke his leg. The reason was he
lost his grip on the cli¨.'
Some have even considered that parataxis, or more properly asyndetic
conjunction, should be interpreted simply as elided conjunction (Lang
1984: 80). But the range of possible invited inferences is much wider with
parataxis as shown by (57c) with the assumed temporal order reversed
from the clauses, an interpretation not available to the same sentence with
an explicit conjunction as in (57d) (Gazdar 1979: 44). In fact, sheer
adjunction of clauses is often su½cient to invoke a diversity of complex
interclausal relations, as in Bloom®eld's ``It's 10 o'clock, I have to go
home'' (see Matthews 1981: 32¨; Hobbs 1979). Indeed, many languages
make do without the complex set of intersentential connectives available
in English, using parataxis instead. For example, the Australian language
Guugu Yimithirr lacks sentential disjunctions, indicative conditionals,
and sentential conjunctions, which are all signaled paratactially (in the
case of disjunctions and conditionals, with the addition of dubitative or
emphatic particles). The following examples indicate how one would
translate English sentences with connectives: 55
(58) English target: `Either he died or he didn't.'
Nyulu biinii nguba, nguba gaari.
he died perhaps, perhaps not
`Perhaps he died, perhaps he didn't.'
(59) English target: `If you go, you'll get meat.'
Nyundu budhu dhadaa, nyundu minha maanaa bira.
You maybe go you meat get for sure
`Maybe you will go, (then) you will certainly get meat.'
(60) English target: `I came and he left.'
Ngayu gaday; nyulu dhaday.
I came he left
`I came. He left.'
Although it may be unusual for the main way of expressing conditionals
to be parataxis (plus or minus emphatic particles), many languages can
optionally express conditionals through parataxis. In English, we do this
by, for example, conjoining or adjoining an imperative and an indica-
tive as in (61a±b), but in Vietnamese (as well as Chinese, Hindi, Hua, and
126 Chapter 2

in many other languages) two ordinary indicatives will do, as in (61c)

(Haiman 1985a: 44±5).56
(61) a. Touch that chest and I'll scream.
b. You're so smart, you ®x it.
c. KhoÃng co maÁn khoÃng chiu noà i.
Ç
not be net not bear can
`If there's no net, you can't stand it.'
The basis for an I-implicature from an utterance of the form ``p. q'' to `if
p, then q' is not clear. But Haiman suggests that in both cases there is a
relation between ordering and the way in which the second sentence is
informationally processed (according to Greenberg's (1966a: 84) universal
14, conditionals in all languages tend to have a canonical order with the
antecedent ®rst). Thus in conditionals, the antecedent provides the con-
text for the assessment of the consequent.57
Parataxis is an important instance of the tendency to ®nd from minimal
speci®cations maximally cohesive, rich interpretations. Some theorists,
like Hobbs (1979), suggest that this kind of inference is possible because
there is in fact a ®xed taxonomy of possible clause relations, and the
recipient need only select between the half dozen or so major relations in
order to ®x an interpretation.
We have noted that sentences joined by parataxis often invite the same
inferences as conjunctions. But we have also noted that paratactic infer-
ences are a broader class. In some languages, there is a clear opposition
between explicit conjunction and sheer parataxis, with parataxis suggest-
ing a closer more integrated sequence of events, and sentences explicitly
conjoined suggesting that the events either took place at a di¨erent time,
or were not connected, or involved di¨erent protagonists, and so forth
(Haiman 1985a: 111±115). An account of this can be given in terms of the
I- and M-principles, in a pattern that will shortly become familiar: mini-
mal forms invite maximally rich interpretations, maximal forms defeat
them.58 Haiman goes on to suggest that there is a cline of the form
conjunction > parataxis > gapping=reduction where the more reduced the
form, the greater the suggestion that the two events are actually a single
complex one.59
The classic cases of ``bridging inferences'' (Clark and Haviland 1977)
occur in paratactic (asyndetic) constructions (but the inferences go
through in explicit conjunctions too), and they seem to be forced by the
attempt to ®nd coherence across the two conjoined or adjoined clauses.
127 The Phenomena

Note that although they may utilize background knowledge as in (62a)

(where we all know that every car has a steering wheel), or activate frames
as in (62b) (all dining rooms have windows and some dining rooms have
French windows), they are completely independent of such information.
(62) a. ``Harold bought an old car. The steering wheel was loose.''
I ‡> `The steering wheel of the car.'
b. ``Patience walked into the dining room. The French windows
were open.''
I ‡> `The dining room had French windows'60
c. ``Hilda climbed into the capsule. The proton-thruster was
attached to the console.''
I ‡> `The capsule had a proton-thruster.'
d. ``Hilda climbed into the capsule. A proton-thruster was attached
to the console.''
I ‡> `The capsule had a proton-thruster.'
It is the presumptive quality of the inference, even in the absence of
stereotypical information, that is a hallmark of I-implicature. Notice that
the presumption doesn't seem to reside in the de®niteness of the NP (as
accounts in terms of accommodation presume), because both (62c) and
(62d) carry the same I-implicature (but di¨er in inferences as to the
uniqueness and normality of proton-thrusters in capsules; see Hawkins
1978, 1991).

2.3.2.2 Negative Strengthening

2.3.2.2.1 Gradable Antonyms There are a large range of implica-

tures associated one way or another with negative statements (see Horn
1989). Negative scales, as we have already seen, are regular sources of
Q-implicatures (thus ``Not all of them came'' suggests `Some of them
came'). Double-negatives, litotes (like a not insigni®cant achievement),
and their implicatures fall under my third maxim, the M-principle.
Sometimes, as we shall see in chapter 3, implicatures fall under the scope
of negation (so-called metalinguistic negation).
But various kinds of negative strengthening are clearly attributable to
the I-principle. As a ®rst example, consider the gradable antonyms
(Lyons 1977: 271¨ ) like fhot, coldg, flarge, smallg, fgood, badg, and so
on, which allow explicit grading (as in this is hotter than that). Although
these are clearly semantically contraries (something can be neither hot nor
128 Chapter 2

Figure 2.4
Contradictories implicating contraries

cold, neither good nor bad, etc.), it has long been noted that the negation
of a gradable adjective tends to be interpreted as if the antonyms were
themselves contradictoryÐthat is, as if not good meant bad excluding
middle ground (see Lyons 1977: 278, who suggests a Gricean account;
and Horn 1989: 205¨ ). This amounts to the following (potentially con-
fusing) conclusion: the negative of a gradable, which is in fact a weak
contradictory including the middle ground, implicates the contrary, its
antonym, excluding the middle ground. Figure 2.4 may help to clarify and
show how this implicature informationally restricts, and thus strengthens,
the negative predicate exempli®ed in (63).
(63) ``It was not a good thing to do.''
I ‡‡> `It was a bad thing to do.'
The implicature appears to be restricted to the denial of the unmarked
antonym (good, tall, large etc., as shown by the neutral question form,
e.g., How tall is he? which does not presume that he is tall). The process is
fairly, if not entirely, general. It may be partly motivated by the polite
avoidance of disparaging antonyms or by euphemism. Thus Greenberg
(1966b, cited in Horn 1989: 336) notes that many languages have no word
for `bad' because it is so routinely implicated by `not good'. Note that, as
discussed above under the Q-principle, gradable predicates of these sorts
sometimes form scales, with associated implicatures from the assertion of
the weaker expressions and from the denial of the stronger ones. One might
therefore expect complex interactions between Q- and I-implicatures here,
but on the whole they do not occur. This is because (a) the scales are split
and never encompass both antonyms (there is no scale  hgood, badi as
there is hexcellent, goodi) and (b) the unmarked pair of the gradable
predicate (e.g., good ) is usually the weaker item in a scale with an extreme
value as the stronger (e.g., hexcellent, goodi with negative scale hnot-
129 The Phenomena

good, not-excellenti) and because weaker items become the stronger item
in a negative scale, they fail to generate Q-implicatures (``not good '' has
no scalar implicatures).61 Additionally, gradable predicates like like
appear to yield similar inferences, here to the contrary `dislike', as in (64).
(64) ``I don't like garlic.''
‡> `I dislike garlic.'

2.3.2.2.2 NEG-Raising (NR) There is a tendency for negative main

sentences with subordinate clauses to be read as negations of the sub-
ordinate clause:
(65) a. ``I don't think he is reliable.''
I ‡‡> `I think he is not reliable.'
b. ``I don't think that the suspect wanted to believe it was a crime.''
I ‡‡> `I think that the suspect wanted to believe it was not a
crime.'
The historical ruminations on this phenomenon go back at least to
St. Anselm, as Horn (1989: 308¨ ) describes. The remarks in this section
are based on Horn's excellent account. In the period of transformational
grammar when it was assumed that there should be a direct relation be-
tween meaning and deep structure, it was proposed that this interpreta-
tion, whereby the sentence is read as if the negation occurred in the lower
clause, should be explained by supposing that the negative was downstairs
in deep structure and derivationally raised (hence the epithet Neg-raising
[NR], which the modern reader may prefer to think about in terms of
interpretive Neg-lowering). In that case, it was noted, the negative could
be moved from arbitrarily low down in the sentence, providing the pred-
icates over which it jumped were of a certain type (see (65b)). The relevant
predicates governing this phenomenon appear to be verbs of mental
activity or desire (think, believe, want) or impersonal verbs embedding
sentences like It doesn't seem likely or It is not to be expected that he'll
come. But why?
The syntactic analysis no longer ®ts with today's syntactic theories, and
in any case it runs into trouble with the fact that allegedly raised negatives
can be lexicalized as in ``Nobody believes he'll do it'', which implicates
`Everyone believes he won't do it' (Horn 1978). This leaves the possibil-
ities of explanation in terms of semantic ambiguity or pragmatic enrich-
ment. The essential semantic observations are as follows:
130 Chapter 2

1. NR sentences have two interpretations:

(66) ``I don't believe that he came''
a. `I believe that he did not come'
(the NR-reading or narrow-scope interpretation)
b. `It is not the case that I believe that he came'
(the `literal' L-reading or wide-scope interpretation)
2. The NR-reading (or interpretation) entails the literal L-reading,62
but not vice versa; the NR-reading is thus more informative than the
L-reading. Or to put it another way, the NR-reading is the stronger con-
trary, whereas the L-reading is the weaker contradictory (see ®gure 2.5):
(67) ``It's not probable that the Dow Jones will recover''
‡‡ > `It's probable it won't recover.'
3. Only some verbs permit the NR-reading; many verbs of proposi-
tional attitude, or other verbs allowing internal negations of subordinate
clauses, like know, realize, be certain do not (e.g., ``I don't know that he
came'' cannot be read `I know he didn't come'). In general, the class of
verbs permitting NR seems to be a subset, in any language, of verbs of
opinion, volition, obligation, probability, and perception (Horn 1978,
1989: 322±323).
(68) a. I don't believe the USA will put another man on the moon
(contradictory: @Bp)
b. I believe the USA will not put another man on the moon
(contrary: B @ p)
Let us set aside the possibility that negated NR predicates could just be
ambiguous in some way (in e¨ect the claim that, e.g., not believe has
an idiomatic, noncompositional reading equivalent to the meaning of
believe not). There are intrinsic impediments to such a claimÐfox exam-
ple, Nobody believes we can win a nuclear war would require, on an idiom
theory, another idiom linking Nobody believes p with the meaning of

Figure 2.5
Neg-raising as implicated contraries (see example 67)
131 The Phenomena

Everybody believes that not-p, and so on for every appropriate incorpo-

rated negative and every NR predicate. Second, there are generalizations
to be captured by taking the alternative, pragmatic approach that are
unavailable to the idiom/ambiguity approach. The assumption taken
here is that NR-interpretations are in fact I-implicatures (not, e.g., Q-
implicatures) for reasons that will become clear.
The central problem for understanding the phenomenon is under-
standing the third point above, why just some predicates allow the NR-
reading. The categories of predicates responsible have resisted any simple
semantic generalizations (indeed have ba¿ed generations of linguists),
and the processes are indeed not entirely predictable. But there is a good,
if not exceptionless, overall generalization about which predicates allow
the NR-reading, which can best be stated in Horn's arithmetic scalar
model. Take the set of predicates indicating epistemic states, states of
knowledge, opinion, and the like: in English, these include know, realize,
be certain, be evident=clear=sure, be odd=signi®cant, believe, suppose,
think, be likely=probable, seem, appear, and so on. Let us extract as a
coherent set that form a pair of positive and negative scales the epistemic
adjectives hcertain, likely, possiblei with corresponding negative meta-
scale himpossible, not likely, not certaini. Then we can map these onto the
arithmetic square of oppositions as shown in ®gure 2.6 (from Horn 1989:
325), with the strong items in a contrary relation, the weak items in a
pragmatic subcontrary relation, and the diagonal lines across the square
(e.g., from certain to not certain) constituting contradictories.
By LoÈbner's test, likely and not likely are intolerant: *It's likely she'll go
and likely she won't, so are placed above the 0.5 midline, whereas possible
and not certain are tolerant (It's possible but not certain she'll go) and
therefore belong below the midline. By the suspension tests likely is below
certain (It's likely, and indeed certain . . . but ??It's certain, and indeed
likely . . .) and so should be placed just above the midline. Now it turns
out that it is predicates in this position, just above the midline like likely,
that permit NR-interpretations: thus ``It's not likely that she'll come''
suggests `It's likely that she'll not come'. Stronger intolerant predicates
generally do not (``It is not certain that she'll come'' fails to implicate `It is
certain she will not come'63). And weak, tolerant ones like possible like-
wise do not have NR-readings (``It's not possible she'll come'' fails to
implicate `It's possible she will not come').
Horn's generalization is that NR-readings can only occur with (a rela-
tively weak) intolerant predicate P if both the contrary P@ and the
132 Chapter 2

Figure 2.6
Epistemic predicates in the arithmetic square

contradictory @P are close to each other and to the 0.5 midpoint, as

illustrated in ®gure 2.7.
Another way of putting the same generalization is to say that NR-
interpretations are, as noted above, from the statement of a contradictory
(e.g., @P) to the inference of a contrary …P@†. Because only tolerant
predicates can have contraries (intolerant ones below the 0.5 line will have
only subcontraries), it follows that only intolerant predicates (scalar items
above the 0.5 line) yield NR-readings. Besides, the contradictory of a
weak scalar like possible (namely, impossible) is informationally stronger
than its subcontrary ( possible not), so there can be no implicature to a
stronger subcontrary. But why should strong intolerant predicates (e.g.,
know, be certain) be barred from NR-interpretations? The answer that
Horn gives (1989: 326¨ ) is that the NR- interpretations are only possible
when the di¨erence between them and the literal L-interpretations carries
minimal functional load: ``a lower-clause understanding for higher-clause
negation tends to be possible only when the two readings for the outer-neg
133 The Phenomena

Figure 2.7
Restricted locale for Neg-raising within the square

sentence that would result are almost, but not quite, truth-conditionally
identical'' (Horn 1989: 329). Clearly this condition is not met in the case
of a strongly intolerant, strong scalar item like know: there is an enormous
informational gulf between the contradictory (not knowing that p) and the
contrary (knowing that not p): NR-interpretation here would magically
convert our ignorance into knowledge (would it were so easy)!
I return now to the question: what type of implicature is responsible for
NR-understandings? It is immediately evident from the ®gure above that
NR-interpretations cannot be due to the Q-principle. In fact, they work
in exactly the reverse direction: they run upwards along the negative
scale (here hknow-not, not-believei). But this of course is just what the Q-
principle in general disallows: if there is a stronger alternate expression on
a scale, then the Q-principle enjoins you to use it if it is applicable. That is
why there is no implicature from ``I don't know that p'' to `I know that
not p': if you had meant the latter, you should have said it. Inferences that
do work in the reverse direction are I-implicatures, and NR-inferences
must be I-implicatures. They have the necessary hallmarks: (a) they are
inferences to a more informative proposition that would entail what is
said (the implicature as usual for I-inferences not being representable as
an independent proposition from what is said); (b) they do not impose an
additional negation on the already negative understanding of the utter-
ance; (c) they are not metalinguistic (there is no necessary reference to
other linguistic itemsÐthe inference is from a contradictory to a con-
trary); (d) as in many other kinds of I-inference that invite a stronger
134 Chapter 2

interpretation from a weaker statement, they have extensive euphemistic

uses (Leech 1983: 101f; Horn 1989: 330±337, 350±352).
In fact, NR-inferences are the one systematic class of exceptions that I
know of to the generalization that that Q-implicatures block inconsistent
I-implicatures. But they are the exceptions that prove the rule: they are a
kind of minor, highly delimited creep up a Q-scale:
1. NR-interpretations are restricted to cases where they can carry only
the smallest functional load, where the semantic di¨erences between
the asserted contradictory and the implicated contrary are at a mini-
mum.
2. NR-interpretations, although they are not unmotivated idioms, are
idiomatic or governed by conventions of use (see Searle 1975: 76 for
the distinction; also Morgan 1978):
a. One can ®nd pairs of almost synonymous predicates where one
allows NR-interpretations and the other doesn't: compare ``I don't
want to go'' I ‡‡> `I want to not go' with ``I don't hope to go''
with no implicature `I hope to not go'.
b. NR-interpretations are often signaled by the inclusion of negative
polarity items that occur as if the negation was syntactically in the
lower clause. Thus one can say I don't think he knows anything at
all, but not *I think he knows anything at all. Similarly I don't want
to go at all is ®ne, but ?*I don't hope to go at all is odd. These are
the hallmarks of partial conventionalization or ``short-circuited
implicature'' (Horn and Bayer 1984; Linebarger 1991).
In short, the systematic crosslinguistic tendency for NR-interpretations
can only be explained in terms of a general pragmatic process like I-
implicature.64 No theory of nonce implicature can explain the regular
tendency. On the other hand, its domain of application is strictly bounded
by the other informational maxim, the Q-principle: in fact, the very
occurrence of NR is a leakage in the blocking mechanism whereby the
Q-principle bounds the I-principle. We can be grateful for the leakage
because it makes clearly visible the antinomic character of the two prin-
ciples (which some like Carston (1995) have doubted), even if it is a bit of
a theoretical embarrassment. The embarrassment is lessened by noting
that NR-inferences are protected from extinction by partial convention-
alization. The full path of conventionalization can also be noted, whereby
not likely becomes lexicalized as unlikely, which can then only have the
contrary, `likely not' reading associated with NR.
135 The Phenomena

2.4 M-IMPLICATURES AND HORN'S DIVISION OF LABOR

Grice (1967, 1989: 27) introduced his ®nal maxim thus: ``Under the cate-
gory of Manner, which I understand as relating not (like the previous
categories) to what is said but, rather, to how what is said is to be said, I
include the supermaximÐ``Be perspicuous''Ðand various maxims such
as:
1. Avoid obscurity of expression
2. Avoid ambiguity.
3. Be brief (avoid unnecessary prolixity).
4. Be orderly.''
Most modern accounts reduce these in one way or another, by rejecting
some and sweeping the remainder into a maxim of Relevance (Sperber
and Wilson 1986) or by reassigning the ®rst two submaxims to the Q
principle and the last two to an R principle (Horn 1984, 1989: 194). As
far as the present account goes, in so far as the second submaxim (``avoid
ambiguity'') is interpreted to include semantic generality, it will be par-
tially accounted for under the Q-principle (because it will in general be
less informative to express a general interpretation where it could have
been avoided by choice of another expression). Genuine ambiguity is
otherwise rather hard to avoid: the simplest sentences tend towards mul-
tiple ambiguities, as work in natural language processing has demon-
strated. The fourth submaxim, ``be orderly,'' relates to the larger issue of
iconicity in language: because speech is linear, the main dimension of
iconicity is order. I take this to be a natural cognitive propensity, whose
implications have been explored elsewhere (Haiman 1985a,b), and I have
some doubts that it has maxim status, because its e¨ects may be largely
achieved by the I-principle.65 The third submaxim, ``be brief,'' has both
an informational and a formal aspect. The informational aspect is sub-
sumed under the I-principle (maxim of Minimization), whereby semanti-
cally general expressions pick up speci®c reference, and I noted that by
Zip®an principles, semantically general expressions also tend to be more
frequent and thus briefer in form. Nevertheless, the formal aspects of
``avoid unnecessary prolixity'' (M3) and the ®rst submaxim (M1), ``avoid
obscurity of expression,'' survive in my scheme as a single Manner prin-
ciple. Manner elements survive in rival schemes, too, but not as indepen-
dent principles. In Sperber and Wilson's (1986) Relevance theory, marked
expressions require more processing e¨ort and thus contribute to the
136 Chapter 2

estimation of Relevance; in Horn's (1984, 1989) scheme of Q- and R-

principles, unmarked expressions are encouraged by R and marked ones
signal that Q is in force. Horn thus con¯ates informational minimality
and expression brevity within R, and informational maximality with
expression markedness within Q. The underlying reason to resist this
con¯ation in my scheme is that the three principles Q, I, and M are
here considered heuristics, which are invoked by the kinds of linguis-
tic expressions employed: just as the use of an item from a contrast set
suggests that the contrastive items would be inappropriate, or the use
of a minimal expression invokes a maximal interpretation, so the use of
a marked expression signals an opposing interpretation to the one that
would have been induced by the use of an unmarked expression. Each
heuristic has its mode of invocation, and in the case of the M-heuristic, it
is the form of the expression rather than its meaning that performs this
service.
The version of Grice's Manner maxim that I favor is one that reduces it
to a single simple principle that has been much commented on by analysts
with other axes to grind, and in fact Haiman's (1985a: 147) formulation
(grinding his own axe, the iconicity underlying much grammar) is quite
similar to my (shamefully) independent formulation of the same era. He
suggests, as a ®rst approximation, that ``Formal complexity corresponds
to conceptual complexity.'' But a better rival formulation, ``Morphologi-
cal markedness corresponds to semantic markedness,'' he points out, is
not quite a corollary because semantic markedness encompasses not only
conceptual complexity but also unfamiliar, infrequent concepts. It is, then,
this relation between formal markedness and semantic markedness that
is the relation in question.
The M-principle was introduced in chapter 1 as a general heuristic, and
needs only a little further elaboration here.
(69) The M-Principle
Speaker's maxim: Indicate an abnormal, nonstereotypical situation
by using marked expressions that contrast with those you would
use to describe the corresponding normal, stereotypical situation.
Recipient's corollary: What is said in an abnormal way indicates an
abnormal situation, or marked messages indicate marked situations,
speci®cally:
Where S has said ``p'' containing marked expression M, and
there is an unmarked alternate expression U with the same
137 The Phenomena

denotation D which the speaker might have employed in the

same sentence-frame instead, then where U would have I-
implicated the stereotypical or more speci®c subset d of D, the
marked expression M will implicate the complement of the
denotation d, namely d of D.
The notion of markedness employed here is a generalization of the Prague
School concept (as in Jakobson 1939; Greenberg 1966b; Horn 1989: chap.
3). On the formal side, marked forms, in comparison to corresponding
unmarked forms, are more morphologically complex and less lexicalized,
more prolix or periphrastic, less frequent or usual, and less neutral in
register. On the meaning side, such forms suggest some additional mean-
ing or connotation absent from the corresponding unmarked forms. The
notion of same sentence frame, which needs further explication along
structuralist lines, is intended to restrict the substitute U versus M ex-
pressions, so that, for example, the opposition between actives and pas-
sives will not come under this principle, which is of restricted scope.

2.4.1 Horn's (1984) Division of Pragmatic Labor

The M-principle as here stated builds in Horn's (1984, 1989: 197¨ )
``division of pragmatic labor,'' on which indeed the present idea is based
(Horn in turn borrowing from McCawley 1978). For the record, Horn
states the division of pragmatic labor in terms of his Q- and R-principles:
his Q is pretty much equivalent to both my Q plus my M, whereas his
R principle is roughly coextensive with my I-principle. Thus where the
implicatures of unmarked alternates pick up the central, stereotypical
subset of denotations by my I and his R, the marked expressions implicate
the complementary extension by my M but his Q. (I hardly need to point
out that the roots of these ideas are as old as structural linguistics: clear
formulations of similar principles can be found in BreÂal 1900, Saussure
1916, and Bloom®eld 1933).
The central observation is that M-implicatures seem to be essentially
parasitic on corresponding I-implicatures: whatever an unmarked expres-
sion U would I-implicate, the marked alternative (denotational synonym)
M will implicate the complement of U 's denotation. The Gricean reason-
ing, of course, is that the speaker seems to have gone out of his way to
avoid using the unmarked expression and so must be trying to avoid
whatever the unmarked expression would suggest. The following exam-
ples illustrate the potential range of the phenomenon.
138 Chapter 2

(70) a. ``John could solve the problem.''

I ‡> `and he did.'
b. ``John had the ability to solve the problem.''
M ‡> `but he didn't.'
(71) a. ``John turned the switch and the motor started.''
I ‡> `John intentionally caused the starting.'
b. ``John turned the switch and almost immediately thereafter the
motor started.''
M ‡> `the two events may have been coincidental.'
(72) a. ``Sue smiled.''
I ‡> `Sue produced a nice happy expression.'
b. ``The corners of Sue's lips turned slightly upwards.''
M ‡> `Sue produced a smirk or grimace.'
(73) a. ``Sally was knitting. Occasionally, she looked out the window.''
I ‡> she refers to Sally.
b. ``Sally was knitting. Occasionally, the woman looked out the
window.''
M ‡> the woman does not refer to Sally.

2.4.1.1 Lexical Doublets and Rival Word Formations Let us begin by

considering how this division of pragmatic labor might play out within
the lexicon. Consider lexical pairs of the kind drink versus beverage: they
belong to di¨erent registers (one colloquial, one primally written); the
®rst is the more familiar, frequent, normal, friendly word. But drink is no
longer neutral: ``He bought a drink'' suggests an alcoholic drink, it
has become specialized by I-implicature towards the stereotypical extension
(perhaps egged on by euphemism).66 To avoid the connotations, chocolate-
drink manufacturers extol the virtues of their beverage (historically, bever-
age was at least as contaminated!). Some further pairs of this sort follow:
(74) a. ``He was reading a book''
I ‡‡> `He was reading an ordinary book.'
b. ``He was reading a tome''
M ‡‡> `He was reading some massive, weighty volume.'
(75) a. ``Her house is on the corner''
I ‡‡> `Her house, of the normal variety, is on the corner.'
b. ``Her residence is on the corner''
M ‡‡> `Her immodest, pretentious house is on the corner.'
139 The Phenomena

(76) a. ``A missile struck the house''

I ‡‡> `A rocket struck the house.'
b. ``A projectile struck the house''
M ‡‡> `Some brickbat other than a rocket struck the house.'
(77) a. ``They had an argument''
I ‡‡> `They had a disagreeable disputation.'
b. ``They had a discussion/disputation''
M ‡‡> `They had a gentlemanly debate.'
Cross-register doublets like these are numerous in English (horse, steed;
letter, missive; gift, donation), but their interest is mainly historical: it is
fairly clear that the process of narrowing by I-implicature (as with missile
to `rocket') not only allows a new life for worn-out (marked) words but,
by M-implicature (to the complement of the I-implicature associated with
the unmarked term), trims their meaning too. Horn (1984: 32±37) makes
further remarks on the relation of diachronic studies of lexicography to
Gricean principles.
A more interesting area for the synchronic study of the lexicon is deri-
vational morphology. Often, there are two or more rival derivational
routes, as from verb to noun: thus we have from inform both informant
and informer. As Kiparsky (1982) points out, this can occur only if the
two forms pick up di¨erent extensions, but otherwise an existing deriva-
tion preempts a new one: to cook is the source of the noun cook, `one
who cooks', and this blocks cooker with that sense, but not with the sense
`thing that cooks'. (Often blocking is complete: there is no braker `thing
that brakes' from to brake given the preexisting blocking noun brake,
except in child language where it occurs; q.v. Clark 1993: 118.) Horn
(1984: 26¨ ) relates this to his Q and R (here corresponding to my M and
I) principles; the regular, unmarked formation picks up the stereotypical
extension, often narrowed in the typical wayÐso informer becomes not
only `one who informs' but also `one who informs against his own';
informant is the marked derivation and picks up an exclusive (if not
complementary) denotation, as in the linguistics usage. (Historically, the
OED entries suggest that informant may once indeed have picked up
the complement, before itself becoming associated with betrayal, an
association now lost.) By Kiparsky's generalization, ``avoid synonymy,''
the outputs of derivational processes must be semantically contrastive,
but they may be assigned rather di¨erent extensions: thus sealer is one
who seals (documents or woodwork), but it may also refer to a layer of
140 Chapter 2

varnish; a sealant is then restricted to stu¨ that seals of some more gooey
consistency, in short a mastic (Webster's Dictionary, 1991 edition).
This division of labor in word formation, between regular formations
with stereotypical meanings, and irregular ones with specialized mean-
ings, is of special interest in the study of child language acquisition. The
gradual increase of lexical knowledge requires constant adjustment in
the meanings of the already learned lexicon. If a learner doesn't know the
verb to row, then on the analogy with denominal verbs like sail, he or she
may coin the verb to oar; later discovery of the verb to row may not result
in immediate blocking of to oar. Rather, the child may try to maintain the
hypothesis of a contrast in meaning, by perhaps assigning the new verb to
row a superordinate meaning over to oar, to paddle, or to sail (Clark 1993:
94). Similarly, children produce regular in¯ections like goed at the same
age (or even after) they are correctly saying went, because they appear to
have assumed that went is a di¨erent verb with a very similar, but contras-
tive, meaning, and this hypothesis may be maintained for years (Clark 1993:
103f ). The prediction from Horn's division of pragmatic labor is that
where the child has two words like to oar and to row not only will they be
assigned distinct meanings by the Saussurean or Clarkian principle of
contrast, but also that the more regular formation may initially (and often
incorrectly) be assigned the more stereotypical, central denotation.67

2.4.1.2 Lexicalized Forms versus Periphrasis Often, an existing lexical-

ization may be in opposition to a periphrastic alternative. Take a simple
example like pink versus pale red (McCawley 1978: 245±246, after House-
holder). Pink preempts the central values of that color; thus although one
might well de®ne pink as `pale red', an utterance ``The dress is pale red''
would suggest (by M-contrast with pink) that the dress is not prototypical
pink but somewhere between pink and red. Such observations have often
been taken to be an argument against lexical decomposition, but of course
they merely establish the existence of pragmatic contrasts. Note that pale
blue carries no such implicature to `not the prototypical value of whitish
blue', because there is no contrasting lexicalization. Modal verbs often
allow periphrastic alternatives, and the distribution of interpretations is
then often in the predicted direction:
(78) a. ``You may leave.''
I ‡> `and please do so.'
b. ``You are permitted to leave.''
M ‡> `but you may stay.'
141 The Phenomena

It is presumably for this reason, Horn (1984) points out, that indirect
speech acts with the lexicalized forms (like Can you pass the salt?) are
idiomatic, whereas the periphrastic forms (like Are you able to pass the
salt?) resist the richer interpretation.
The causatives are a cause ceÂleÁbre.68 McCawley (1978, in part follow-
ing Shibatani 1972) pointed out that there is a regular association of the
lexicalized causative with direct causation and the periphrastic one with
indirect causation as in:
(79) a. ``Larry stopped the car.''
I ‡‡> `Larry caused the car to stop in the normal way, by
using the foot pedal.'
b. ``Larry got the car to stop.'' or ``Larry caused the car to stop.''
M ‡‡> `Larry caused the car to stop in a nonstereotypical way,
e.g., using the emergency brake.'
(80) a. ``The Spanish killed the Aztecs.''
I ‡‡> `The Spaniards slaughtered the Aztecs directly.'
b. ``The Spanish caused the Aztecs to die.''
M ‡‡> `The Spaniards killed the Aztecs indirectly, e.g., by
disease and hard labor.'
(81) a. ``Sue moved the car.''
I ‡‡> `Sue moved the car by driving it, by using the engine.'
b. ``Sue made the car move.''
M ‡‡> `Sue moved the car in some abnormal way, e.g., by
pushing it.'
The Gricean argument in these cases is clear enough: if the speaker has
gone out of his way to avoid the simpler expression, he must be intending
to refer to an event that contrasts to that which would have been describ-
able with the simple lexicalized causative. But the proof of the pudding,
McCawley noted (1978: 250), is that when there is no lexicalized causative,
there is then no implicatural contrast and the periphrastic construction
carries no extra meaning. Thus there is no causative verb `to make laugh'
or `to make x drop y' (e.g., no to *laughen, or to *droppen), and the peri-
phrastic forms here pick up stereotypical extensions:
(82) a. ``Bill made Mary laugh.''
No implicature `by some unordinary means, e.g., tickling her toes'
b. ``Bill caused Mary to drop her parcel.''
No implicature `by some extraordinary means'
142 Chapter 2

McCawley (1978: 257) concludes by noting the iconic relation, whereby

``the syntactically unmarked causative constructions (i.e., lexical causa-
tives) will be interpreted as referring to minimally marked causal relations
(i.e., some kind of direct causation).'' Thus the unmarked interpretation is
not de®nitively coded in the lexical causative; rather the inference to the
stereotype is itself implicated. Thus in (83), although (83a) suggests the
stereotype and (83b) suggests the complement, there is no anomaly in
(83c) despite the fact that the by-phrase cancels the assumption in (83a).
(83) a. ``The outlaw killed the sheri¨.''
I ‡> `by direct means, e.g., gunning him down'
b. ``The outlaw caused the sheri¨ to die.''
M ‡> `by some unusual means, e.g., spiking his gun, half
cutting his stirrup-leather, poisoning his aperitif.'
c. ``The outlaw killed the sheri¨ by putting poison in his water
bottle.''
This phenomenon is widely attested across languages; for example, in
Japanese (Shibatani 1972), Inuktitut (Allen 1994), and numerous lan-
guages cited by Haiman (1985a: 108¨ ). Many languages have lexical
causatives and regular causative derivational morphology: in this case,
the longer derivational form is restricted to human (or animate) causes,
presuming the indirect causation of a verbal command: one says in e¨ect
`He stopped the ball' but `He made them come to a stop' (Haiman 1985a:
109±110). The marked derivational or periphrastic constructions may
even carry magical connotations, roughly as in `He raised the plate'
versus `He made the plate rise up.'

2.4.1.3 Litotes: When Two Negatives Don't Make a Positive Another

area where lexicalized versus periphrastic constructions produce produc-
tive oppositions is the contrast between p and @@p. To understand this
area, we need to understand the apparent irregularities of English a½xal
negations; this may seem to be a digression, but it also has pragmatic
implications (see Horn 1989: 273±308, who discusses much antecedent
work). We have already seen the pattern whereby ``I don't like garlic'' I-
implicates `I dislike garlic'; that is, where the contradictory gets strength-
ened to the contrary. There are many lexical items incorporating a½xal
negatives in English, like impossible, unhappy, inhuman, or noncommunist.
The lexicographer's problem is to explain why these sometimes mean the
contrary but sometimes the contradictory of the positive: unhappy means
143 The Phenomena

the contrary `sad', not the contradictory `not happy (i.e., either sad or
in-between'), but unimpoverished means the contradictory `not impov-
erished', unworthy means just `not worthy', and so on.
Horn's account of this forest of inconvenient and troublesome sememes
goes as follows. There are constraints on the I-implicature from the con-
tradictory of lexeme W to the contrary of W. Simplifying somewhat, the
inference only goes through if:
1. the base expression W is a relatively weak positive scalar item, thus
one notes the absence of *dislove but the presence of dislike, and
2. W is a¨ectively positive, so we have unhappy implicating `(nearly) sad',
but ??unsad if allowable at all, not implicating `happy' (and so on for
impolite vs. ??unrude, unhealthy vs. ?unsick, unwise vs. ??unfoolish,
etc.),69 and
3. the negative form is relatively unproductive, i.e., not-W is relatively
lexicalized. This is something of a gradient a¨air, along the scale from
most productive to least productive: non- > un- > in- > dis- etc.
These conditions predict that we will get pairs of a½xally negative lexemes
like the following, where only the ®rst of each pair can have the strong
contrary reading: immoral vs. non-moral, irrational vs. non-rational, un-
Christian vs. non-Christian, inhuman (or unhuman) vs. non-human.
The third constraint, the more lexicalized the more likely to have a
contrary reading, is actually a side e¨ect of the fact, described in chapter
1, that the O corner of the square of oppositions tends not to lexicalize but
is instead typically instantiated as the contradictory negation of the A
corner (thus we have not all rather than the hypothetical *nall, on the
model of none). If the O-corner expressions do get their negatives fused,
then they have to be interpreted as having moved up the negative scale
towards the E corner.
We are now in a position to tackle double negations. Not every nega-
tive of a negative is equivalent to a positive. In particular, if W is a posi-
tive item, and W is the contrary of W, @…W † will not be equivalent to W,
as ®gure 2.8 makes clear. Thus not unhappy is logically compatible with
both being happy and being neither happy nor sad, although it pragmat-
ically suggests the latter. The two negatives in this case are not equivalent
to the positive (happy) for logical reasons. On the other hand, saying
``It's not true that he's not happy'' is indeed logically equivalent to the
statement that he is happy. For this reason, no pragmatic motivation is
necessary for the use of negated contraries of the kind not unhappy, not
144 Chapter 2

Figure 2.8
Logical range of di¨erent negatives

infrequent, or not unintelligent: speakers may use these forms to say what
they logically mean (however, there is in fact a systematic pragmatic
suggestionÐmore in a moment).
But it is quite di¨erent with the frequent use of the double negative
@(@W where @W is a contradictory of W. For example, things are ei-
ther possible or impossible; there's no middle ground. So, when someone
asserts ``It's not impossible'' (i.e., it's not not-possible), logically he might
as well have said ``It's possible.'' Note that some of these negated con-
tradictories have super®cial similarity to the negations of contrariesÐfor
example, not unuseful is logically equivalent to useful, while we have seen
how not unhappy is not equivalent to happyÐso the analyst must beware.
This frequent ®gure of speech (the negation of the contradictory) needs a
pragmatic motivation, and it is of course given by the Hornian division of
pragmatic labor: if two expressions have the same denotation, the marked
one will pick up the interpretation complementary to the unmarked
expression's. The pragmatic functions are intuitively clear in examples
like the following, where the speaker appears to be withholding the posi-
tive term, because the situation does not quite match that extreme:
(84) a. ``She's not incapable of exertion.''
M ‡‡> `She's less capable of exertion than ``capable'' would
I-implicate.'
b. ``His record is not unblemished.''
M ‡‡> `His record is perhaps not exactly blemished, but nearly
unblotted.'
Alternatively: `His record is not just blemished, it's atrocious.'
c. ``It took a not inconsiderable e¨ort.''
M ‡‡> `It took a close-to-considerable e¨ort.'
145 The Phenomena

d. ``She was not unpleased by his e¨orts.''

M ‡‡> `She was less than fully pleased or ecstatic about his
e¨orts.'
e. ``He's a not unjust man.''
M ‡‡> `He's a reasonably just man.'
f. ``A not unreliable service.''
M ‡‡> `A sort of reliable-ish service.'
Nevertheless, as Horn (1991a) documents, rhetoricians have long argued
about both the meaning and the desirability of litotes, some (like Erasmus)
suggesting that it is in fact an appropriately discreet and gentlemanly way
to assert the extreme positive, as in ``Your letter was no small joy.'' But
the consensus (or at least that extracted by Horn) is that the inference
may go either way: (a) it may suggest that the property described falls
short of the positive or (b) that it exceeds the positive or that the positive
needs emphasizing (as in ``I was not unaware of the problem'' suggesting
that the speaker was indeed acutely aware). The learned commentary
suggests indeed that the (b) use is parasitic on the (a) use: by asserting
``I am not an ungenerous man'' I might implicate that I am something less
than fully generous, but because modesty would be reason for that, I may
in fact implicate something less modest. Thus polite understatement may
be a factor in the extreme value readings. The Hornian division of labor
is of course neutral to the swing of interpretation one way or the other
(which may well be conditioned by politeness or other factors; see Horn
1991a: 91), but it predicts that whatever is meant is complementary to
what the plain positive would have meant.
Let us return now to the cases ®rst raised, the negation of the con-
traries, which I set aside as justi®ed in use by their logical properties. Now
the pragmatic division of labor suggests that even when justi®ed by the
logic of a negated contrary, such cumbersome forms as not unhappy, with
all the compounded processing di½culties of negatives of (special) neg-
atives, may also be justi®ed by pragmatic specialization. Why not say the
simple positive disjunction ``He's happy or content''? It seems that ``not
unhappy'' is used rather to suggest that something less than fully or ex-
tremely `happy' holds: although not quite equivalent to `content', it seems
to exclude the extreme value of happy. For this reason, the rhetoricians
seem to have swept these cases in with the denial of the contradictories: in
many cases, they have a similar weakening force.
146 Chapter 2

2.4.1.4 Grammaticalized/Morphologized versus Lexical Expressions

Consider the child's picture versus the picture of the child. The ®rst sug-
gests an intimate relation; the child perhaps painted the picture, or at least
owns it. The second strongly suggests that it depicts the child (otherwise
why not say the picture by the child or even, dialect permitting, the picture
of the child's?). The of construction of course doesn't encode `depiction
of ': the best picture of the Louvre clearly allows the possessive reading,
and the guns of the warship has nothing to do with pictures at all. The
generalization is just that the briefer, morphologized su½x of possession
suggests by I-implicature a closer, more intimate relation between the
NPs, and the choice of the lexical of construction discourages this by M-
implicature. Note here a fact of some interest, which we can ®nd else-
where too: we sometimes ®nd a double-marking, as in the picture of the
child's, and this can have the e¨ect of returning us to the interpretation
associated with the unmarked form. If expression ``X'' I-implicates re-
striction to denotation D, and expression ``Y'' M-implicates restriction to
the complement D, then expression ``Z'' which M-contrasts with ``Y'',
suggests the complement of D, which of course is D.
Possessive constructions in many languages o¨er two forms, one asso-
ciated with ``inalienable'' possessions (kin, body parts, etc.) and another
associated with ``alienable'' possessions (houses, horses, etc.). Haiman
(1985a: 130±136) points out that the inalienable relation is almost invari-
ably expressed with a less prolix form, so that there is an iconicity be-
tween the reduced bulk of the expression and the intimacy of the relation.
This might be taken to be a consequence of the I- and M-principles, now
conventionalized and frozen in grammar.

2.4.1.5 Zero versus Nonempty Morphology Finally, there is an inter-

esting, rather miscellaneous set of circumstances under which there is an
opposition between a zero-morph, or a gap, and an explicit morpheme.
Such an opposition, I pointed out, cannot arise under the Q-principle,
where matching linguistic expressions of di¨erent informativeness are
required to generate an implicature. Thus gaps and zero-morphs are open
to I-interpretation and tend to engender rich interpretations to speci®c
referents, stereotypical situations, or connected events. And these inter-
pretations may be defeated or blocked by the insertion of lexical material
instead.
Consider, for example, the opposition between the N and N in the fol-
lowing sentences (Horn 1993: 41):
147 The Phenomena

(85) a. She went to school/church/university/bed/hospital/sea/town.

Conventionalized I-implicature: `She went to do the associated
stereotypical activity.'
b. She went to the school/church/university/bed/hospital/sea/town.
M ‡> `She went to the place but not necessarily to do the
associated stereotypical activity.'
Notice that not all the relevant nouns in English permit this zero-article:
one can't say for example *She went to o½ce. Although this might tempt
one to claim that the whole pattern is purely a matter of idioms, one
should note that the underlying pattern of pragmatic interpretations is
essentially regular but constrained by the availability of the contraction.
Thus She went to the o½ce I-implicates `to do the associated activity (i.e.,
to work)' because there is no possibility of the shorter expression *to
o½ce, and thus no M-implicature can arise from the o½ce to block the
I-enrichment.
Nominal compounds in English, and in many languages, have an
unmarked N-N form. Assuming that the semantic relation between the
nouns is no more than an existentially quanti®ed variable over relations,
the exact relation must be inferred. Grammars of English give a dozen or
more patterns, with relationships like N1 powers N2 (windmill ) or N1 has
N2 , (doorknob), and so forth (see, e.g., Quirk, Greenbaum, Leech, and
Svartvik, 1972: 1021±1026). Note that the interpretation is not ®xed: a
wood container could be a container for wood or of wood. A glance at any
technical manual shows that this is indeed a productive system, and the
relationship must be inferred in such novel compounds as ferrule ring,
antenna coupler, standby memory, key protect indication, power lock, and
so on. The I-induced interpretation, we have seen, is usually to a rich
relationship between the nouns, as ®ts most plausibly with stereotypical
assumptions. But if one tries to spell this out lexically, the resulting
utterance often has distinct implicatures: compare matchbox (the box),
box of matches (box containing matches), box for matches (nonproto-
typical box specially made for containing matches), and so on. Intuitively,
the N-N form suggests a dubbing or naming of a speci®c relation that the
recipient should be able to infer; any spelling out of the relation suggests
(by M-implicature) some other relationship between the nouns.
In the discussion of the I-interpretations of conjunction and parataxis,
we have seen that in English many of the same interpretations are avail-
able for sentences of the form ``A and B'' and ``A. B'', but not all. Thus
148 Chapter 2

``He broke his leg. He fell'' I-implicates `He broke his leg because he fell',
an interpretation that is blocked by the insertion of and in ``He broke his
leg and he fell.'' In English, the systematic opposition between and and
zero is perhaps undermined by two factors, (a) the often obligatory re-
quirement for and within a clause (*He came went, *Sue Mary came),
(b) the tendency for the phonetic reduction of and to a single assimilated
nasal as in ``He came'n' got it.'' In other languages, more systematic
oppositions have been observed: in Turkish, Ono, Fe?fe? Bamileke, and
other languages, the absence of the conjunction between two actions by a
single actor carries the meaning that they constitute one single complex
action: saying in e¨ect ``He went to the market he bought yams''
implies that the yams came from the market, whereas ``He went to the
market and he bought yams'' may imply that the yams came from else-
where (i.e., there were two independent actions) (Haiman 1985a: 111±
115). In some languages, the e¨ect is weak but discernible: in Sissala, for
example, ``When [the conjunction] a is left out, . . . one would expect the
event to happen exactly according to the stereotype or scenario. . . . The
addition of a . . . might alert the hearer to some minor departure from
stereotypicality'' (Blass 1990: 250).
English conjunction reduction, though, ®ts the same pattern of inter-
pretation. In this case, the gap is a grammatically conditioned reduction
subject to constraints of various sorts, most obviously the identity of the
subject across two clauses. Example (86a) strongly suggests a single com-
plex action of store-going-in-order-to-whiskey-buy.70 The sentence (86b)
suggests some contrast with such a single-minded whiskey expedition; he
may, weak soul, have gone to do the weekly shopping and been tempted
by a bottle of whiskey. When the explicit conjunction is reinforced with
both as in (86c) or with in addition as in (86d), we seem to be dealing with
a list of actions and not a single expedition of any kind at all.
(86) a. ``He went to the store and bought some whiskey.''
I ‡‡> `He went to the store and, there and then, he bought the
whiskey.'
b. ``He went to the store and he bought some whiskey.''
M ‡‡> `He went to the store and he (perhaps elsewhere,
perhaps incidentally) bought some whiskey.'
c. ``He both went to the store and he bought some whiskey.''
M ‡‡> `He did two separate actions: he went to the store and
he bought some whiskey.'
149 The Phenomena

d. ``He went to the store and in addition he bought some whiskey.''

M ‡‡> `He did two separate actions: he went to the store and
he bought some whiskey.'
Conjunction reduction, of course, involves some sort of deletion of the
subject and thus raises the issue of alternative referring expressions.
Under the I-principle, I have already brie¯y reviewed evidence from con-
versation analysis for a maxim of the sort ``choose the minimal referring
expression that will achieve identi®cation'' (after Sacks and Scheglo¨
1979; Scheglo¨ 1972, 1979, 1996). In the same sort of way, many linguists
have noted a scale of referring expressions zero > pronoun > name > de-
scription, so that the more toward the left an item is, the more likely an
anaphoric interpretation is preferred. Although sometimes these choices
are grammatically speci®ed, and thus their interpretations constrained,
sometimes there is a syntactic freedom of choice, as in English:
(87) a. ``Susan went to the library and continued to read Wuthering
Heights.''
b. ``Susan went to the library and she continued to read Wuthering
Heights.''
c. ``Susan went to the library and the woman continued to read
Wuthering Heights.''
d. ``Her mother wanted to stay at home and ®nish her novel, and
urged Susan to leave. Susan went to the library and she
continued to read Wuthering Heights.''
e. ``Susan went to the library and the dreamy girl continued to
read Wuthering Heights.''
In (87a) we are forced to the coreferential reading; in (87b) we prefer it,
unless the context biases an alternative as in (87d). But in (87c) we
strongly prefer a disjoint reading, although a coreferential reading can be
restored with additional material as in (87e). These issues, perennially
controversial, will be covered in depth in chapter 4. Here I simply note
that this pattern ®ts the general scheme of the pragmatic division of
labor: minimal forms will pick up maximally informative readings by the
I-principle, whereas contrastive maximal forms (the woman vs. she) sug-
gest a complementary interpretation by the M-principle.

2.4.1.6 Repetition and Reduplication One of the most straightforward

predictions from the M-principle is that repetitions and reduplications
150 Chapter 2

should suggest interpretations distinct from the use of the unduplicated

form. Any principle of contrast, however, will predict some meaning dif-
ference between the use of a form ``X'' and ``X-X.'' The prediction from
the division of pragmatic labor, however, is that where ``X'' I-implicates
restriction to the subset of denoted entities D, ``X-X'' will pick out the
complement D; where ``X'' carries no clear I-induced restriction, or non-
stereotypical exemplars are salient in the context, ``X-X'' should implicate
a narrowing to, for example, an extreme or prototypical variant of D.
Hence we have both Oh we're just living-together living-together (room-
mates, not lovers), and I want a salad-salad (a real lettuce salad, not a
tuna salad, etc.) (Horn 1993: 48±51).
Two preliminaries: First, not every instance of a repeated form ``X-X''
is to be understood as a repetition or reduplication for pragmatic e¨ect; it
can be determined by good structural principles. For example, the phrase
a rich, rich man is structurally ambiguous, as usually perhaps indicated in
the prosody:
(88) a. ``a [[rich][rich man]]''
ˆ `a man rich among the rich'
b. ``a [[rich rich] man]''
Implicates: `a very wealthy man'
It is only the ¯at structure in (88b) for which we need a pragmatic
account. Second, the linguistic terminology is here confused, with one
author's duplication, another's reduplication, and so on. I follow Naylor's
(1986) recommendation to follow Sapir (1921: 76) on a wide notion of
reduplication:
Nothing is more natural than the prevalence of reduplication, in other words, the
repetition of all or part of the radical element. The process is generally employed,
with self-evident symbolism, to indicate such concepts as distribution, plurality,
repetition, customary activity, increase of size, added intensity, continuance.
The term reduplication can nevertheless be restricted to a morphological
process of word formation, repetition thus denoting word and phrase-
level duplication. As Sapir (1921: 77) noted, ``it can hardly be said that
the duplicative process is of a distinctively grammatical signi®cance in
English . . . we must turn to other languages for illustration'' (cited in
Naylor 1986: 177). However, before leaving English, it is worth noting the
following patterns of repetition (partly from Quirk, Greenbaum, Leech,
and Svartvik 1972: 618).
151 The Phenomena

(89) a. Modi®ers and intensi®cation

a very, very, very, steep path
a deep, deep pool
an old old church
Run quickly quickly!
a blue, blue sea
b. Conjoined comparatives and increasing value
He grew angrier and angrier.
It became more and more apparent.
c. Conjoined verbs and continued/repeated action
He ran and ran and ran up the hill.
She knocked and knocked on the door.
d. Conjoined/adjoined nouns and distributed plurality
He could see stars (and) stars (and) stars.
There were kids (and) kids all over the place.
e. Conjoined nouns and distinctions within the class
There are doctors and doctors; some you can trust, some you can't.
All of these repetitions seem to have a similar structure to nonrepetitions,
so that he ran and ran would have the same structure as he ran and
walked. Semantically, the repetitions thus appear to be redundant. It is
possible, as Naylor (1986) suggests, that one could adopt an analysis
where the sequence of forms introduces a sequence of reference points,
giving us an aspect-like characterization of he ran and ran or he grew
angrier and angrier or, alternatively, an interpretation as a sequence of
reference points along a time-line as suggested by Partee (1984) for dif-
ferent cases. But this will not give us an analysis of blue, blue sea, or
doctors and doctors, and so on. Because this is a highly productive process
(unreduceable to a small list of idioms), I am inclined to a pragmatic
interpretation.71 Note that the interpretations are richer than, for exam-
ple, Quirk et al. (1972) suggest: he ran and ran and ran suggests not only
continued action but intensity: he ran hard.
An M-implicature analysis suggests that the X(and)X form should pick
up a complementary interpretation from the X form. In many cases, this
seems correct. ``He's clever'' states that he is intelligent, but ``He's clever-
clever'' implicates that he is trying too hard to be clever and is thus less
than fully bright. Or consider the following contrasts:
(90) a. ``He went to bed and slept''
I ‡> `he had a normal night's sleep'
152 Chapter 2

b. ``He went to bed and slept and slept''

M ‡> `he slept longer than usual'
c. ``In Outer Mongolia, there are doctors.''
I ‡> `in Outer Mongolia one can ®nd useful general
practitioners of medicine.'
d. ``In Outer Mongolia, there are doctors and doctors.''
M ‡‡> `in Outer Mongolia, one can ®nd useful general
practitioners of medicine, on the one hand, and on the other
so-called doctors who could not be called useful general
practitioners of medicine.'
In the intensity readings of cases like big, big dog, the repetition seems to
pick out the extreme value along the following lines: if big dog denotes the
set of big dogs D, and I-implicates a stereotypical subset (e.g., Alsatian-
size dogs), then big, big dog M-implicates that the speaker has in mind a
distinct subset (e.g., St. Bernard-size dogs), within the set D.
Turning to other languages, one can ®nd repetition in productive word
formationÐthat is, full reduplication. In Tzeltal and Tzotzil, for example,
whereas the use of nonreduplicated color terms picks out the prototypical,
stereotypical hues, the reduplicated terms indicate that the color in ques-
tion deviates from the relevant focal or prototypical hue (Berlin 1963:
216). In Australian languages (Dixon 1980), reduplicated verbs may have
aspectual meanings (as in Dyirbal gunbal-gunbal `cut repeatedly'), but we
also ®nd clear M-inferences to the nonprototype extensions as in Western
Desert wati-wati `kids pretending to be men' (vs. wati `men'), or kintha-
kintha-la `little dog' (vs. kintha `dog') in Dyari. Perhaps the most extended
study of this kind of reduplication is that by Botha (1988) for Afrikaans,
where root reduplication is a highly productive kind of derivational
morphology. The traditional grammarians of Afrikaans list about 13 dis-
tinct kinds of meanings that result (Botha 1988: 92±95), many in opposed
directions: thus bakke-bakke `lots of bowls', but ruk-ruk `a few times,
occasionally'; or dik-dik `very thick', but skop-skop `tentatively kick'; or
even troppe-troppe `scattered ¯ocks' versus tien-tien `ten together'. Botha's
valiant attempt to specify a single semantic rule of interpretation involv-
ing the notion of increased value is therefore less than fully convincing.
Rather, some fairly complex pragmatic process of interpretation may be
suspected, and the M-principle is certainly a candidate: if dik snye brood
means `thick slice (of ) bread' and I-implicates a stereotypical generous
slab, then dik-dik snye brood M-implicates something di¨erent from that
153 The Phenomena

(because dik rules out di¨erence in the wafer direction, the suggested
denotation is thicker than the normal generous slab). If skop `kick'
tends to I-implicate `kick hard', then skop-skop will suggest `less than
hard'. Whether the inference goes in the intensity/increase direction or
the attenuation/limited dimension seems to depend on the direction of
I-inference from the unreduplicated form, the reduplication then picking
up the complement.
There is much literature on partial reduplication, too extensive for re-
view here. But it is worth pointing out that even in this area, often deeply
built in to morphological processes and subject to all the vagaries of
meaning associated with derivational morphology, there are some cases
worth the pragmaticist's attention. Many languages have a means of
giving a word sheer extra bulk. In Tamil, for example, there is a kind of
reduplication known as echo-word formation, where the ®rst syllable of a
multisyllabic word is repeated after the word with the initial consonant
altered by rule. The interpretation is always to shift the interpretation
away from the prototype extension of the word (cf. Yiddish English
problem shmoblem).
(91) Echo-words in Tamil
a. kooma-sooma
M ‡> `medical coma or something like that'
b. paittiyam giyttiyam
M ‡> `almost insanity'
c. pooliis kiiliis
M ‡> `police or someone like them'

2.5 THE JOINT EFFECT OF Q-, I- AND M-IMPLICATURES

At the risk of redundancy, it is worth pointing out how these three prin-
ciples converge to predict the inferences in slightly di¨erent sentences.
Consider, for example, the coordinating connectives once again. We have
already met all of the following kinds of inference:
(92) ``John went to the library or he worked on his book.''
Q-scalar ‡> `John did not both go to the library and work on his
book.'
Q-clausal ‡> `I don't know whether John went to the library or
whether he worked on his book.'
No I-implicature from p or q [because Q-implicatures block
154 Chapter 2

I-implicatures; see section 2.5.1]

No M-implicature from either p or q [bcause Q implicatures block
M implicatures; see section 2.5.1]
(93) a. ``John went to the library and he worked on his book.''
I ‡‡> `John went ®rst to the library, and then, as intended,
worked there on his book.'
No Q-implicature from p and q because and is the strong
member of the scale hand, ori
No M-implicature from p and q because in English zero and and
are not systematically opposed.
b. ``John went to the library and in addition he worked on his
book.'' or ``John both went to the library and he worked on his
book.''
M ‡‡> `contrast with (a)., e.g., John did two independent
actions, library-going and book-writing.'
Or consider the modals:
(94) a. ``You may/can use the car.''
Q scalar ‡> `It's not the case that you must use the car.'
I ‡‡> `I o¨er you the use of the car.'
b. ``You are allowed to use the car.''
M ‡> `Contrast with I implicature, e.g., this is not an o¨er, but
a statement of fact.' [M-implicatures cancel I-implicatures, see
section 2.5.1]
Q-scalar ‡> `It's not the case that you are obliged to use the
car.' [M-implicatures do not cancel Q-implicatures, see section
2.5.1]
Or consider the many implicatures due to negation:
(95) a. ``Not all the girls are sporty.''
Q-scalar ‡> `Some of th girls are sporty.' [negative meta-scale
hnot-some=none, not-alli]
b. ``I don't believe all the girls are sporty.''
I ‡‡> `I believe not all the girls are sporty.'
Q-scalar ‡> `I believe some of the girls are sporty.'
c. ``It's not unlikely that all the girls are sporty.''
M ‡‡> `It's just slightly less than likely that all the girls are
sporty.'
155 The Phenomena

We have also considered the potential of the application of these principles

to closed-class contrast sets like the prepositions or the articles (see
Hawkins 1991 for close discussion):72
(96) a. ``I bought a good novel. I talked to an author about it.''
Q-scalar ‡> `I talked to someone other than the author of the
good novel.' [nonuniqueness is implicated by Horn-scale
hthe, ai]
b. ``I bought a good novel. I talked to the author about it.''
I ‡‡>: `I talked to the author of the good novel.'
c. ``I bought an old car. A wheel was loose.''
I ‡‡>: `A wheel of the old car was loose.'
Q-scalar ‡> `A nonunique wheel was loose, i.e., not the steering
wheel but a road wheel' (otherwise the speaker should have said
The wheel ).
These examples merely illustrate that it is the intersection of the principles
that yield the richest predictions, and it is this intersection that is exploited
in the account of anaphora o¨ered in chapter 4. For these predictions to
be reasonably precise, however, it is essential to consider how the princi-
ples interact when inconsistent implicatures are generated. It is to that
issue that I turn next.

2.5.1 The Projection Problem

Conversational implicatures are, by de®nition, volatile, defeasible infer-
ences. Generalized implicatures are, of course, just as defeasible in the
context of statements to the contrary, as they are to contextual assump-
tions that rule them out of court:
(97) ``If you come, he'll forgive you, and in fact he'll forgive you
anyway (regardless of whether you come).''
Potential I-implicature of conditional perfection: `If you don't come,
he will not forgive you.' (or: I ‡‡> `if and only if you come will he
forgive you.')
Entailment of the second conjunct and the whole sentence: ``and in
fact he'll forgive you anyway (regardless of whether you come)''.
The entailment defeats the I-implicature.
But what happens if, by dint of multiple principles, we have multiple
con¯icting inferences or inconsistent potential implicatures?73 We have
156 Chapter 2

seen that Q-inferences and I-inferences appear to work in opposite

directions: I-implicatures are strong inferences, encouraged by minimal
forms, to the likely intended stereotypical or more speci®c interpretation;
Q-implicatures are bounding inferences from the use of a particular ex-
pression to the inapplicability of another, stronger one that has not been
uttered. The antinomic tendencies of these two principles are well brought
out in Horn's (1984) version of the story. Q-implicatures thus appear to
induce the negation of the very sort of stronger interpretation that the I-
principle encourages. Atlas and Levinson (1981), on whose principles
Horn drew, suggested that this clash is in part resolved by constraining
the Q-principle appropriatelyÐin particular, by setting up two further
constraints on scales: expressions can only form a ranked Horn-scale if
they are (i) equally lexicalized and (ii) are ``about'' the same semantic
relations, as reviewed above. Nevertheless, where genuine inconsistencies
appear, it was suggested that Q-inferences have priority over I-inferences
that are inconsistent with them.
Note, though, that I-inferences often come about through reduced
forms, which are then outside the scope of the Q-principle. For example, I
suggested that hthe, ai form a scale, such that use of the inde®nite sug-
gests that the speaker is not in a position to use the de®nite. There is also
a zero-article form for certain places and institutions, which is associated
with an (at least partially conventionalized) I-inference to stereotypical
associated activities. But because this has a zero realization, it fails the
lexicalization constraint on scales and can play no role in the de®niteness
scale, and has an independent existence:
(98) a. ``John went to university.''
(conventionalized) I ‡> `to the institution.'
b. ``John went to the university.''
can be read as `went to the campus' or `went to that particular
institution.'
c. ``John went to a university.''
Q ‡> `not the university', thus suggesting the institutional rather
than the location interpretation.
In this sort of way, clashes between the I- and Q-principles are often
averted.
There is also a potential clash between the I-principle and the M-
principle: the I-principle generates inferences to the rich, speci®c sub-
157 The Phenomena

case, whereas the M-principle generates inferences in the precisely

opposed direction. In this case, the M-implicatures take priority over the
I-implicatures that would have arisen from the use of a corresponding
unmarked expression, as described in Horn's division of pragmatic labor
discussed earlier.
Thus we have three principles potentially producing incompatible
inferences from a simple utterance: in general, utterances generate multi-
ple potential implicatures. As sentences become complex, these inferences
may arise from di¨erent clauses, and tra½c rules will need to be estab-
lished. In short, we have a projection problem: we need to be able to state
the conditions under which a potential implicature generated from a part
of a sentence is in fact inherited by the whole. In previous publications
(Levinson 1987a,b), I suggested the following ``resolution schema'':
(99) a. Genuine Q-implicatures from tight Horn scales (and their
meta-rule derived negative scales) and similar contrast sets of
equally brief, equally lexicalized expressions ``about'' the same
semantic relations, take precedence over I-implicatures;
b. In all other cases, the I-principle induces stereotypical
interpretations, unless:
c. A marked expression has been used where an unmarked
one could have been employed instead, in which case the
M-implicature defeats the relevant I-implicature, by inducing
the inference to the complement of the I-implicature that would
have arisen from the unmarked expression.
This schema simply captures the two observations already made: that
Q-inferences defeat I-inferences inconsistent with them, and that M-
inferences do likewise (as in the division of pragmatic labor). It leaves the
relation between M- and Q-inferences unresolved. There is some evidence
that in fact Q-inferences defeat M-inferences inconsistent with them, and
this would allow the simple statement that inconsistencies between
potential implicatures are resolved by assigning priority ®rst to Q-, then
to M-, and only ®nally to I-inferences:
(100) Resolution of inconsistent potential implicatures
Priority is assigned to inferences according to the principle under
which they are generated:
Q-implicatures > M-implicatures > I-implicatures
Examples:
158 Chapter 2

a. Q > I see (101)±(103)

b. Q > M see (105)±(106)
c. M > I see (83)±(86), (104)
This still leaves much detail unresolvedÐfor example, the resolution of
potential implicatures of di¨erent subtypes arising under the same princi-
ple, or of inferences coming from di¨erent clauses. But ®rst let me illus-
trate how this resolution schema works.
First, I want to illustrate the way in which Q-inferences seem to defeat
I-inferences.74 Consider a conjunction like ``Bill entered the room and
Harry left.'' As I have shown, we tend to interpret this, as licensed by the
I-principle, along the lines `Bill entered the room, and then, perhaps as a
result, Harry left the room'.75 Now the disjunction ``Bill entered the room
or Harry left'' carries no such temporal and causal inferences. That is
because the use of the disjunction carries two Q-implicatures: the scalar
inference `It's not the case that both Bill entered the room and Harry
left' and the clausal inference `For all the speaker knows, Bill may or may
not have entered the room, and Harry may or may not have left it'. These
Q-inferences are clearly inconsistent with the I-inference, which requires
the conjunction of events, and it is an observable matter of fact that the
I-inference simply is not associated with the disjunction.
Conditionals give rise (according to the account in Gazdar 1979) to
(potential) clausal Q-implicatures, and it is these Q-implicatures to the
speaker's epistemic uncertainty about the embedded clauses that account
for the very general tendency for pragmatic inferences of a consequent to
be suspended by mention in an antecedent. This pattern holds for scalar
implicatures (as in If not all of the boys came, some of them did ) and even
presuppositions (as in If there is a king of France, he is probably bald ),
as reviewed in Levinson 1983 (pp. 142±145, 223±225). It also holds for
I-inferences, as illustrated by the following (note that I revert to the
explicit epistemic modi®cation of implicatures, with all but the clausal
implicatures embedded under a belief operator K, in order to bring out
the inconsistencies):
(101) ``John drank three beers and drove home, if not in the reverse order.''
Q-clausal inference from second clause: `Possibly he did it in the
reverse order', i.e., P(John ®rst drove home then drank three beers)
Potential I-inference from ®rst clause: `John drank three beers and
then drove home', i.e., K(John drank three beers and then drove
home)
159 The Phenomena

Implicatures of the whole utterance: only Q-clausal inference from

second clause, not I-inference from ®rst clause.
A similar pattern holds for disjunctions: the whole disjunction lacks an
implicature arising from one disjunct if the other disjunct entails its con-
tradictory. Just as the potential scalar implicature in Some of the boys
came or all of them did is suspended by the clausal implicature from the
second disjunct, so we have the same phenomenon with I-inferences can-
celed when they are inconsistent with Q-clausal ones:
(102) a. ``John drank the three beers and drove home, or ®rst drove
home and then drank the three beers.''
Potential I-implicature of ®rst disjunct: K(`John ®rst drank the
three beers and then drove home')
Q-clausal implicature from the disjunction: P(`John ®rst drove
home then drank the three beers.')
b. ``John and Marty bought a piano or they bought one each.''
Potential I-implicature from ®rst clause (Atlas and Levinson
1981, after Harnish 1976):
K(`John and Marty bought a single piano together')
Q-clausal implicature from the disjunction: P(`they bought one
each.')
Consider now a slightly more involved case, where Q-inferences defeat the
I-inference to ``conditional perfection'':
(103) ``If they try again, they'll beat the landspeed record, or they've
beaten it already and won't try again.''
Potential I-implicature of conditional perfection: `If they don't try
again, they won't beat the landspeed record', or K(`if they will not
try again at t2 , they have not beaten the record.')
Q-clausal implicature from second disjunct: `Possibly they've beaten
the record already and will not try again.'
P(`they've beaten the record at t1 and will not try again at t2 .')
Q-implicature defeats the potential I-implicature.
The generalization over this and many other cases thus appears to be that
Q-inferences have priority over I-inferences.76 One class of systematic
exceptions, I have noted, appear to be the Neg-raising I-inferences, which
seem to be partially conventionalized and are restricted so that the infor-
mativeness of the inference only slightly exceeds the content of what is
said (see section 2.3.2).
160 Chapter 2

The priority of M- over I-inferences was detailed in the discussion of

the M-principle and the division of pragmatic labor. For example:
(104) a. ``Cortes killed Montezuma.''
I ‡‡> `Cortes directly caused the death of Montezuma, e.g.,
strangled him with his own hands.'
b. ``Cortes caused the death of Montezuma.''
M ‡‡> `Cortes indirectly caused the death of Montezuma, e.g.,
ordered him to be put to death.'
Now what happens if an M- and a Q-implicature are inconsistent?
Consider the following example:
(105) a. ``Cortes caused the death of Montezuma, or indeed he killed
him outright with his own hands.''
b. ``Cortes caused the death of Montezuma.''
M ‡> K (`Cortes brought about the death of Montezuma
indirectly, not with his own hands.')
c. p or q Q-clausal ‡> fPq; P @ qg; instantiating: Q ‡>
P(`Cortes killed him outright with his own hands.')
Example (105a) clearly lacks the potential M-implicature of (105b) arising
from the ®rst clause, and this is obviously due to the content of the following
disjunct. The M-inference would equally have been defeated by a condi-
tional clause: ``Cortes caused the death of Montezuma, if indeed he did not
kill him outright with his own hands'', because conditionals give rise to the
same type of Q clausal implicature as disjunctions. Or another example:
(106) a. ``It's not unlikely that Giant Stride will win the Derby, and
indeed I think it likely.''
b. ``It's not unlikely that Giant Stride will win the Derby.''
M ‡‡> `It's less than fully likely that Giant Stride will win the
Derby.' [from use of double negative not unlikely]
i.e., K(`it's not likely that Giant Stride will win the Derby.')
c. ``and indeed I think it likely.''
Q-clausal ‡> `It is possible it is likely.' [from use of think which
does not entail its complement]
i.e., P(`it is likely that Giant Stride will win the Derby.')
d. The Q-inference in (c) defeats the inconsistent M-inference
in (b).
Again, other clausal inferences will do the same job: consider, for exam-
ple, ``It's not unlikely that Giant Stride will win the Derby, if indeed
161 The Phenomena

there's any uncertainty at all'', which will Q-implicate `It is possible that it
is certain.'
As far as I can see, therefore, the proposed priority of Q > M > I
appears to work. But why? It is fairly clear why both the Q- and the M-
principles should have priority over the I-principle: they are both meta-
linguistic principles that work by reference to what the speaker might
otherwise have said but hasn't. In short, the speaker has gone out of his or
her way to produce an utterance that suggests otherwise than the straight
inference to the stereotypical or canonical subcase. The relative priority of
the Q-principle over the M-principle is presumably attributable to the
relative importance of informational content over expression modulation.
This line of explanation is based on the idea that the I-principle legiti-
mates the working assumption ``enrich!'' in the normal case, and the other
principles serve to constrain that enrichment. Another way of looking at it
is to compare the phenomena to blocking in morphology and syntax: the
more speci®c rules block the application of more general rules. Thus infer-
ences based on highly constrained sets of lexemes (Q-inferences) block
those based on wider ranging contrasts in markedness (M-inferences),
which in turn block those based on stereotypes about the world and the
inference of maximal cohesion (I-inferences). The need for such blocking
rules is clear in default reasoning systems: as Lascarides and Asher (1993:
444) note, ``we want a defeasible logic to validate the inference in which
the more speci®c rule is preferred.''
Let us now turn to the question of subspecies of inference types. Gaz-
dar (1979), as I showed earlier, made the distinction between clausal and
scalar Quantity implicatures. One reason for this is that they appear to
have di¨erent projection properties (but see Soames 1982 for critique, and
Kay 1992 for response).77
(107) a. ``If all of my friends don't earn more than me, then some of
them earn more than me.''
b. ``Some of them earn more than me.''
Q scalar ‡> K(`Not all of my friends earn more than me.')
c. ``If p then q'' Q-clausal ‡> fPp, P @ p, Pq, P @ qg
d. ``If all of my friends don't earn more than me.''
Q-clausal ‡> P(`all of my friends earn more than me.')
[from P@ (`all of my friends don't earn more than me.')]
e. The clausal-implicature in (d) overrides the potential scalar
implicature in (b).
162 Chapter 2

Whether there are other such subspecies with ordered priorities has simply
not been established at this stage. Indeed the whole issue of implicature
projection has not been taken seriously hitherto, with the exception of
Gazdar (1979). I have already noted Gazdar's (1979: 56±59) claim that
implicatures should be blocked by any logical operators whose scope they
fall within, and Hirschberg's (1985: 73) restriction of this blocking rule to
negation (accepted by Kadmon 1987 and Horn 1989: 234). I pointed out
that in fact no such blocking under negation is necessary, once the sys-
tematicity of negative scales is appreciated: thus, an utterance of the form
``. . . not . . . some . . .'' invokes the scale hnot-some=none, not alli, in
which the o¨ending expression is the strongest member and therefore
generates no scalar implicatures.78
In fact, I shall take Gazdar's (1979) projection mechanism as a ®rst
approximation to what is required. Let us idealize a conversation in such
a way that each utterance contributes its information to the context or the
common groundÐthat is, that which is taken for granted by the speaker
as shared assumptions. Now let each utterance be processed in such a way
that its encoded and implicated content is added incrementally to the
context in a ®xed order.79 The order suggested by the remarks above
should be:
(108) Ordered incrementation of the context by the content C of an
utterance U
a. Each aspect of utterance content should be added to C only if
it is consistent with what has already been added to C by U or
preceding utterances, or is part of C by virtue of being
assumptions taken for granted;
b. Aspects of utterance content should be added to C, providing
they are consistent, in the following order:
i. the entailments of U
ii. (1) the potential clausal Q-implicatures of U
(2) the potential scalar Q-implicatures of U
iii. the potential M-implicatures of U
iv. the potential I-implicatures of U
In such a system, all the potential implicatures are generated and are then
®ltered out by a process of incremental addition to the context only if
consistency is satis®ed. This process of generation plus ®ltering captures
the essential notion of default inference. The process can be viewed simply
as a technical solution, but it can also be viewed as psycholinguistic pre-
163 The Phenomena

diction of some interest. The important concept of defeasibility is thus

modeled in terms of ®ltering by inconsistent, higher ordered assumptions.
There would be other, more complex reconstructions of that concept, as
discussed in chapter 1 under the rubric of nonmonotonic inference sys-
tems, and there are perhaps reasons to think one may have to escalate in
that direction especially when we move outside the domain of GCIs (see,
e.g., the ®fth point in the next paragraph).
In various other ways that hardly need to be spelled out, this ®ltering
system is an idealization. First, Gazdar's (1979: 130) model is a model of
speaker's commitment, not actually a model of mutually accepted prop-
ositions as appropriate to conversation (in this respect, it borrows from
Hamblin's (1971) model of a speaker's commitment slate), although that
modi®cation looks feasible. Second, the model cannot handle retraction,
and more importantly it cannot handle implicature cancellation at arbi-
trary distance. The notion of utterance in Gazdar's model follows the
Bar-Hillel (1954) concept of a sentence-context pairing, but in fact it is
clear that implicatures can be canceled from across sentences, as in ``I ate
some of the biscuits. In fact, I'm dreadfully sorry, they were so delicious, I
ate them all.'' For this reason, I propose in chapter 3 that implicatures
must be kept distinct from other assumptions so that they can be later
canceled from theoretically arbitrary distance.80 Third, in the case of
tropes (e.g., ironies and metaphors), what is entailed may sometimes be
canceled by what is implicated (e.g., ``Chomsky is never inscrutable'', said
ironically, might implicate `he is always inscrutable'). Fourth, there are
probably constructional di¨erences in the order in which clauses are
incrementedÐfor example, in theories such as DRT the antecedents of
conditionals contribute to the context before the consequent (Kamp and
Reyle 1993: 147¨ ). Such construction-speci®c cumulative rules were
employed in earlier models of presupposition projection (see Soames 1982
for a summary), and as I show in chapter 3, they may play a role in how
GCIs are interpreted. Fifth, the central concept of addition under consis-
tency runs into trouble with particularized implicatures attributable to the
maxim of Relevance, which I assume to be centrally concerned with the
relation of utterances to the presumed goal-structure of the ongoing
activity to which the discourse is contributing.81 Consider for example:
(109) A: ``Well, what did you achieve today?''
B: ``Oh I went downtown and dealt with some bills.''
B's utterance potentially I-implicates: `I went downtown and
then dealt with some bills.' or `I went downtown in order to
deal with some bills.'
164 Chapter 2

Here it seems that because the requirement set by local conversational

goals is just to provide a list, the I-inferences to order and teleology are
not ®rm (in contrast, for example, to the same response to ``Where did
you go?'').
A further major di½culty will become clear in the chapter 3. General-
ized implicatures can play a role in ``®xing'' the propositional content of
statements and the entailments of that content. This of course threatens
any such scheme as the one sketched above, where entailments cancel
implicatures inconsistent with them, unless such pragmatic intrusion into
truth conditions has a systematic basis (which, in fact, seems to be the
case).
Despite these insu½ciencies, the proposed model is a useful ®rst ap-
proximation for the purposes of modeling just the kind of inferences we
are concerned with hereÐnamely, the generalized implicatures. And I
shall presume that something along these lines is at least a productive
research strategy to follow. Some theorists interested in GCIs seem to
suggest a pessimism that any system of this kind will work.82 Nonce-
implicature theorists, who refuse to countenance the existence of GCIs,
may look at these projection complications attending GCI theory as a
giant reductio ad absurdum, the last epicycles of a Ptolemaic system.
Surely a theory with less maxims or principles, a Copernican alternative,
is to be preferred? It may seem that the fewer the principles, the fewer the
clashes between rival inferences there are likely to be. But that is not the
case: there will simply be fewer ways to classify the rival inferences, and
so fewer ways to specify tra½c rules between them. Similarly, nonce-
implicature theorists (those who only grant the existence of PCIs) fail to
attend to the problems of inconsistent inferences arising from di¨erent
clauses. It may indeed be a general assumption in our conscious phe-
nomenology of communication that ``it is the ®rst interpretation to occur
to the addressee that is the one the communicator intended to convey''
(Sperber and Wilson 1986: 169), but this uniqueness can only be achieved
by a massive pruning, a veritable massacre, of lexical and structural
ambiguities, pragmatic resolutions of deixis and anaphora, and the myri-
ads of pragmatic inferences that arise. GCI theory, by suggesting default
heuristics, greatly restricts the dimensions of the task, at least in this cor-
ner of utterance-meaning.
Chapter 3
Generalized Conversational Implicature and the Semantics/Pragmatics
Interface

3.1 BACKGROUND

In the previous chapters, I outlined a theory of generalized conversational

implicatures, sketched a tentative catalog of the phenomena that come
under the theory, and explored some of the ways in which they may
interact with one another. Grice developed a theory of implicature not
because he had any overriding interest in, for example, conversational
understanding, but because he saw correctly that a theory of preferred
interpretations would make possible a reassessment of what aspects of the
meaning of particular linguistic expressions are really semantically speci-
®ed and which aspects come about through a process of interpretive
enrichment. This program, when pushed through in the sort of way we
have explored in the previous chapters, renders problematic and ``up for
grabs'' the correct division of labor between semantics and pragmatics in
the explanation of many aspects of meaning. These possibilities have
hardly been explored. We have already seen, for example, that GCIs may
be responsible for the assignment of scopeÐthat was the import of the
discussion of the Neg-raising examples in chapter 2 (see also Bach 1982,
1994: 129±130; Hobbs 1996). Similarly, GCIs seem to play a role in the
fundamentals of reference, as noted in the discussion of the scalar impli-
catures associated with de®nite versus inde®nite determiners; this is a
theme further explored in this chapter,1 but especially in the next. Many
speci®c alleged ambiguities, of both a potentially structural and lexical
kind, might succumb to an account in terms of monosemy (semantic
generality) plus or minus implicatures (Ruhl 1989). There are perhaps
speci®c historical reasons, but no good intellectual ones, for why the
Gricean program in this direction has not been pushed further through.2
166 Chapter 3

In this chapter, though, rather than develop the analysis of GCIs fur-
ther, I wish instead to use the modest harvest gathered so far as a lever on
the general theory of meaning. It turns out that GCIs have fundamental
implications for the way in which we should construct an overall theory
of meaningÐin particular, for what we may call (somewhat grandiosely)
its architecture. I shall assume that ``meaning'' is essentially compositeÐ
that is, that we need to carve up the semiotic pie much in the way that
Grice envisaged, into the component entailments, conventional implica-
tures, presuppositions, felicity conditions, GCIs, and PCIs associated with
particular expressions on particular occasions of use.3 To describe and
analyze these di¨erent aspects of meaning, we will need distinct bodies
of principles (each constituting a component in the overall theory); for
example, semantic inference is based on monotonic principles, but many
aspects of pragmatic meaning are defeasible and nonmonotonic in char-
acter. We then have to ask: How do these elements or components of
meaning interact? That I take to be a question about the architecture of a
theory of meaning. Chomsky (1965) introduced a highly in¯uential way
of thinking about this: he suggested that the traditional levels of linguistic
analysis (syntax, phonology, semantics) should be thought of as indepen-
dent components or modules, with the interaction between components
being conceived of in terms of logical priorityÐthat is, the interaction
being determined by what kind of input each module needs.4 He was thus
able to ®nesse the central interesting question of the control and ¯ow of
information between components by building a solution into the compo-
nents themselves (each component transforming a speci®c type of input
into a speci®c kind of output appropriate as input to another component,
such that the logical ordering of components does not need to be stipu-
lated).5 In the theory of meaning, we seem to have inherited this thinking:
semantics is normally held to be logically prior to pragmatics, and thus
semantics is autonomous with respect to pragmatics in just the way that
Chomsky imagines syntax to be autonomous with respect to semantics.
This chapter attempts to demonstrate that this is simply not a tenable way
of thinking in the theory of meaning.6
The crucial fact that I will try to establish is that generalized conver-
sational implicatures seem to play a role in the assignment of truth-
conditional content.7 This may seem not only a distinctly odd idea but even
de®nitionally impossible, because implicatures are often partially de®ned
in opposition to truth-conditional content. Some scholars have therefore
argued that any such inferences should be called something else (e.g.,
167 Implicature and the Semantics/Pragmatics Interface

``explicatures'' or ``implicitures''), but calling them something else will not

change the fact that they are the very same beasts. In the end, it will be
seen, I think, to be an inevitable conclusion that GCIs play a role in truth-
conditional content. Now if there is such a role for generalized implica-
tures, then it will have major implications for the architecture of the
theory of meaningÐwe will need to drastically revise our understanding of
the pragmatics/semantics interface and the ¯ow of information between
the components in a general theory of meaning.
A revolution of that order may not be prompted by the role of gener-
alized implicatures aloneÐafter all, energetic theorists may ®nd some
other way of accounting for the facts. However, there are convergent
strands of thinking that are heading toward the same general conclu-
sionÐnamely, that a rethinking of the pragmatics/semantics interface is
in orderÐso that my particular thesis (that generalized implicatures can
contribute to truth-conditional content) is merely one of a number of
contributory arguments. Other arguments heading in the same direction
can be found, on the one hand, in new pragmatic theories like Sperber
and Wilson's Relevance theory and the arti®cial-intelligence work that
goes under the rubric of ``local pragmatics'' (e.g., Hobbs et al. 1987); and,
on the other hand, in new semantic theories like Situation Semantics
(Barwise and Perry 1983) and the context-change theories of semantic
representation (as in the Discourse Representation Theory [DRT] of
Kamp 1981 or the File Change Semantics of Heim 1982). These trends,
together with the evidence adduced here, do indeed suggest that the entire
theory of meaning requires a radical new architecture that will integrate
old insights in new ways. My purpose is to be brash about the implica-
tions, to try and force them onto the agenda of those engaged in formal
semantics.
Although the arguments here can be taken as contributory to these
general trends, they are also intended to favor some of them against
others. Thus I believe the results favor a DRT conceptualist approach to
formal semantics, where there is a speci®c kind of level of mental repre-
sentation for meanings (Kamp 1981) over the realist position favored in
Situation Semantics (Barwise and Perry 1983). I also believe that the
results argue against Sperber and Wilson's (1986) Relevance theory by
showing that generalized implicatures exist and play a systematic role
in truth-conditional content. The facts also argue against the AI work in
local pragmatics where scalar implicatures, for example, play no role in
embellishing logical forms. In this sort of way, I hope to advance the
168 Chapter 3

search for the new architecture for the theory of meaning by placing fur-
ther constraints on the solution.
In the end, the picture I advance is much less radical than some alter-
native possibilities (some of which will be considered below). To spill the
beans: essentially, I argue that the theory of meaning has the components
or levels we always thought it hadÐit is just that we have to reconstrue
the kinds of relations that hold between them. Thus, semantics remains
the familiar beastÐit consists of two rather separate processes: one that
cranks out a level of semantic representation (albeit much more abstract
and underdetermined than used to be thought) from surface structure,
and another process of semantic interpretation that interprets a richer
level of representation (where truth conditions, entailment relations, or
sense relations can be captured). Both levels of representation are closely
related to linguistic structure, and they are intimately connected because
the ®rst level of representation is functionally constrained by the role it
has to play in semantic interpretation. Pragmatics too encompasses more
than one process: on the one hand, default inferences can be calculated on
the ¯y, given fragments of semantic representation; on the other hand,
further pragmatic inferences (of both a speci®c and default character) can
be calculated given the results of semantic interpretation. The novel sug-
gestion here is that semantic and pragmatic processes can interleave, in
ways that are probably controlled by the constructional types in the
semantic representation. Where pragmatic inferences end up embedded in
semantic representations or their interpretations, I shall talk about intru-
sion; some kinds of intrusion are relatively uncontroversial, but the sys-
tematicity of GCI intrusion has not been properly appreciated, and it is
this that requires a rethinking of the semantics/pragmatics interface.
Although this picture is relatively conservative, the interaction between
pragmatic and semantic factors raises many terminological problems:
what has happened to the distinctions between semantics and pragmatics,
sentence-meaning and utterance-meaning, what is said and what is impli-
cated? What do we mean by ``semantic representation'' when even the
®rst level may involve pragmatic disambiguation and the second level is
shot through with pragmatic contributions? I attempt no terminological
reformation here, and this is largely because many of the old notions are
good enough, provided we focus on processes or operations on repre-
sentations rather than on the representations themselves. The representa-
tions are going to become strange, hybrid semantico-pragmatic entitities,
but the processes that construct them and interpret them are perhaps just
169 Implicature and the Semantics/Pragmatics Interface

the familiar old semantic and pragmatic processes, with the addition of
presumptive meanings. The picture will become clearer, I hope, as we
proceed, so I ask the reader to bear with me.
One further caveat: because both pragmatic theory and semantic theory
are undergoing rapid change at the moment, any proposals about the
exact location of an interface are bound to be theory-dependent. There
are enough new varieties of semantic theory, each constantly evolving, to
make it well nigh impossible to predict how each would maneuver around
every pragmatic shoal. It will often be the case that some semantic
reanalysis might be able to absorb the particular example of what I am
arguing to be implicatural content. My strategy is thus to try and foresee
and deal with the counterarguments as far as this is possible but also to
simply ¯ood the opposition by showing that there is a vast range of data
that would be simply explained by general pragmatic processes, but that
would require an insuperable amount of ad hoc semantic reanalysis (if
that were possible at all).
Once again, I will concentrate on just one, very large set of pragmatic
factsÐthe generalized conversational implicatures (GCIs) that are the
subject of this book. In certain respects, this may amount to an arti®cial
restriction because the general theme to be explored is the extent to which
pragmatic inferences in general contribute propositional content. But
there are particular reasons to concentrate on GCIs. One is that (under-
explored though they are) they are well understood in comparison to any
other kind of pragmatic inferenceÐthere are clear predictions about what
interpretations should arise in what conditions. Second, it is the default
nature of GCIs that causes them to be deeply entangled in grammar and
semanticsÐthey are both hard to distinguish from encoded content, and
they exert functional pressure on syntax, lexicon, and semantics (we have
already seen for example, that there is a redundancy constraint: certain
kinds of GCIs block grammatical and lexical encodings; see Horn 1989).
Third, it is not clear that particularized conversational implicatures play
the same kind of role in ®xing propositional content (but see section
3.5.1.1 and, e.g., Bach 1994).
Let me summarize the kind of theory of GCIs that has emerged in the
preceding chapters:
1. A rationale is provided, in terms of ways of overcoming the constraints
of the narrow bandwidth of human speech, for three cross-cutting
inferential heuristics that more or less amount to Grice's maxims of
170 Chapter 3

Quantity 1, Quantity 2, and Manner (Relevance inferences can't be

GCIs).
2. This partially motivates a typology of GCIs, into Q, I, and M types
with subtypes.
3. A single utterance often gives rise to con¯icting subtypes of GCI, and
the clash (or projection problem) is resolved by a set of priorities,
Q > M > I (with ordered subtypes).
4. GCIs can up to a point be formalized, and their defeasible behavior in
complex sentences modeled, either in terms of an incremental addition
of ``all the news that ®ts'' or perhaps in terms of some nonmonotonic
logic (e.g., default or autoepistemic) of the kinds currently being
explored in AI.
5. Finally, I made some observations that will be developed in this
chapter: GCIs play a role in fundamental linguistic processes like ref-
erence resolution, generality narrowing, scope assignments, and so on.
I use the details of this typology of GCIs in the material that follows,
but little depends on the details of my schemeÐI simply use the typology
to demonstrate the extensive range of pragmatic intrusions into semantics.

3.2 THE RECEIVED VIEW: SEMANTICS AS INPUT TO PRAGMATICS

Grice (1989: 25, 87f ) made the distinction between saying and implicat-
ing, but his de®nition of what is said is complex and by no means clear
(see Harnish, 1976: 332¨; Bach 1994). Roughly, though:8
(1) U said that p by uttering x i¨:
a. x conventionally (``timelessly'') means p
b. U speaker-meant p (this condition serves, e.g., to select one of a
number of ambiguous readings)
c. p ˆ the conventional meaning of x minus any conventional
implicatures (i.e., any non-truth-conditional but conventional
aspects of meaning that `indicate' but do not contribute to `what
is said').
There is a philosophical literature on Grice's notion of ``what is said''
that I will refer to in passing (but see especially Bach 1994, Recanati
1993). But linguistic commentators have on the whole been happy to take
Grice's notion of ``what is said'' as mapping more or less onto truth-
conditional content (an interpretation encouraged by Grice's treatment
of the logical connectives). (Due allowance has to be made of course for
171 Implicature and the Semantics/Pragmatics Interface

the satisfaction conditions, rather than truth conditions, of nonassetoric

speech acts, but here I abstract away from this additional problem.)
However, as Atlas and Levinson (1981: 1¨ ) note, two utterances with the
same truth conditions (e.g., It's done and It's done and if it's done, it's
done) can have quite distinct implicatures (see also the discussion in of
litotes in section 2.4.1.3, and why, in particular, @@ p does not have the
same import as p). Atlas and Levinson (1981) proposed therefore that
what is said, in Grice's sense, should be related to the level of semantic
representation or logical form, and this is the position I take here.9 (Such
a level is almost certainly much more abstract and underspeci®ed than
was previously thought, but more of that later.)
Grice (1975: 44) also notes some essential prerequisites for an identi®-
cation of what is said, and these will prove crucial:
``for a full identi®cation of what the speaker had said one would need to know (a)
the identity of [the referents], (b) the time of utterance, (c) the meaning on the
particular occasion of utterance, of the phrase [uttered].''
In short, we need to have resolved reference, ®xed indexicals, and dis-
ambiguated expressions before we can identify what a speaker has said on
an occasion of utterance.
In contrast to what is said, and on the basis of what is said, a speaker
may conversationally implicate further propositions. By saying p, utterer
U conversationally implicates q if:
(a) U is presumed to be following the maxims,
(b) the supposition q is required to maintain (a), and
(c) U thinks the recipient will realize (b).
``To work out . . . a particular conversational implicature the hearer will rely on . . .
the conventional meaning of the words . . . together with the identity of any refer-
ences involved'' (1975: 50).

The central idea is simple and clear. What is said is the input to the
pragmatic reasoning responsible for output of implicatures: what is
implicated is calculated on the basis of what is said (together with aspects
of how it was said, in the case of Manner implicatures). ``What is said''
seems to be designed to be equivalent to the proposition expressed by the
use of a sentence or the truth-conditional content of the utterance, and
is in turn dependent on reference resolution, indexical ®xing, and dis-
ambiguation. Although it is impossible to provide a diagram that is neu-
tral over all the di¨erent theoretical positions, I believe that the received
172 Chapter 3

view of an abstract (competence) model of meaning assignment is more or

less accurately captured in ®gure 3.1 (this picture clearly doesn't capture
the view in Situation Semantics or DRT [more on this later] nor does it
incorporate theory-speci®c components of grammar). Here, there is a
logical ordering of operations, so that the syntax provides strings with
structural analyses, selected between by a disambiguation device of some sort
(usually not directly addressed in linguistic theory). The disambiguated
structure can then be associated with a semantic representation or logical
form, which in turn can then be associated with a model-theoretic inter-
pretation, but only after the ``®xing'' of indexicals with the aid of a highly
restricted kind of pragmatic inputÐnamely, the values of the pragmatic
indices obtaining in the speech situation. The interpretation is then taken
to provide a proposition expressed by the utterance on an occasion of use.
The output of the semantics is then the input to the pragmatics, where
Gricean mechanisms provide augmented (and occasionally, as in ``¯outs,''
altered) interpretations.
In this chapter, I examine two major kinds of problem for the received
view. The ®rst problem, covered in section 3.2.1, is that even if we accept
Grice's rather restricted account of what is involved in ®xing what is said,
namely the factors already mentioned (reference identi®cation, deixis, and
disambiguation), implicatures can be seen paradoxically to play a role in
the establishment of what is said. This leads to an interesting chicken-and-
egg problem about the priority of what is said versus what is implicated.
At this point, it will prove useful to pause to consider possible responses.
The second problem, examined in section 3.3, is that the content of gen-
eralized conversational implicatures can fall within the scope of logical
operators and other higher level processes of semantic composition. In
section 3.4, I turn once again to reference ®xing, now in detail, and show
that GCIs interact with some of the crucial philosophical distinctions in
reference. These three lines of evidence converge in requiring a rejection
of the received view of the semantics/pragmatics interface and point in the
direction of a new architecture for the theory of meaning.

3.2.1 Grice's Circle: Implicatural Contributions to ``What Is Said''

An initial problem for the received view is just this: Grice gives three
preconditions to determining what was said: (a) identifying the referents,
(b) ®xing the deictic parameters, and (c) disambiguating the linguistic
string in question. To these we may add at least the further preconditions
to determining the proposition expressed: (d) unpacking ellipses and (e)
173 Implicature and the Semantics/Pragmatics Interface

Figure 3.1
Received view of the semantics/pragmatics relation
174 Chapter 3

narrowing generalities. But each of these turn out to involve exactly those
inferential mechanisms that characterize Gricean pragmaticsÐthat sup-
posedly postsemantic (and perhaps nonlinguistic) process that maps literal
meanings (the output of semantic interpretation) into speaker-meanings.
In the following subsections, I brie¯y illustrate this claim that Gricean
inferences are involved in determining what proposition has been ``liter-
ally'' expressed. Note that no claim is being made that Gricean infer-
ences are exclusively responsible for reference ®xing, disambiguation,
and indexical resolutionÐjust that they sometimes, and quite plausibly
often, are. Later I will return and substantiate some of these claims in
much greater detail. The aim here is simply to make a prima facie case in
order to elucidate the structure of the argument.

3.2.2 Disambiguation
It is uncontroversial that arbitrary amounts of world knowledge might be
involved in disambiguation, as the following examples should make clear
(example (2a) is after Hirst 1987 and (2c) is after Winograd 1972).
(2) a. The view could be improved by the addition of a plant out there.
(lexical ambiguity of plant)
a 0 . The view would be destroyed by the addition of a plant out there.
b. Mary left [the book] [on the bus.]
b 0 . Mary left [the book on the atom.] (structural ambiguity of PP
attachment)
c. The administrators barred the demonstrators because they
advocated violence.10
0
c . The administrators barred the demonstrators because they feared
violence.
d. He looked at the kids [in the park] with a telescope.
d 0 . He looked at the kids [in the park with a statue].
It's a more interesting claim that Gricean principles would be involved.
For example, a moment's re¯ection will convince that the maxim of
Relevance will be involved in selecting those readings that make pos-
sible ``sequiturs'' to a prior utterance (see also Sperber and Wilson 1986:
183¨ ). The following is a simple case (from Hirst 1987; note though that
prosody would often disambiguate this sort of case):
(3) A: i. ``What are they doing in the kitchen?''
ii. ``What kind of apples are those?''
B: ``They are cooking apples.''
175 Implicature and the Semantics/Pragmatics Interface

It is much less obvious that GCIs might be involved in disambiguation,

but a few examples should make that plausible:
(4) a. ``He's an indiscriminate dog-lover; he likes [some cats] and dogs.''
b. He likes [[some cats] and dogs]
Interpretation: `he likes some cats and dogs (in general)'
GCI: `he likes some-but-not-all cats'
c. He likes [some [cats and dogs]]
Interpretation: `he likes some cats and some dogs'
GCI: `he likes some-but-not-all cats and some-but-not-all-dogs'
Example (4a) makes the point this way. Suppose we take the last NP of
the sentence to have the structure: [[some cats] and dogs]. Then he likes
some cats has the scalar (Q) GCI `he doesn't like all cats'; this interpreta-
tion is compatible with the ®rst clause, which implies that he likes all
dogs. Now suppose instead that we take the NP to be structured this way:
[some [cats and dogs]], with the interpretation `he likes some cats and he
likes some dogs'. Then there will be a GCI `he likes some-but-not-all cats
and some-but-not-all dogs'. But such an implicature is inconsistent with
the ®rst clause He's an indiscriminate dog-lover; if we then reject the syn-
tactic analysis that gives rise to the inconsistent implicature, we'll arrive at
the right bracketing of the NP.
This account has some plausibility and some generality. Consider:
(5) a. ``He has a whole pack of dogs. He has three cats and dogs.''
b. He has [[three cats] and dogs].
Interpretation: `he has three cats and an indeterminate number
of dogs'
GCI: `he has no more than three cats'
c. He has [three [cats and dogs]]
Interpretation: `he has three cats and three dogs'
GCI: `he has no more than three cats and no more than three dogs'
Just as before, the structural analysis that has a semantic interpretation
that induces an implicature consistent with the context will be preferred
over another analysis that induces an inconsistent implicature. Note that,
on the assumption that the ``exactly'' interpretation is only implicated, the
semantic content alone cannot account for the favored readings.
Another example with structural ambiguity might be:
(6) a. ``She heard some of the ®recrackers in the kitchen.''
b. She heard [some of the ®recrackers] [in the kitchen] (but others
while she was in the sitting room).
176 Chapter 3

c. She heard [some of the ®recrackers in the kitchen] (but some of

them must have been quieter).
d. Gloss of literal content: `She heard at least some of the ®recrackers
in the kitchen.'
e. GCI: `She heard some but not all of the ®recrackers in the kitchen.'
f. She heard [some but not all of the ®recrackers] [in the kitchen]
(but others while in the sitting room).
A PP like in the kitchen in (6a). can often be attached high (here to the
VP) or low (here to the object NP), as clari®ed in (6b) and (6c). The literal
content of (6a), roughly glossed in (6d), does not predispose us one way or
the other. But the GCI in (6e) biases us toward the (6b) reading, as in (6f ),
because otherwise it will take special circumstances (e.g., variable hearing
loss, uneven quality of ®recrackers) to explain why the protagonist only
heard some but not all of the (presumably equally loud) ®recrackers that
went o¨ in the kitchen.
GCIs may participate in the disambiguation of lexical ambiguities,
especially perhaps through the I-inference to stereotypical extensions.
Thus the kind of disambiguations often attributed to brute encyclopedic
knowledge may be mediated by Gricean procedures: the speaker must
presume that the addressee will use the same salient stereotype (regardless
of its factual probabilities) to resolve the ambiguity (as in He took another
drink, where drink is interpreted as the alcoholic variety).
If the observation that GCIs can select between ambiguities is correct
and fully general, it is perhaps not self-evident why. GCIs are cancelable;
thus given a context inconsistent with them, they quite routinely evapo-
rate. The answer must be that the recipient works on the assumption that
the speaker would so design his utterance that, other things being equal,
the predictable GCIs would go through.11 Thus given the option of pre-
serving the implicatures, an interpreter generally chooses that interpreta-
tion that will do so.
These examples are less than fully natural because they attempt to iso-
late just one kind of pragmatic inference at a timeÐin many, more natu-
ral, examples, GCIs of many di¨erent kinds collaborate, as it were, to
favor one reading among a set of ambiguous ones:
(7) a. ``She submitted an uneven article. Some of the paper was
excellent.'' (I, Relevance, Q)
b. article1 : `thing'
article2 : `short academic treatise'
177 Implicature and the Semantics/Pragmatics Interface

paper1 : mass noun `stu¨ for writing on'

paper2 : count noun `academic treatise'
uneven1 : `bumpy'
uneven2 : `variable quality'
c. GCIs:
i. ``Some of the paper'' Q ‡> `not all of the paper'
ii. the referent of an article2 ˆ the referent of the paper2
(I-preference for local coreference)
iii. uneven article should be read `variable quality academic
treatise' rather than `bumpy academic treatise' (attributable to
an I-inference to the stereotypical interpretation)
Without specifying all the inferential mechanisms, it seems likely that a
global process of inference to the best interpretation will seek an inter-
pretation consistent with all the GCIs (Atlas and Levinson 1981). Such an
interpretation will minimize the entities and properties postulated, as
suggested by the I-principle. In many cases, that interpretation will select
between ambiguous meanings and structures.

3.2.3 Indexical Resolution

It is uncontroversial that certain pragmatic factorsÐnamely, pragmatic
indices or deictic parametersÐare input to semantic interpretation. But it
has been supposed that these indexical factors are of a limited kind and
may be swept within a truth-conditional apparatus.12 There are a number
of distinct assumptions here, each of which can be challenged:
1. Indexical expressions are few in number and limited in kind. (But in fact
there are many di¨erent kinds of covertly indexical expressions; see
EncË 1981, Barwise and Perry 1983, Mitchell 1986);
2. The indexical parameters or indices are a ®xed, small set for e.g.,
speakers, addressees, times of utterance, places of utterance, and indi-
cated objects (see, e.g., Lewis 1972). (But crosslinguistic studies show
that many more parameters may be requiredÐfor example, Kwakiutl
forces a choice between ``visible/nonvisible to participants'' on most
NPs in the same way that English forces a choice of tense on most
verbs; see the catalogs of exotic deictic parameters in Fillmore 1975,
Levinson 1983, Anderson and Keenan 1985).
3. The assignment of an indexical value to an indexical expression is a
simple process of a direct mapping between context and expression.
There is a look-up table as it were, that assigns the expression I to the
178 Chapter 3

current speaker, the expression here to the place of utterance, now to

the time of uttering, and so on.
It is this last assumption that I challenge here: that indexicals require no
inferential resolution and, in particular, that the assignment of indexical
values has nothing to do with pragmatic inference in general, and thus
nothing to do with GCIs. The assumption is undermined by examples
like the following, organized under the major traditional dimensions of
deixisÐnamely, place, person, and time (Levinson 1983: chap. 2):
(8) Place
a. ``Take these three drinks to the three people over there; take these
four to the four people over there.''
Q ‡‡> `Take the drinks that are exactly three in number over
there, take the drinks that are exactly four in number over there.'
b. ``Some of those apples are ripe; those ones are for you, the other
ones are for Bill.''
Q ‡‡> `Some but not all of those apples are ripe; just the ripe
ones are for you . . .'
c. ``This sofa is comfortable; come over here'' vs. ``California is
beautiful; come over here.''
I-principle: delimits here by anaphoric resolution.
(9) Person
a. A: i. ``He said you did it?''
ii. ``He confessed?'' (alternative possible question)
B: ``He said I did it'' (maxim of Relevance)
b. ``Some of you know the news; I'm not talking to you; I'm talking to
the rest of you.''
Q-principle: delimits to you to `to some but not all of you'
(10) Time
a. ``I used to have a car like that.''
Q-implicates: `I no longer have a car like that'
b. ``The meeting is on Thursday.''
Q-implicates: `not tomorrow' when tomorrow is Thursday.
Consider the use of example (8a) in the absence of gestural indication
(on the phrase these three drinks) and in a context where the speaker is
pouring wine into two sets of glasses: one set has exactly three glasses, the
other exactly four. Given that the numeral words in English can be
argued to only implicate an upper bound, three glosses as `at least three'
179 Implicature and the Semantics/Pragmatics Interface

at the level of semantic content, although it normally carries a GCI `at

most three'. Thus the expression these three glasses has the semantic con-
tent `these glasses of cardinality no less than three'. It follows that, in the
absence of gesture, the expression might equally designate either set of
glasses. Yet the utterance would be perfectly felicitous in the circum-
stances sketched, because any interpreter would use the GCI `at most
three' to resolve the indexical reference. (Without such pragmatic resolu-
tion, the utterance ought to be anomalous; because it isn't, we must
assume that GCIs play a role in ®xing indexical reference.)
Consider now example (8b). The expression some of those apples is
consistent with all of those apples but is delimited by a GCI to `some but
not all of those'. It is the pragmatically delimited set, a proper subset of
the apples, that is introduced as a discourse referent by some of those
apples. This is shown by the fact that the following demonstrative pro-
noun those ones must refer to a subset of the apples (cf. similar observa-
tions for nonindexicals by Kadmon (1987)). If the phrase those ones is
both anaphoric and deictic, which I believe is the right analysis, then the
truth conditions for those ones are for you are dependent on the GCI that
restricts the set of apples to less than all. The point may be sharpened
by the subsequent contrastive referring phrase: without the pragmatic
restriction on those ones, the other ones has no clear denotation.
Example (8c) makes a simple but very general point. Indexical param-
eters are not alone su½cient, on the whole, to give one a reasonably cir-
cumscribed denotation (the one clear exception is the pronoun I, which is
individuated by the speaker parameter alone; but see (9a)). For example,
here denotes some region inclusive of some part of the speaker's imme-
diate location; bring it here might mean `bring it to my ®ngertips'
(imagine the surgeon, in need of a scalpel, to the nurse) or `bring it to Los
Angeles' (perhaps we are talking about something that needs no exact
location, like an advertising blimp). Suppose A is in Los Angeles and B is
in New York and the following exchange takes place:
(11) A: ``Where's the conference being held?''
B: ``It's being held here.''
If here means `some space inclusive of the speaker's location', then would
B have told the truth if the conference is being held in Princeton? Surely
not. But without further pragmatic restriction of the extent of here, what
B said was true (a proper account of such restrictions can be found in
Scheglo¨ 1972). In the cases in (8c), it is clear that pragmatic restriction of
180 Chapter 3

the extent of here is made by anaphoric linkage to antecedents, a process

that comes under the I-inferential category: here is essentially indexical
but anaphorically restricted.
Person deixis might be thought to be immune from pragmatic reso-
lution, but this is not the case. After all, even I can be interpreted as
referring to the speaker of directly quoted speech (as in (9a), where a
Relevance implicature would be instrumental in selecting interpretations
according to alternate prior utterances). You clearly may refer to more
than one addressee, or parties not addressed at all, or even some subset of
the parties just previously addressed as in (9b). Here only if some of you
implicates `not all of you' will the rest of you have denotation.
Temporal deixis is in fact especially permeated by Gricean mechanisms,
because temporal reference is normally made more precise by opposition
to other temporal expressions that were not usedÐthat is, by Q- and
M-implicature. Thus used to V in (10a), where V is verb, will normally
implicate `no longer V' by opposition to the simple present (yet there's
no inconsistency in ``I used to love swimming, and in fact I still do''; see
chap. 2). Or more simply, Thursday said on Wednesday will Q-implicate
`not tomorrow', because tomorrow would be the more informative expres-
sion and was not used (there are many Thursdays, but only one tomor-
row, given a today!).
There can be little doubt then that GCIs, along with other pragmatic
processes, are involved in the resolution of indexical expressions, without
which, of course, no exact proposition will be expressed.

3.2.4 Reference Identi®cation

There are two ways in which reference might be said to fail: (a) cases
where there is nothing in the domain of discourse that ®ts the descriptive
content of the referring expressionÐvariously analyzed as yielding sen-
tences that are false (Russell 1905) or unspeci®ed for truth or falsity
(Strawson 1950); and (b) cases where there are more than one entity in the
domain of discourse (or more than one discourse referent in the sense
of Karttunen 1976) that ®t the descriptive contentÐcases of referential
ambiguity as it were.13 Both are critical to the establishment of ``what is
said'': we can hardly be sure what proposition is expressed if we are
unsure what the arguments of a predication denote (or, sometimes, which
other discourse referents they are identical to).
In section 3.4 below I provide a detailed argument for the role of GCIs
in nonanaphoric reference determination. Here I indicate brie¯y how
181 Implicature and the Semantics/Pragmatics Interface

GCIs may be claimed to have a role in anaphoric reference determina-

tion. First, recollect that it is uncontroversial that pragmatic factors are
involved in the resolution of discourse anaphora (see, e.g., the case of the
anaphoric reference to administrators or demonstrators in (2c)).
Anaphoric expressions are regularly less prolix and more semantically
general than the expression with which the discourse referent was ®rst
introduced (by an inde®nite or sometimes a de®nite descriptive phrase).
This, I have argued (Levinson 1987a,b) is a natural by-product of the
I-principle (Grice's ``Don't say more than necessary''). Thus if you intend
coreference, you should make the referring expression short or semanti-
cally general. The recipient's corollary is: ``From a semantically general,
minimal reference form choose the interpretation that maximizes infor-
mative content,'' which will generally amount to a preference for local
coreference. On this account, example (12a) has a preferred reading (more
or less regardless of how many competing prior antecedents for he there
may be) in which he is coidenti®ed with John.
(12) a. ``John came in and he sat down.''
I-induced coreference preference
b. ``John came in and the man sat down.''
M-induced disjoint preference
c. ``Felix voted for him.''
Q-induced disjoint reference
d. ``Only Felix voted for him.''
Can be coidenti®ed, because ``Only Felix voted for himself ''
would mean something di¨erent.
The identi®cation can be avoided by choosing a de®nite referring expres-
sion that is not maximally semantically general, as in (12b). Here the
speaker's use of the longer expression the man will only be warranted
by an assumption that the speaker wishes to avoid what the natural
interpretation associated with the pronoun would be, as in (12a). By the
M-principle, there is a preferred interpretation under which the man and
John are identi®ed with di¨erent individuals.14
More controversially, Reinhart (1983) has suggested that (contrary to
Chomsky's binding conditions) the object pronoun in (12c) is only inter-
preted as disjoint from the subject because of a Manner GCI (see also
Horn 1984). My own analysis (Levinson 1987a,b) is that this is a Q-
opposition: given the choice of the coindexing re¯exive form or the free
pronoun, the use of the latter suggests that the former is not intended
182 Chapter 3

via a metascale hREFLEXIVE, NONREFLEXIVE PRONOUNi. This

predicts that disjoint reference in sentences like (12c) should be defeasible
(cancelable in context). Thus in the right circumstances, the pronoun may
be coidenti®ed with the subject as indicated in (12d). Reinhart's account is
controversial because it yields up much of what has been supposed to be
grammar to pragmatics; nevertheless the defeasibility of noncoreference
assignments seems to make better empirical predictions. I return to these
issues in chapter 4.
But whether this pragmatic account of the disjoint reference of non-
re¯exive clausemate NPs is correct or not, the I- and M-implicatures
would seem to play a crucial role in reference determination, as sketched
in (12a,b). By the same token, I-inferences would seem to be responsible
for ``bridging'' inferences, legitimating the inference to the best interpre-
tation in cases like the following:
(13) a. ``The bridge collapsed. The wood was rotten.''
I ‡‡> `The wood of which the bridge was made was rotten.'
b. ``The bridge collapsed. A wooden strut was rotten.''
I ‡‡> `A crucial wooden strut of the bridge was rotten.'
c. ``Every old car needs to have the battery checked.''
I ‡‡> `Each car needs to have the battery which belongs to
that car checked.'
The ®rst example seems to require that we suppose that the bridge was
made of wood. But we only need to make that supposition in order to ®nd
a coherence relation holding between the two clauses. The supposition
that there is a causal relation requires the supposition that the bridge was
made of wood; a speci®c mass of wood is then available for de®nite ref-
erence. The inference is supported by the I-induced submaxim ``Don't
state what can be inferred'' and is line with our stereotypical expectations
of the kinds of materials bridges can be made of. (Some theorists [Heim
1982, Kadmon 1987] have supposed that the inference is motivated only
by the ``accommodation'' of the de®nite referring expression (the wood )
which requires a familiar or given antecedent; but the bridging inference
in the inde®nite version in (13b) shows that more global I-principle pro-
cesses are involved.)
The second kind of example in (13c) (due to Kempson 1986) is inter-
esting because it shows that de®nite expressions acting as variables under
explicit quanti®cation may still require pragmatic bridging: the assump-
tion must be made that each car has one battery. The general form of the
183 Implicature and the Semantics/Pragmatics Interface

I-inference in these cases is: given a de®nite description with no obvious

antecedent, make that minimal assumption that will tend to the best
(maximally informative) interpretation (often, that `the x' is part of, or
associated with, a previously mentioned `the y').

3.2.5 Ellipsis Unpacking

If A says ``Who came?'' and B replies ``John,'' has B said something that
can be true or false? Surely; he has expressed the proposition that John
came. Just as the second maxim of Quantity (``Don't say more than nec-
essary'') or my I-principle motivates minimal expressions like pronouns,
so, together with a maxim of Relevance, it motivates ellipsis. The corre-
sponding pragmatic principle of construal is (very crudely): supply the
missing predicates (here marked between angled brackets, as in (14a))
from the preceding discourse context. Ellipsis and anaphora are of course
closely linked, with the interpretation of the ``unsaid'' governed by the
I-principle, in tandem with a Relevance maxim (as discussed in chap. 1),
as in example (14b).
(14) a. A: ``Who came?''
B: ``John'' hcamei (Relevance and I-implicature)
b. ``Over to 41, there's a nice move. Gets fouled by number 35 of
the Spartans.''
c. A: ``Which side got three goals?''
Q-implicates: `Which side got three-and-no-more goals?'
B: ``Tottenham Hotspurs'' hgot three-and-no-more goalsi
d. ``Our team got three goals and theirs did hget three-and-no-
more-goalsi too; so it was a draw, three all.'' (Q)
e. A: ``They won't visit Mary's parents.''
B: ``Old grudge.'' (from Barton 1988)
Now notice that the interpretation of a gapped or elided constituent
carries the implicatures of the antecedent. Thus in (14c), if the predicate
of A's utterance carries a scalar implicature, B's ellipsis over the same
predicate does too. As for (14d), does the conclusion follow from the
®rst sentence and the rules of soccer? Only if our team got exactly three
goals and no more, and their team got exactly three too. But that is the
proposition, ampli®ed by implicatures, that is indeed expressed (see
section 3.3.).
In many structural loci, ellipsis is a fundamentally linguistic, rule-
governed process, so that recovery of elided material can in fact be thought
184 Chapter 3

of as guided by syntactic and semantic principles. But it is clearly not true

for examples like (14e), where complex reasoning and not rule application
is involved in proposition recovery. It would take us too far a®eld to
explore ellipsis in any depth, but it is self-evident that the recovery of
the proposition expressed will depend in these cases often on substantial
Gricean inference (see Barton 1988 for detailed explorations). The con-
sequence is this: the pragmatics is involved in the recovery of the elided
linguistic material, which must then be semantically interpreted, at which
point we apparently need another pragmatic processing stage to recover
the implicatures of the elided material.

3.2.6 Generality Narrowing

A central and persuasive tenet of Radical Pragmatics has been that what
had previously been interpreted as ambiguity was better treated as gener-
ality plus implicaturesÐhence the claims that despite appearances or, not,
possible, some, and so on are univocal, not ambiguous. If correct the cor-
ollary is that many natural language expressions are maximally general,
to the point of bordering on vacuity (see Ruhl 1989). Consider the fol-
lowing examples:
(15) a. ``Fixing this car will take some time.''
b. ``I've eaten breakfast.''
c. ``The ¯ag is white.''
d. ``The British ¯ag is white.''
e. ``If the ¯ag is white, then it can't be blue.''
f. ``John is tall.''
g. ``Since the surgeon came quickly, he cannot be blamed.''
Without pragmatic restrictions (Sperber and Wilson 1986: 189f point
out), (15a) will border on the tautological, whereas (15b) will always be
true of any ordinary breakfast-eater, even before he gets out of bed.15
Similarly, if (15c) only means `the ¯ag is at least partially white' (Harnish
1976), then (15d) will be true but just grossly misleading and (15e) will be
false. At an extreme, it is not totally implausible to suggest that gradable
adjectives only predicate of their arguments that the entity in question has
the relevant property to some (unspeci®ed) degree (Sadock 1981). Thus
the content of (15f ) might be little more than `John has height'! But the
truth of (15f ) is judged against some implicit category that John is
assigned to (tall for a seven-year-old is not tall for a basketball player).
(Alternatively, all degree adjectives may be treated as implicit com-
185 Implicature and the Semantics/Pragmatics Interface

paratives containing a covert indexical-like argument place.) As for

manner adverbs like quickly in (15g), they too will hardly restrict the
interpretation of their predicates unless some implicit standard is taken
into account. Suppose we have an emergency in the hospital, and the
surgeon started jogging from his home when he should have jumped in his
car; no judge would judge (15g) true.
One should distinguish two possible views here. On the original
Gricean or standard Radical Pragmatics view, the lexical content of
such expressions is taken to be their weakest exhibited meaning, with the
assumption that the pragmatics supplements the expressions to give their
stronger interpretations. Thus, for example, one takes or to have as lexical
content the speci®c meaning of inclusive disjunction, strengthened to
exclusive disjunction by the GCI based on the scale hand, ori, which
yields a `not both' implicature from the failure to use and. Let us call this
the weak univocality view. On another view, energetically purveyed by
Atlas (1979, 1989), the lexical content of expressions may be indi¨erent,
or semantically general over, both the weak and strong interpretations.
On this view, the lexical content is also weak, but it is not speci®c. For
example, Atlas (1993a) maintains that the semantic content of three
encompasses both the weak interpretation (`at least three') and the strong
one (`exactly three'), and is specialized in context. He dubs this the
semantic generality view. However, because both views involve semantic
generality (either based on the weak interpretation or on an overarching
abstraction), I use the phrase to cover both kinds of analysis.
GCIs play a systematic role in generality narrowing. Horn (1984) notes
in particular that both Q-narrowing and I-narrowing are frequent enough
to play a systematic role in semantic change. Here are some examples of
implicature-based narrowing:
(16) a. ``I ate a few of the cookies''
‡> `not all of them.' (Q-narrowing)
b. ``He hurt a ®nger''
‡> `not a thumb.' (Q-narrowing)
c. ``It's a big city''
‡> `not enormous.' (Q)
d. ``I put the warm pot on the table and left the hot one on the
stove.''
‡‡> `I put the warm-but-not-hot pot on the table.' (Q)
e. ``He had a drink''
I ‡> `alcoholic drink.' (I-narrowing)
186 Chapter 3

f. ``This is a bread knife''

I ‡> `for slicing bread.' (I)
g. ``This is a steel knife''
I ‡> `made of steel.' (I)
h. ``Larry's book is about negation''
I ‡> the book Larry wrote. (I)
(cf. ``Larry's book is a thriller he got from the library.'')
If pragmatic narrowing of semantically general predicates is not taken
into account, the proposition expressed by many sentences will be general
to the point of vacuity. Intuitions about truth conditions are very much
more speci®c: surely when we say Larry's book is about negation the
proposition expressed depends on the implicit pragmatic narrowing of the
generality associated with the possessive. The possessive encodes that
there is some intimate relation between Larry and the book, but it is an
inference guided by Gricean principles on the one hand and contextual
information on the other that gives a speci®c content to that relation (see
also Sperber and Wilson 1986), and thus plays a role here in reference
determination. It is worth pointing out that all the oblique cases in Case-
marking languages have this kind of semantic generality (as with Latin
Ablative, which expresses source, cause, agency, instrument, deprivation,
or direction from), so that no reasonably circumscribed proposition will
be expressed without pragmatic resolution of the semantic content.

3.2.7 Some Interim Conclusions: Responses to Grice's Circle

The argument so far is this. Grice's account makes implicature dependent
on a prior determination of ``the said.'' The said in turn depends on dis-
ambiguation, indexical resolution, reference ®xing, not to mention ellipsis
unpacking and generality narrowing. But each of these processes, which
are prerequisites to determining the proposition expressed, may them-
selves depend crucially on processes that look indistinguishable from
implicatures. Thus what is said seems both to determine and to be deter-
mined by implicature. Let us call this Grice's circle.16
It should be clear that this is not a minor point in Gricean exegesis. It is
a circle that equally a¿icts any theory that seeks to make a semantics/
pragmatics distinction play a crucial role in the general theory of mean-
ing. The ``said'' can be taken to be truth-conditional contentÐthe prop-
osition expressed, the output of the process of semantic interpretation; the
proper domain of a theory of linguistic meaning. The ``implicated'' can be
187 Implicature and the Semantics/Pragmatics Interface

taken more generally than I am taking it here, to include all the processes
of pragmatic inference; it is the proper domain of a theory of communi-
cation. Then truth-conditional content depends on most, perhaps all, of
the known species of pragmatic inference; or the theory of linguistic
meaning is dependent on, not independent of, the theory of communica-
tion. Starkly: general pragmatic inference (not just theÐsupposedlyÐ
limited phenomenon of indexicality) is crucial input to semantic interpre-
tation. Those who hope that a theory of linguistics will be a theory of a
closed mental module (aÁ la Fodor) will have to yield up semantic inter-
pretation to another module or mental process. These conclusions have
already been reached and stoically accepted by Sperber and Wilson (1986,
1995: 257±258) and their followers on the grounds of partially di¨erent
but parallel arguments.17
Grice's circle seems to require at least two applications of Gricean
pragmatics. Not only will we need the Gricean inferences to determine
truth-conditional content, but we will also need the inferential mecha-
nisms in just the way that Grice imagined, namely to give us postsemantic
pragmatics. This is because we still need an account of the double-
barreled ¯avor of ironies, indirections, and other tropesÐtwo propositions
are expressed: the so-called literal meaning and the full speaker-meaning.
And the only way to get two levels is to have the literal proposition ®xed
prior to a second batch of pragmatic processing. Figure 3.2 makes the
point graphically, with two layers of semantic processing interleaved with
two layers of pragmatic processing. (Once again, such a diagram is not to
be mistaken for a processing modelÐit just indicates which kinds of
information are prequisites to which kind of assignment of meaning. The
two boxes labeled ``pragmatics'' are here presumed to constitute one and
the same component or kind of process.18). Grice's circle is the dilemma I
wish to focus on here. Does it require a rethinking of the relations between
semantics and pragmatics, perhaps even of the whole nature of modular
interaction as currently envisaged in linguistics? Or are there ways to
resist the conclusion that semantics and pragmatics are interleaved?
Let us consider possible ways to resist the conclusion, and if that fails,
to contain the damage.

3.2.7.1 Denial of the Evidence The ®rst line of defense is the challenge:
the evidence is simply not good enough. Let us create an imaginary
(although not entirely ®ctional) character whom we may call the Obstinate
Theorist and who will recur throughout what follows (he may be identi-
188 Chapter 3

Figure 3.2
Presemantic and postsemantic pragmatics

®ed with a number of eminent theoreticians, e.g., the author of Kripke

1977; see also Bach 1994). The Obstinate Theorist (a¨ectionately OT) is
principally an old-fashioned Gricean, assuming postsemantic pragmatics,
although he is willing (we'll assume) to reanalyze pragmatic phenomena
as semantic wherever they threaten to upset the cozy picture by intruding
on the semantics. So how will he respond to our threats to the traditional
picture?
The traditional picture can be maintained in part by simply letting the
semantics provide highly general meanings (as envisaged originally in
Radical Pragmatics), later pragmatically restricted. I will call this the
189 Implicature and the Semantics/Pragmatics Interface

Generality-plus-implicatures move (covering both the ``weak univocality''

and ``nonspeci®c semantic generality'' views, as distinguished in Atlas
1989). Implicatures of course have truth conditions too; and (it may be
claimed) the intuitions about truth conditions of sentence-tokens are
simply being mixed with the intuitions about truth conditions of utter-
ance-tokens. Thus in the cases of generality narrowing and indexicality
narrowing reviewed above (sections 3.2.6 and 3.2.3, respectively), the
Obstinate Theorist can simply claim that the proposition expressed has
indeed a near-vacuous semantic content, triggering pragmatic narrowing.
In the case of disambiguation, OT can simply opt to resolve ambi-
guities postsemantically; that is, for each n-ways ambiguous string, the
semantics must provide n propositional contents, which can be selected
between by a later pragmatic process. As for ellipsis, we need to distin-
guish two kinds of cases: (a) the cases that are rule-governed (as in VP
deletion) and for which a semantic interpretation can be algorithmically
provided; and (b) the cases that are pragmatically construed. For the
latter, we need a theory of subsentential phrasal semantics that will give
us satisfaction conditions for sentence fragments. The satisfaction con-
ditions can then be the input to the pragmatic component, which will
yield a full proposition, the one elliptically expressed.
But the Obstinate Theorist has an Achilles heelÐnamely, the cases
where reference is determined by implicature (these will include some of
the I-narrowings, as in Larry's book, and some of the indexical reso-
lutions, as in some of you, as well as the cases in section 3.2.4). For if the
extensions of the parts are partially determined by implicature, the ex-
tension of the whole expression cannot be calculated without pragmatic
input. Thus in these cases, insistence on postsemantic pragmatics will yield
no propositional content at all, and thus no grist for the postsemantic
pragmatic mill: without something to work on, the pragmatics cannot
rescue the semantics (in the way that it may be held to in the cases of
semantic generality approaching semantic vacuity).
The main purpose of this chapter is to show that the Obstinate Theorist's
position is simply not tenable however attractive it may seem at ®rst sight.
Thus the bulk of my arguments will be aimed at his retort that the evidence
for Grice's circle is simply not good enough, with special attention paid
to the Obstinate Theorist's Achilles heel, the argument from reference.

3.2.7.2 Damage Limitation: Restricted Presemantic Pragmatics A sec-

ond move we can make is to admit that there is pragmatic intrusion but
190 Chapter 3

claim that the kind of pragmatic intrusion prior to semantic interpretation

is of a special limited kind, even though it may be derived by Gricean
mechanisms. Thus Sperber and Wilson (1986) and Kempson (1986) make
a distinction between explicatures and implicaturesÐexplicatures expli-
cate logical forms, which are then input to interpretation yielding propo-
sitional contents, which in turn are then input to the implicature-deriving
process. This distinction will only be of interest if the range of explicatures
is somehow of a special restricted kind, a disjoint type of inference or at
least a distinct subset of the processes called implicature. For example, it
would be an interesting (although, it will transpire, a probably false) claim
that only GCIs, rather than particularized implicatures, can play a role in
proposition determination. This move limits the damage, as it were, by
entertaining the hypothesis that I call restricted presemantic pragmatics.

3.2.7.3 Semantic Retreat The London School (Sperber and Wilson,

Kempson, Carston, and colleagues, so dubbed by Horn (1984)) couple
this claim about two distinct kinds of pragmatic inference to a logically
independent claim of Fodorean modularity: a semantics (or more gener-
ally linguistics) module is by de®nition a closed processing system. It fol-
lows that semantics stops before any pragmatic processing begins. Thus
linguistic semantics has nothing directly to do with propositions: it char-
acterizes a level of logical form or semantic representation that is algo-
rithmically derived from syntactic information. Semantic representations
are syntactical objects; their interpretation belongs to the Fodorean cen-
tral processor, where a theory of communication belongs.19 Let us call
this move semantic retreat.
Notice that semantic setreat has characterized Chomsky's thought all
along apart from a brief and uncharacteristic ¯irt with the feature-based
semantics of Katz and Fodor in the mid 1960s. The level of LF is not only
syntactical (uninterpreted) but actually just a slightly modi®ed level of
syntax. Curiously, even Katz's position is also a kind of semantic retreat
because his markerese is an uninterpreted syntactical level of representa-
tion, supposedly algorithmically derived from syntactic input. The greater
richness of Katzian Semantic Representation over Chomskyan LF, and
the correlated claim that semantic inferences can be captured by mark-
erese, is just an expression of the greater optimism on Katz's part that
pragmatic intrusion is minimal! In fact, the new evidence of pragmatic
intrusion will be su½cient to make Katz's position untenable: one simply
191 Implicature and the Semantics/Pragmatics Interface

cannot fully capture the semantic relations of synonymy, antonymy,

entailment, and so on at a preintrusive level.
The strategy of semantic retreat leaves semantic representations as
incomplete formulae without determinate interpretation. Consequently,
none of the traditional notions of semantics, like entailment relations,
sense relations, or any other kind of necessary meaning relations, hold at
this level at all. There can be no principled objection to the setting up of
such a level of representation if it solves important problems (and in fact
I assume a similar level below). But clearly, the burden is then on the
theorist to complete the picture and reconstruct the notions of traditional
semantics on another level.

3.2.7.4 Enlarged Indexicality A fourth move we can make is to treat

pragmatic intrusion of the type illustrated above as a kind of enlarged
indexicality. For example, Sag (1981) outlines a way in which metonymic
shifts of reference (e.g., from Plato to `Plato's works') may be captured
using Kaplan's character/content distinction. In essence, an expression has
an intension only relative to context, which ®xes the intension (and thus
the extension) in that context. I discuss this proposal in more detail later,
but I believe it cannot work in the cases presented here, given the defea-
sibility of implicatures.
A similar strategy would seem to be the preferred move in Situation
Semantics (Barwise and Perry 1983), although in this framework in addi-
tion to the utterance situation providing indexical parameters, there may
also be resource situations (i.e., other contextually pertinent situations)
that serve to ®x contextual variables. In contrast to the Generality-plus-
implicatures approach, this approach must assign to apparently general
predicates some kind of implicit but built-in placeholders for indexical or
contextual material. For example, just as behind in ``John is behind the
tree'' is, despite appearances, really a three-place predicate (BEHIND
(John, tree, reference-point)) so tall in ``John is tall'' is really a two-place
predicate (TALL (John, reference-value)). Because meaning is, on the
Situation Semantics view, a relation between the utterance-situation and
the described-situation, the utterance-situation (along with resource-
situations) provides the implicit reference values.
On this view, many natural language expressions are covertly quasi-
indexical or, to coin a term, are contextuals. They contain within their
semantic representation variables requiring contextual identi®cation.
192 Chapter 3

There is no doubt that such covert indexicality is more widespread than

normally recognized, as in the de®nition of local in local pub (see, e.g.,
EncË 1981). It is also clear that constructions that are often treated in terms
of semantic-generality-plus-implicatures, like the genitive in Larry's book,
may be reanalyzed as encoding an indexical variable that has to be con-
textually speci®ed (see Recanati 1993: 235¨ ). But the approach being
discussed here, best represented by Situation Semantics, is to assume that
all cases of pragmatic intrusion are in fact kinds of indexical resolution in
a broad sense. It is thus hoped that a proper treatment of semantics along
these lines will accommodate pervasive context dependency while erad-
icating the threat of more disorderly kinds of pragmatic intrusion. It is
important to see, though, that this framework has no place for other kinds
of pragmatic determination of propositional content and therefore that
the arguments developed in this chapterÐto the e¨ect that implicatures
help to ®x propositional contentÐwould appear to undermine the Situa-
tion Semantics approach to context dependency.

3.2.7.5 ``Pragmantics'' One response to pragmatic intrusion into se-

mantic interpretation is to take it as evidence against the possibility of any
principled distinction between semantics and pragmatics in the ®rst place.
Lako¨ (1972, 1987) has been a consistent advocate of the abolition of the
distinction, and others too (e.g., Baker and Hacker 1984, Jackendo¨
1990: 18) have seen the semantics/pragmatics distinction as an unfortu-
nate corollary of the tactic of treating natural languages with the mecha-
nisms developed for the analysis of formal languages.
Here, we can only make a few elementary points against this response.
First, in the perspective advocated here, pragmatic inference is funda-
mentally nonmonotonic, semantic inference is monotonic. Thus it is simply
not possible to con¯ate two fundamentally di¨erent kinds of reasoning
without doing violence to the essential nature of one or the other of them.
The ``pragmantics'' cocktail that results would lose the predictive power
that the modular account provides (entailments will end up defeasible, or
implicatures will end up nondefeasible, or both). Second, although the
abolitionists may gain succor from the evidence for pragmatic intrusion,
the intrusive phenomena discussed here do not directly support their
position: all they show is that semantics and pragmatics must be more
intimately related than had been thought. Our arguments will provide no
evidence for ``pragmantics'' as long as there are alternate models that will
193 Implicature and the Semantics/Pragmatics Interface

preserve modularity and accommodate the facts (one of which we now

turn to).

3.2.7.6 Discourse Representation Theory: A Common Slate for Semantic

and Pragmatic Contributions The ®nal position we consider here just
bites the bullet: let pragmatic information intrude into semantic repre-
sentations and consequently into their interpretation. Pragmatic intru-
sionism can be construed in di¨erent ways. Some theories of semantics
already countenance it. In many ways, this is the route taken by Dis-
course Representation Theory (Kamp 1981) and its relatives (e.g., Heim's
(1982) File Change semantics). For, despite talk of deriving discourse
semantic representations (DRSs) by algorithm from syntactic input,
DRSs in fact incorporate the results of pragmatic resolution (most obvi-
ously, anaphoric linkages). This is clearest in a series of works where it is
explicitly proposed that implicated material is entered into DRSs or File
cards (Landman 1986, Kadmon 1987, Roberts 1987). To adopt a meta-
phor, in these proposals there is a common slate, a level of propositional
representation, upon which both semantics and pragmatics can writeÐ
the contributions may be distinguished, let's suppose, by the color of the
ink: semantics in black, pragmatics in red. Semantics and pragmatics
remain modular ``pens'' as it were: they are separate devices making dis-
tinctively di¨erent contributions to a common level of representation. The
slate thus represents the semantic and pragmatic content of accumulated
utterances, and it is this representation as a whole that is assigned a
model-theoretic interpretation. We shall explore this model below (see
especially section 3.5.4), and although it raises many questions it also
appears to o¨er a simple, perfectly coherent way of conceiving of modu-
larity with intrusion.

3.2.7.7 Taking Stock In this chapter, I can only respond to some of

these moves. First, I shall try to pile on the evidence for pragmatic intru-
sion to make the denial of the evidence as di½cult as possible. Second, I
argue brie¯y against the enlarged indexicality and restricted presemantic
pragmatics positions. Third, I consider the di¨erent kinds of moves that
are open to the Pragmatic Intrusionist and o¨er a proposal about how to
think about the relation of between semantic representations, their prag-
matic enrichments, and truth conditions (which incidentally is a con-
structive alternative to semantic retreat).
194 Chapter 3

Let us take stock. It appears from the evidence (and more will be pre-
sented shortly) that semantic representations (i.e., semantic structures
algorithmically decodable from syntactic structures) are indeed sub-
propositional: it takes pragmatic resolution of the full Gricean sort to ®x
the proposition expressed. The proposition expressed is the entity that
participates in the normal semantic relations of entailment, sense relations
with other expressions, and so on. But that proposition may then receive
further pragmatic enrichments of the standard kind, including additional
implicatures. This leaves us with terminological di½culties: what is seman-
tics, what is pragmatics, what is said, what is implicated, and so on?
Although recognition of the problem is relatively new, the di½culty has
already been compounded by rival recommendations for terminology.
Some want to retain Grice's label ``what is said'' with a narrow applica-
tion, so that it corresponds to a semantic representation, disambiguated
with indexicals resolved, but not necessarily yet expressing any determi-
nate proposition (Bach 1994: 144). Others want to expand the notion to
cover something much broader, closer to the everyday concept of ``what is
stated''; on this account, any kind of pragmatic resolution or implicatural
strengthening up to a level corresponding to a conscious, commonsensical
notion of what has been claimed counts as ``what is said'' (Recanati
1989). On this account, GCIs belong entirely to the domain of ``what is
said'' (for response, see Horn 1992a,b). Some have tried to isolate out a
middle ground between what is given by the semantic representation,
perhaps after disambiguation and indexical resolution on the one hand,
and full-blown Gricean implicatures on the other hand. Thus Sperber and
Wilson (1986) call explicatures the informational enrichments of semantic
representations necessary to achieve not only minimal propositions but
also ones of su½cient informational content to count as conversational
contributions (thus ``Fixing your watch will take some time'' expresses a
proposition, but one too vacuous as it stands to count as the proposition
expressed). On this account, GCIs are explicatures and lie outside the
scope of a theory of implicature (which was I think the intention).
Carston (1988) has tried to clarify this view, suggesting that the notion
``implicature'' should be reserved for propositions that are not in any way
enrichments of what is said but rather ``functionally unrelated'' (see
Recanati 1989 for critique). Bach (1994) o¨ers the new term ``impliciture''
(with an i) for the middle ground but he views this in a di¨erent way,
covering only completions and expansions of the semantic content that
are in line with the structure of the sentence (and, as mentioned, prag-
195 Implicature and the Semantics/Pragmatics Interface

Table 3.1
Terminologies in the domain between ``what is said'' and ``what is implicated''

Deictic &
Semantic reference Minimal Enriched Additional
Author Representation resolution proposition proposition propositions
Grice 1989 ``What is said'' ``Implicature''
Sperber & ``Semantics'' ``Explicature'' ``Implicature''
Wilson 1986
Carston 1988 ``Semantics'' ``Explicature'' ``Implicature''
``What is said''
Recanati 1989 ``What is said''
``Sentence ``Explicature''
meaning''
Levinson ``What is said''
1988b
``The coded'' ``Implicature''
Bach 1994 ``What is said'' ``Impliciture'' ``Implicature''

matic resolutions of indexicals and ambiguities count for him not as

implicitures but as what is said). Table 3.1 attempts to display in a crude
fashion the mismatching alignments of terms and concepts (it should be
interpreted as showing the maximal extent of notions like ``implicature''
and ``explicature'' according to di¨erent authors).
Let us discuss the explicature/implicature distinction in a bit more de-
tail. Both explicatures and implicatures are pragmatic inferences, derived
by the same mechanisms, with the same context dependence and defeasi-
bility (Carston 1988: 158). The idea is that there is a sharp distinction
between the ¯eshed-out logical form attributed to an utterance (its expli-
cature), and fully implicit, unrelated but pragmatically conveyed propo-
sitions (the implicatures; see Sperber and Wilson 1986: 181±182). Carston
(1988: 173) hopes that explicatures will correspond with the set of prag-
matic inferences that fall within the scope of logical operators or intrude
into the truth conditions of ``what is said.'' The problem is then to identify
by independent criteria what an explicature is (see Recanati 1989, 1993:
chap. 13±14 for helpful discussion). The following criteria have been
suggested: (a) the representations of explicatures contain the semantic
196 Chapter 3

representation attributed to the utterance as a proper subpart (Sperber

and Wilson 1986: 181); (b) explicatures express unique and complete
propositions (Sperber and Wilson 1986: 179; Carston 1988); (c) explica-
tures express the minimal complete propositionÐthat is, they are the
minimal enrichment of a semantic representation su½cient to make it
truth-evaluable (rejected by Sperber and Wilson 1986: 189; Carston 1988;
Recanati 1993); (d) explicatures have a ``functional independence'' from
implicatures and must support distinct inferences (Carston 1988: 158); (e)
explicatures fall within the scope of logical operators (Recanati 1989:
321±325); and (f ) explicatures correspond to a pretheroretical intuition
about what a speaker said (Recanati 1993: 246±250).
These criteria in fact fail to make any clear distinction between expli-
cature and implicature. Criterion (a) is partly just a matter of how the
analyst phrases the pragmatic inference: after all, any implicature can be
added as a conjunct to ``what is said.'' Take the utterance John's three
children came to the partyÐthe corresponding scalar inference can be
phrased as `The totality of John's children of cardinality 3 came to the
party' or as a separate proposition `John has no more than three children',
and so on. Criterion (b) is taken as a necessary but not su½cient criterion,
and in particular the rather promising criterion (c) is rejected, so that an
explicature is not just the minimal expansion of an utterance until it
reaches propositional explicitness but rather an expansion until it reaches
su½cient informational speci®city to play a role in further pragmatic in-
ference. (Criterion (c) would have left the bulk of GCIs as implicatures,
whereas Recanati (1993: 249) and Carston (1995: 237) presume that these
are essentially explicatures.) This leaves any distinction between explica-
tures and implicatures completely unresolved. Hence the suggestion of
criterion (d), functional independence, which in e¨ect requires that impli-
catures do not entail explicatures, or what is implicated does not entail
what is said. But as we saw concerning criterion (a), this is essentially a
matter of how the analyst phrases the implicature (see also Recanati 1989:
317 n 11).20 Moreover, Recanati (1989: 318±320) points out that this
criterion cuts across the proposed distinction, because there are clear cases
where an implicature entails what is said. There are also cases where what
is explicated fails to entail the sentence meaning (as in ``I haven't eaten
breakfast'' with the explicature `I haven't eaten breakfast today'; see Bach
1994: 134). Recanati therefore proposes to shift the burden to two further
criteria (e) and (f ). The ®rst states that implicatures cannot fall within the
scope of negation or other logical operatorsÐany such pragmatic infer-
197 Implicature and the Semantics/Pragmatics Interface

ences that do so are really explicatures. But Recanati (1993: 269±274)

came to recognize that there are substantial di½culties with such a scope
testÐas we shall see in section 3.3.3, the very same scalar inference may
or may not fall within the scope of negation (cf. ``John doesn't have three
children, he has two'' vs. ``John doesn't have three children, he has four''),
and many other aspects of an utterance may be ``metalinguistically''
negated. He therefore places the burden on the pretheoretical intuition
about what has been said (Recanati 1993: 246±250), claiming that this is
always consciously accessible. A consequence, he notes, is that GCIs will
normally be explicatures, part of what is said (yet, as Horn (1992b: 176±
178) notes, the history of linguistic thought tends to demonstrate the
contrary intuition). This is shaky ground indeed: as Bach (1994: 137±139)
points out, judgments are likely to depend on how one elicits the intu-
itions. Moreover, in noncooperative circumstances like cross-examination
in a court of law, what would normally taken to be part of what the
speaker is committed toÐbut has not literally saidÐis routinely queried,
indicating that our pretheoretical notions of what has been said are de-
pendent on the mode of talk (Levinson 1983: 121). Thus at the moment
we have no reliable distinguishing criteria for the explicature/implicature
distinction even for terminological purposes, and there is certainly no in-
dication that any such distinction would substantially help us understand
the nature of the inferences in question.
Bach (1994) o¨ers us a di¨erent way to cut the pie: we should recognize
a middle ground between ``what is said'' (including indexical resolution
and reference ®xing) on the one hand, and what is implicated on the
other. In the middle stands ``impliciture'', what is implicit in what has
been said. What is implicit involves both ``completion'' (getting us from
``what is said'' or proposition-radical to a minimal proposition) and ``ex-
pansion'' (which gets us from the minimal proposition to what is implic-
itly meant). An example of a completion would be the inference from
``This suitcase is too heavy'' to `This suitcase is too heavy for cabin bag-
gage', and an example of an expansion would be the inference from ``I
have nothing to wear'' to `I have nothing appropriate to wear to the
wedding'. GCIs will, at least in many cases, be implicitures (with an i) on
this account (see Bach 1994: 135).21 Although the distinctions seem
clearer and perhaps more helpful than the Relevance theory distinctions,
Bach (1994: 140) equally fails to give us a clear boundary between
impliciture and implicature, saying only that ``an implicatum is com-
pletely separate from what is said and is inferred from it.'' But as we have
198 Chapter 3

seen, this is in large part a matter of how the analyst phrases the infer-
ences.22 Bach is not proposing any special kind of inference exclusive to
implicitures, and thus the issues are essentially terminological.
My own view is that all such terminological e¨orts are fruitless: Grice's
circle ensures that there is no consistent way of cutting up the semiotic
pie such that ``what is said'' excludes ``what is implicated.'' None of the
authors here consider that the processes involved in deriving explicatures
or implicitures and so on are essentially di¨erent in kind from those used
in deriving implicatures.23 Merely relabeling pragmatic inferences that
play a role in ®xing the intended proposition something other than
``implicature'' hardly alters the nature of the problem. Rather, I shall
presume that we want to de®ne the types of content by the processes that
yield them and the important semantical properties they have (e.g.,
default presumption, defeasability under distinct conditions).24 The
search for ``what is explicated'' or ``what is implicited'' betrays a failure
to grasp the dimensions of Grice's circle; nevertheless I will return to the
issue of the middle ground between the ``said'' and the ``implicated'' at
the end of the chapter.

3.3 INTRUSIVE CONSTRUCTIONS

We have established a prima facie case for the existence of Grice's circle,
the idea that implicatures may play a role in ®xing what is said. But we
will need more evidence before we will be willing to reorganize our whole
way of thinking about semantic theory. What we seek is some incon-
trovertible evidence that implicatures contribute essentially to truth con-
ditions. One method would be to search for evidence that, in some cases,
assignment of truth conditions without prior assessment of implicatures
will give us the wrong truth conditions. In particular, I shall attempt here
to ®nd a set of constructions, dubbed intrusive constructions, that have the
following property: the truth conditions of the whole depend in part on
the implicatures of the parts (for an early collection of such observations,
see Gazdar 1979: 164±168). Cohen (1971) was the ®rst to pursue this
strategy; however his goal was the reverse of oursÐhe sought to establish
that Grice's arguments about the implicatural overlays on the logical
connectives had to be wrong by showing that the consequence of accept-
ing them was that pragmatics was intrusive. If we no longer take this as a
reductio ad absurdum, the arguments have a new interest.
199 Implicature and the Semantics/Pragmatics Interface

3.3.1 Comparatives
Wilson (1975: 151) provides the following kind of example:
(17) ``Driving home and drinking three beers is better than drinking
three beers and driving home.''
I ‡‡> `Driving home and then drinking three beers is better than
drinking three beers and then driving home.'
Sentences of the form ``A is better than B'' will be anomalous, indeed
contradictory (necessarily false), unless the proposition expressed by A is
distinct from that expressed by B. But if A has the form `p and q' and B
has the form `q and p', and and is equivalent to the truth-functional con-
nective & (and thus p & q 1 q & p), then ``A is better than B'' will be
necessarily false. But example (17) has this form; and such an assertion
seems, let us agree, just plain true!
What makes ``p and q is better than q and p'' noncontradictory is the
inference from and to `and then'; ``p and then q is better than q and then
p'' is a perfectly sensible proposition capable of being true. But if the
inference to `and then' is pragmatic (and the Gricean arguments seem
perfectly sound), then our assessment of such comparatives involves
pragmatic intrusion: the truth conditions of the whole construction are
calculated taking into account the implicatures of the parts. The inference
from and to `and then' is on Grice's scheme a Manner implicature, and on
my scheme an I-implicature.
But perhaps Cohen (1971) is right that this is just evidence for the
semantic nature of temporal sequentiality. For example, given a sequence
of tensed clauses like John came in and sat down, Grice's position with
respect to and could be maintained, while the truth-conditional nature of
the `and then' inference could be attributed to the implicit indexical
sequence of reference times in the two verbs.25 This account is more dif-
®cult to maintain with untensed gerunds or other in®nite verb forms as in
example (17), but nevertheless it raises the possibility that some reanalysis
along semantic lines (some semantic escape hatch, as it were) might be
found for these cases of apparent pragmatic intrusion.
So the critical question is: how general is the phenomenon? In particu-
lar, can other implicatures (perhaps less amenable to semantic reanalysis)
intrude in the same way? It is at this point that our typology of implica-
tures (GCIs) comes back into play; what we can attempt to do is run the
di¨erent types of GCI through each type of intrusive construction to see if
there is a general tendency towards intrusion in these constructions.
200 Chapter 3

The evidence will show, I believe, that there is such a tendency. To

reduce the need for discussion, I will build into the examples some mini-
mal context suggestive of the intended interpretation. Let us turn ®rst to
Q (scalar) implicatures within comparative constructions. Consider the
following, which on a purely semantic basis should be self-contradictory;
if they are not felt to be self-contradictory and seem rather to express a
plausible proposition, then that can only be the result of a strengthening
of the underlined expressions by a Q implicature.
(18) a. ``AIDS vaccination will cause signi®cant mortality: but losing
some of the population is better than losing at least some and
perhaps all of it.''
a 0 . ``AIDS vaccination will cause signi®cant mortality: but losing
some of the population is better than losing at least some of it.''
b. ``You shouldn't continue to gamble; I know you've already lost
two hundred dollars, but losing two hundred dollars is better than
losing at least two hundred dollars and possibly more.''
c. ``You should get out of stocks: I know you've already lost a lot,
but a large loss is better than a large and possibly enormous
one.''
d. ``It's best to know where you stand: to be charged with reckless
driving or manslaughter is better than being charged with one or
the other and possibly both.''
Some brief notes: in (18a) the semantic content of some (roughly, `at least
one' or `some, not excluding all') is perfectly compatible with all; thus the
two expressions on each side of the comparative (some, at least some and
perhaps all ) are approximately synonymous or extensionally equivalent.26
What makes the sentence noncontradictory is simply the scalar implica-
ture from some to `not all' (an implicature explicitly canceled in the phrase
some and maybe all ).27 Thus taking the sentence to be of the form `A is
better than B', A is distinct from B only in having that additional impli-
cature. Similarly for (18b): two hundred has the semantic content `at least
two hundred' (two hundred and possibly more), but Q-implicates `no
more than two hundred'; only on the assumption of that implicature is
(18b) other than contradictory. In the case of (18c), we have a scale
henormous, largei such that use of large implicates `not enormous' but is
semantically compatible with enormous; again only if the implicature is
used to strengthen the ®rst clause is the sentence as a whole capable of
being true. In (18d), there is a scale hand, ori such that use of p or q
201 Implicature and the Semantics/Pragmatics Interface

implicates `not both p and q' (for discussion see Gazdar 1979; Levinson
1983: 133¨ ); again, only if the implicature goes through will the content
of A and B be su½ciently distinct to avoid a contradiction.
Intrusive Manner implicatures are somewhat harder to ®nd, but never-
theless the following examples seem to be reasonable candidates:
(19) a. ``It's better to eat your meat than to ingest it.''
b. ``It's better to be a long way from one's friends than a long long
way.''
c. ``You need to train harder: running and running and running is
better than running.''
d. ``Legible handwriting is better than not illegible handwriting.''
(or: ``You are more likely to get the job if your handwriting is
legible than if it's not illegible.'')
e. ``Linguists should use clearer examples: citing a grammatical
example is better than citing a not ungrammatical one.''
Possible cases of I-intrusion (other than those based on and ) in com-
paratives are the following:
(20) a. ``Ivan and Penny driving to Chicago is better than Ivan driving to
Chicago and Penny driving to Chicago.''
b. ``If Steve and Jay write a paper there's a better chance of truth
than if Steve writes a paper and Jay writes a paper.''
c. ``Hammering the nail into the wood is safer than hammering the
nail some way into the wood.''
There's a prima facie case, then, that pragmatic intrusion in comparatives
is not limited to cases (like those based on the implicatures of and ) where
an analysis in semantic terms is an alternative.
However, readers may ®nd some of the examples given in (18)±(20)
somewhat strained. This is a point then to reconsider the structure of the
argument and see whether it can be made more general. The case we are
making is based on the assumption that to establish pragmatic intrusion
in the construction ``A is better than B'' we must ®nd two clauses A and B
which are logically equivalent but pragmatically distinct, such that sub-
stitution in the construction yields an intuitive truth where a necessary
falsehood would be predicted on purely semantic grounds. The reasoning
is based on the assumption that, whatever the semantics of the compara-
tive construction are exactly, the semantics must at least meet the follow-
ing condition: the comparative relation f-er is necessarily irre¯exive, thus
202 Chapter 3

in A is f-er than B, A and B must be distinct. Now, our examples are open
to objection because it is a little hard to come up with cases where A and
B are semantically equivalent but pragmatically distinct, and yet will still
®t happily into the construction. But given the irre¯exivity of the con-
struction, I think weaker requirements can be met and still be su½cient to
establish pragmatic intrusion. As far as I can see, the semantics of the
comparative A is f-er than B requires that not only must the two clauses
be semantically distinct, but also that the two clauses must not be related
in such a way that B encompasses what A denotes (e.g., A and B should
not be privative opposites). More exactly, if B entails A, and B is ``about''
the same semantic relations as A (e.g., B is not a conjunction of the form
`A and C'), then a comparative of the form ``A is f-er than B'' will be
necessarily false, unless rescued by pragmatic intrusion. For example,
``Having a child is better than having a son'' seems false or nonsensical
(unless we can somehow pragmatically intepret child as excluding `male
child'). So if we can ®nd examples of this kind where in fact the utterance
is intuitively true or potentially true, we should be able to ®nd that the
acceptability is due to the implicatural strengthening of A so that it is no
longer entailed by B. For example, if I assert ``Eating some of the cake is
better (for my health) than eating all of it'', what I say is false unless some
is construed as `some and not all' (i.e., is strengthened by the scalar
implicature).28 Without the pragmatic strengthening, I have asserted that
one state of a¨airs (eating at least some) is better than another state of
a¨airs (eating all) that guarantees the very state of a¨airs I claim is pref-
erable. Moreover, the state of a¨airs I claim to be preferable is completely
compatible, in no way precludes, and may indeed be embedded within the
state of a¨airs I claim to be less preferable.
Thus if example (21a) has the semantic content in (21b), we can see that
(21a) should (on semantic grounds alone) be nonsensical: any situation in
which one has four children will also be one in which one has at least
three children, so how could one prefer a situation (having at least three
children) to another which guarantees that same situation? It is a bit like
saying ``It is better to be a woman than to be a queen'' (which if it makes
sense at all, requires forcing an unwomanly interpretation of queen). We
can then go on to make parallel arguments for (21c) and (21d) and indeed
for cases parallel to all those above.
(21) a. ``Having three children is better than having four children.''
b. `Having at least three children is better than having at least four
children.'
203 Implicature and the Semantics/Pragmatics Interface

c. ``Eating some of the cookies is better than eating all of them.''

d. ``Drinking warm co¨ee is better than drinking hot co¨ee.''
Clearly, rather a lot depends here on the correct analysis of comparatives,
on which we will momentarily digress. Existing accounts (to the best of
my knowledge) only deal with comparison of individuals, not proposi-
tions as in these cases. In the existing accounts, a is taller than b is ana-
lyzed variously as follows: a sentence of this form asserts that there is
some comparison class relative to which a is tall and b is not tall (Klein
1980); or the sentence asserts that there is some measure m of tallness such
that a exhibits m and b does not (Seuren 1974); or there is some measure
of tallness m exhibited by individual a that is greater than all the measures
of tallness exhibited by b (Atlas 1984b: 362¨; n. 11). By analogy, A is
better than B might be analyzed as asserting that there is some situation in
which A is good and B is not good; or (on Atlas's account) that there is
some degree of goodness that A has that is greater than all degrees of
goodness that B has.29 On that analogy, a comparative ``A is f-er than
B'' where there is an entailment relation between A and B (speci®cally,
where B ` A and the representation of A explicitly encompasses the pos-
sibility of B) should self-destruct (be incapable of being true), because
any situation in which B holds will contain the situation described by A.
Thus the following sentences are predicted correctly to be semantically
anomalous:30
(22) a. ?``Being a bachelor is better than being an unmarried person.''
b. ?``Having a vehicle is better than having a car.''
c. ?``Having a child is better than having a male child.''
If this is correct, then all the examples in (21a±d), for example, which do
not seem anomalous, must be rescued by the pragmatic intrusion of
Q-implicatures.
Perhaps the issue can be clari®ed by reverting to cases where individ-
uals are being compared. Suppose we assert a sentence of the form `a is
better than b' in conditions where as a matter of linguistic fact any b-
individual must also be an a-individual, then on any reasonable account
of comparatives we should, I think, get an anomaly (cf. ?``A woman is
better than a queen'', ?``A vehicle is better than a car'', ?``A child is better
than a boy'', etc.).31 It follows that examples like those below ought to be
anomalous. For example, the literal content of (23a) amounts to (23a 0 )
but (23a 0 ) is indeed clearly anomalous, because any woman with at least
four children will also be a woman with at least three children. However,
204 Chapter 3

contrary to the semantic prediction, sentence (23a) is perfectly felicitous

and capable of being trueÐbut it can only express a coherent proposition
if three is strengthened to `exactly 3'. Indeed, (switching to a clearly
extensional predicate like smaller) I take (23b) to be unequivocably true,
even though the semantics alone will render it false (since the family with
three children is only semantically speci®ed as having at least three chil-
dren!). Lest this be thought to be a property of the numerals (and thus
reliant on a possibly mistaken scalar analysis there), note that the same
argument will go through for the other cases:32
(23) a. ``Any mother with three children is happier than any mother
with four.''
a 0 . Any mother with at least 3 children is happier than any mother
with at least 4 children.
b. ``A nuclear family with three children is smaller than a nuclear
family with ®ve children.''
c. ``A student who cheats on some exams is better than one who
cheats on all.''
d. ``A teacher who is sometimes late is preferable to one who is
always late.''
There are still some semantic points that need at least some minimal
clari®cation. For example, I take it that these sentences are, disregarding
implicatural enrichments, clearly literally false (assuming the `at least'
semantic analysis of the relevant scalar predicates) on the assumption
that there is at least one entity in the domain of discourse that satis®es the
stronger conditions (e.g., in (23a) that there is a mother with four chil-
dren); that will ensure that these implicitly universally quanti®ed inde®-
nite NPs range over some individual b such that one is asserting `b is f-er
than b', which will be necessarily false. It is a separate question (one about
which I am at present unclear) whether a sentence of the form `a is f-er
than b' where every b is an a actually entails `a is f-er than a' and thus
entails a necessary falsehood.33 But we cannot pursue these matters here.
The point of this digression is to try to establish that the argument for
pragmatic intrusion into comparatives can be made on a much wider
basis than on examples of the form ``A is better than B'' where A is
semantically equivalent to B. However, regardless of the success of my
foray into the semantics of comparatives, and thus my extended list of
examples, there is no doubt that examples of the more restricted kind can
be found, and that these appear to constitute knock-down evidence for
205 Implicature and the Semantics/Pragmatics Interface

the role of pragmatic strengthening in semantic interpretation. It is a truth

that the reader may appreciate that reading a book is easier than reading
a tome.

3.3.2 The Conditional

The conditional provides some of the strongest apparent evidence for
pragmatic intrusion. It is generally conceded that there is still no adequate
semantic analysis of the natural language conditional, but that existing
accounts agree on the following conditions on such an analysis:
If A then B is true if and only if ``every one of a number of ways in
which A can be true constitutes, or carries with it, a way of B's being
true'' (Kamp 1981: 11). On Kamp's speci®c version of this analysis, if A
then B will be true i¨ each way of verifying A carries with it a veri®cation
of B (i.e., each embedding of the discourse representation of A in a model
M can be extended to an embedding of the discourse representation of B
in M as well). Thus we only assent to the truth of the whole conditional if
every way in which A can be true guarantees the truth of B.
With this semantic analysis in mind, we can now turn to consider cases
where the semantics will predict a falsity, but our intuitions nevertheless
declare a plausible truth. In these examples, the conditional as a whole is
assessed on the assumption that the implicatures (as well as the entail-
ments) of the antecedent are satis®ed by the model: thus in these cases
it seems that if A then B is intuitively true i¨ every way of verifying
the entailments and the implicatures of A carries with it a veri®cation of
B.34 The clearest examples are once again provided by Q-implicature
intrusions:
(24) a. ``If each side in the soccer game got three goals, then the game
was a draw.''
a 0 . Gloss of the semantic content: If each side in the soccer game got
at least three goals, then the game was a draw.
b. ``If you ate some of the cookies and no one else ate any, then
there must still be some left.''
(or ``If the USA won some of the Olympic medals, other
countries must have got the rest.'')
c. ``If the chair sometimes comes to department meetings that is not
enough; he should come always.''
d. ``If you earn forty thousand dollars and have no capital, you
can't buy a house in Palo Alto.''
206 Chapter 3

e. ``If John owns three cars, then the fourth outside his house must
belong to someone else.''
f. ``If you think John has three children, you're wrong; he has
four.''
If you feel inclined to assent to any of these, then you are interpreting
them in such a way as to strengthen the antecedent by a Q-implicature
(e.g., in (24a) three must be interpreted as `3 and no more', in (24b) some
must be interpreted as `some and not all', in (24c) sometimes as `some-
times and not always', and in (24d) forty thousand as `forty thousand and
no more'). Otherwise, there are no grounds for acceding to the con-
sequent. To see this, let us consider just (24a) in detail: if three has the
semantic content `at least 3', then (24a) has the literal content glossed in
(24a 0 ). But given the rules of soccer, (24a 0 ) is clearly (contingently) false:
there are ways of verifying the antecedent (e.g., one team got exactly three
goals and the other got exactly four goals) that do not guarantee the truth
of the consequent. If however the antecedent is strengthened by the scalar
implicature so that the antecedent reads `each side got 3 and at most 3
goals', then (given the rules of soccer) the whole will be predicted to be
true. Because our intuitions are quite clear that the sentence is true, it
seems evident that the truth conditions of the whole depend on taking into
account the implicatures of the antecedent.
We can also ®nd I-implicature intrusions: thus the following are only
reasonably interpreted as truthful claims if and is strengthened to `and
then'.
(25) a. ``If you have a baby and get married, then the baby is strictly
speaking illegitimate.''
b. ``If he turned on the override switch and the reactor overheated,
he's responsible for the disaster.''
c. ``If they declared a Republic and the old king died of a heart
attack, then they're responsible for his death; but if the old king
died of a heart attack and they declared a Republic, they cannot
be blamed for the old king's death.'' (after Cohen 1971)
d. ``If you reconnect the battery and turn the ignition key, the car
will probably start.''
To see that and is not a special case, consider other I-inferences:
(26) a. ``If Bill and and Penny drive to Chicago, they can discuss
sociolinguistics in the car for hours.''
207 Implicature and the Semantics/Pragmatics Interface

a 0 . ``If Mandy and Rick drive to New York, they'll end up with two
cars and no parking in Manhattan.''
b. ``If you hammer the nail into the door, you can hang things on
it.''
b 0 . ``If you hammer the nail into the door, no one will scratch
themselves on it.''
c. ``If Bill's book is good, he will get tenure.''
c 0 . ``If Bill's book is good, he'll never return it to the library.''
In each of these cases, it has been argued, there is a strengthening impli-
cature attributable in our scheme to the I-principle. Thus a predication of
two agents usually carries a `together' implicature (Harnish 1976, Atlas
and Levinson 1981) as in (26a); the GCI can be overridden as illustrated
in (26a 0 ). In (26b) the spatial relation in merely speci®es at least partial
inclusion; the interpretation of partial versus full inclusion is a matter of
stereotypical knowledge and inference, as sketched in (26b) versus (26b 0 )
(see, e.g., Herskovits 1986). In (26c), the relation indicated by the posses-
sive is semantically general; the semantics of Bill's book presumably only
speci®es that there is some (more or less intimate) relation between Bill
and a de®nite bookÐthe relation might be a relation of writing, reading,
owning, borrowing, editing, and so forth as given by an I-inference in the
context (see Kay and Zimmer 1976, Sperber and Wilson 1986). Without
having previously talked about Bill's book, we can indicate the relation
we have in mind, as in (26c) versus (26c 0 ). But (26c) will only be true, let
us suppose, if the book Bill wrote is good. (Clearly, the inference here
involves a matter of reference determination; see section 3.4.)
Finally, M-implicatures may also perhaps intrude in conditionals. The
following may be cases, but I concede that they are less than wholly clear:
(27) a. ``If it's a long way to your place, we can make it; but if it's a
long, long way, we may not be able to.''
b. ``If he hit and hit and hit the victim then he's more culpable than
if he hit him.''
b 0 . ``If the surgeon tried and tried and tried to save the patient, then
he did more than try.''
c. ``If the President is not ignorant of the facts, he doesn't really
know them either.''
d. ``If we're not unwelcome, then we are hardly welcome.''
e. ``If he caused Marlene's death, he'll get a manslaughter charge; if
he killed her, he'll get second-degree murder.''
208 Chapter 3

Thus in (27a), if it is conceded that the reduplication of long is not

semantic modi®cation, then long, long merely suggests (M-implicates) a
greater distance than long; in which case (27a) will be contradictory unless
the M-implicature is used to strengthen the antecedent of the second
conditional. Similarly, the repetition in (27b) arguably adds no semantic
content, in which case there will be ways of verifying the antecedent that
do not verify the consequentÐso the conditional should be false, even
though it is arguably true. In this case, as noted above, if past tenses are
not treated as quanti®cations over past times but rather as enforcing a
sequence of reference times (as in Partee 1984), then perhaps there is a
semantic account of the possible truth of (27b). Where repetition suggests
intensity, as in (27b 0 ) not repeated action as in (27b), then the case for
semantic intrusion would be better, but perhaps (27b 0 ) is not too clear. In
(27c) the double negative not ignorant logically implies knows (because
ignorance and knowledge are contradictories) but suggests (M-implicates)
something less than full cognizance. Thus what ought to be a contra-
diction comes out as an intuitive truth; we can account for this if the
M-implicature is taken into account in the assessment of the whole con-
ditional. Similarly for (27d). In (27e) the periphrastic causative cause
x's death suggests indirect or unintentional causation by reference to the
complementary I-implicatures from the use of the less-prolix lexical
causative kill, which suggests direct or intentional causation. Assuming
that manslaughter and murder are inconsistent charges, the two condi-
tionals in (27e) should be contradictory with respect to one another, be-
cause the two antecedents with identical semantic conditions are paired
with inconsistent consequents. But intuitively, that is not the caseÐ(27e)
seems like a perfectly consistent and sensible assertion. Again, the problem
evaporates if the M-implicature (of indirect causation) of the ®rst ante-
cedent and I-implicature (of direct causation) of the second antecedent
play a role in the truth-conditional assessment of each conditional.
Unfortunately, as noted above, a good semantic analysis of condi-
tionals still eludes us. So it may be possible to claim that somehow the
semantics of conditionals makes special allowance for pragmatic factors.
For example, Barwise (1986) argues (following Goodman) that condi-
tionals have a built-in ceteris paribus condition, such that if you say, for
example, ``If I strike this match it'll light'' we want to maintain that the
utterance is true even though various ceteris paribus conditions (e.g., the
matches are not damp) have not been spelled out. So it may be said that
our intrusive implicatures are merely part of the ceteris paribus con-
209 Implicature and the Semantics/Pragmatics Interface

ditions: ``If each side has three goals, then it is a draw'' is true relative to
the condition `each side has at most three'. I do not think Barwise's ac-
count could as it stands handle this kind of intrusion, because such intru-
sions are systematic inferences that are conveyed by the utterance itself,
not independently existing background assumptions (not, for example,
antecedent ``resource situations''). Nevertheless, we should attend to the
possibility that some better account of the semantics of the conditional
might eliminate these apparent pragmatic intrusions.
To deal with this objection it is worth making a few further observa-
tions. First, some of the same intrusions take place in disjunctions where
there are no ceteris paribus conditions on any analysis ((28a) from Wilson
1975: 151):
(28) a. ``She either got married and had a child, or had a child and got
married; but I don't know which.''
b. ``According to the prosecution, the victim was either shot and
thrown in the lake or thrown in the lake and shot, but not both
(i.e., not shot, thrown in the lake, and shot again).''
c. ``If you believe his weak alibi, he cannot have been responsible
for all the crimes. But one thing is certain: he either committed
some of the crimes or he committed some and possibly all of
them. I don't know which.''
d. ``Marcos either owned three Swiss bank accounts or at least three
Swiss bank accounts; only on the ®rst assumption can we be sure
that all the embezzled money has been recovered.''
Some brief notes: in (28a) the speaker's assertion of ignorance over two
possibilities will be inconsistent or anomalous with the disjunction if each
disjunct has the same semantic content. Plausibly the inconsistency is
pragmatic rather than semantic, because it amounts to Moore's paradox:
the speaker of (28a) is asserting ``p or p and I don't know that p.''35 But
the point is that we have no intuitions of any such oddity. The only
way to account for the lack of such intuitions is to suppose that the I-
implicature from and to `and then' has strengthened the disjuncts. The
second example is similar except that here a contradiction rather than a
pragmatic paradox would result if the semantics were not strengthened
by the same pragmatic intrusion from and. In (28c), if some has semantic
content `at least some' and only Q-implicates `not all', then some and possi-
bly all is an approximate gloss of the semantic content. Again, we can
only account for the intuitive absence of the pragmatic paradox if the
210 Chapter 3

Q-implicature plays a role in the truth conditions. In (28d) we have an

assertion of the form ``p or q; only if p and not q, r'' where p is semanti-
cally equivalent to q; again only a Q-implicature from three to `no more
than three' will rescue the speaker from an absurdity that is not felt in fact
to arise.
A second argument against the possibility of accounting for pragmatic
intrusion in conditionals by a Barwise-type semantics is that the kind of
pragmatic intrusion in conditionals is not limited to informative GCIs
that might conceivably play a role in ceteris paribus conditions. Tropes of
various sorts can also intrude:
(29) a. ``If you appoint a little Chomsky, all the sociolinguists will
resign.''
b. ``If the smog is so adorable, that explains why the mansions in
Beverly Hills are deserted in the summer.''
c. [Sun shining] ``If it continues to rain like this, I'll come to
England more often.''
Finally, most of the intrusions that can be found in conditionals can also
be found in because-clauses, which are (I assume) semantically remote
from the antecedents of conditionals since because-clauses are entailed.36
(30) a. ``Because he drank three beers and drove home, he went to jail.''
b. ``Because he turned on the override switch and the reactor melted,
he's responsible for a thousand deaths.''
c. ``Because he earns forty thousand dollars, he can't a¨ord a house
in Palo Alto.''
d. ``Because he has one child, Huang is permitted to have an
academic job in Beijing.''
e. ``Because the police recovered some of the missing gold, they will
later recover it all.''
f. ``Because he's such a ®ne friend, I've struck him o¨ my list.''37
g. ``Because they appointed another little Chomsky, the
sociolinguists all resigned.''

3.3.3 Metalinguistic Negation and Other Negatives

It has been noted for some time (see Wilson 1975: 149¨; Horn 1985;
Kempson 1986) that there are uses of negation where what is negated
is only an implicature. Here are some cases where what is negated is a
Q-implicature:
211 Implicature and the Semantics/Pragmatics Interface

(31) a. ``John doesn't have three children, he has four.''

b. ``I'm not happy, I'm ecstatic.''
c. ``He didn't lose a ®nger, he lost an arm.''
d. ``It's not possible, it's certain.''
e. ``It's not true that either the President or the Vice-President must
sign the treaty; they both must.''
f. ``He's not good-looking, he's downright handsome.''
g. ``I didn't do it once or twice, I did it once only.''
h. ``Some men aren't chauvinists, all men are.''
The analysis of these cases is controversial. Kempson (1986; see also
Kempson and Cormack 1981, 1982) claims they provide evidence that
truth conditions must be stated over explicated logical formsÐthat is,
that these cases prove that some limited pragmatic intrusion (of explica-
tures) takes place. Thus Kempson's claim is that a sentence like (32a) can
be thought of as negating two positive contentsÐthe implicated and the
entailed, as detailed in (32b). Thus negation ranges over both the impli-
cated and the entailed as in (32c), with the consequence that (32d) (where
the positive implicature is false and the positive content true) is one set of
conditions under which (32a) will be true.
(32) a. ``John doesn't have three children, he has four.''
b. Positive content p: `John has at least 3 children.'
Positive implicature q: `John has at most 3 children.'
c. Content of whole: @… p & q† 1 @p4@q
d. Verifying circumstance: p & @q
Instantiating: `John has at least three children but not at most
three children.'
In contrast to Kempson's claim, Horn (1985, 1989: chap. 6) argues that
these are not true logical negations; what is involved is a metalinguistic
operator, a denying of the appropriateness of an utterance. Thus (32a) is
shorthand, as it were, for ``You can't appropriately describe John as
having three children; you should describe him as having four.'' He pro-
vides persuasive evidence for the existence of a metalinguistic negationÐ
for example, no logical analysis is possible of utterances like ``It's not a
[tPmeItPU], it's a [tPmA:tPU].'' So the only question is whether these prag-
matic intrusions are cases, in which case we are not dealing with prag-
matic intrusions into semantic negations.
Horn provides a set of diagnostics for metalinguistic negations
and concludes that these implicatural cases conform to the type: (a)
212 Chapter 3

metalinguistic negations typically occur with a special intonation contour;

(b) they are typically followed by a rectifying clause (``he has four'' in (32a);
(c) they do not permit negative polarity items (cf. ``He didn't eat some of
the cookies; he ate all of them'' vs. *``He didn't eat any of the cookies, he
ate all of them''); and (d) they do not allow negative incorporation (cf.
``John doesn't have three or four children; he has four'' vs. *``John has
neither three nor four children; he has four.'') However, most if not all of
these properties follow from the fact that metalinguistic negations are
retorts to prior utterances, and as we shall see, these implicature-denying
negations must also be responses to prior positive assertions.
Horn's (1989: 384) motivation for defending the metalinguistic account
of examples like (32a) is that (following Kempson) he thinks that other-
wise scalar inferences would have to be part of the semantic content of the
scalar expressions. By assimilating these cases to a nonlogical use of
negation, the pragmatic account of the scalar inferences is preserved.
Horn's position would be worth defending if these were the only other-
wise convincing examples of pragmatic intrusion into truth-conditions.
But if, as argued here, pragmatic intrusion can occur in many construc-
tions, there may be little motivation to resist the intrusion account here
too. Besides, Kempson (1986) provides some evidence that the negation
in this kind of sentence doesn't have the properties associated with
unequivocably metalinguistic negations.
These cases are truly puzzling for a number of reasons. First, Kempson
is right to insist that unlike some uses of metalinguistic negation (as in
``It's not a [vA:z], it's a [veõz]'') where no proposition is being denied, these
implicature-denying negations are logical in character: a proposition (the
implicature) falls under the scope of a one-place truth-functional operator
indistinguishable from the familiar one we call negation! On the other
hand, Horn is right to insist that what is being denied is the assertability
of the positive sentence that would have carried the implicature that is
being explicitly canceled. But there is a further wrinkle to the problem
that does not seem to have been noticed, because of the prevailing inade-
quate theory of implicature projection. This is simply that, under the
prevailing theory, implicatures do not arise under negation (Gazdar 1979,
Hirschberg 1985, Horn 1989). Thus, at least if the negation in question is
the ordinary truth-functional operator, there could be no implicature to
deny! This theory of the blocking of implicatures under negation is in fact
not correct (as pointed out in chapter 2), but the correct theory yields
equally odd results: the implicatures that will arise from negative sen-
213 Implicature and the Semantics/Pragmatics Interface

tences will not be the same as the ones that arise from their positive coun-
terparts. Thus, for example, ``John doesn't have three children'' cannot
induce the implicature `John has no more than three children', which is
therefore not available for cancellation by the negation, metalinguistic
or logical. Below we will see that if we adopt a general account of prag-
matic intrusion, all these puzzles evaporate together, and this conceptual
mopping-up itself recommends both the thesis of pragmatic intrusion in
general and its role in these implicature-denying negations in particular.
Note that implicature-denying uses of negative expressions are not
limited to explicit negations. Consider the rather than construction as
below:
(33) a. ``He has three children rather than four.''
Entails: `He has three children and he does not have four.'
Conventional implicature: `It has been suggested that he has
three-children-and-no more.'
b. ``He has four children rather than three.''
Here (33a) would naturally be used to correct an erroneous presumption
in the discourse, and p rather than q may be held to have a conventional
implicature that p, together with its conversational implicatures, has been
previously erroneously asserted or assumed. Now note that we can invert
p and q as in (33b); in this case what is negated can only be the conver-
sational implicature of three, viz. `only three'. Or in Horn's terms, what is
being negated is the felicity of asserting three given that it carries the `only
three' implicature. This ``metalinguistic'' use of the construction must be
involved in the following examples:
(34) a. ``He is brilliant, rather than clever.''
b. ``She had a baby and got married, rather than got married and
had a baby.''
c. ``The meal was delicious, rather than edible.''
d. ``His reprieve is not impossible, rather than de®nitely possible.''
e. ``It is a long way, rather than a long, long way.''
f. ``John prefers cars with six cylinders, rather than those with
four.''

3.3.4 Conclusions Regarding Intrusive Constructions

I have argued that there are a number of constructions, dubbed intrusive
constructions, where the truth conditions of the whole expression depend
214 Chapter 3

on the implicatures of some of its constituent parts. Perhaps the case is

better made on some of these constructions than on others. There will
always be doubts about whether a better semantic analysis of the relevant
construction might not accommodate the apparent pragmatic intrusions
in some other way. Further, there are unclarities in the data: for example,
some of the comparative examples may seem marginally acceptable be-
cause of the unnatural repetition of the implicature-inducing expression.
(More natural-sounding paraphrases would beg too many questions
about the synonymy of the paraphrases; but if we shift the argument to
cases where A entails B in utterances of the form ``A is better than B'' we
can get clearer examples, although we raise further queries about the
semantics of the comparative.) Others examples, especially those based on
Manner intrusions, may seem like metalinguistic or quotative uses that
ought to be treated exceptionally. Nevertheless, despite these di½culties,
the clearest casesÐlike the Q- and I-implicature intrusions in conditionals
Ðseem close to knockdown cases for the intrusive role of pragmatic fac-
tors in truth-conditional content. It would seem quite counterintuitive to
deny the truth of utterances like ``If John ate some of the cookies but no
one else ate any, there must still be some left.''
Yet the argument from these intrusive constructions (good as it has
seemed to Cohen 1971, Wilson 1975, Gazdar 1979, and Levinson 1983) is
in fact vulnerable to a simple alternativeÐnamely, the traditional account
in terms of a postsemantic Gricean pragmatics. Here is an example of
how the alternative account might go. Suppose the speaker has said
``Having to study and take an exam is better than taking an exam and
having to study.'' What she has said is of the form `A is better than A',
which is necessarily false, a contradiction. But to utter a contradiction is
not only to ¯out the maxim of Quality, it is also to ¯out the ®rst maxim of
Quantity (`Provide su½cient information'). Therefore, on the assumption
that the speaker is being cooperative at some underlying level, that is not
what she means; for the utterance to have some informative content it
must amount to `A is better than B' where B is distinct from A. To obtain
such an interpretation all that needs to be supposed is that the speaker
means `Having to study and then take an exam is better than having to
take an exam and then belatedly do some studying'. So that is what the
speaker intends the hearer to infer.38 In a similar way, Bach (1987a: 71±
74, 78; 1994: 153±154) takes the line that, taken literally, metalinguistic
negations like ``John doesn't like Mary; he loves her'' are just contradic-
tions, so cooperative interlocutors understand them charitably, as non-
215 Implicature and the Semantics/Pragmatics Interface

literal utterances short for related, expanded propositions (like `John

doesn't merely like Mary, he loves her').39
In short, for every case that we put forward as a case of semantic in-
trusion in a complex construction, the traditional (postsemantic prag-
matics) theorist (our Obstinate Theorist) can always counter that what is
literally said is contradictory, false or su½ciently empty that it will con-
stitute a ¯outing of the maxims of Quantity and Quality. The infelicity
of what is said is su½ciently blatant that it will induce an implicature
to preserve the underlying assumption of cooperation. What could be
a better trigger to a Gricean inference than (for example) a categorial
falsehood?
But against this, one should note that not all of our cases provide literal
contradictions as potential triggers for a ¯outing interpretation. Although
case (35a) below provides a necessary falsehood, case (35b) provides on
this analysis merely a contingent falsehood given the rules of soccer (but
still, arguably enough to constitute a ¯agrant violation of Quality).
(35) a. ``Having to study and take an exam is better than taking an exam
and having to study.''
Anti-intrusion hypothesis: this is patently false, and a ¯outing of
the maxim of Quality.
b. ``If each side gets three goals, it'll be a draw.''
Intrusion hypothesis: this is a very dubious ¯out.
c. ``If it costs twenty dollars, I have enough money to get in.''
Content: `If it costs twenty dollars or more, I have enough money
to get in.'
d. ``In the ancient Toltec sacred ball game, if each side got three
goals, the game was a draw.''
But (35c) is more di½cult: there is no particular reason to think the sen-
tence is false even on the literal analysis (i.e., without the implicature
stengthening the antecedent); the speaker may have a thousand dollars in
his pocket. Still, it may be objected, the conditional will be false under
some veri®cations of the antecedent (say, if it costs not only $20 but
another $1000), and that is su½cient to render it generally false and thus
capable of triggering a ¯outing interpretation. This seems highly dubious.
But, supposing we concede, consider (35d). We have no idea what the
rules of the ancient Toltec ball game were; we may be quite open to the
idea that the aim was to stop the other side getting three goals, after
which goals were no longer determinative of victory. But that is not how
216 Chapter 3

one reads the assertion: the speaker of (35d) is surely committed to stating
that a draw results if each team gets exactly three goals. Yet this inter-
pretation cannot be based on a ¯outing of the maxim of Quality, because
no known falsehood has been stated. Thus there are serious problems for
a ¯outing analysis of these intrusive cases. An account in terms of the
intrusion of default GCIs, on the other hand, predicts that the implica-
turally strengthened reading should be available even when it is not
required or coerced in order to obtain a cooperative interpretation.40 In
contrast, the ¯outing analysis suggests that the implicaturally strength-
ened interpretations should only arise to rescue an utterance from absur-
dity, falsity, or some other uncooperative infelicity. It seems to me that
the GCI account is correct in cases like (35d) and the ¯outing account
makes incorrect predictions.
But perhaps the main objection to the ¯outing account is that it is
simply not in line with our intuitions. The comparative and conditional
cases simply do not give rise to intuitions of exploitation in the way that,
say, ironies or fresh metaphors do; there's no sense of violation and in-
ferential repair. To refresh the intuitions, consider what is being claimed:
comparatives like (36a) must be treated as overt ¯outs of Quality in the
same way that obvious contradictions like (36b) are. Similarly, condi-
tionals like (36c) must be treated as obvious contingent falsehoods exactly
on a par with the ironic interpretation of (36d). But intuitively there is no
such parallelism. Although we may have few intuitions about most lin-
guistic processes, the mechanisms involved in processing ¯outs do seem
capable of being made consciousÐwitness the great body of ancient
learning on the tropes. Thus (36b) and (36d), but not (36a) and (36c),
intuitively require double processing, as it were.
(36) a. ``Getting drunk and driving home is much worse than driving
home and getting drunk.''
b. ``Clinton is both the President and not the President.''
c. ``If each team gets three goals, it's a draw.''
d. ``If you want a theory-neutral education in linguistics, you should
go to MIT.''
A ®nal response I shall save in detail for later, but in a nutshell it is this.
The ¯outing account is invoked in our intrusive cases simply to obviate
the claim of pragmatic input to semantics. It amounts to insisting on
prepragmatic semantics. But if the account is pursued, it will turn out to
have the very consequence it is trying to avoidÐnamely, a presemantic
217 Implicature and the Semantics/Pragmatics Interface

pragmatics (or pragmatic input to the compositional semantics of the

complex constructions we have been focusing on).
Despite these severe di½culties for a ¯outing account of these intrusive
constructions, we nevertheless have to concede that these kind of exam-
ples are simply not as knockdown as has sometimes been thought. And
we can assume that our Obstinate Theorist, determined to preserve a
pragmatics-free semantics, will obstinately hang on to the ¯outing analy-
sis (and for a spirited defense of this position, see Bach 1987a: 71±74, 78±
82, 1994: 135±136). Let us therefore pursue a di¨erent tack that is not so
easily countered, and in particular let us attack the Obstinate Theorist's
Achilles' heelÐreference determination.

3.4 THE ARGUMENT FROM REFERENCE

3.4.1 How Implicatures Can Determine De®nite Reference

The idea that pragmatic factors play an important role in reference de-
termination has been quite widely explored in the philosophy of language
(one has only to think of Strawson's (1950) presuppositions of de®nite
descriptions, of Searle's (1969) account of referring as a speech act, of
Donnellan's (1966) referential/attributive distinction between uses of re-
ferring expressions, of Nunberg's (1978) indirect ostension). Nevertheless,
the ®ction that (Carnapian) intensions, given just a world and a time, de-
termine extensions has been quite largely maintained in formal semantics
(and this despite Putnam's demonstration of the implausibility of the idea;
see Putnam 1975: chap. 12). Still, the ground is well prepared for the idea
that Gricean inferences might play a decisive role in reference, although
the idea has scarcely been explored.
I have already argued brie¯y that anaphoric resolution is guided in part
by the play between I-inferences (favoring local coreference and other
cohesive connections) and M-inferences (favoring the complementary
interpretations; see also Levinson, 1987a, b, and chapter 4 in this book).
Thus anaphoric pronouns, whether these are thought of as referring
expressions or variables, are linked to antecedent referring expressions at
least partly by implicature.
In this section I argue that GCIs can play a systematic role in the de-
termination of the reference or extension of de®nite descriptions, includ-
ing deictic or exophoric uses thereof, and I shall go on to make some
further remarks about inde®nite referring expressions. The theoretical
signi®cance of this claim that implicatures can determine extensions will
218 Chapter 3

depend on one's theories about the nature of reference and the role that it
plays in meaning. There are certainly general implications for the theory
of referenceÐfor example, this claim, if substantiated, seems to me to be
one of the best empirical arguments against the just-mentioned view that
intensions determine extensions.
However, we are only interested here in the theory of reference as it
bears on the main task: showing that pragmatic factors are prerequisites
to semantic interpretation. So here is what I am going to attempt to show.
Suppose we can ®nd those very Gricean inferences, which are supposed to
operate only postsemantically, involved in the determination of the refer-
ence of canonical referring expressions. Then, given the principle of the
compositionality of reference (as in, say, Martin 1987: 88), it follows that
the extensions of the maximal expression cannot be determined without
determining the extensions of each of its constituent expressions, which in
turn may depend ex hypothesi on implicature. Thus we can cannot obtain
a truth value (a maximal extension) without ®rst doing the pragmatics. In
short, if we can ®nd cases of implicaturally determined reference, then we
show that the whole semantic apparatus that recursively de®nes intensions
and extensions is reliant on pragmatic input.
It may be objected in advance that theories divide on what happens
when reference is semantically ill-de®ned (as in cases where implicature
determines reference). In Strawsonian theories (now in the ascendant, in
the modern guises of Situation Semantics and DRT), without reference
determination semantic interpretation cannot get o¨ the ground (we get
null interpretations). In Russellian theories, we can get an overall exten-
sion in such circumstances, namely a truth value False. Thus, on the
®rst kind of theory, pragmatic determination of reference will be essential
input to semantics; but on the second kind of theory, the ability to pro-
duce truth values for these cases will, I shall demonstrate, be more of an
embarrassment than an asset. So the demonstration that implicatures play
a role in the determination of reference is as near to a knockdown argu-
ment for intrusion that we are likely to ®nd.
Here is a sample of the argument. Consider the following, where (37a)
is said in a context where (37b) is mutually evident to speaker and
addressee:
(37) a. ``The man with two children near him is my brother; the man
with three children near him is my brother-in-law.''
219 Implicature and the Semantics/Pragmatics Interface

b. Context: Two men are clearly visible, one has two children near
him (and he's the speaker's brother) and the other has three
children near him (and he's the speaker's brother-in-law).
c. Content: `The man with at least two children near him is my
brother; the man with at least three children near him is my
brother-in-law.'
d. ``You should buy the car with four doors rather than the one with
two; it's more useful and the price is good.''41
Has the speaker successfully referred? Surely. Yet what he has `said'
amounts only to (37c); thus the referring expression the man with two
children near him applies equally to the speaker's brother and to the
speaker's brother-in-law. Thus on any theory of de®nite descriptions
(from Russell 1905 to Kadmon 1987) that builds in a uniqueness condi-
tion, the expression will either fail to denote (Strawson), or render the
utterance false (Russell).42 But, of course, in practice it would felicitously
denote. The reason is that a Q-scalar-inference will, by perfectly regular
process, amplify the content to `the man with at least two children and no
more than two children near him'.43 To belabor the point: what could be
a more felicitous referring expression and a truer utterance than the use
of sentence (37a) in the context (37b)? Yet on a Strawsonian account,
the speaker will have failed to make any (true or false) statement at all,
whereas on a Russellian account the statement he made will be false.
It takes only a moment to see that there is nothing special about the
exampleÐ(37d) above is perhaps a more natural example. Nor does the
observation rest on a particular analysis of ordinary language numerals
(which, as we have already seen in chapter 2, is contentious). Consider the
following examples of referring expressions whose referents can only be
individuated by Q-implicatures based on other scales:
(38) a. ``I don't like sitting in the hot tub, but I like sitting in the warm one.''
(Scale: hhot, warmi)
b. ``The student who cheated on some of the exams should be
pardoned, but the one who cheated on all of the exams should be
expelled.''
(Scale: hall, somei)
c. The man who tried to assassinate the President was arrested, but
the man who succeeded in assassinating the President was never
apprehended.''
(Scale: hsucceed, tryi)
220 Chapter 3

d. ``The senator who is possibly implicated in the scandal is the bald

man over there; the senator who is certainly implicated is the one
who looks like Al Capone.''
(Scale: hcertainly, possiblyi)
e. ``Drosophila mutants are of two sorts: The mutant that has a
left wing or a right wing is not viable; but the mutant that has
both a left and a right wing is perfectly viable.''
(Scale: hand, ori)
f. ``For once there were two reasonable ®nalists. But true to form,
the committee gave the prize not to the excellent one but to the
good one.''
(Scale: hexcellent, goodi)
(39) Generalization: given a scale hS, Wi, where S is the ``strong''
expression, W the ``weak'' one, the W cannot (by virtue of W alone)
be distinct from the S, since anything that satis®es S will satisfy W.
But in fact the W is felicitous in these cases, because the W is
pragmatically strengthened to exclude the S.
To run through these examples brie¯y. In (38a) the warm one [tub] won't
single out a referent given its semantic content alone; there are two tubs,
one hot and one warm, but the meaning of warm is only `at least warm
(and possibly hot)'; but the use of the weaker expression scalar-implicates
that the stronger expression (expressing the more extreme property, hot-
ness) would have been inappropriate. Thus the warm tub implicates `the
warm-and-not-hot tub'. In (38b), there are two students mentioned who
would satisfy the condition cheated on some of the exams, for someone
who cheats on all certainly cheats on some. Again, only if some is
strengthened to `some-and-not-all' will the student who cheated on some of
the exams pick out a unique referent. In (38c), the property of being the
one who tried to assassinate the President must equally apply to the one
who succeeded (assassination is an intentional, e¨ortful activity). But
normally, when we use the expression try we implicate lack of success;
that this is only an implicature is clear from the noncontradictoriness of a
sentence like `He tried and ®nally succeeded'. Again, the inference is due
to a scalar opposition between try and succeed; thus, the man who tried to
assassinate the President is uniquely described only if tried is strengthened
by Q-implicature to `tried-and-didn't-succeed'.44 In (38d), one who is
certainly implicated in a scandal is also possibly implicated in a scandal;
and one who is possibly implicated in a scandal may indeed be certainly
221 Implicature and the Semantics/Pragmatics Interface

implicated as shown by the noncontradictoriness of a sentence like `He is

possibly, and in fact certainly, implicated in the scandal'. Thus again the
expression with the weaker scalar item, the senator who is possibly impli-
cated, will fail to uniquely denote unless it is strengthened by scalar
implicature to `the senator who is possibly-but-not-certainly implicated'.
In (38e), the description with a left or a right wing will apply equally to an
entity with both wings; for the semantics of or is inclusive, only strength-
ened to the exclusive interpretation by a scalar Q-implicature. So only
when we have the implicature `with a left or a right wing and not both' in
place, as it were, will the description succeed in denoting. Finally, in
(38f ), the description good ®nalist does not semantically preclude that the
®nalist is excellent; however, given the scale hexcellent, goodi, there is an
upperbounding implicature that resolves the description to `the good but
not excellent ®nalist', thus establishing a unique description. In all of
these cases, a default conversational implicature will arise by perfectly
general mechanisms to render the referring expression felicitous, but the
reference can only succeed with the help of such pragmatic strengthening
(see the generalization in (39)).
We can ®nd similar examples from the other two major categories of
GCI. Here are some examples with I-implicatures:
(40) a. ``The men who drank beer and drove home are in jail; the men
who drove home and drank beer are free.''
b. ``The couple who started saving and bought a house are now
worse o¨ than the couple who bought a house and started saving:
real estate values just took o¨.''
c. ``The students who don't like math are going to fail this course;
the students who like it and even the students who [are not fond
of it but] don't mind it are going to pass.''
d. ``The boss came in and said hello to the secretary; she/the
un¯appable woman smiled.''
e. ``The woman at the desk is the secretary; the woman near the
desk is her boss.''
f. ``The man reading is my brother-in-law; the man working is my
cousin.''
g. ``John admires the book he's reading; but John's book is in fact
better.''
In (40a±b) the implicated asymmetry of and comes up again: only on
the pragmatically strengthened `and then' interpretation will de®nite
222 Chapter 3

Figure 3.3
Contradictories before pragmatic strengthening to contraries

references of the form ``The X who did U and V '' be distinct from ``The X
who did V and U.'' In (40c), not liking math is compatible with not
minding it; it is, literally, simply the state of not being fond of it, which
encompasses both the haters and those indi¨erent to it. But by a perfectly
general I-mechanism whereby contradictories are strengthened to contra-
ries (Horn 1989), the students who don't like math is interpreted as `the
students who positively dislike math' (cf. ``He doesn't like garlic''). Only
when strengthened in that way (to the set of dislikers) is the set of not-
liking students distinct from (not overlapping with) the set of students
not-fond-of-but-not-minding math (the indi¨erent ones), as required by
the inconsistent predicates fail and pass. Figure 3.3 shows the overlapping
nature of the sets as literally stated and may help to show how in inter-
pretation the sets have to be pragmatically delimited. Note that if we
change the example slightly so that the material in square brackets (are
not fond of it but) is omitted, we then have a three-way nonoverlapping
distinction set up by implicature: (a) the don't-likers are pragmatically
restricted to the dislikers by I-strengthening from contradictory to con-
trary, as before; (b) the students who like math (stated); and (c) the stu-
dents who don't mind math, which is strengthened by a Q-implicature to
the students who don't-like-but-don't-mind (due to the scale-like opposi-
tion between the strong `like' and the weak `not mind' there's an inference
from not mind to `not positively like' even though `not minding' is con-
sistent with `liking'.)45
Examples (40d±e) are perhaps less clear, but they are candidate cases
where I-inferences to the stereotype (see chapter 2 and Atlas and Levinson
1981) might play a role in referential determination. In (40d), assuming
that she (or an anaphoric epithet like the un¯appable woman) is under-
stood to be anaphoric to one of the prior NPs, there is a tendency to
interpret it as coreferring with the secretary based on our stereotypical
223 Implicature and the Semantics/Pragmatics Interface

assumptions about the sex of secretaries and bosses. Given this, the
speaker ought to have used a di¨erent set of NP expressions if the speaker
had intended otherwise, and this gives su½cient warrant for the inference
(thus the inference is an implicature, warranted by cooperative assump-
tions, and not a mere probability about the world that aids one to
understand a misleadingly ambiguous utterance). In (40e), the two
descriptions at the desk and near the desk are almost truth-conditionally
equivalent (that is, anything that satis®es the one, might be claimed to
satisfy the other); but I shall assume that at speci®es a closer relation (i.e.,
a smaller boundary area) than near. But if the terms are distinct in this
way, they are certainly overlapping in descriptive content: one who is at x,
is certainly near x. Thus the de®nite description the woman near the desk
will be equally satis®ed by either woman (provided there are two women
in the immediate vicinity of the desk). Yet if one woman is sitting down
with knees under the desk and the other standing in front of it, even
though either woman might be described truthfully as near the desk or at
the desk, there is little doubt who, given pragmatic principles, should be so
described. The reason is that at (no doubt owing to its `very immediate
vicinity' meaning) tends to induce additional stereotypical assumptionsÐ
someone who is at a desk is typically sitting down and working on it,
someone who is at the church is typically worshipping in it, someone who
is at the door is typically waiting to be let in, and so on (see Herskovits
1986). Thus given these stereotypical I-enrichments, the de®nite descrip-
tions in (40e) will be unambiguous and felicitous. Similarly in (40f ), given
the generally prevailing assumption that reading is more likely to be done
for pleasure than pro®t, and given an opposition between the man reading
and the man working which will only be resolved by assuming that the
man reading is not thereby working, a speaker will I-implicate `the man-
reading-&-not-working'. Finally, (40g) will be false unless the book he's
reading is distinct from John's book (recall sentences of the kind `The
book John is reading is better than John's book'). Thus we tend to assume
that the relation between John and the book in John's book is the other
stereotypical person-to-book relation, namely authorship.
Finally, we can construct (perhaps somewhat less clear) examples
where a de®nite referring expression will only succeed in denoting if an
M-implicature strengthens the descriptive content. Consider:
(41) a. ``The tall man is my brother; the tall, tall man is my uncle.''
b. ``The large building is CSLI; the large, large building is the
medical school.''
224 Chapter 3

c. ``She doesn't mind the teacher who teases her; but she hates the
teacher who teases and teases and teases her.''
d. ``True to form, instead of appointing the clever candidate, they
appointed the clever clever candidate.''
e. ``She made the fruit salad; I made the salad salad.''
f. ``There are two routes to the summit: the possible one and the not
impossible one.''
g. ``The boss always used a hit man. In the dock, the man who
caused her death is the man to the left, the man who killed her is
the man to the right.''
h. ``The California Code makes a distinction between ordinary
vehicles and special purpose ones. So, in the lot there, the car has
to conform to vehicle emissions, but the vehicular device doesn't
have to.''
In (41a±c) repetition suggests a heightened quality to the predications
in question, su½cient to achieve unique reference by the referring expres-
sions. Clearly, the assumption here ( justi®ed in chapter 2) is that the kinds
of repetition used in these examples is not semantic modi®cation. Thus,
on this assumption, the repetitive descriptions add no truth-conditional
content, but the use of the more prolix expressions implicates a distinction
from the unrepetitive descriptionsÐnamely, one in which the action or
property is intensi®ed (as suggested iconically).46 In (41d±e) the redupli-
cations clearly take on special meanings by M-implicature (as suggested
to me by Larry Horn). In (41f ), the opposition between the simple posi-
tive the possible one and the double negative the not impossible one sug-
gests that one route is markedly more passable than the other, and thus
again an M-implicature allows two referring expressions with identical
semantic content to refer uniquely to di¨erent referents. In (41g), the
use of the periphrastic causative (as before) suggests indirect causation by
M-implicature calculated in contrast to the lexical causative kill, which
suggests direct causation by I-implicature; thus we can identify the in-
direct killer with the boss, the direct killer with the hit man. Similarly
in (41h), the periphrastic vehicular device suggests some very un-car-like
vehicle.
I have already mentioned further examples of M-determination of ref-
erence that are of a rather di¨erent kind and involve a preference for dis-
joint reference invoked by the use of a more prolix or informationally
nonredundant NP (Levinson 1987a, b and chapter 4 of this book). This
225 Implicature and the Semantics/Pragmatics Interface

has to be understood against the I-heuristic for anaphoric linkage: wher-

ever a following NP has a semantic content properly included in the con-
tent of a preceding NP (and thus the following NP is more semantically
general than the preceding NP), there is a strong preference for corefer-
ence or, more strictly, an anaphoric relation. This is illustrated in (42a±b).
Note that this preference for coreference is discouraged where a subse-
quent referring expression contains semantic conditions not contained
within the potential antecedent as illustrated in a 0 (where not all ships are
ferries) and b 0 (where not all Johns are adult males).
(42) a. ``The ferry1 hit a rock. The ship1 capsized.''
a 0. ``The ship1 hit a rock. The ferry2 capsized.''
b. ``John1 entered the room. He1 walked over to the window.''
b 0. ``John1 entered the room. The man2 walked over to the
window.''
This pattern will be explored further in chapter 4, but the point here
is simply that it appears that implicature is often the linking principle
that makes one NP referentially dependent on a speci®c prior referring
expression.
If this account of all these cases of pragmatic determination of refer-
ence is roughly on the right lines, then it would seem that our job is done.
Recollect that our aim was to ®nd knockdown arguments to the e¨ect
that Grice's circle is a genuine dilemma: The Gricean `what is said', upon
which basis implicatures are supposed to be calculated, is itself partly
determined by implicatures that help ®x reference. The theoretical impli-
cation is that Carnapian intensions (together with worlds and indices) do
not alone determine extensions. Instead, there is crucial pragmatic input
to semantic interpretation, and the pragmatic principles in question are
identical to those thought to operate only as postsemantic principles of
Gricean utterance interpretation. Yet the battle with the Obstinate Theo-
rist is still not quite over.

3.4.2 Implicaturally Determined Reference and Donnellan's Referential/Attributive

Distinction
The examples I have given are meant to establish that implicature calcu-
lation is often a prerequisite to establishing reference. Can our Obstinate
Theorist escape this conclusion? I don't believe so, but we know he's
going to try. Here is what he might say:
226 Chapter 3

You've made an implicit distinction between implicaturally determined reference

and pure (semantically) determined reference. Actually, this is just the identical
distinction to Donnellan's (1966) referential versus attributive distinction. Quite
respectable people (Donnellan (1978) himself, Barwise and Perry (1983)) think
that this distinction is semantic. Other, even more respectable people (like Kripke
(1977)) think that the distinction is one between semantic correctness on the one
hand and pragmatic success with semantic error on the otherÐmore precisely, that
it's a distinction between a proper semantic use and an improper but occasionally
successful use of referring expressions. Either way, there's a perfectly good account
of the referential uses (alias implicaturally determined references) that is quite
compatible with the received view that all Gricean pragmatics is postsemantic.
Could the Obstinate theorist be right, that our distinction between pure
reference and implicaturally ®xed reference is the very same distinction as
Donnellan's? Recollect Donnellan's examples of the kind:
(43) a. ``The man over there drinking champagne is happy tonight.''
b. Context: the man the speaker has in mind is holding a
champagne glass in fact full of seltzer.
c. Kripke's alternate context: same, but there's also another man
over there (whom the speaker does not have in mind) who has a
nonvisible glass of champagne and is gloomy.
d. ``Smith's murderer is insane.''
Referential: That chap, Smith's murderer, is insane.
Attributive: Whoever is Smith's murderer is insane.
The puzzle associated with the utterance of (43a) in the context (43b) is: if
the fake-champagne drinker is indeed happy, we're inclined to think what
was said was true, even though on a Russellian analysis (given context
(43b)) it should be false, and on a Strawsonian analysis of de®nite
descriptions it should be neither true nor false. If we change the context,
as suggested by Kripke (1977), to that in (43c), then on either analysis
what is said ought just to be false.
Donnellan's suggestion: in some cases, we use the description to get to
the referent and then throw away the description as it were; then if the
predication holds of the referent, what was said was true. Thus there are
often two interpretations of de®nite NPs as sketched in (43d), where the
referential interpretations treat the description merely as a route to the
referent, but the attributive interpretations require that whatever meets
the description has the predicated properties.
Are our cases instances of the very same referential use of de®nite
descriptions? Consider again the prototype example:
227 Implicature and the Semantics/Pragmatics Interface

(44) a. ``The man with two children near him is my brother; the man
with three children near him is my brother-in-law.''
b. Context: Two men are clearly visible, one has two children near
him and the other has three children near him. The ®rst is the
speaker's brother, the second his brother-in-law.
c. Russellian analysis: (a) is false
Strawsonian analysis: (a) is truth-valueless
Intuitively: (a) is true! (stronger than Donnellan's cases)
On a Russellian analysis, the ®rst de®nite description applied to its pred-
ication will be just plain falseÐit fails to meet the uniqueness speci®ca-
tion. But again, just as in the Donnellan cases, reference may succeed
nevertheless, and we'd be loath to call the sentence false in context (44b).
So this suggests an immediate identi®cation of implicaturally ®xed refer-
ence and speaker reference (Donnellan's referential uses). The other term
of the correspondence would identify implicature-free reference determi-
nation with Donnellan's attributive uses. The parallel between the two
distinctions may seem quite exact, as comparison of examples (43) and
(44) above should make clear.
But there's an insuperable problem with this identi®cation of the two
distinctions. Implicaturally ®xed references exhibit within themselves the
very same Donnellan ambiguity (if that is what it is). Consider:
(45) ``The man who has two children is prudent; the man who has three is
a fool.''
Referential use: in a context where two men are visible, one with
two children near him, the other with three.
Attributive use: gloss `whoever has two children and no more is
prudent; whoever has three is a fool.'
In the referential use, note that the man who visibly has two children may
in fact have three (one hiding behind a tree); the fertility of the two men
has nothing to do with the predications because the number of the chil-
dren is simply a route to the referents. But in the attributive use, the
description ``occurs essentially'' as Donnellan puts it, just as in his proto-
type cases.
Thus we cannot assimilate the implicaturally ®xed versus pure reference
distinction to the Donnellan distinction; it is orthogonal. So the Obstinate
Theorist is just wrong. But, he may retort, the distinctions may be related,
even if not identical, in which case the semantic machinery or prag-
matic arguments developed to handle the Donnellan distinction might be
228 Chapter 3

redeployed, mutatis mutandis, to handle our distinction. So let us examine

Kripke's pragmatic account and Barwise and Perry's account to see what
succor there may be for the Obstinate Theorist.
Kripke points out that, with a clear distinction between speaker-
meaning and linguistic-meaning, it seems perfectly possible to just main-
tain that the Donnellan referential uses are, where strictly false, just plain
false at the level of linguistic-meaning while being true at the level of
speaker-meaning. That is, Donnellan's distinction is not a semantic one
and we may distinguish between:
1. Semantic reference: ``Where the speaker has a designator in his idio-
lect, certain conventions of his idiolect (given various facts about the
world) determine the referent in the idiolect.''
2. Speaker's reference: ``That object which the speaker wishes to talk
about, on a given occasion, and believes ful®lls the conditions for
being the semantic referent of the designator.''
Note one mismatch between our implicature-determined reference and
Kripke's speaker-reference: in our cases, the speaker has no intention of
conforming to the semantical rules. That is, the speaker is under no mis-
taken illusions about the relation between his words and the world; he
knows perfectly well that his semantic reference will be inadequate or
inapplicable without implicatural strengthening. When a speaker refers to
the man as the man with two children, even though there are two men who
®t the semantic content by having at least two children and even though
the addressee can see that the semantic content is not individuating,
he does so with intention aforethought. And even though a Russellian
analysis requires that what our speaker said was false, the speaker has
no false impressions and by any reasonable account (I would think) is
speaking truly (assuming the predication holds). That is what gives our
cases interestÐthey are not odd quirks of odd circumstances. They are
intentional and routine uses of expressions where the semantic conditions
for use are not met.47
Kripke concedes that there's just one circumstance under which he
might admit that Donnellan's distinction is semanticalÐnamely, if the
successful (but condition-failing) reference was clearly intuitively true.
Donnellan's cases arguably fail this test, but the implicaturally deter-
mined cases seem to pass it.48 Nevertheless, I am not advancing the dis-
tinction as a semantic ambiguity; it's a truth-conditional ``ambiguity''
only because semantic conditions are routinely enriched by pragmatic
229 Implicature and the Semantics/Pragmatics Interface

conditions prior to truth-conditional assessment. Thus we can agree that

our implicaturally determined references involve pragmatic factors while
still insisting that the truth conditions of the utterance are as they intu-
itively are.
At the risk of repetition let me sum up so far: our distinction between
pure and implicaturally determined reference is orthogonal to Donnel-
lan's referential versus attributive distinction. Nor is it accommodated by
Kripke's semantic-reference versus speaker-reference, because of the mis-
match between speaker-reference and implicaturally determined refer-
ence: (a) In speaker-reference there must be a (possibly erroneous) belief
that the semantic conditions are met, but in implicaturally determined
reference this is not so (uniqueness conditions for example may be obvi-
ously infringed); (b) In speaker-reference it is su½cient that the speaker
believes that the semantic conditions are met; in implicaturally deter-
mined reference that is not necessary but in any case not su½cientÐthe
implicated conditions must also be met; (c) Implicaturally determined
reference meets at least the prima facie condition Kripke puts on a genu-
ine semantic ambiguity, namely that one has the clear intuition that such
a statement is true (even when, in our cases, qua semantic-reference it is
false)Ðyet nevertheless it is obviously not a semantic ambiguity because
it is introduced by Gricean mechanisms. The solution to that dilemma is
simply to admit that implicatures contribute to truth conditions.
Kripke's (1977) tack is to treat the referential/attributive distinction as
corresponding to a pragmatics/semantics distinction and thus to dismiss
the cases where there is con¯ict as usable falsities. Another approach is to
take the distinction as indeed a semantic distinction as Barwise and Perry
(1983) do in Situation Semantics. Perhaps that treatment might be able
to accommodate our distinction between (pure) semantic reference and
implicaturally determined reference, and we should consider it brie¯y. In
that theory, the meaning of an expression is a relation between various
contextual factors and described situations. Therefore, the Donnellan
distinction can be captured by letting the referential use of de®nite de-
scription the F (x) (to use an informal notation) only have the denoted
individual a in the described situation, whereas attributive uses have the
describing condition F itself (indirectly) in the described situation (Barwise
and Perry 1983: 146¨ ). Thus in the referential use of, say, the man drink-
ing champagne, the describing condition is used to obtain the referent,
but is thrown away as it were. It does its job by constraining another
situation, the resource situation, which is some accessible situation (e.g.,
230 Chapter 3

the immediate view, in which there is a man drinking what appears to be

champagne). The meaning of the referential use of de®nite NPs can then
be thought of as a function that has as argument the `resource situation'
and as value the referent (here the salient apparent champagne drinker,
let's call him Joe). Thus the interpretation of The man drinking champagne
is happy is a described situation in which Joe is happy.
Can we apply the same analysis to implicaturally determined references
like the man with two children (in the context of a man with exactly two
and a man with exactly three children), and if so would it help? The
answers are yes and no, respectively. The idea that the resource situation,
once it has done its individual-selecting job, can be discarded has some-
thing right about it for these cases too. But the mechanisms of resource
situations will not help us. That would only shift the problem to another
situation (from the described to the resource situation), and in that
other situation we still have the identical conundrum, namely that these
de®nite descriptions fail to describe uniquely without implicatural strength-
ening. Thus, without that strengthening, the function taking us from re-
source situations to individuals will be ill-de®ned (it will have two valuesÐ
the man with exactly two and the man with exactly three childrenÐfor one
argument). A second reason for not looking to this analysis for succor is
simply that if the same mechanism gives a good analysis of the Donnellan
cases, then it cannot do so for this distinction because, as we have seen, it
is orthogonal. Indeed, our cases of implicaturally determined reference
can have both the attributive and the referential readingÐwhich shows
that the resource situations simply can't be involved in our distinction
(because they are invoked on one reading, gone on the other). In any case,
although the general apparatus of Situation Semantics has many impres-
sive ways of handling those aspects of contextual determination of
interpretation that are indexical or quasi-indexical, there is no place for
elaborate pragmatic reasoning of the Gricean kind.
I conclude that implicaturally determined references are not some
semantic distinction in disguise (as they might be on some analyses if they
con¯ated with the Donnellan distinction). They are what they appear
to be: cases where implicatures play a crucial role in determining truth-
conditional content.

3.4.3 The Obstinate Theorist's Final Retort on Reference

In the case of intrusive constructions, we admitted that the Obstinate
Theorist who insists on a postsemantic pragmatics can just require what
231 Implicature and the Semantics/Pragmatics Interface

seems felicitous to be false or contradictory and then treat the intuitive

content as a Gricean inference from a ¯out. Can't the Obstinate Theorist
do that in these cases, too? The man with two children G's, said where
there are two individuals one with two and one with three children, is
false (Russell) or unde®ned (Strawson); to get some cooperative proposi-
tion, the speaker must intend the hearer to push the sentence through the
pragmatics (cf. Bach 1987a: 82±85)!
Two objections:
1. Methodological: if this tactic is pursued willy-nilly, in violation of
our intuitions about truth and falsity, why not claim that any other sen-
tence for which the proponent's semantic theory makes the wrong pre-
dictions is in fact patched up by the postsemantic pragmatics and thus is
after all correctly analyzed by his unlikely theory?
2. Substantive: on some accounts of de®nites (e.g., Russell's), the neg-
ative will be true just when it is intuitively false, and the positive false just
when it is intuitively true, which is too absurd to be rescued by Gricean
maneuvers!
To see this consider the informal Russellian analysis in (46d) of the
negative sentence in (46a), which states that one or more of three con-
ditions are not met. In the context (46c), there's an individual satisfying
the conditions of being my brother and having at least two children; but
the middle condition requiring that there be no other individual with at
least two children is not metÐtherefore the denial of the positive sentence
will be true. Thus the semantic prediction is that (46a) will be true just
when it is intuitively false, and (46b) will be false just when it is intuitively
true. The Obstinate Theorist's attempt to hang on to a postsemantic
pragmatics will thus have quite absurd consequences.
(46) a. ``It's not the case that the man with two children is my brother.''
b. ``The man with two children is my brother.''
c. Context: There is a man with exactly 2 kids who is my brother,
and another with exactly 3 kids who is not my brother.
d. @[bx HAS (x, at-least±2-children) & @by (HAS (y, at-least±
2-children) & y 0 x) & my-brother(x)]
e. On a Russellian analysis: (a) ˆ True (b) ˆ False
Intuition: (a) ˆ False (b) ˆ True!
The conclusion, I would hope, is that implicaturally determined reference
must be accepted, and we must turn to consider the consequences for an
overall theory of meaning. But ®rst we must give the Obstinate Theorist
the coup de grace.
232 Chapter 3

3.4.4 Presemantic Pragmatics versus Postpragmatic Semantics

What (I hope) we have established so far:
1. We need some presemantic Gricean pragmaticsÐto do disambigu-
ation and indexical resolution, not to mention ellipsis unpacking and
generality-narrowing.
2. We seem to need substantial presemantic Gricean pragmatics to
handle the intrusion examples in complex constructions like the compar-
ative and the conditional.
3. Even in simple sentences, we ®nd Gricean intrusion in reference-
determinationÐif we resist the Gricean analysis, we predict bizarre truth
conditions (e.g., the man with two children case, where on the Russellian
analysis we get a falsehood where there's an intuitive truth and a truth
where there's an intuitive falsehood).
All these di½culties disappear if we accept that Gricean processes can
be input to truth-conditional content (although further puzzles arise,
naturally).
But let's return to our Obstinate Theorist, the one who insists on
maintaining a pure semantics and a wholly postsemantic pragmatics.
None of the conclusions listed above seem totally insuperable at ®rst sight
to such a die-hard. Here would be his responses:
1 0 . We don't attempt contextual disambiguation prior to semantics; we
just give a single distinct semantic interpretation for each reading. Like-
wise, we don't do presemantic indexical resolution or any other generality
narrowing; we just end up with extremely general truth-conditional con-
tent (``Bill was here'' will be true just in case Bill was within some unspe-
ci®ed area, possibly of limitless extent, which includes the speaker!).
2 0 . In the intrusion constructions, we Obstinate Theorists may just take
these to be ``Gricean ¯outs:'' thus our intuition that If each side got three
goals, it was a draw is true is simply based on assessing the truth of the
statement-plus-implicaturesÐit is an assessment of speaker-meaning, not
sentence-meaning.
3 0 . In the reference cases, we need to distinguish di¨erent kinds of cases:
a. In the cases where what is said is literally false according to the
Obstinate Theorist (as on a Russellian analysis of our case of the man with
two children) we can adopt a ¯outing analysis, which gets us a truth for
the statement-plus-implicature (the speaker-meaning).
b. Where reference is unde®ned, the account is more tortured. (For
example, the Strawsonian analysis of the man with two children; the
``bridging'' anaphora cases where no unique referent has been previously
233 Implicature and the Semantics/Pragmatics Interface

established, as in The bridge collapsed. The wood was rotten.) Here the
semantics gives us nothing (null interpretations); so the pragmatics can
give us nothing. Here the Obstinate Theorist will have to gesture to some
theory of accommodation (cf. Lewis 1979) whereby the use of an un-
de®ned reference is reinterpreted as if it had been the use of some other
referring expression that would have been well de®ned had it been used.
The important point here is that this invoking of accommodation con-
stitutes an admission of presemantic pragmatics in these cases.
c. In the cases where we have competing antecedentsÐas in my ac-
count of preferred interpretations of pronouns versus lexical NPsÐwe
can just have the semantics give the interpretation for each potential
antecedent (i.e., treat the sentence as n-ways ambiguous).
Let us indulge the Obstinate Theorist and see what his theory would be
likeÐwhat would the consequences be? Well, one surprising consequence
has been pointed out to me by Ivan Sag: namely, the result of having no
presemantic pragmatics will be to have a postpragmatic semantics! In
which case, we come back to more or less the same picture: a semantic
component with Gricean input.
Let us trace through Sag's argument. Our Obstinate Theorist will deny
that the example (47a) below is trueÐbecause without intrusion it is
necessarily false. But the semantics will compute the pragmatics-free
interpretation, namely the contradiction, by (say) treating worse than as
a two-place predicate which can be true only if the extensions of the
two arguments are distinct. The semantics now sends the results to the
pragmatics.49 The pragmatics will now reject the interpretation as a
cooperative assertion, and seek to patch it up, e.g., by computing the
`and subsequently' implicature of each and in the schema `(event1 ) and
(event2 )'. But now what? We need to compute the meaning and extension
of the whole sentence over again, but this time with an alteration in the
meaning and extension of two of its parts. To do that kind of recompu-
tation for any arbitrary sentence we need the whole apparatus of compo-
sitional semantics. But by hypothesis, the semantics is free of pragmatic
input: so this recomputation cannot be done by the ``pure'' semantics.
Equally, the recomputation can't be done in the pragmatics: the prag-
matics is part of a theory of communication, built on nonmonotonic
principles quite di¨erent in character than those operative in semantics.
So we will need a second semantics component, a postpragmatic seman-
tics, which, unlike the ``pure'' semantics does accept pragmatic input but
is in all other ways identical!
234 Chapter 3

So we need to send our sentence with its pragmatically retooled parts

back to a second semantics component to get the semantic content of the
whole: we apply the predicate worse than to two altered arguments (as
in a 0 ) and derive a set of truth conditions and an extension in our world
(say, a truth).
(47) a. ``Having a baby and getting married is worse than getting
married and having a baby.''
a 0 . `Having a baby and then getting married is worse than getting
married and then having a baby.'
b. ``If each side gets three goals, then it's a draw.''
b 0 . `If each side gets three-and-no-more goals, then it's a draw.'
In more or less exactly the same way, the conditional in (47b) will come
out false (contingently, given the rules of soccer), and the semantic con-
tent (the proposition expressed) will be sent to the pragmatics for ®rst-aid
treatment. If the addressee knows the rules of soccer, the falsehood will be
self-evident; the content of three (namely, at least three) can undergo the
normal scalar strengthening. But now we want to know the truth con-
ditions (and the truth) of the whole, recomputed by combining the inten-
sions and extensions of the retooled parts. We need a semantics again, a
postpragmatic semantics.
Thus the only alternative to a presemantic pragmatics is a post-
pragmatic semantics. Let us now assess these two alternatives. The fol-
lowing subsections detail what is against the Obstinate Theorist's scheme.

3.4.4.1 Violation of Intuitions The Obstinate Theorist's scheme makes

the wrong predictions about intuitions of truth and falsity (as in all the
intrusive construction examples); we are told to discount these intuitions
as really intuitions about the truth value of the proposition expressed at
the level of speaker-meaning, not at the level of sentence-meaning. Addi-
tionally, it discounts our intuitions about the distinction between ¯outings
or exploitational implicatures and ordinary GCIs. To repeat an earlier
point, the intrusive comparative constructions are supposed to have the
¯outed quality we associate, for example, with patent contradictions (thus
there should be a close parallel between (48a) and (48b)), and the condi-
tional intrusive examples should feel parallel to ironies.
(48) a. ``Getting drunk and driving home is much worse than driving
home and getting drunk.''
235 Implicature and the Semantics/Pragmatics Interface

b. ``Clinton is both the President and not the President.''

c. ``If each team gets three goals, it's a draw.''
d. ``If you want a theory-neutral education in linguistics, you should
go to MIT.''
The sentences in (48b,d) seem transparently to require the reprocessing
one associates with ¯outs; example (48a) is certainly less clearly in this
category, and (48c) intuitively doesn't belong at all. It's hard to say how
much weight should be put on such intuitions, but surely some.

3.4.4.2 Exceptions There seem to be exceptions to the Obstinate

Theorist's position that he must concede. These are the cases where reference
is unde®ned, so no overall extensions can be computed by the semantics.
If the semantics provides nothing at all in the way of interpretation, the
pragmatics will have nothing to work on. Cases include the de®nite ref-
erence examples on a Strawsonian analysis, the bridging examples and
other cases of so-called accommodation; also perhaps examples of index-
ical resolution where without narrowing we can get no extension (per-
haps, examples like Some of you come here. You can be on this team, you
others on that.) In these cases, we need the pragmatics to give us some-
thing in the way of an extension as input to the semantics, in order to get
any interpretation at all.

3.4.4.3 Lack of Economy The Obstinate Theorist makes us recalculate

the semantic content after pragmatic resolution; thus we get double the
semantic processing. We also need, by hypothesis of the purity of seman-
tics, two semantic componentsÐthe pure one, and one that takes input
from the pragmatics to recalculate more sensible intensions and exten-
sions. And it follows as an additional oddity of this position that we will
need yet another pragmatics! Consider the following:
(49) a. ``In China, if you have a small family with two children, you are
dei®ed.''
b. [Said of inebriated lecturer] ``At least we are all agreed that in
future we'll have the reception before the lecture: having plenty of
port and giving a lecture produces better results than giving a
lecture and having plenty of port.''
Here, after passing through the pure semantics, we need to do the Gricean
pragmatics, so that in (49a) two is restricted to `at most two' and in (49b)
236 Chapter 3

the two ands get ampli®ed to `and then'. Then we need to recompute the
semantics (in the postpragmatics semantics) with these new extensions.
But now that we have the propositions expressed by the utterances, we
®nd they are obviously false (given what we know about China and
dei®cation, and what we know about giving lectures on plenty of port).
So we send them o¨ again for pragmatic processing, to the pragmatics
that follows the postpragmatic semantics!
Thus the overall picture proposed by the Obstinate Theorist has the
character that obstinacy unfortunately tends to have: unattractive repe-
tition. His theory has two more or less identical components doing
semantic interpretation, and he has at least two applications of Gricean
pragmatics (three if he acknowledges the force of the accommodation
examples mentioned above). We end up with a simpler picture with less
exceptions on the view that allows a presemantic pragmaticsÐor more
generally does not insist on a position where semantic interpretation
accepts no pragmatic input. But above all, when we examine the con-
sequences of the Obstinate Theorist's position, we see that he cannot
avoid the very result that motivates his theoretical contortionsÐnamely,
an interleaving of pragmatic and semantic processes. Thus in the critical
respect, both our theory and the Obstinate Theorist's do not in the end
di¨er!

3.5 SOME IMPLICATIONS

3.5.1 Disposing of the Existing Responses

I outlined above a number of existing or potential responses to Grice's
circle (see section 3.2.7). In light of the further evidence adduced we may
return to them and see if we can now dispose of at least some of them.

3.5.1.1 Lack of Evidence This response, personi®ed in our Obstinate

Theorist, has been the direct target of most of the remarks above. If
doubts still remain about intrusion phenomena, then that may be because
we have restricted ourselves to the intrusion of GCIs only. If particular-
ized implicatures are also taken into account, the range of evidence
widens immensely, because many cases of disambiguation, narrowing,
and other kinds of pragmatic resolution depend on Relevance implica-
tures (as argued in Sperber and Wilson 1986). As well, we have already
noted the case of metonymic transfers (Nunberg 1978, Sag 1981, Fau-
237 Implicature and the Semantics/Pragmatics Interface

connier 1985). The following examples make the point that semantic and
syntactic intrusion occurs in these cases:
(50) a. ``James Joyce liked to read himself.''
b. ``If this painter [pointing at a painting] was more abstract, he
would sell well.''
c. ``If you can't ®nd a copy of Chomsky over in Philosophy, I will
lend you mine.''
d. [Pointing at old car] ``That was when I was a kid.''
If the net is widened to take into account the intrusion of background
knowledge generally (as licensed by the I-principle or a maxim of Rele-
vance), then all the pragmatic resolution that goes under the rubric of
local pragmatics in AI work provides directly relevant data (see Hobbs
et al. 1987). This includes the role of knowledge in resolving metonymies,
generality narrowing in nominal compounds, pronominal resolution,
quanti®er scoping, disambiguation, and so forth. Detailed studies like
that by Herskovits (1986) of the pragmatic resolution of spatial descrip-
tions are further grist.
Also, we now need to go back and look at many of the apparently
trivial facts that were adduced in favor of generative semantics and that
were given burial without autopsy (the literature on indirect speech acts,
negative polarity items, and so on all come to mind). For example, we
reviewed in chapter 2 Horn's (1978, 1989) argument that Neg-raising is
actually an I-inference (R-implicature in his terms) whose availability is
systematically determined by scalar strength of predicate (among other
things); because the distribution and interpretation of negative polarity
items depends on the availability of this inference, we have a systematic
intrusion of pragmatics into semantics and grammar (although this
clearly involves ongoing conventionalization). Also, work on nonce uses
of nouns as verbs, and the like (as studied by Clark and Clark 1979) are
directly pertinent (see Sag 1981). A renewal of interest in encoded sub-
jectivity and point of view in language (Lyons 1981, Kuno 1987)Ð
whether in logophoric re¯exives (Sells 1987, O'Connor 1987), free indirect
style and metapragmatics (Lucy 1993), evidentials (Chafe and Nichols
1986), or more generally in the lexicon (Mitchell 1986)Ðsuggests that the
signaling of point of view exploits generalized pragmatic inference (this is
explored in chapter 4). All in all, there appears to be a wealth of data (in
need, it is true, of detailed reassessment) that demonstrates not only the
238 Chapter 3

intrusion of pragmatic inference in semantic interpretation but how that

intrusion can also in¯uence the distribution of syntactic elements.

3.5.1.2 Restricted Presemantic Pragmatics This approach allows that

Gricean pragmatics might provide input to semantics but requires that it
be of a limited and restricted kind. Thus, as mentioned, Relevance theo-
rists propose that there is a special kind of implicature, an explicature,
that embellishes logical forms in limited ways. We have reviewed the cri-
teria for explicatures that have been suggested by Sperber and Wilson
(1986), Carston (1988), and Recanati (1989, 1993), and we found that
they remain unclear. The hypothesis is not well circumscribed, but the
general idea is clear enough: contextual input should ®rst be used to
extract a su½ciently informative proposition. It is an idea shared with
the AI community, who make much the same assumption and call the
restricted ampli®cations of logical form ``local pragmatics'' (see, e.g.,
Hobbs et al. 1987).
For the existence of explicatures to be more than a terminological issue,
they must have some speci®c, identifying properties. The di½culty is to
see how and why certain restricted kinds of pragmatic inference should
alone have presemantic application and, if so, why they should be derived
by the identical apparatus that derives postsemantic implicatures (the prin-
ciple of Relevance in Relevance Theory). In fact, the best candidates for such
presemantic inferences might be GCIs. By hypothesis, they are default
inferences driven by (potentially fragmentary) semantic representations,
but Relevance theory recognizes no such special class of default inferences.
One way to test the hypothesis that explicatures are essentially di¨erent
in kind from implicatures is to take those implicatures that look most
postsemanticÐnamely, the tropes and ¯outs that seem to presuppose the
prior establishment of a literal meaningÐand see whether they can in-
trude in semantic interpretation or play a role in reference determination.
Unfortunately for the hypothesis in question, they do seem to have that
ability. Let us just focus on the ability of such tropes to determine refer-
ence (because from that ability, it will follow that tropes will intrude in
our intrusive constructions).
(51) Metonymic particularized transfers:
a. ``John took the book to Philosophy.''
Context (i): in a bookstore, Philosophy might refer to a certain
stretch of shelving;
239 Implicature and the Semantics/Pragmatics Interface

Context (ii): in a university, Philosophy might refer to a

department)
b. ``Pearl Harbor remembers Pearl Harbor.'' (newspaper headline)50
(52) Ironic reference:
a. ``If you need a car, you may borrow my Porsche'' (referring to
speaker's VW).
b. ``I had the choice between a career in linguistics and a career in
domestic economy; and bless me, I went and chose the easy
option.''
(53) Metaphoric reference:
a. ``Come and meet our resident Greenberg'' (said of Bernard
Comrie).
b. ``Pass the vintage vinegar'' (said of a bad wine).
(54) Reference by indirection:
a. ``I need to go to the you-know-where.''
b. ``I was most sorry to hear of Fred's passing.''
The same kind of reference-by-trope can then be embedded in an intru-
sion construction, as in the now familiar ``If you want a theory-neutral
education in linguistics, you should go to MIT'', or ``Even if you take my
Porsche, you'll still get there in time'', or ``If you like this vintage vinegar,
you will be pleased to hear you can buy it for one dollar a bottle at your
local Tesco Superstore.'' Note that ironies, on the Sperber-Wilson theory
of Relevance, are implicatures interpreted as ``echoes'' of what someone
might have said: they are distinctly not explicatures. Yet they may play a
role in determining the proposition expressed by the antecedent of a con-
ditional, against which the consequent should be assessed.
The idea that some implicatures are essential to the determination of
the propositional content of an utterance, whereas others are further
embellishments or exploitations, may well be correct. But it does not
follow that these two roles are necessarily occupied by pragmatic infer-
ences of di¨erent natural classes. As we have seen, the evidence does not
directly point in that direction. However, caution is in order. It could be,
for example, that there are distinct types of intrusion: some, like those
associated with our GCIs and perhaps the metonymic transfers above,
may be genuinely presemantic (in the sense that they can be calculated
without access to the full propositional content of the utterance), whereas
others (like the cases of ironic reference) may require double processingÐ
240 Chapter 3

being sent back for pragmatic readjustment after the ``literal meaning'' is
computed (much as our Obstinate Theorist has argued for all the intru-
sion cases). There is some intuitive support for this (see Horn 1992b: 184±
188). But single versus double processing would seem to be a dimension
orthogonal to the explicature/implicature distinction: we may be forced
to compute an irony by double processing to obtain a proposition (an
explicature) relevant enough for the kind of processing envisaged in
Relevance theory.
The position adopted here is therefore thatÐfor reasons spelled out
belowÐGCIs constitute a genus of pragmatic inference especially able
and likely to intrude into semantic representations, but whether they do
so will depend in part at least on the position of GCI-inducing expressions
in constructions. In short, on particular occasions, GCIs may or may not
be explicatures: they crosscut the Relevance-theory distinction between
explicatures and implicatures, as do other kinds of pragmatic inference.

3.5.1.3 Semantic Retreat: Semantics as Syntax The strategy of seman-

tic retreat is to yield over matters of (what is normally called semantic)
interpretation entirely to pragmatics (or other general reasoning abilities),
while assigning to semantics only the job of computing an LF or semantic
representation from the syntactic input and lexical content. As mentioned,
a number of theorists who hold very di¨erent ideas about the nature of
semantic representations may be tempted by such a position. For exam-
ple, not only Relevance theorists (Sperber and Wilson 1995: 257±258)
may ®t in here, but also those who hold that linguistic semantics is
exhausted by the level of LF in Chomskyan theory (a level that consists of
``those aspects of semantic structure which are expressed syntactically''
(May 1987: 306)). Even those conceptualists like Katz, who think of
semantics as essentially computations over an extremely rich syntax of
``markerese'' or the like, may wish to hand over matters of the interpre-
tation of such representations to pragmatics.
On such a strategy, we escape semantic intrusion by yielding up all
aspects of utterance interpretation to pragmatics. Few theorists would, I
think, deny that such an abstract, underspeci®ed level of semantic content
should play an important role in a semantic theory (see e.g., Ruhl 1989,
Pustejovsky 1995, van Deemter and Peters 1996). The question is whether
this is all there is to linguistic semantics. It is therefore important to see
that what one is left with is an extremely impoverished level of represen-
tation. Kempson (1986) recognizes this: on her account, logical forms will
241 Implicature and the Semantics/Pragmatics Interface

have undecided scope, meta-variables for pronouns, vacuously general

terms, in short, subpropositional status. The consequence is that such a
level of semantic representation or logical form is no longer a level
over which the traditional sense relations can be statedÐit simply cannot
be as rich as Katz (1987) imagines. That is because, if we are correct
about pragmatic intrusion, the intuitive concept of entailment, in terms of
which all the sense relations (synonymy, antonymy, converseness, etc.)
are de®ned, cannot be stated over semantic representations alone. If one
thinks of entailment in terms of truth conditions, the reason for this is
obvious: truth is a property of utterances (complete with intrusive impli-
catures), and entailment is the concept of inference-while-preserving-
truth. But equally, if one thinks of sense relations as primitive, such
relations must still relate well-de®ned propositions, not underspeci®ed
propositional schema of the kind that we may be able to algorithmically
extract from syntactic structures. (Given indexicality, Katz's position is
already threatenedÐfor example, synonymy relations can't really be
stated except by holding indices constantÐbut given pragmatic intrusion
the position is overwhelmed.51)
Semantic retreat therefore has the consequence that semantics does not
look like either of the familiar, conventional, rival enterprises: it is not
about truth-conditional content on the one hand, nor about the relations
of sense that hold between sentences on the other hand. Instead, it is
exclusively about a new, strange level populated by semantic wraithsÐa
level of fragmentary structures, underspeci®cation and half-information,
even archi-sememes (Atlas 1989: 146). The recognition of the existence of
this level is one of the important sea changes in the history of semanticsÐ
it is real enough, but it is relative terra incognita.
But one can recognize the actuality of this level of underspeci®ed
semantic representations without subscribing to semantic retreat. There
are two reasons to resist such a limited scope for linguistic semantics. The
®rst is conservative: what little we know about the properties of semantic
representations is based on their abilities to support logical inferenceÐ
indeed without taking inferential properties into account, there are more
or less no known constraints on semantic representations. Underspeci®ed
semantic representations do not constitute a level on which notions of
semantic inference can be based. To retain the insights of the long tradi-
tion of research about semantic inference, we need to ®nd another level at
which these insights hold, and it will be a level to which pragmatic infor-
mation has contributed. Second, the position of semantic retreat presumes
242 Chapter 3

that there is nothing especially linguistic about the inferences derived from
utterances or the processes by which they are derivedÐit is just in e¨ect
thinking (or processing general conceptual structures). This view hugely
underestimates the gap between linguistic and nonlinguistic cognition:
linguistic representations (whether pragmatically enriched or not) have
many special properties like linearity, generality, indexicality, and culture-
speci®c lexicons which would seem to be quite remote from a ``language
of thought'' (Levinson 1997). There is every reason then to try and
reconstrue the interaction between semantics and pragmatics as the inti-
mate interlocking of distinct processes, rather than, as traditionally, in
terms of the output of the one being the input to the other.

3.5.1.4 Enlarged Indexicality This approach, typi®ed by Situation

Semantics but attractive also to more conservative logical models of
semantics, seeks to treat contextual dependence as limited to the resolu-
tion of preexisting variables, built in to the semantics of speci®c expres-
sions (``contextuals'' in the terminology suggested above). The radical
pragmatics approach of generality-plus-implicatures (i.e., lexical or con-
structional weak univocality) is often in direct opposition to this indexical
approach to the analysis of particular constructions and lexical items.
Sophisticated tests are needed to distinguish between such implicit vari-
ables (®xed either by the utterance situation or resource situations) and
generality (narrowed by implicatures). The study of indexicality by EncË
(1981) provides some tools for this. For example, once indexicals are
®xed, they are ``rigid'' under negation, intensional contexts, and ana-
phora; if the enlarged class of contextuals also must have this property
(which I think it must), then we have a whole battery of tests for
contextuals versus generality-plus-implicature.52 With such tests, we can
expect the phenomena to divide: we can expect some cases to be better
analyzed one way and some the other. For example, it is conceivable
that the narrowing of degree predicates or the asymmetric temporality
of and may ultimately be better analyzed in this way than as I-induced
generality-narrowings (although the existing tests strongly suggest other-
wise). But we can be assured, I think, that GCI intrusion as a whole and
the intrusion of general background knowledge as explored in AI local
pragmatics cannot be handled by this enlarged indexicality approach.
There is another approach that might also come under the rubric of
``enlarged indexicality''Ðnamely, Sag's (1981) proposal that we treat
lexical and phrasal categories as generally capable of behaving quasi-
243 Implicature and the Semantics/Pragmatics Interface

indexically (see also EncË 1981). He suggests that Kaplan's machinery for
handling indexicals can be naturally extended to deal with many kinds of
pragmatically determined shifts of intension and extension. I discuss (and
reject) the applicability of this approach to GCI intrusion in section 3.5.3
below.

3.5.1.5 ``Pragmantics'' As noted previously, some may take pragmatic

intrusion as evidence against the possibility of any semantics/pragmatics
boundary, and positive evidence for what I have called pragmantics, a
horrid cocktail of semantics and pragmatics, popular in the days of Gen-
erative Semantics (Lako¨ 1972) and still popular with many Cognitive
Linguists who see no role for a semantics/pragmatics distinction (Lan-
gacker 1987: 154; Jackendo¨ 1983: 105±106, 1992a: 32). This is of course
not the conclusion I wish to draw, and although such a view may seem
consistent with the intrusion data, it doesn't follow from any of the evi-
dence we have collected. What is lost on such a view is the crucial dis-
tinction between semantic entailment and pragmatic defeasible inference,
which itself is part of the critical data.
Pragmantics can be avoided if we can ®nd a way of accounting for
pragmatic intrusion into truth conditions while maintaining the modu-
larity of a distinct pragmatics (built on nonmonotonic principles) and
semantics (built on monotonic principles). Conceptually this is not as
di½cult as it seems at ®rst sight. But we might ask for moreÐnamely,
evidence that by maintaining modularity, and by de®ning exactly how the
semantics and pragmatics interact, we can predict exactly how, when, and
why pragmatic intrusions take place; and especially if we can show that
intrusive constructions have the properties they do just because of a prin-
cipled interaction between two distinct components. It is to this we must
now turn.

3.5.2 Modularity and Control53

Linguistic theory has, through its presiding genius, come to have a par-
ticular architecture. A theory, in this tradition, consists in a set of com-
ponents (A, B, C, etc.), each of which can be thought of as a function that
relates an input to an output. Component A can thus be thought of as a
function A that takes an a and gives us a b as value; component B as a
function that takes a b as argument and gives a c as value; component C
as a function that takes c as an argument and gives us d as value; and
so on. (Thus A might be syntax issuing surface structures, B might be
244 Chapter 3

phonology taking surface structures into phonetic representations, etc.)

Within components we can have levels that meet the same requirements.
Then the overall structure of the theory is given to us by the nature of its
parts: the output of A feeds B, whose output goes to C, and so forth. The
components or modules (in a non-Fodorean sense) themselves de®ne a
logical order of the passing of ``control'' (to use programming parlance).
This Chomskyan picture is not though in any way inevitable or intrinsic
to the subject matter. It is elegant precisely because it frees us from the
need to have a theory of ``control''Ðan independent theory that speci®es
how components may interact. But if pragmatic intrusion is indeed
established, this picture cannot easily be maintained (at least for the rela-
tion between the components or modules that do semantic interpretation
and pragmatic processing). The obvious proposal is to change the unidi-
rectional monologue between components into a bidirectional dialogueÐ
to let the components talk more freely to one another. For example, se-
mantic processing could proceed up to some point, pragmatic processing
then be called in the way that a subroutine can be called in a program,
and semantic processing can then proceed further. The result: the e¨ects
of pragmatic intrusion. This picture is decisively not the same as the
pragmantics cocktail rejected above. The reason is that each component
continues to be built, exactly as on current assumptions, as a homoge-
neous system of rules or principles of an essentially di¨erent kind from
those in other components (that's what justi®es the split into compo-
nents). Thus semantics might continue to operate on the basis of logical
principles, whereas pragmatics might operate on an system of abduction
and default rules. (This methodological justi®cation of ``modularity'' has
nothing to do with the psychological justi®cations of Fodorean ``vertical''
modularity, or faculty psychology; it is a method of making a simpler
theory by a policy of divide and conquer, in the ultimate hope of course
that a simple theory of the mind is more likely to be correct than an
unmotivatedly complex one.)
But once we have abandoned the Chomskyan architecture for a theory
of meaning, now we do indeed have a new task: namely, a theory of
modular control. Indeed, much of the theoretical interest in a theory of
meaning now moves to that theory. What might it look like? Let us brie¯y
explore some ways in which we can get some constraints back in by con-
sidering two di¨erent kinds of existing proposal for incorporating prag-
matic information in semantic representations.
245 Implicature and the Semantics/Pragmatics Interface

3.5.3 Sag's Proposal and Possible Ampli®cations

First, let us consider a very tightly constrained theory of the control be-
tween the semantics and pragmatics components. Suppose that each con-
stituent of a sentence is processed from the bottom up. Suppose too that
for each constituent we retain semantic priorityÐthat is, for each syntac-
tic unit, the semantics is called ®rst, the pragmatics second. Thus for the
sentence The man with two children is a rocket alarm specialist, we might
start with the semantic interpretation of lexical items and begin to build
them into the interpretation of phrases. Whenever we process a lexical
item, we check in a semantic network or lexicon for possible stronger
scalar items; if we ®nd one, we call the pragmatics to calculate the scalar
implicatures, which then get combined with the lexical content, which
then contributes to the interpretation of the phrase. Similarly, whenever
we hit a construction like nominal compounding as in rocket alarm, the
pragmatics will be called to calculate an I-implicature to the most proba-
ble inference. Thus on this account, the semantic processor or the control
structure must locate triggers that invoke the pragmatics, and the seman-
tics also has the overall responsibility, as it were, of constructing the
framework for the overall interpretation, using the compositional mecha-
nisms for building intensions and extensions as classically assumed in
formal semantics. On this view, which I take it is close to, or at least
consistent with, Sag 1981, the main change from classical assumptions is
simply that intensions (and correspondingly extensions) may be partially
determined by pragmatic information.
Let us brie¯y review Sag's original proposal. He considered cases orig-
inally brought up by Nunberg (1978) where metonymic transfers of sense
and reference are involved (on (55g) see Fauconnier 1985, Jackendo¨
1992b).
(55) a. ``The ham sandwich is getting restless.''
b. ``Every ham sandwich at table 9 is still waiting.''
c. ``I like this painter'' [pointing at a picture]
d. ``The Financial Times is saying sell IBM and buy Apple.''
e. ``Stanford swims to win.''
f. ``Chomsky takes up three feet of my bookshelf.''
g. ``Norman Mailer likes to read himself.''
Sag argues that rather than treat these as postsemantic Gricean infer-
ences from absurdities, we should consider ways of allowing denotations
to be shifted within the semantics. For example, we can use the Kaplan
246 Chapter 3

Constant character Variable character

Constant content

Variable content

Figure 3.4
The possible place of pragmatic reconstruals in Kaplan's scheme

distinction between character and content (intension), where a Kaplan

character is a function from contextual indices to intensions or contents,
which are in turn thought of in the usual way as functions from world/
time pairs to extensions. As EncË (1981) points out, Kaplan's scheme gives
us a typology as in Figure 3.4. Sag's point is that expressions used with
Nunbergian metonymic shifts could have both variable character and
variable content (a cell unused in Kaplan's scheme), so that context could
®x the intension (from, e.g., ham sandwich to `ham sandwich orderer'),
and the intension could then pick out the relevant denotation in the rele-
vant world.
That way we preserve the compositional e¨ect, whereby having
``coerced'' (in the parlance of arti®cial intelligence) ham sandwich into
`ham sandwich orderer' in example (55a), we get the normal composi-
tional e¨ects of preposing the or every (as in (55b)). The phrase ham
sandwich is just a normal N 0 with a shifted denotation; and the shifted
denotation behaves compositionally just as the denotation of any N 0 . It is
hardly surprising that the classical machinery of formal semantics could
absorb such changes in the meaning of constituentsÐafter all, it is largely
indi¨erent to lexical content, from which indi¨erence it derives the gen-
erality of the compositional mechanism. Hence augmenting or changing
the lexical content does not essentially damage the mechanism.
The question now arises, could our GCI intrusions be treated in exactly
this same way? For example, could scalar lexical items be treated as
having systematically variable Kaplan charactersÐsometimes contri-
buting implicatures, sometimes not. (Thus three could have the intension
`at least 3' or `exactly 3', as the context dictates.)
Although the Kaplan-style analysis may be ®ne for the metonymic
conversions that Sag was considering, it will be di½cult to adapt for
247 Implicature and the Semantics/Pragmatics Interface

implicatural restrictions on denotation for the simple reason that impli-

catures are subject to cancellation by systematic projection rules as shown
by (56a±d), and by contextual assumptions, as in (56e).
(56) a. ``The man with two or more children over there is the Dean.''
b. ``The man with ®ve if not six children over there . . .''
c. ``The man with two and perhaps three children . . .''
d. ``The man with two, in fact three, children . . .''
e. ``Only the single man with three children should bother to apply
for food stamps'' (attributive `at least' interpretation).
f. ``The man with three children is my brother; actually he has four
but the other is hidden behind him from this angle.''
The cancellation may be done at arbitrary distance in the discourse (as
illustrated in (56f )), so our Kaplanesque noncancelable system will be in
trouble. The problem is that once expressions plus their implicatures have
been converted into Kaplan contents (intensions) they would be just as
monotonic (nondefeasible) as entailments. What we need instead is some
model where we can keep a record of implicatural contribution so that,
unlike entailed content, it can be annulled by later parts of the same sen-
tence (or other parts of the discourse, or general background assump-
tions). This is an argument instead for the intermediate kind of semantic
representation used in, for example, Discourse Representation Theory
(Kamp and Reyle 1993); we may need to be able to get back to implicated
conditions and erase them.
Another problem is the size of constituent that is processed at a time; if
the pragmatic processing operated automatically and irrecoverably at the
lexical level it would make the wrong predictions: a GCI cannot convert
every occurrence of two into `two and no more' because immediate col-
locations may make the decision one way (as with at least two) or the
other (exactly two); and indeed while all is not an implicature inducing
item, the collocation not . . . all is (where all may be at arbitrary distance
from not but must be within its scope). But if we retreat to the next level
up, say the two-bar constituent level, collocations at the next level up may
be immediately relevant (as in a minimum sentence of two years vs. costs
two dollars), and so on, all the way up to the clause, the sentence, and
ultimately the discourse. If we hold o¨ until we have all the information
possibly required, we end up with a postsemantic pragmatics (which, we
have argued, cannot alone handle intrusion and pragmatic resolution of
reference failure).
248 Chapter 3

This is of course a problem for any suggestion for incorporating impli-

catures into truth conditions; what is di½cult is that we sometimes seem
forced to incorporate implicatures by intrasentential context (as in the
intrusive constructions), but as a general strategy need to be able to hold
o¨ commitment, or at least need to be able to cancel them. Here sugges-
tions within the Discourse Representation Theory framework by Kadmon
and others are of special interest.

3.5.4 Kadmon's DRT Proposal and Possible Extensions

3.5.4.1 Background Discourse Representation Theory (Kamp 1981,

Kamp and Reyle 1993) is one of a family of recent semantic theories that
build on the insight that utterances are interpreted incrementally in the
context of prior utterances. Although its protagonists often talk as if there
was an algorithm that converts syntactic strings into semantic repre-
sentations of a discourse (Kamp and Reyle 1993: 85), in fact contextual
parameters intervene in various ways, most obviously in resolving ana-
phora (p. 67) or specifying background information (p. 85). Thus DRT is
in fact a hybrid pragmatico-semantics, where truth conditions are stated
not directly on representations derived solely from syntactic structures
but on a level of representation that is only partially determined by the
syntactic structure of a single sentence. To start with, a DRS (discourse
representation structure) is a representation of a sequence of assetoric
sentences (a so-called discourse), so that prior sentences can o¨er ante-
cedents for anaphoric items in a sentence whose representation is under
construction as it were. Kamp's (1981) theory has nothing to say about
how choices between competing antecedents are resolved: this is clearly
a pragmatic matter except for gender constraints, or grammatical con-
straints like the Binding Conditions (Chomsky 1981; pragmatic versions
of these are described, e.g., in Reinhart 1983 and explored in chapter 4).
But a DRS records a choiceÐand in that way incorporates pragmatic
resolution.
Heim (1982, 1983a) independently developed a similar system, File
Change Semantics, which contains more explicit suggestions about the
incorporation of pragmatic information.54 Very sketchily, here are the
central ideas. Each discourse referent introduced is assigned a new ®le
card, upon which any predications of the discourse referent are added as
conditions. A ®le is a set of ®le cards, and it represents the state of the
discourse at any point; each ®le can be assigned truth conditions. The
249 Implicature and the Semantics/Pragmatics Interface

content of an utterance amounts to the ®le-updating it achieves; sentences

(or the logical forms associated with them) do not have truth condi-
tionsÐtheir truth-conditional content is captured only by their potential
to update ®les. De®nite NPs carry felicity conditions, namely that a ®le
card for their referents already exists (this captures the existence pre-
suppositions of de®nites). A corollary of the claim that a de®nite can only
be used felicitously if it is familiar, is that de®nites licensed by common
knowledge (the President) or by the visual ®eld (that cat over there) must
be admitted into the ®le (Heim 1982: 309¨ ). And to deal with the use of
de®nites not overtly licensed in any of these ways (as in I am sorry I'm
late; the car wouldn't start), Heim has to appeal to accommodation: they
are processed as if they had a preexisting ®le (1982: 370¨ ). These cases
include ``bridging'' de®nites. Thus, pragmatic processes like bridging
inferences (I-implicatures in our scheme) are incorporated in ®les, and it is
upon ®les that truth conditions are stated. Thus pragmatic intrusion is
explicitly accepted in this model.55 Because de®nite-description felicity
conditions behave like classical presuppositions with regards to pre-
suppositional projection, Heim has to relativize the existence implications
to particular constructions: in e¨ect, temporary (``local'') ®le cards are
introduced in some cases to handle presuppositional denials (The king
of France didn't come; there isn't one; but see Heim 1983b). Both DRT
and File Change Semantics thus have an explicit representation of ``the
common ground''; but unlike Stalnaker's (1972) original proposal, the
common ground contains not only a set of propositions but also a set
of designated variables, discourse referents.

3.5.4.2 Kadmon's Proposal Kadmon (1987) however takes the matter

of pragmatic intrusion somewhat further. She takes the DRS to be a
complete record of the common ground including all that is implicated.
She assumes that implicatures are generated and canceled by a mecha-
nism separate from DRS processing, but that the content of implicatures
are recorded in just the same way as entailed content, but in a designated
``color'' as it were (my metaphor rather than hers), so that they are
marked as cancelable.56
Let us brie¯y review her suggestions about scalar implicatures. She
suggests that inde®nites with numeral determiners and common nouns
(n CNs) like three boys in Three boys ran away should have the same form
of semantic content as a boy in the Heim/Kamp proposalsÐthat is, the
phrase should introduce a new discourse referent with the conditions of
250 Chapter 3

being a set of boys of cardinality 3. But because discourse referents in

general only have to have their conditions satis®ed by some entity in the
model,57 it does not follow that there are no other boys. Thus the semantic
content amounts to that assumed in standard pragmatics, namely `at least
three boys' (Kadmon defends this position against Kamp's `exactly three'
proposal).58
Kadmon is particularly concerned with arguing, against Heim's con-
ditions on de®nites based exclusively on familiarity, for a resuscitation of
the uniqueness condition. Accepting her argument for current purposes, it
follows that anaphoric de®nites should also require unique antecedents,
and thus that n CN inde®nites (e.g., three dogs) should only be possible
antecedents if the `at most' implicature has strengthened them. Her
examples are telling (p. 82¨ ), but I try some of mine:
(57) a. A: ``The toll plaza is up ahead. Do you have three dollars?''
B: ``Sure, I've got over ten.''
A: ``Okay, have it handy.''
b. ``Three men walked in. They were a rough lot. ??A fourth was
wearing a beret.''
c. ``At least three men walked in. They were a rough lot.''
Here in (57a) I suggest that, despite a preference for local coreference, it
refers most naturally to the exactly three dollars that we assume is
required; above all, it could not refer to the over ten. It might refer to
another unique entity, such as the wallet containing the collection of more
than ten dollars; or even just to that collection in its entirety, however big
it is, but not just to the ten that were included in it. Example (57b) illus-
trates that anaphora seems to force the scalar implicature (which being a
GCI would of course tend to arise anyway); whereas the interpretation of
they in (57c) must apply to the total collection, not the three mentioned.
Kadmon gives some indication of the role of implicature cancellation in
this account. Consider the following example (Kadmon 1987: 117, 167f,
172¨ ):
(58) ``I have to show this document to three colleagues. They are in a
meeting right now. I also have to show it to two other colleagues, but
they have already left.''
Here the ®rst they will force a scalar-implicature interpretation of three as
`exactly three'. The admission in the third sentence that there are two
other relevant colleagues will cancel this implicature. Therefore the felic-
251 Implicature and the Semantics/Pragmatics Interface

ity condition on they, which requires a unique maximal set as antecedent,

now needs to be satis®ed some other way. Kadmon suggests (p. 174) that
an antecedent is accommodatedÐnamely, the set of three colleagues in a
meeting (rather than, as prior to the cancellation, the maximal set of three
colleagues who must be shown the document).
To Kadmon's arguments for pragmatic intrusion of scalar implicatures
into anaphoric reference, we may add examples of di¨erent kinds. Con-
sider for example the following (from what has been said above, readers
may proliferate examples for themselves):
(59) a. ``Some of the students passed the exam. They were the lucky
ones. The others have to retake it.''
b. ``If the Princess had a baby and got married, then doing so is
clearly now more fashionable than getting married and having a
baby.''
c. ``If some of the students passed the exam, then that is not as good
as all of them passing the exam.''
Thus in (59a) the truth conditions for They were the lucky ones must
identify they as referring to a proper subset of the students (as required by
the scalar implicature `some but not all'); otherwise the others will be
unde®ned. In (59b) the VP anaphora clearly picks up the I-implicature
from and to `and then', whereas (59c) shows that sentential anaphora with
that must be construed as picking up the implicatures of the antecedent
sentence.
DRT is speci®cally set up to capture anaphoric relations, so if Kadmon
is correct about these observations, then DRSs not only may be thus
enriched with scalar implicatures, they must be if the original motivation
is to be met. Thus there is a strong argument from anaphora for the
intrusion of GCIs into truth conditions, and DRT seems to be the theory
best able to handle this intrusion.

3.5.5 Some Future Directions: DRT and Intrusive Constructions

This way of representing pragmatic intrusion seems extremely promis-
ing.59 But DRT would be especially attractive if it provided predictions
about just which constructions would naturally be interpreted with intru-
sive implicatures. In fact it does indeed seem to make some predictions in
the right direction. I discuss here just two problems: the intrusion in con-
ditionals and in metalinguistic negation.
252 Chapter 3

Most obviously, the conditional intrusions may have a natural account

in terms of the sequence of interpretation. Thus Kamp's (1981) scheme
for the truth conditions of ``If A then B'' (where m(A) is the DRS for A
and m(B) that for B) reads thus: the conditional If A then B will be true i¨
every proper embedding of m(A) can be extended to a proper embedding
for m(B). Thus A is assessed prior to B, and B is understood as a condi-
tional assertion dependent on A: hence the intrusions of the implicatures
of A into the conditions under which one can expect the consequent to
hold.
Now, Heim (1983b) has already noted that this treatment of the con-
ditional gives a natural account of presupposition projection: the sequen-
tial updating of the context, ®rst by A, and then by B, predicts that where
B has a presupposition entailed by A, the presupposition is itself condi-
tional as it were. Because presupposition projection is intrinsically linked
to negation, before turning to see how DRT might give a natural account
of the metalinguistic negation cases (i.e., intrusion in negative sentences),
it will be helpful to examine brie¯y the problem of implicature projection.
Indeed the viability of the DRT approach to pragmatic intrusion is likely
to depend on the projection properties of conversational implicatures,
which have hardly been studied and about which there are currently
erroneous assumptions (owing to under-critical acceptance of Gazdar
1979).
A ®rst point to note is that the projection properties of scalar Q-
implicatures look surprisingly similar to those of classical presuppositions
(e.g., factives, de®nite descriptions), as pointed out by Levinson (1983:
223±225). Compare, for example, the lifting of a scalar implicature (here
`no more than three') to the suspension of the factive presupposition due
to regret (namely, that its complement is true) in the following contexts:
(60) a. ``John has three children, if indeed not more.''
a 0. ``John doesn't regret doing it, if indeed he did it.''
b. ``Either John has three children or he has more.''
b 0. ``Either John doesn't regret doing it or he didn't do it.''
c. ``John doesn't have three children.''
c 0. ``John doesn't regret doing it.''
Thus the ``®ltering'' in conditionals and disjunctions looks exactly parallel,
suggesting that whatever the correct treatment is for presuppositions, it
should carry over to implicatures. Where the di¨erences emerge is under
negation: example (60c) does not share with its positive counterpart the
253 Implicature and the Semantics/Pragmatics Interface

implicature `John has at most three children', whereas (60c 0 ) does indeed
(by de®nition of course) share the presupposition `John did it' with its
positive counterpart. Because there is a natural account of at least some
aspects of the presuppositional projection problem in DRT and File
Change Semantics (see Heim 1983b, Landman 1986), there is every rea-
son to pursue this parallelism and see whether we can't ®nd an indepen-
dent explanation for the failure of implicatures under negation.
With this background, let us now turn back to the other kind of intru-
sive constructionÐnamely, the metalinguistic-negation intrusions, as in:
(61) a. ``John doesn't have three children, he has four.''
b. ``John has three children.''
c. NOT (John has at least three children & John has at most three
children).
The intuition is that the kind of sentence in (61a) could only be used
where it is supposed in the context that John has three children. One way
of construing this is that an assertion has been previously made in the
discourse, viz. (61b), ``John has three children''. In that case, the scalar
implicature `at most three' would have been added distinctively `in red
ink' into the DRS. (As suggested before, let us imagine that semantic
conditions are written into the DRS with black ink, but defeasible prag-
matic conditions with red ink, so that they can be canceled at arbitrary
distance.)
Against this background, (61a) is construed as a denial of the assertion
of (61b). In the DRT framework there is a natural way to represent that
denial, namely by treating example (61a) as quasi-anaphoric to the pre-
ceding positive assertion in (61b) so that all the semantic conditions of
(61b) are copied over into the DR of (61a).60 Thus (61a) and (61b) have
identical representations except that (61a) has an external negation that
ranges over both the entailed content (black ink) and the implicated (red
ink) as informally represented in (61c) as italics. The overall DRS already
contains entailed commitment to John having at least three children (due
to the prior assertion of (61b)), so only the implicated commitment in
(61c) (carried over from (61b)) is cancellable. Hence the metalinguistic
negation interpretation of (61a) follows naturally, and there is no contra-
diction (at the level of entailments).
The elegance of the DRT analysis may not immediately strike the
reader. However, it combines the best parts and avoids the worst parts of
two incompatible accounts of metalinguistic negations of implicature,
254 Chapter 3

those by Horn (1985) and Kempson (1986). On Horn's account the

negation in these cases is not logical (not reducible to a truth-functional
one-place connective with the familiar truth table), but ``voices an objec-
tion to a previous utterance (not proposition) on any grounds whatever.''
I take this to be a bad part of the analysis, because the negation is in fact
in this case behaving exactly like the normal truth-functional operatorÐit
simply ranges over di¨erent (nonentailed) content. The correct part of
Horn's analysis is that these utterances do indeed ``voice an objection to a
previous utterance'' (i.e., to the utterance-meaning, replete with implica-
tures), and not just to the sentence-meaning, of some previous utterance in
the discourse. It is because of that quasi-anaphoric relation between the prior
positive sentence (61b) and its later denial in (61a) that these implicature-
denying negations share certain properties with nonlogical metalinguistic
negations (as in ``It's not a [vA:z], it's a [veõz]''). As mentioned, meta-
linguistic negations in general do not permit negative polarity items nor
the incorporation of negatives, because they echo the utterances to which
they are retorts (Horn 1989: 392±402; see also Carston 1996). Horn mis-
takenly assimilates implicature-denying negations to the nonlogical class
of metalinguistic negations because of these shared properties, whereas in
fact the properties follow from the shared function as retorts.61
Turning now to Kempson's account, the good part is the insistence that
these negations just are our familiar one-place truth function. The bad
part is the assumption that negation could ordinarily range over explica-
tures. The only case in which implicatures come within the scope of
negation is where the negative sentence is a denial of a prior assertion (as
Horn insists). The reason that this must be so is that the implicatures in
question would never arise directly under negation: the implicatures being
denied belong to the prior positive assertion.
Let us expound this point carefully (at the risk of redundancy with the
exposition in chapter 2). According to the pragmatic theory developed
here so far, for each and every positive Q-scale of the form hS, Wi where
S is a strong scalar item entailing (in an arbitrary simplex sentence frame)
the weak scalar item W (in the same sentence frame), there is a cor-
responding negative scale hnot-W, not-Si where `not-S' (or `not-W ') is
a shorthand for S (or W ) occurring within the scope of negation, and
not-W entails not-S (in the same sentence frame).62 Thus for example,
given the positive scale hall, somei, there is a corresponding negative scale
hnot-some (i.e., none), not-alli, and the following kinds of implicature are
therefore predictedÐwhere only the weak items induce scalar implicatures:
255 Implicature and the Semantics/Pragmatics Interface

(62) a. ``Some of the men came'' ‡> `Not all of the men came.'
b. ``All of the men came'' (no scalar implicature).
c. ``Not all of the men came'' ‡> `Some of the men came'.
d. ``None of the men came'' (no implicature).
If we now return to the relevant example, repeated in (63a), we can see
that where the plain negative (63d) is asserted, and asserted without the
background of a prior assertion of the positive (63b), we do not have the
positive scale h . . . ; four, three, two, onei to induce the implicature `not
four'. Indeed `not four' is entailed by the assertion not three (which means
`not three or more'). Instead, we have a negative scale running the other
way: hnot-one, not-two, not-threei. The prediction is that when you say
(63d), you implicate that John has at least one child and you assert
(entail) that he has less than three.63
(63) a. ``John doesn't have three children, he has four.''
b. ``John has three children.''
c. NOT [John has at least three children & John has at most three
children].
d. ``John doesn't have three children.''
Now we see also the answer to the one apparent di¨erence between the
projection properties of presuppositions and implicatures. According to
Horn (1972, 1985, 1989), Gazdar (1979), and Hirschberg (1985), scalar
implicatures are simply blocked (i.e., do not arise) under negation. This is
an erroneous assumption based on the correct observation that the same
implicatures are not shared by positive and negative counterparts of the
same sentence (utterance-type). As we have seen, the apparent blockage
is due to the fact that negatives reverse scales (because negation e¨ects
the strength of scalar items) and so we get di¨erent implicatures, which
themselves survive negation just as presuppositions do. This opens the
way for a fully general account of presupposition and implicature projec-
tionÐand if Heim (1983b) is right that File Change/DRT±type semantics
gives a natural treatment of the former, it should also give a natural
treatment of the latter.
To return to the claims about metalinguistic negation: the DRT treat-
ment can now be appreciated as unifying prior accounts. The negation is
ordinary negation (Kempson), the utterance is a denial of a previous
positive assertion (Horn), and the implicature that is canceled by the
negation is not an implicature of the negative sentence itself (contra
256 Chapter 3

Kempson), it is the implicature of the prior positive assertion. All of this is

obtained by existing DRT machinery.
This kind of result lends substantial prima facie support to the DRT
treatment and requires further exploration. Apart from getting the can-
cellation aspects of the projection problem for implicatures right, the
DRT treatment must also be able to explain the fact that just like some
presuppositions, implicatures pass through ``holes'' that would block
entailments. Thus in nonentailed clauses, like the antecedent and con-
sequent of conditionals, both implicatures and presuppositions can get
entered permanently into the DRS as if asserted, even though neither the
antecedent nor the consequent leave unconditional information.
(64) a. ``If John has any conscience, he'll regret he did it.''
Presupposes: `John did it'.
b. ``If John has a low income, his three children will su¨er.''
Implicates: `John has three and no more than three children'.
Just as presupposition projection has been seen to be one of the cardinal
attractions of DRT, so implicature projection may be an interesting test
ground for that theory and its ilk.

3.5.6 A Residual Problem: How to Get from Semantic Representations to

Propositions
On all theories of semantics, pragmatic resolutionÐat least the ®xing of
indexical reference, and disambiguationÐis required before a semantic
representation expresses a proposition or describes a situation. In the
newer semantic frameworks, like DRT and Situation Semantics, the
range of kinds of pragmatic resolution is enlarged: anaphora resolution,
the ®lling in of contextual parameters and ceteris paribus conditions, all
need to be taken into account. None of these models sketch any mecha-
nism for pragmatic resolution; indeed they are often misleadingly phrased
in terms of an algorithm deriving semantic representations from syntactic
representations.
But the new Gricean pragmatic theories appear to be in no better
shape. In fact Radical Pragmatics programs, which reassign much that
had been thought semantic to pragmatic resolution (Cole 1981: xi), only
serve to increase the essential problem. Atlas (1989) persuasively argues
for the Strawsonian dictum that sentences express no propositionsÐonly
statements, or pragmatically interpreted utterances, can do so. The same
sentiments have been expressed and argued for by the London School
257 Implicature and the Semantics/Pragmatics Interface

(Kempson 1986, Sperber and Wilson 1986, Carston 1988). For these
theories the essential problem, quite unacknowledged by the way, is this:
a Gricean mechanism is supposed to get us from a level of semantic repre-
sentation, now impoverished to a subpropositional schema of some sort,
to a representation capable of specifying propositional content, a level of
logical form as it were. But Gricean mechanisms are reasoning mecha-
nisms, and they presuppose propositional input as well as propositional
output. In this respect, owing to its explicit commitment to deductive
reasoning, Relevance Theory appears especially incoherent: ``no assump-
tion is simply decoded [from an utterance], and . . . the recovery of any
assumption requires an element of inference'' (Sperber and Wilson 1986:
182). But how are we to get from nonlogical forms to further contextual
implications that will enrich them by a deductive process that can only
handle logical forms? Theories that are abductive or inductive in charac-
ter, or that are phrased in terms of an inference to the best explanation,
may not face this theoretical incoherence, but of course that does not
erase the computational problem. Thus Atlas (1989: 146) states ``Gricean
inference is not restricted to mapping propositions (`what is said') into
propositions (`what is implicated'). It also maps Lockean, `abstract'
sentence-meanings into propositions (utterance-meanings).'' He goes on
to point out that the intrusion phenomena cataloged here (or rather in
Levinson 1988b) are evidence for precisely this claim.
What kind of a computation is it that would take us from under-
speci®ed semantic representations to logical forms capable of expressing
propositions? Much depends on the form of the input semantic represen-
tation, and here Radical Pragmaticists have o¨ered us only the sketchiest
of ideas. We might think in terms of logical formulae with unbound
variables, or semantic representations with contextual parameters to be
instantiated. But clearly, Radical Pragmaticists of all persuasions have in
mind representations that are much further divorced from logical forms
than this. Even Fodor (1983: 135 n. 29, cited in Atlas 1989: 148) concedes
that scope may be indeterminate at this level and that ``perhaps the spe-
ci®cally linguistic processes in the production/perception of speech deploy
representations that are shallower than logical forms.'' Atlas (1989: 146)
views the indeterminacy in terms of semantic generality and o¨ers various
analogies, including the archi-phoneme. Sperber and Wilson (1986: 181)
talk of the ``development'' of a logical form ``by using contextual infor-
mation to complete and enrich this logical form into a propositional
form.'' The notion of explicature, or Bach's notion of impliciture, might
258 Chapter 3

have been employed to denote the inferences that map a semantic repre-
sentation into a proposition, but in fact both Sperber and Wilson's expli-
cature (and Carston's (1988) or Recanati's (1989) versions) and Bach's
impliciture take us well beyond the minimal proposition expressed, closer
to some intuitive notion of `what is said' as what is e¨ectively stated.
The notions developed in this book have some bearing on these issues,
which constitute it seems to me by far the most pressing problem in an
overall theory of meaning. First, we have presumed that an ampliative
reasoning system must be involved (i.e., a kind of reasoning that, unlike
deduction, expands the information given). Just as we may from a few
shards guess the outlines of an ancient pot, so we seem to make the
ampliative enrichment from a subpropositional to a propositional repre-
sentation. Deduction can only map determinate logical forms into logical
forms. But abduction and induction can map any data format into prop-
ositions (e.g., shards into the date they were made). Second, the theory of
GCIs o¨ers us a glimpse of how we may constrain this inference to the
best interpretation (Atlas and Levinson 1981), namely by using heuristics
to warrant presumptive solutions.
Let me develop this last point in a little more detail. It has long been
noted that humans can coordinate on a single joint solution where the
problem involved in fact o¨ers no unique solution, and they can do this in
the absence of communication. For example, two people told that they
will each get $100 if they think of the same number will beat the in®nite
odds and think of the same number, often enough at least to make this an
expensive experiment (Schelling 1960). Humans solve these problems by
considering what each would think the other would think that each would
think was the most salient solution. On the Gricean account of commu-
nication (Grice 1957), this is exactly the intentional context in which an
utterance is o¨ered for comprehensionÐas a clue to a salient solution to a
coordination problem (see Clark 1996: chap. 3). The mutual knowledge
that communicator and recipient are mutually oriented to heuristics of the
kind proposed hereÐthat is, the Q-, M-, and I-principlesÐprovide extra,
highly salient constraints to the solution of the problem. The problem of
course is: how should the speaker frame his utterance so that the propo-
sition the recipient attributes to it has exactly the properties that the
speaker intended it to have? Our heuristics recommend to the speaker the
selection between forms that might invoke the relevant relations, so that,
for example, minimal forms will pick up stereotypical or otherwise more
speci®c interpretations, maximal forms will discourage this, and forms
259 Implicature and the Semantics/Pragmatics Interface

that are weaker than others on a scale of informativeness will implicate

that the stronger forms may not hold.
The Gricean intentional umbrella, coupled with the Schelling saliency
principle, provides inferences arising under our heuristics with their de-
fault, presumptive character. Now, the importance of GCIs is that they
o¨er some partial solution for how to get from fragmentary semantic
representations to fulsome, proposition-bearing representations or logi-
cal forms. Because they are presumptive, and because they depend in the
®rst instance on the recovery of the meaning and form of the particular
expressions that give rise to them, the inference can be made without
complete access to the logical form of the whole utterance. This has both
procedural and psychological aspects. Procedurally, the expression some
of the boys G'd can be immediately associated with the default assumption
some but not all of the boys G'd even when some indeterminate aspect of
the predicate G has not yet been resolved. Psycholinguistically, in the
stream of incoming speech during comprehension, the phrase some of the
boys can invoke the default assumption `not all of the boys' even before
the predicate has been heard. The default presumption must of course be
held separately from the other recovered content because once the predi-
cate is resolved it may be defeated by higher-ordered assumptions (e.g.,
entailments in the discourse or antecedent assumptions).
Thus the default properties of GCIs account for their intrusive abilities:
such inferences can and will be made on the ¯y and, if not canceled by the
mechanisms sketched in this book, will end up within the propositional
content of complex sentences whose parts are semantically evaluated with
respect to each other (as in comparatives and conditionals, and in the case
of denials, evaluated with respect to the corresponding a½rmative).

3.6 CONCLUSIONS

In this chapter, I have pursued an idea that at ®rst seems nonsensicalÐ

namely, the notion that implicatures could play a role in truth conditions.
I hope to have made a prima facie case for that view. Indeed, I hope to
have shown that the view is inevitable.
Our natural reluctance to accept a view that entails such far-reaching
overhaul of the current theory of meaning is unlikely to be overcome
without some prospect of some alternative, no less-coherent overall pic-
ture. I believe that a theory like DRT, which provides a level of repre-
sentation to which both pragmatic and semantic processes can contribute
260 Chapter 3

and which is the level at which semantic interpretation is de®ned, does in

fact o¨er such a picture. I will have to leave it to others with more com-
petence in that area to develop or reject these suggestions. Meanwhile, I
should record that Grice (1987: 375) claimed in his ``retrospective epi-
logue'' to have foreseen from the beginning the problems we have here
been grappling with: ``It certainly does not seem reasonable to subscribe
to an absolute ban on the possibility that an embedding locution may
govern the standard nonconventional implicatum rather than the con-
ventional import of the embedded sentence.'' He gives conditionals and
negations as examples of such embedding locutions and goes on to add:
``But where the limits of a license may lie which allows us to relate
embedding operators to the standard implicata rather than to conven-
tional meanings, I have to admit I do not know.'' Knowing the answer to
that question will be to have a new and profoundly di¨erent understand-
ing of the semantics/pragmatics interface.
Chapter 4
Grammar and Implicature: Sentential Anaphora Reexamined

4.1 GRAMMAR AND IMPLICATURE

In the previous chapter we explored the quite far-reaching consequences

that the phenomena of generalized conversational implicature may have
for our conception of the semantics/pragmatics interface. In this chapter,
we shall explore another interaction, namely the relation between gram-
mar and pragmatics, and ask how we may be forced to reconstrue this
division of labor given the phenomena of GCIs.1 In both cases, the un-
derlying questions are the same: given the existence of preferred inter-
pretations, and as a corollary, preferred ways of putting things (i.e.,
expression types associated with preferred interpretations), aspects of in-
terpretation that we have in the past thought of as stipulated by the lexi-
con or the grammar may in fact be attributable to pragmatic principles of
preferred interpretation. This chapter focuses on just one area, anaphora,
where plausibly such misattributions have been made.
It may seem that a theory of implicature or preferred interpretation
would be remote from any theory of grammatical well-formedness. In-
deed, the relation between syntax and pragmatics is of a fundamentally
di¨erent kind than the semantics/pragmatics interface, for it is indirect. It
is only through e¨ects on meaning, broadly construed, that implicature
can impose distributional or collocational constraints. But we should
remind ourselves that distinctions between syntactic, semantic, and prag-
matic phenomena are not given to us by direct intuition in advance of
theoretical development: ``On the contrary, the problem of delimitation
will clearly remain open until these ®elds are much better understood than
they are today'' (Chomsky 1965: 159). With the development of a theory
of generalized, default inference, these border discriminations become
considerably harder. For example, we may think we are blessed with
262 Chapter 4

access to speci®cally syntactic intuitions that rule out *The boys is funny;
but consider the sentence The french fries is impatient for his bill, which at
®rst seems to violate the same basic agreement rule of English, but
becomes acceptable on the construal that the french fries metonymically
refers to a singular customer who ordered french fries at a restaurant
(Nunberg 1978). The theoretical moves over such cases are not so easy,
but they do suggest that verb agreement is after all determined by some-
thing more complex and contextual than the inheritance of syntactic fea-
tures of plurality.2 Thus we cannot accept, in lieu of argument, hunches
and presumptions that such and such phenomena have an ``unmistakable
syntactic ¯avor.''3
If we are willing to keep an open mind, it does seem fairly clear that
there are intimate connections between distributional constraints and
implicature or, more generally, between utterance-form and utterance-
type meaning (i.e., the third level of meaning, concerned with favored
means for expressing certain meanings, as discussed in chapter 1). There
are connections both on the synchronic and diachronic axes. On a syn-
chronic basis, the connections are in principle straightforward: a particu-
lar utterance-type or form has a range of semantic interpretations, which
are in turn correlated with default implicatures of various types. Such
implicatures then preempt alternative expressions, which would instead
implicate something else by contrast (as in the familiar play between I-
and M-implicating expressions, e.g., in the direct vs. indirect suggestions
from kill vs. cause to die).4 Implicatures thus restrict collocations of
expressions relative to interpretations, as many syntactic principles (e.g.,
the Binding Conditions) are also held to do. They are much less likely to
yield impossible collocations on any interpretation (as other syntactic
principles, e.g., movement constraints, may do) because of their defeasible
nature (e.g., immediately, intentionally, directly collocate better with kill
than cause to die, but they can override the implicature of indirect causa-
tion associated with the periphrastic expression). Where such strong col-
locational constraints exist, they are probably due to an additional
`standardization' (Bach and Harnish 1979: 192±195; Bach 1995) or
`short-circuited implicature' (Morgan 1978, Horn and Bayer 1984), as in
the licensing by implicated Neg-Raising of negative polarity items in
positive subordinate clauses (e.g., ``I don't think he has any money''; see
Linebarger 1991).5
Once we turn to a diachronic perspective, the relation between gram-
mar and implicature can be shown to be rather fundamental. Implicated
263 Grammar and Implicature

meanings are a major source of grammatical change, as shown by the

literature on grammaticalization, which explores the diachronic processes
whereby grammatical elements and constructions are recruited from pre-
viously available expressions and constructions. For example, many
modals and auxiliaries seem to have acquired their special grammatical
status through the implicatures of their prior main-verb usages (Bybee,
Pagliuca, and Perkins 1991). Thus the grammatical reanalysis of be going
to V, from a verb of motion together with a purposive clause to a future
auxiliary, may plausibly be claimed to be based on the combined tempo-
ral implicatures of go and the purposive construction (Hopper and Trau-
gott 1993: 82¨ ). There is a main-verb reading of go in (1a) below,
followed by a purposive clause, the whole glossable as in (1b). On this (the
original) interpretation, implicatures arise that the subject will indeed
reach the destination in the future, and that he will carry out the intended
course of action. It is these strong future implications that have been
grammaticalized into a new structural pattern with a ®xed interpreta-
tionÐnamely, be going to analyzed as an auxiliary constituent with future
time reference, roughly with the gloss in (1d) (intentional implications
perhaps persisting).
(1) a. ``He is going to join the army.''
b. `He is on his way to join the army.'
c. I ‡> `at some later point he will have arrived and will have
joined the army.'
d. `He will join the army.'
Examples of such diachronic processes can be found throughout the
grammaticalization literature (see Traugott and Heine 1991; Heine,
Claudi, and HuÈnnemeyer 1991; Bybee, Pagliuca, and Perkins 1991; Hop-
per and Traugott 1993), although they are rarely analyzed strictly in neo-
Gricean terms (but see Traugott and KoÈnig 1991, Traugott 1998; see also
Levinson 1979, where a sequence from particularized through generalized
to conventional implicature was proposed). One major source for new
grammatical constructions is what we have called I-implicatures; for ex-
ample, the evolution of causal since from an earlier purely temporal con-
nective with the meaning `after' is nicely documented by Traugott and
Heine (1991). Similarly, M-implicatures seem to play a systematic role in
deriving, and rederiving, emphatic re¯exives (Langacker 1977, Faltz
1989), a point to which I will return. This cyclical process, whereby im-
plicatures become conventionalized and then acquire further implicatures,
264 Chapter 4

which can in turn conventionalize, gives us a deep diachronic layer-

ing of ``fossilized'' and partially conventionalized implicature that can be
attested in various domains, especially perhaps in areas like euphemism
and honori®cs, where there is a functional pressure on innovation (Brown
and Levinson 1987: 258¨ ). In this chapter, this diachronic layering will
play an important role in our attempt to understand crosslinguistic pat-
terns of anaphora in implicatural terms. In short, the historical evidence
points to systematic relations between grammar and implicature that
ought to be more carefully explored as an interesting test ground for dif-
ferent theories of implicature.
If we may take as granted that grammaticalization often involves the
codi®cation of a new construction out of an expression plus its gen-
eralized implicatures, then there are some general consequences for the
analysis of grammar:
1. That a pattern of interpretation is now rule-governed and fully gram-
maticalized does not preclude the relevance of a pragmatic analysis as
an hypothesis about diachronic origins.
2. We might have very similar patterns in two languages, but in one lan-
guage ®nd the pattern grammaticalized and in the other merely a pre-
ferred pattern of interpretation.
3. We might ®nd that such a match between a grammatical pattern and
its implicatural origins plays a role in language learning and language
change through reanalysis: what adults treat as implicature-based,
children may treat as rule-based (and vice versa).
4. We might catch the process halfway, as it were, and ®nd rule-governed
aspects and implicatural fringes to a construction-in-the-making.
A useful symptom of grammaticalization in progress is the existence of
morphosyntactic re¯exes of an inference, as explored in the literature on
``short-circuited'' implicature (see Horn and Bayer 1984) . Thus we must
embrace the possibility of pragmatic explanation without pragmatic
reduction.
We may ®nd ourselves genuinely in doubt about whether we are deal-
ing with what is fundamentally a grammatical or a pragmatic pattern.
Where we have rival accounts of the same phenomenon in terms of
grammatical principles or in terms of pragmatic principles, both kinds of
account could have substance. But in a number of fundamental ways,
they may be incompatible:
265 Grammar and Implicature

1. Insofar as the grammatical account is supposed to follow from uni-

versal principles that are too abstract to be learned, the existence of a
pragmatic account in terms of general principles of usage under-
mines the alleged unlearnability essential to this type of grammatical
hypothesis.
2. And insofar as the constructional meanings are indefeasibleÐthat is,
the interpretations are in¯exible and can be speci®ed by exceptionless
rule, we may con®dently attribute the interpretation to a grammatical
source; but insofar as they are defeasible and show all the hallmarks of
nonmonotonic inference that we associate with pragmatic inference, we
should attribute the pattern of preferred interpretation to pragmatics.6
In this chapter, we shall focus on the pragmatics of intrasentential
anaphora. This choice is not fortuitous. Anaphora has been the focus of a
tremendous amount of work over 30 years of the generative tradition, and
especially within the Government and Binding (GB) or Principles and
Parameters framework (Chomsky 1981, 1986, 1995: chap. 1), and thus
also within its direct competitors. In the GB framework, the Binding
Conditions governing the interpretations of di¨erent kinds of overt refer-
ring expressions were also intended to hold for covert nominals (NP
traces, pro, PRO, and variables). Thus the Binding Conditions had a
central position within syntactic theory, e¨ectively controlling both ana-
phora and movement. Since then there has been a tendency to handle
movement constraints in other ways (by chain theory) and restrict the
Binding Principles to the handling of anaphora. Still, in Chomsky's
Minimalist Program, the Binding Conditions still play a crucial role, and
remain essentially the same, except that they are recast as interpretive
conditions holding at the level of LF (Chomsky 1995: 211). Intrasenten-
tial anaphora thus continues to be a focus of current work within the
generative framework (see, e.g., Reinhart and Reuland 1993, Koster
and Reuland 1991, Hornstein 1995, Harbert 1995). Here then are a series
of closely argued proposals about syntactic constraints on the interpreta-
tion of anaphoric expressions. If we could successfully show that at least
some of these constraints are in fact not syntactic in origin, but follow
from patterns of preferred interpretation, it would amount to a substan-
tial redistribution of the division of labor between grammar and prag-
matics. I shall make tentative proposals along these lines, trying to
attribute to our existing apparatus of inferential heuristics some of the
patterns of interpretation that have been thought to be syntactically
speci®ed. Again, this is the classical Gricean strategy, now applied to the
266 Chapter 4

syntax/pragmatics relation rather than the semantics/pragmatics one:

suspect any alleged grammatical stipulation that ®ts a Gricean pattern to
be a pattern of preferred interpretation (possibly under varying degrees of
grammaticalization).
Such attempts at pragmatic reduction can be misunderstood, and we
shall need to steer away from possible misinterpretations of the goal of
the enterprise. But brie¯y, the goal has nothing to do with wholesale
functional reductionism. It has rather to do with introducing an addi-
tional toolÐa rival line of explanation that has been grossly under-
estimated, largely because pragmatic inference has been thought to be
largely contextual and thus not systematic. Once the very notion of a
preferred interpretation based on highly general principles is accepted, it
opens up the possibility of a radical redivision of labor between grammar
and pragmatics. The possibility ought to be welcomed by grammarians,
because once recalcitrant phenomena with the hallmarks of preferred
defeasible interpretations are removed from what is to be explained, much
cleaner syntactic theories may become available. Because GCI theory is in
its infancy, worked on by only a handful of scholars, we cannot expect
these ®rst attempts to handle all of the complexities. If ®rst indications are
even at all promising, however, it ought to indicate that more work
should be diverted in this direction.
Often, the reactions of syntacticians to such attempts to renegotiate the
division of labor are of the form ``a miss is as good as a mile,'' with the
provision of a few apparent counterexamples; but all syntactic theories, in
the domain of anaphora at least, are also replete with counterexamples
(and indeed Chomsky (1995: 19f ) himself sees the need for fairly radical
idealization). Because the pragmatic theory, in a way to be explained, is in
crucial cases essentially parasitic on the syntactic theory, the apparent
counterexamples for the pragmatic theory may actually re¯ect inadequate
syntactic analysis. Indeed, one of the special properties of pragmatic
analyses is that they are simple, direct instantiations of enormously gen-
eral, crosslinguistic tendenciesÐyou cannot tinker with them to get the
right result in a particular language in the same way that you can ®ne-
tune a syntactic analysis. In the long run, when such pragmatic principles
are better grounded, we may even be able to use the pragmatic theory
predictions to help in the ®ne-tuning of the syntactic analysis: because the
pragmatics operates on the syntactically speci®ed interpretations to yield
further preferred interpretations, if the latter are incorrectly predicted it
may indicate that the former have been misanalyzed.
267 Grammar and Implicature

4.2 IMPLICATURE AND COREFERENCE

4.2.1 The Pragmatics of Local Anaphora

Anaphora is the phenomenon whereby one linguistic expression (the
anaphor), lacking clear independent reference, can pick up reference or
interpretation through connection to another linguistic expression (usu-
ally an antecedent). (Incidentally, the special Chomskyan sense of Ana-
phorÐre¯exives, reciprocals, and their likeÐwill be distinguished by a
capital letter; the unmodi®ed term pronoun will be used in the Chomskyan
manner to exclude re¯exive pronouns.) Not every anaphoric expression
has a referential interpretation. In Everybody has his foibles, for example,
the pronoun his is usually thought to be equivalent to a logical variable,
whereas in John has his foibles, the pronoun hisÐwhen its referent is
coidenti®ed with JohnÐis usually thought to be parasitically referential,
through linkage to John. Some of these decisions are theory-dependent,
and I shall loosely talk about coreference and disjoint reference without
necessarily presuming that the anaphoric elements in question are strictly
referential. I am concerned here only with NP-anaphora, but there are
many other kinds (see Webber 1979: chap. 1 for a catalogue).
Despite the preoccupation with anaphora in the syntax and semantics
journals, there are good prima facie reasons to think that anaphora is
primarily a pragmatic matter. Anaphora is, after all, basically a cross-
sentence (``discourse'') phenomenon and thus quite largely outside the
domain of intrasentential processes. Now it would be odd if cross-
sentence patterns of interpretation were entirely unrelated to intra-
sentential patterns; and of course they aren't. Consider for example the
parallel between (2) (so-called discourse anaphora) and (3) (sentential
anaphora).
(2) a. ``John came in. He sat down.'' (preferred coreference)
b. ``John came in. The man sat down.'' (preferred disjoint
reference)
(3) a. ``John likes his father.'' (preferred coreference)
b. ``John likes the man's father.'' (disjoint reference)
There is an obvious pattern here (let us call it the general anaphora pat-
tern): reduced forms (less prolix, more semantically general) favor co-
referential readings, prolix full forms favor noncoreferential (disjoint)
readings (the pragmatic vs. grammatical status of the disjoint reading of
268 Chapter 4

(3b) will exercise us) . Now it is clear that the pattern does indeed go
across the pragmatics/grammar borderÐthat is, that what is in essence a
pragmatic preference for local coreference induced by minimal forms can
become a matter of grammatical stipulation leaving no room for other
interpretations:
(4) a. ``John came in and the man sat down.'' (disjoint reference
preferred)
b. ``John came in and he sat down.'' (coreference preferred)
c. ``John came in and sat down.'' (coreference required)
But that, as we have already noted, is part of the importance of the
pragmatic preferred interpretationsÐthat they tend to become gramma-
ticalized, providing a diachronic source for grammatical stipulations of
interpretation. Not all grammar is ``frozen pragmatics'' of course, but
some undoubtedly is. From a diachronic perspective, there is thus no
necessary incompatibility between a grammatical and a pragmatic analy-
sis. From a synchronic perspective, on the other hand, we clearly want
just one analysis. In trying to decide whether we are dealing with gram-
matical stipulation or mere preferred interpretation, we inevitably appeal
to our litmus test of defeasibilityÐin (4b) coreference is preferred but not
necessary, and in (4c) there is only one interpretation.
Another, more fundamental reason for thinking of anaphora primarily
in pragmatic terms has to do with the essence of anaphoric expressions:
they are semantically general, and require additional contextual resolution
or inferential speci®cation. Anaphora consists in the referential dependence
of one NP on another, rather than simply in the matter of coreference.
(One might have accidental coreferenceÐas when a speaker forgets that
he has already mentioned some protagonist or is unaware that Cicero was
the same man as Tully.) That referential dependence can only arise if
an NP fails on its own to achieve reference, generally through under-
speci®cation. That is why pronouns are the primary type of anaphoric
expression in many languages; they are minimal nominal expressions with
maximal semantic generality (usually encoding only some subset of per-
son, number, gender, and case features). Although languages may gener-
ally (but apparently not universally) have specialized expressions that are
necessarily referentially dependent, like himself in English, just about any
semantically general expression can function anaphorically. To see this,
consider the following pairs:
269 Grammar and Implicature

(5) a. ``The ferry hit a rock. The ship capsized.''

b. ``The ship hit a rock. The ferry capsized.''
(6) a. ``The bus overturned. The vehicle was completely wrecked.''
b. ``The vehicle was completely wrecked. The bus overturned.''
(7) a. ``The professor rubbed his eyes. The man looked weary.''
b. ``The man looked weary. The professor rubbed his eyes.''
(8) a. ``Their Boeing 727 is relatively new, but the aircraft is of an old
design.''
b. ``The aircraft is relatively new, but their Boeing 727 is of an old
design.''
(9) a. ``I like London. My wife likes it there.''
b. ``I like it there. My wife likes London.''
In these pairs, the two referring expressions in subject position in the (a)
examples can easily be read as coreferential, but it is much harder to get
the coreferential interpretations in the (b) examples; instead we tend to
read them as lists (e.g., a list of shipping disasters, or vehicle crashes, or
tired gentlemen, or aircraft; see Sanford and Garrod 1981, Garnham 1989
for experimental veri®cation of these tendencies). Yet the only di¨erence
is that in the (a) sentences a semantically more speci®c (less general) de-
scription precedes a compatible one that is semantically less speci®c (more
general), whereas in the (b) sentences the order is reversed. This shows, I
think, that semantic generality is at the heart of anaphora; it is the essen-
tial property for an anaphoric expression. Now notice that semantic gen-
erality is also the essential property inducing I-implicaturesÐfor example,
it is the semantic generality of ``I don't like garlic'' that invites the infer-
ence `I dislike garlic', and the semantic generality of conjunction that
suggests causality in ``He threw the switch and the engine stopped.'' There
is a prima facie case that the same principle is involved.
Thus although it is tempting to think of pronouns as semantically spe-
cialized anaphoric devices, it seems likely that they play such a prominent
role in anaphora simply through their semantic generality; in English, for
example, the third-person pronouns merely restrict their referents to third-
person entities that are, as required, singular or plural, and if singular (and
human or closely associated animal) male, female, or neuter. In many
languages, even those restrictions are missing (e.g., in many Australian
languages, a single third-singular pronoun will cover any animate entity
indiscriminately). Pronouns are often held to be ambiguous between
270 Chapter 4

variable, referentially dependent, and deictic uses (and yet other senses
have been proposed); but the fact that in language after language all three
functions can be performed by the same pronominal expressions suggests
that their semantic character simply encompasses all three.7 The examples
in (5)±(9) make the point that anaphoric potential is a matter of semantic
generality relative to the antecedent. Anaphoric potential is, on the whole,
a property not of expressions but of uses of expressions. Nevertheless, the
more semantically general an expression is, the greater the likelihood that
it will be descriptively inadequate for reference identi®cation, and thus the
greater the chance that it must be referentially dependent (hence the ten-
dency to the anaphoric use of pronouns even though they are capable of
supporting in many circumstances a directly referential use).
In perhaps the majority of languages (including the Australian ones
again) NP-anaphora is also expressed through an NP-gapÐby dropping
the NP altogether. Such a device gives us the limiting caseÐNP-gaps are
of course at the extreme limit of semantic generality, with no intrinsic
content features of their own. Many languages thus provide their speakers
with a scale of semantically reduced NPs that includes at least the fol-
lowing kinds of expressions:
lexical NPs > pronouns > NP-gaps
where a choice towards the right increases the likelihood of an anaphoric
reading.8
Thus an anaphoric NP is dependent by virtue of being semantically
general to the point that it fails to adequately refer; it is only by being
linked to some other, usually prior, NP that has fuller semantic speci®ca-
tion (or is linked to one that does) that reference is achieved. But that
linkage is nearly always pragmatic and inferential and scarcely ever a
matter of grammatical speci®cationÐit is generally agreed that (besides
any gender restrictions) the only information that a grammar provides (if
any) is some speci®cation of the domain in which a search for an ante-
cedent can be made. (Sometimes it will turn out, of course, that there is
only one possible antecedent in the speci®ed domain, but that is a matter
of accident.) Although the grammars of some languages (but not all, as
we shall see) provide forms that are necessarily referentially dependent
within some domain (Anaphors in the GB sense, or re¯exives and recip-
rocals in particular), the linkage process itself nearly always remains
®rmly outside grammar in pragmatics and is often, as we know, a matter
271 Grammar and Implicature

of probabilities, of likely interpretations, or as I shall prefer to call it ``an

inference to the best interpretation'' (speci®cally an I-inference).9
That pragmatic principles are involved in the picking out of ante-
cedents for anaphoric expressions is uncontroversial; what principles are
involved, on the other hand, is much disputed. But it is clear that much
``encyclopedic'' knowledge and other contingent factors are involved, as
the following sort of example (after Winograd 1972: 33) shows:10
(10) a. ``The police1 barred the demonstrators2 because they2 advocated
violence.''
b. ``The police1 barred the demonstrators2 because they1 feared
violence.''
Here the reference of they depends crucially on assumptions about who
would most likely be advocating or fearing violence. In the same way,
re¯exives too can often have competing antecedents, where a choice be-
tween them is made on the grounds of background assumptions:
(11) a. ``The benefactors1 gave the faculty2 a clubroom for
themselves2 .''
b. ``The benefactors1 gave the faculty2 a portrait of themselves1 .''
However, in addition to these ad hoc, case-by-case factors, I want to
claim that general pragmatic principlesÐand in particular, GCIsÐplay a
role in restricting and delimiting the search for antecedents. They do this
by suggesting that antecedents can be found locally, or alternatively, by
contrast, outside a local domain. To be succinct: the use of a relatively
semantically general expression (like a pronoun) will I-implicate local
coreference, and the use of a more semantically speci®c expression (like a
de®nite description) will M-implicate the complementary interpretation
(i.e., disjointness from the reference of local NPs).11 In the case of cross-
sentential anaphora, the relevant concept of a local domain will be given
to us by a theory of discourse (as for example sketched in Fox 1987,
where the local domain amounts to a single conversational sequence). In
the case of intrasentential anaphora, local domains are given to us by the
syntax. In addition to the role of the I- and M-principles, the Q-principle
will suggest that if a less informative, less restrictive expression (like a
pronoun) is used just where a contrastive more informative expression
(like a re¯exive) could have been used instead, then the reading that
would have come about through the use of the more informative ex-
pression is not intended. These are very simple, switch-like operators:
272 Chapter 4

M-inferences cancel the I-tendency to local coreference, Q-inferences

cancel restrictive re¯exive interpretations (or logophoric ones to be dis-
cussed later). It should be immediately clear that GCIs could at most be
just one of many factors determining the choice of particular antecedents
for an anaphor. My claim will be only that GCIs provide some additional
simple, but nevertheless crucial, semiotics that help the addressee compute
the intended anaphoric link-up.
To summarize: there are at least three reasons to think that anaphora is
fundamentally pragmatic. First, the essential patterns are cross-sentential,
as well as intrasentential. Second, the necessary semantic property for an
anaphoric expression is not some specialized semantic featureÐany
semantically general expression will do the job (i.e., any expression that is
more general than the intended antecedent). Third, virtually all anaphora
involves pragmatic resolution between alternatives, which may themselves
be constrained by many other factors, including syntax.
Let us return to what we called the general anaphora pattern, illustrated
in (2)±(3) and repeated here:
(12) a. ``John came in. He sat down.'' (preferred coreference)
b. ``John came in. The man sat down.'' (preferred disjoint
reference)
(13) a. ``John likes his father.'' (preferred coreference)
b. ``John likes the man's father.'' (disjoint reference)
It should already be more or less clear how our preexisting pragmatic
apparatus can give an account of this pattern. The explanation will go like
this: I-implicatures predominate in the absence of countervailing Q or M
implicatures, so in the (a) sentences there will be a general preference for
local coreference encouraged by the use of semantically general forms. On
the other hand, where, as in the (b) sentences, a marked, more prolix ex-
pression has been used, we obtain an M-implicature that is complemen-
tary to that which would have arisen by I-implicature from the use of the
simple expressionÐin other words, we will now obtain the preference for
noncoreference in a local domain. (Note that the reading in (13b) is nor-
mally assumed to be stipulated by grammatical principle, but it is also
predicted by the pragmatic pattern.) The reader is owed an account of
exactly how the I-principle yields this preference, which is provided in the
next section, 4.2.2.
These principles can be seen at work in discourse (see Levinson 1987a),
but we cannot expect to ®nd such preferred interpretations ¯awlessly
273 Grammar and Implicature

predicted, for there are many con¯icting processes involved in the inter-
pretation of texts or conversation. Our principles are just one component
in a full account of how reference tracking takes place. Much work has
been done on the empirical study of tendencies in various kinds of dis-
course, and it is clear that many complex factors are at work (see, e.g.,
Fox 1996). One should note that the GCI contribution to a full account of
reference tracking in discourse is intended to be modest (the critique in
Ariel 1994 mistakes it for an exhaustive theory of anaphora). A full ac-
count will certainly involve topic and ``paragraph'' units and other mea-
sures of textual distance to antecedents, thematic and syntactic saliency,
cognitive saliency and mutual accessibility, encyclopedic associations, and
so on (see, e.g., Kibrik 1996). Second, modest though the contribution is,
it is essential: GCIs provide vital semiotic signals about roughly where to
look for antecedentsÐit is, unlike many of the other factors, part of the
signal itself, a kind of meaning that piggybacks on the central coded sig-
ni®cance, as sketched in the Introduction and chapter 1.12

4.2.2 Inferring Coreference13

I have claimed that the basic anaphora pattern is for a relatively seman-
tically general NP to pick up coreference locally and a marked or prolix
one to indicate that that would be inappropriate. It is important to show
that there is some natural way in which the preference for locally co-
referent interpretations of semantically general expressions is a by-product
of the I-principle. What must be shown in particular is that where a
coreferential interpretation is available it will generally provide an infor-
mationally richer, more informative, stereotypical, and ``better'' interpre-
tation in line with the general direction of that principle.14 It is possible to
show that, in general, coreferential interpretations will indeed be more
informative than noncoreferential ones. There are a number of distinct,
but mutually compatible, arguments.
First, there's an argument at the level of referenceÐan argument to the
e¨ect that the presumption of a minimal domain of reference increases
the informational load an utterance may bear. Let us adopt, for lack of
serious alternatives, the analysis of informativeness due to Bar-Hillel and
Carnap (1952; see also chapter 1 above), which is one way of operation-
alizing the Popperian idea of informativeness-relative-to-falsi®ablity: the
greater the number of possible state descriptions that are ruled out by (are
incompatible with) a proposition, the more informative it is. This will
have as a consequence (Bar-Hillel and Carnap, 1952: 227¨ ) the simple
274 Chapter 4

entailment analysis of informativeness that underlies the Horn scales

involved in the Q-implicatures.
Now, the method of state descriptions suggests one way in which we
might derive, or theoretically predict, the empirically observable prefer-
ence for coreferential interpretations that was stated as condition (c) of
the I-principle in (39) in chapter 2. (Without such a theoretical argument,
linking the preference for coreferential interpretations to the I-principle,
we will have to state condition (c) as an independent pragmatic principle.)
The argument would go as follows: a coreferential interpretation reduces
the number of entities required in the domain of discourseÐthat is, the
interpretation can proceed on a smaller domain of discourse than it
could on a noncoreferential interpretation. But if the set of entities in the
domain of discourse is smaller, so too will be the number of state
descriptions compatible with the assertion in question.15 Thus the asser-
tion will be, from an extensional point of view, less ``general'' (i.e., com-
patible with fewer possible states of a¨airs) and thus more speci®c and
informative on the assumption of the minimal possible domain of dis-
course. Assuming coreference rather than introducing further entities into
the domain of discourse will thus increase the informativeness of a state-
mentÐand hence, it may be argued, the preference for coreferential
interpretations can be seen to be properly a part of the I- principle.16
Popper (1959: 68¨ ) independently o¨ers another argument leading in
the same direction. Existential statements are weakÐthey are not falsi®-
able by any single observation; consequently, the fewer existential com-
mitments the stronger the ``theory'' (read ``assertion'' for our purposes).
Consider, for example, the two interpretations of the same sentence indi-
cated below:
(14) a. ``John1 loved his1 father.''
b. ``John1 loved his2 father.''
On Popper's view the logical forms would be something like the follow-
ing, respectively:
(15) a. bx [LOVE(x, hz (F(z, x))) & x ˆ John]
There's an x such that x loves the unique z which is father of x,
and x is John.
b. bx by [LOVE(x, hz (F(z, y))) & x 0 y & x ˆ John]
There's an x and a y such that x loves the unique z who is father
of y, and y is distinct from x, and x is John.
275 Grammar and Implicature

Given the principle, ``the more existential quanti®ers, the less informative
the statement,'' the interpretation in (15a) is clearly to be preferred by the
I-principle to the interpretation in (15b). Thus the preference for corefer-
ential interpretations can also be seen to follow naturally from the I-
principle on this alternative Popperian account.17
Some di¨erences between these two accounts are worth pointing out.
Notice that the Carnapian argument is at the level of reference; it is an
argument favoring the assumption of parsimony in referenceÐit might be
glossed ``for interpretation, choose a domain (D) or world (compatible
with all one knows) that makes a statement most speci®c and thus maxi-
mizes D-informativeness.'' The Popperian argument assumes that the
domain of discourse is held constant, and o¨ers an evaluation for alter-
native statements that a sentence might have been used to makeÐit might
be glossed ``choose the logical form (or set of truth conditions, or propo-
sition expressed) that minimizes existential commitment and thus maxi-
mizes P-informativeness.'' It is thus an argument essentially at the level
of logical form, although because it involves the referential apparatus of
existential quanti®cation, it is also an argument at the level of reference.
There may be other ways in which to derive the preference for corefer-
ential readings from the Principle of Informativeness. If both of the above
arguments are partially at the level of reference, another argument is at
the level of sense. In `John came in and he sat down', regardless of the
domain of discourse, the coreferential reading is obviously more infor-
mative about John. Indeed the coreferential reading is more informative
about John than the noncoreferential reading would be about either ref-
erent if distinct; it is likely therefore to be a better, more informative
interpretation.18
Recently, and largely independently of the philosophical tradition,
notions similar to these have been explored in AI approaches to ``com-
mon-sense reasoning.'' For example (as mentioned in chapter 1), some
nonmonotonic reasoning systems employ the notion of circumscription
(see Ginsberg 1987, sect. 3), which limits the model in which formulae
are proved. The idea is to capture common-sense reasoning of the kind
``Assume the car will start unless one knows it is broken'' by adding spe-
ci®c conjectures about the data base to the reasoning system. One such
conjecture is: if the system can't prove that certain entities satisfy a pred-
icate (e.g., the battery is ¯at), then assume they do not satisfy it (i.e., the
battery is not ¯at). Another is: if the system doesn't know about certain
entities, then assume they don't exist (e.g., don't even consider that the a
276 Chapter 4

might be broken, if one doesn't know whether the car has an a). Assume
then that, for the domain in question, ``the known entities are all there
are.'' McCarthy (1980) calls this domain circumscription and notes that it
is closely related to communicational presumptions, which turn such a
conjecture into something more like a reliable rule of inference. Mini-
mizing the domain of discourse is informative, on this account, because it
allows us to make many additional inferences.
Yet another argument would make the preference for coreference a
mere side e¨ect of the search for strong I-induced connections across
clauses. For example, by identifying John and he in (16) we can interpret
the two VPs as describing a single complex action, here the action of
entering in order to sit down:
(16) a. ``Bill was waiting. John came in. He sat down.''
b. Single complex action: coming in & sitting down
Such an interpretation would not be available if less local coreference
(e.g., of he with Bill ) was assumed. There is a great deal of crosslinguistic
evidence, nicely surveyed by Foley and Van Valin 1983, that supports the
view that coreference assignment is intimately connected to cross-clause
linkage. They claim that, looking at multiclause constructions in general,
the tighter the encoded semantic linkage, the greater the degree of syn-
tactic bonding, and the more tendency to coreference across clauses (see
also Haiman 1985a, Lehmann 1988). Again, working from the point of
view of computational pragmatics, Hobbs (1979, 1982) argues that it is
the assumption of a speci®c kind of conceptual linkage between clauses
that determines coreference rather than the other way around. The pre-
sumption (by the I-principle in our scheme) is that any two clauses will be
linked by one of a number of restricted conceptual ties (causality, instan-
tiation, etc.) and in general this will require coreference across the clauses.
There is much to be said for this alternative conceptualization of the
problem, but it is surely a complementary perspective. In any case, here
we will tend to idealize away from this more global view of the nature of
coreference assignments.
I shall assume that at least one of these arguments, and most probably
a conjunction of a number of them,19 is sound and that we may thus de-
rive the preference for coreferential readings from the general I-principle,
rather than stipulate it. And in any case, we may take that preference to
be an empirical fact, whatever the success of this theoretical account.
277 Grammar and Implicature

4.2.3 Inferring Disjoint Reference

If the I-principle induces inferences to coreference, such inferences are
blocked in a straightforward manner by the M-principle. As I noted in
chapter 2, the I-principle is triggered by expressions that are either brief or
semantically general, the two generally being correlated by Zip®an prin-
ciples (see Levinson 1987b for further discussion). The corollary is that,
by the M-principle, if a speaker wants the addressee to avoid such a co-
referential interpretation, the NP in question should be marked in form or
semantically speci®c in content. Thus (to partially repeat earlier examples)
we have the patterns in the following contrastive pairs of utterances
(preferred readings marked with the appropriate indices):
(17) a. ``The car ferry1 hit a rock. The vessel1 capsized.''
I ‡> coreference
b. ``The vessel1 hit a rock. The car ferry2 capsized.''
M ‡> disjoint reference
c. ``The car ferry hit a rock. It capsized.''
I ‡> coreference
Although the signal is gradient, the interpretation is a switch between
local coreference or local disjointness: the ``heavier'' (in semantics or form)
a succeeding NP, the more likely an addressee will read it as disjoint.
Now, implicatures being defeasible, there are systematic counterexamples.
If the succeeding ``heavy'' NP can be seen as motivated by expressing an
attitude of the speaker (as in epithets), or expanded to add a piece of fur-
ther descriptive or explanatory information, it may nevertheless be read as
coreferential:
(18) a. ``The car ferry hit a rock. The whole bloody thing capsized.''
b. ``The car ferry hit a rock. The badly overloaded vessel
capsized.''
I take it that this much is straightforward. Less obvious is that the
Q-principle might also be involved in generating disjoint interpretations.
Q-inferences typically (but not exclusively) operate over privative oppo-
sites (Horn 1989: 317¨ ), where two expressions di¨er only in that one of
the pair has additional semantic speci®cation. Arguably, pronouns and
re¯exives have precisely this property: thus himself and him both have
reference to singular, male entities whose identity is recoverable from
context. They di¨er in that the re¯exive has an additional property, which
one may conceive of in di¨erent ways: one could hold that himself must
278 Chapter 4

be bound (syntactically coindexed) in a local domain, whereas the pro-

noun is free; or one could hold that the essential di¨erence is that himself
is necessarily referentially dependent on a prior NP, whereas him may
pick up its reference in other ways (deictically or contextually). In either
case, in a straightforward way, the re¯exive is more informative than the
pronoun: it ®xes reference much more narrowly. Even where the pronoun
is presumed to be referentially dependent, while the re¯exive must ®nd an
antecedent in a limited domain, the pronoun may pick up an antecedent
in a much wider domain. A re¯exive thus e¨ectively rules out many more
referential possibilities and by the Bar-Hillel and Carnap metric must be
judged more informative than a corresponding pronoun. The Popperian
metric, whereby the less existential commitments the stronger the state-
ment, will also give the same result.20
Now, expressions that are paradigmatically contrastive (can occur in
the same syntactic slots), that are in privative opposition within a seman-
tic ®eld, and that are both lexicalized ought to form a Horn scale. If
hhimself, himi forms such a scale, asserting an utterance with the less
informative him in a slot where the expression himself could have oc-
curred should implicate that the speaker is avoiding the more informative
statement. In e¨ect, by utilizing a pronoun where a re¯exive could have
been used, the speaker Q-implicates disjoint reference. This gives us an
alternative, and perhaps at ®rst sight implausible, account of the follow-
ing familiar pattern:21
(19) a. ``John1 admires himself1 .'' (coreferential by grammatical
stipulation)
b. ``John1 admires him2 .'' (disjoint by Q-implicature)
Most readers will feel that both sides of this pattern are grammatically
stipulated and thus that him simply cannot refer to John. This is not
however invariably the case. A crucial fact in the plausibility of this ac-
count is that the reading is, despite being a strong default interpretation,
nevertheless a defeasible inference: as Evans (1980) and Reinhart (1983:
168¨ ) point out, the interpretations indicated by the referential indices
below are entirely natural if unusual:22
(20) a. ``No one else voted for Marcos1 , but Marcos1 voted for him1 .''
b. ``Ronald1 and Nancy2 have one thing in common: Nancy2
adores him1 and Ronald1 adores him1 too.''
c. ``Everyone despises poor William1 . Mary despises him1 . Joan
despises him1 , why, even poor William1 despises him1 !''
279 Grammar and Implicature

The conditions under which the pronoun can be read as coreferential

despite being clause-bound are admittedly limitedÐin the examples at
hand, for example, the discourse sets up as topic a property predicated
about an individual (e.g., voting for Marcos), and then this predicate is
predicated of the same individual. Further kinds of examples are where
the pronoun is employed coreferentially but deictically, or in statements
of identity:
(21) a. ``I didn't hurt you1 , you1 hurt you1 .''
b. ``[The man who did it]1 is him1 .''
c. ``He1 became the ®rst chairman1 of the department.''
Yet another case is where the re¯exive would have a bound-variable
interpretation that the speaker wishes to avoid. Thus (22a) does not mean
the same as (22b). In (22a), everyone may have voted for the chairman,
but no one except the chairman voted for themselves; whereas in (22b) the
chairman got only one vote. Thus (22b) is freed from pragmatic restric-
tions to allow the expression of the referential interpretation.
(22) a. ``Only the chairman1 voted for himself1 .''
b. ``Only the chairman1 voted for him1 .''
These sorts of examples would seem to establish that the strong preference
for disjoint reference of a pronoun where a re¯exive could have appeared
is not a matter of entailment or grammatical stipulation (unless we are
willing to abandon the litmus test of defeasibility discussed above).23 Let
me remind the skeptical reader that it is not given to us by direct intuition
whether a pattern of interpretation is speci®ed by grammar and lexicon or
by generalized implicatureÐwitness the millennia of controversy about
the correct semantic analysis of the Square of Oppositions reviewed in
chapters 1 and 2.
Those readers who nevertheless continue to ®nd implausible the idea
that in John hit him disjoint reference is just implicated, not grammati-
cally stipulated, should bear with me for two reasons. The ®rst reason is
that I noted at the beginning of the chapter that GCIs induce inter-
pretations so reliably associated with expressions that they tend to gram-
maticalize. It may be that we are dealing here with a partially or generally
grammaticalized pattern in English (although the limited defeasibility
then needs some explanation). Note that the diachronic rather than syn-
chronic status of an implicatural analysis would not lessen its importance.
For these patterns would no longer follow from deep, unlearnable and
280 Chapter 4

thus innate principles of Universal Grammar, as Chomsky has proposed,

but rather owe their generality to the universality of usage principles that
are not essentially linguistic, everywhere exhibited, and thus intrinsically
learnable.
The second reason for begging patience is that we shall see in the later
part of this chapter that there are indeed indications that there may be
additional principles behind the presumption that in John hit him the
subject and object are distinct in reference. One piece of evidence (to be
discussed at length in section 4.3.3.1) is that there are languages without
any overt marking of re¯exivity (contrary to the many con®dent claims
that no such languages existÐsee, e.g., Reinhart and Reuland 1993: 662).
If the preference for reading statements of the kind `John hit him' was
only due to Q-implicature from the availability of a re¯exive that was not
used, it would be hard to understand the patterns in these languages,
where the same (if much weaker) preference persists. Later, I shall attri-
bute this to a background presumption that co-arguments of a verb are
distinct in reference. Such a principle might be grounded in various ways.
Indeed, I shall entertain the idea that the presumption of clausemate dis-
tinctness is simply an I-principle induced preference for the stereotype,
where the stereotypical actions that humans are preoccupied with are
irre¯exive. If so, it should be easily canceled by Q- or M-inferences,
which, as we have noted, always take priority if inconsistent (this raises
one set of problems, which I discuss in section 4.3.3). In the English case,
then, this I-inference to disjointness would be overridden by the semantics
of the re¯exive and reinforced by the Q-implicature to noncoreference by
the use of a pronoun where a re¯exive could have been used to refer to
what a co-argument refers to. This double implicature in the case of
nonre¯exive pronouns will not prove totally redundant: often, re¯exives
encode more than just referential identity, so that avoiding a re¯exive
may be motivated by other considerations; but the I-presumption of co-
argument distinctness will help to focus the contrast between pronoun and
re¯exive on reference alone just when antecedent and anaphoric expres-
sion are co-arguments of the same verb.

4.3 BINDING THEORY AND PRAGMATICS

4.3.1 Introduction
Let us turn now to intrasentential anaphora, the area where we expect
to ®nd competing grammatical and pragmatic accounts of constraints on
281 Grammar and Implicature

the interpretation of referentially dependent expressions. We will take

Chomsky's (1981: 188) Binding Theory as the point of departure:
(23) Binding Conditions
A: An Anaphor is bound in its governing category.24
B: A pronominal is free (not bound) in its governing category.
C: An R-expression (lexical NP or variable) is free everywhere.
This set of conditions relies on a series of de®nitions which have been
varied from time to time, but which at their simplest are of this
kind (Chomsky 1981; see Radford 1981: chap. 11 for elementary exposi-
tion):
1. An argument is bound i¨ is it coindexed with a c-commanding argu-
ment, otherwise it is free.
2. A constituent X c-commands Y i¨ the branching node over (dominat-
ing) X dominates Y, but neither X nor Y directly dominate one
another.25
3. The governing category or local domain for an expression X is the
minimal S or NP that contains whatever constituent (e.g., verb or
preposition) governs X (and also contains, in more recent formula-
tions, such as Chomsky 1986: 169, other speci®ed constituents, for
instance a subject).
Despite many e¨orts to improve the empirical coverage of the theory
by altering the conditions or tinkering with the constituent de®nitions,
Chomsky himself still adheres in the Minimalist framework to a set of
conditions that as far as we are concerned are more or less equivalent
(Chomsky 1995: 211):
(24) Interpretive version of the Binding Conditions
A: If a is an Anaphor, interpret it as coreferential with a
c-commanding phrase in the relevant local domain D.
B: If a is a pronominal, interpret it as disjoint from every
c-commanding phrase in D.
C: If a is an R-expression, interpret it as disjoint from every
c-commanding phrase.
What has changed is that the Conditions are now held only to hold at LF,
the syntax/semantic interface, where they ®lter interpretations, and fur-
ther, as mentioned earlier, they no longer play the same role as constraints
on movement.
282 Chapter 4

The empirical coverage of the Binding Conditions is reasonably good,

but there exists a large catalogue of well known counterexamples even
within English to the classical formulations (see the useful summary in
Zribi-Hertz 1989). Particularly worrisome, though, is the crosslinguistic
evidence that points to substantial di¨erences, especially with regards to
Anaphors that can be bound at seemingly arbitrarily long distances. The
response has been a ¯urry of attempts to come up with better formula-
tions. It would be impossible here to review the many alternative for-
mulations that have been proposed. Su½ce it to say that there are a
number of critical ways to ®ne-tune the predictions. First, one can employ
di¨erent notions of c-command or, more radically, seek to replace it al-
together with other notions less directly con®gurational (theta-hierarchies
or obliqueness hierarchies, see, e.g., Jackendo¨ 1972 or Pollard and Sag
1994). Second, one might rethink the basic classi®cation of NPs in three
types (Anaphors, pronominals, and R-expressions): one could deny that
everything that looks like a re¯exive is an Anaphor, or one could insist on
di¨erent kinds of Anaphors, and so on. Third, and this has been the
major focus of research, one can play with the de®nition of the domain D
in rather di¨erent ways: (a) one can, like Chomsky, attempt to hold onto
a universal de®nition but so arrange things that the D for Anaphors is
e¨ectively slightly di¨erent than the D for pronominals, so handling one
class of counterexamples where the general complementary distribution of
pronominals and anaphors breaks down; (b) one can abandon universal
constraints and specify, for example, a domain D according to each lan-
guage; or one can allow that some languages may have more than one
kind of Anaphor, each with di¨erent relevant domains. On the whole,
the preferred solution is to parameterize D or the notion of governing
categoryÐthat is, to handle the variation needed both across languages
and within languages with di¨erent kinds of Anaphors, by expanding or
contracting the local domain D by specifying that it must also contain
some other element, such as a subject or an agreement feature (see Reu-
land and Koster 1991: 2, and other papers in Koster and Reuland 1991,
Harbert 1995). We will return to these complexities later.
The reader will recollect that the Binding Conditions were developed to
predict, inter alia, the following kinds of interpretive constraint (as before,
I use coindexing simply as an expository device to indicate referential de-
pendency where the indices are identical, referential independence where
di¨erent):
283 Grammar and Implicature

(25) a. ``John1 admires himself1 .'' (coindexing stipulated by

Condition A)
b. ``John1 admires him2 .'' (coindexing forbidden by
Condition B)
c. ``John1 admires the professor2 .'' (coindexing forbidden by
Condition C)
d. ``John1 said he1 would come.'' (coindexing permitted by
Condition B)
e. ``John1 said the professor2 would (coindexing forbidden by
come.'' Condition C)
Let me now outline how I shall proceed to suggest that much of this
simple pattern, together with many of the further complex details, is in
great part pragmatic in origin. First, I will outline one kind of approach
that I developed in Levinson 1987a,b. For expositional purposes, we may
consider a series of approximations to the data. Initially, this kind of ac-
count was developed to deal with data in an Australian language, where
the patterns are rather clearly pragmatic, but I shall give only the briefest
indication of that here (see Levinson 1987a). This ®rst line of approach
has the following general properties:
1. Binding Condition A is taken to be grammatically speci®ed, as in the
GB accounts.
2. Conditions B and C are then reduced to pragmatics by showing them
to be the consequence of applying GCI theory to Condition A.
This account relies heavily both on Chomsky's original (1981) formula-
tion of the problems and the data and on Reinhart's (1983) own proposals
for pragmatic reductionism. I dub this the A-®rst account because it
derives the rest of the anaphora patterns from Condition A of the Binding
Theory.
However, it has become clear that the original formulation of the
problem is not quite correct; in particular, it ignores the following aspects
of the problem that have ®lled the linguistics journals between 1982 and
the present:
1. Re¯exives and pronouns are not quite in complementary distribution
(on a given indexing) even in the familiar languages; but in many other
languagesÐespecially those exhibiting long-range re¯exivesÐthe
complementarity breaks down fundamentally.
284 Chapter 4

2. Just where the complementarity in coreference possibilities breaks

down, another kind of contrast can be seen to take its placeÐnamely,
a contrast in logophoricity, or subjective point of view (long-range
re¯exives generally suggesting the point of view of their referents).
3. The whole semantic apparatusÐindexing, the notions of disjointness
and coreferenceÐneeds revising in the light of Evans's (1980) funda-
mental insight that coreference must be distinguished from referential
dependency.
I will extend my original 1987 account to handle some of these facts,
but in the end I concede that for some languages the account remains
unsatisfactory.
So in the second half of this chapter I want to propose a much revised
pragmatic reductionism of the Binding Conditions, which in many ways is
rather more radical because it does not accept the unreducable nature of
Condition A. The crucial ingredients of the new account are:
1. Most of the ideas of the old account;
2. Three additional ingredients:
a. a pragmatic treatment of what GB theorists have assumed are dif-
ferent Binding Domains for Conditions A and B; this is gotten by
introducing a pragmatic analogue of Condition B;
b. a pragmatic account of logophoricity contrasts;
c. a sensitivity to Evans's distinction between coreference and refer-
ential dependency.
3. Some overview of new data from other languages that suggests that a
more wholesale reductionism will in some cases be unavoidable.
I dub this alternative account the B-®rst account (Levinson 1991) because
it gives pride of place to the anaphora pattern sometimes attributed (but
not here) to Condition B of the Binding Theory and makes the rest of the
patterns derivative from that.
Finally, I will suggest that rather than replacing the old account with
the new one, the two should be seen as complementary. For some lan-
guages the A-®rst account works well, for others the B-®rst account is
essential. However, in many languages, it seems that some amalgamation
of the two accounts is required. This is to be explained, I suggest, by
taking a diachronic perspective, in which A-®rst patterns derive from
B-®rst patterns of anaphora; thus many A-®rst languages have B-®rst
residues. This evolutionary (or B-then-A) account seems to provide a uni-
285 Grammar and Implicature

®ed explanation for many facts, including the association of long-range

re¯exives with logophoricity.

4.3.2 The A-First Account: Pragmatic Reduction of Binding Conditions B and C

4.3.2.1 English: A First Approximation We have noted what I have

called the general anaphora pattern: the more semantically general or
reduced an NP is, the greater the tendency to read it as locally corefer-
ential. In those languages that permit zero-anaphora (e.g., the so-called
pro-drop languages like Italian, or many noncon®gurational languages,
like most of the Australian ones) we can extend the scale of reduction,
from lexical NP, to pronoun, to NP-gap. And our pragmatic principles
can now be seen to predict the general anaphora pattern along the fol-
lowing lines.
First of all, the I-principle will tend to favor locally coreferential inter-
pretations wherever feasible, but especially where an NP tends towards
semantic generality. Second, this tendency will be tempered by the M-
principle, for by the Hornian division of labor wherever an I-implicature
would lead to a coreferential interpretation, the use of a marked or pro-
lix expression will M-implicate the complementary disjoint interpreta-
tion. By our projection principles, M-implicatures defeat incompatible
I-implicatures; therefore, the use of a marked expression will overrule the
tendency to coreferential interpretations.
(26) The general anaphora pattern
lexical NP > pronoun > NP-gap
ƒƒƒƒƒƒ M-implicates noncoreference
I-implicates coreference ƒƒƒƒƒ
ƒ!
An interesting prediction here is that where a language permits anaphora
by NP-gap, the use of a pronoun despite its semantic generality will M-
implicate some contrastive interpretation, and quite often a contrast in
local noncoreference. Usage in many languages seems to support this
prediction (e.g., in Chinese, Guugu Yimithirr, Tamil, Korean, Turkish,
Greek)26 although the contrast is by no means invariably interpreted on
every occasion as a contrast in intended reference (but rather, sometimes
in terms of emphasis, topichood, etc.).27
So far, we have employed only part of the pragmatic apparatus devel-
oped here. Let us now bring in the third kind of implicature in our scheme
286 Chapter 4

by suggesting that the opposition between re¯exives and ordinary pro-

nouns is essentially a matter of scalar (Q-) implicature. In section 4.2.3 I
suggested that re¯exives and pronouns are in the kind of privative oppo-
sition typical of Horn scales, of the form hhimself, himi and thus that the
use of the pronoun can Q-implicate that the re¯exive situation does not
obtain. The use of a re¯exive will indicate coreference by semantic stipu-
lation (in the A-®rst theory). But wherever a nonre¯exive pronoun is used
in a locus where a re¯exive is structurally permitted, the I-inference will
be overruled by a strong scalar Q-inference to disjointness (recollect that
Q inferences always take priority over I-inferences). Both M- and Q-
principles thus constrain I-inferences to coreference.
This much seems to follow directly from the theory outlined in chapters
1 and 2Ðnothing more needs to be stipulated. We turn now to the pre-
dictions for intrasentential anaphora in English. Let me illustrate how we
can put these ingredients together to get a ®rst approximation to an
account of the binding patterns in English (a more adequate account will
follow in section 4.3.2.2). Our target, initially anyway, can be taken to be
the simplest version of Chomsky's conditions as below (Radford 1981:
367; Koster and May 1982: 137):
(27) Binding Conditions
A: Anaphors (re¯exives and reciprocals) must be bound in their
governing category.
B: (Nonre¯exive) Pronouns must be free in their governing
category.
C: A lexical NP must be free everywhere.
The constitutive concepts may again for the purposes of argument be
taken to have their simplest speci®cation: bound means coindexed with a
c-commanding argument, free means not bound, and governing category
can be taken to be the minimal domain, S or NP, containing the governor
of the constituent in question.
Here is how a partial pragmatic reduction of the Conditions might go.
First, no attempt is made to reduce the pattern of interpretation in Con-
dition A of the Binding TheoryÐthat is, we take the interpretation of
re¯exives and reciprocals to be stipulated by speci®cation of the local do-
main (S or NP) in which an antecedent must be found (but see section
4.3.2.2). Then, taking Condition A as given, we can derive Condition B
and C along the following lines:
287 Grammar and Implicature

First, where the syntax permits a direct coding of referential depen-

dency, and thus of coreference with another NP in a restricted domain (by
the use of an Anaphor), a speaker intending such coreference should use
such a re¯exive or reciprocal. Otherwise a speaker would be in breach of
the Q-principle that enjoins the speaker to avoid the use of a statement
that is informationally weaker than his knowledge of the world allows.
This warrants the recipient to infer from the use of a nonre¯exive pronoun
that the speaker intended some reference distinct from that which would
have been coded by the use of a re¯exive.
Second, where a semantically general expression (e.g., a pronoun) is
used, the use of the semantically general expression will I-implicate local
coreference unless: (a) a re¯exive could have been used in that locus, in
which case the pronoun will Q-implicate locally disjoint reference (as
explained in the previous paragraph); or (b) a still more semantically
general expression might have been used than was in fact employed, in
which case the use of the more prolix or semantically speci®c form may
M-implicate the complement of the I-implicature to coreference, that is,
disjoint reference with NPs in the local domain (as explained in the dis-
cussion of the general anaphora pattern).
To see how this gives a pragmatic reduction of facts normally attrib-
uted to Conditions B and C, consider the following sentences, where the
preferred interpretations are indicated by referential indices:
(28) a. ``John1 likes him2 .''
a 0 . ``John1 likes himself1 .''
b. ``John1 told her2 that he1 gave her2 a valentine.''
b 0 . *``John1 told her2 that himself1 gave her2 a valentine.''
c. ``John1 told her2 that the man3 gave her2 a valentine.''
Let me spell out the account. First of all, the use of the sentence in (28a)
contrasts with the use of the re¯exive as in (28a 0 ). The alternates hhimself,
himi form a clear contrast set, one member more informative than the
other, as in a Horn scale, such that the use of the weaker him Q-implicates
the inapplicability of the stronger himself.28 Second, Q-implicatures can
only arise where there is a possible alternation between a stronger and
weaker expression, such as a re¯exive versus nonre¯exive pronoun. There
is no such possibility for the subject pronoun in (28b)±(28b 0 ) is ungram-
matical. In the absence of any Q-implicatures to block I-enrichment, the
I-principle will induce coreference in order to achieve parsimony in refer-
ence; any semantically general expressions, minimized in an informational
288 Chapter 4

sense, are (given the I-principle) ripe for such I-enrichment (unless the Q-
principle intervenes). Therefore, where person/number/gender features
permit, pronouns should be interpreted as locally coreferentially where
possible, as in (28b). (There is of course no account of this preference
within the GB framework, but within our pragmatic framework it is cor-
rectly predicted.)
Finally, however, expressions of roughly equal semantic generality may
contrast on the level of markedness of linguistic form (they may be periphras-
tic phrases, complex in derivational morphology, or nonidiomatic terms).
Where there is such a set, and a marked form is employed, an M-implicature
will arise to suggest the complementary interpretation to the I-implicature
that would have arisen from the use of the unmarked alternate. Thus the
use of the man in (28c) induces an M-implicature to noncoreferentiality,
just where the pronoun he would induce an I-implicature to coreferentiality.
It should be clear from the analysis of the sentences in (28) that we now
have a partial pragmatic reduction of the Binding Conditions themselves.
At the risk of redundancy, let me spell it out:
1. We accept Condition A as a rule of (English) grammar.
2. Condition B is then predicted (and rendered unnecessary) by the
Q-principle as follows: wherever a re¯exive (the semantically stronger
expression) could occur, the use of a pronoun (the semantically weaker
expression) will Q-implicate the inapplicability of a coreferential read-
ing. We will thus obtain the complementary distribution of Anaphors and
coreferentially interpretable pronouns, without invoking a separate Con-
dition B. Where re¯exives cannot occur, pronouns, being semantically
general, will I-implicate preferred coreferential interpretations (a pattern
for which there is, quite properly, no GB account, it being clearly a matter
of preferred interpretation).
3. The pattern described in Condition C is the outcome of two prag-
matic principles: the one applied to Condition A, the other to the pattern
normally labeled Condition B. The Q-principle, just as in the case of
pronouns, will induce a noncoreferential interpretation wherever a lexical
NP is used where a re¯exive might have occurred (this alone would pre-
dict a similar distribution for pronouns and lexical NPs). Additionally,
however, the use of the more prolix or marked lexical NP where a re¯ex-
ive could not have occurred, and thus where a pronoun would normally
be subject to I-induced coreferentiality, will M-implicate disjoint reference
(thus distinguishing lexical NPs from pronouns in terms of distribution on
coreferential interpretations). Schematically:
289 Grammar and Implicature

(29) Partial reduction of Binding Conditions

Condition A:
Content: ``Anaphors (re¯exives and reciprocals) are bound
(coindexed with a c-commanding NP) in their minimal governing
category.''
Pragmatic account: Nil
Condition B:
Content: ``Pronouns must be free (not coindexed with a
c-commanding NP) in their minimal governing category.''
Pragmatic account: Given the scale hhimself, himi, use of the
weaker, less speci®c him will Q-implicate disjoint reference (if you
had meant the stronger, more speci®c coreferential subcase, you
should have used the form that encodes coreference, namely
himself ).
Condition C:
Content: ``Lexical NPs must be free (not coindexed with a
c-commanding NP) throughout the sentence.''
Pragmatic account: Two parts:
i. Just as in B, there will be a Q-opposition between hhimself, the
mani, such that wherever you could have used a re¯exive you
should have, if you intend coreference.
ii. Unlike in B, there will be a further opposition between the
potential use of the pronoun and the use of lexical NP:
fhe, the mang are opposed by the Hornian division of labor: He,
being a minimal form, will encourage an I-inference to local
coreference; the man (etc.) will M-implicate disjoint reference,
because the speaker would seem to be avoiding the I-implicatures
induced by the pronoun.
Thus a case can be made for the elimination of both Condition B and
Condition C from principles of (universal) grammarÐtheir e¨ects seem
to be predicted by independently motivated pragmatic principles. Note in
particular how the argument depends not only on the details of our three
interacting pragmatic principles but also on the projection principle for
implicatures given by the hierarchy Q > M > I. It is certainly of interest
that these anaphoric patterns should be predicted by a relatively struc-
tured, independently motivated pragmatic apparatus that is su½ciently
constrained to make sharp predictions.
290 Chapter 4

Such a reduction is, of course, far from wholly original: the mirror-
image relation between Conditions A and B of the Binding Theory invites
a grammatical explanation of Condition A and a pragmatic explanation
of the complementary pattern in Condition B.29 An early attempt was
made by Dowty (1980: 32¨ ) using a maxim of ambiguity-avoidance, and
another more carefully constructed account is made by Reinhart (1983),
which can be combined closely with the present proposals and is thus
discussed below (see also Sadock 1983, Horn 1984). In addition, there has
been a growing body of counterexamples not only to the original Binding
Conditions but also to the successive revisions developed to account for
class after class of problems. This in turn has led some authors to ac-
knowledge that much that had been thought to be grammatically stipu-
lated is in fact better handled by a pragmatic or discourse theory. These
retreats naturally open up an acknowledged space for a pragmatic account.
Now, there are some signi®cant residual problems for such a ®rst ap-
proximation to a pragmatic reduction of Binding Conditions B and C.
Three major problems are the following:
1. First, the pragmatic account appears to be in trouble wherever the
complementary distribution between Anaphors and (nonre¯exive) pro-
nouns breaks down. This in fact happens in English in a number of well-
studied contexts, for example certain prepositional phrases:
(30) a. ``John1 saw a snake near him1 /himself1 .''
b. ``John1 didn't like Sue being promoted above himself1 /him1 .''
Attention to this central problem will be postponed till we have ®rst con-
sidered a second approximation to the English patterns.
2. Arguably, the unacceptability of *He1 thinks John1 is a genius is far
too strong to be accounted for just by the M-principle. We may point
to the fact that the ban on the coreferential interpretation is in fact
defeasibleÐfor example, the sentence becomes acceptable when the con-
text sets up a special relevant property of, say, thinking-John-to-be-a-
genius as in Everybody thinks John is a genius, why, even he1 thinks John1
is a genius. Still, there is a contrast in acceptability between John1 thinks
John1 is a genius (relatively acceptable) and He1 thinks John1 is a genius
(relatively unacceptable), and this is not so easily explained by the M-
principle alone. This issue will be addressed directly in the second
approximation to English below.
3. The pragmatic account given so far relies on an implicit unmarked
order of antecedent and anaphoric expression. This order e¨ect was
291 Grammar and Implicature

demonstrated in section 4.2.1, where two NPs, one semantically speci®c

and one more general, resist a coreferential reading unless they are in the
order: speci®c NP > general NP. Some rule of the sort, often mentioned
in studies of anaphora, seems to be in play: `First introduce your referent
by a full speci®cation, thereafter use a reduced form' (see, e.g., Fox 1987:
18). But in fact here structural position rather than sheer order matters, as
shown by the following pairs, where it appears that it is the c-command
relation between John and him that matters, not the order:
(31) a. ``John1 saw a snake near him1 .''
b. *``He1 saw a snake near John1 .''
c. ``Near him1 , John1 saw a snake.''
d. *``Near John1 , he1 saw a snake.''
Thus at the very least a pragmatic account would have to have some
sensitivity to marked phrase orders. That alone would also be insu½cient
because of the following kind of example, where on the syntactic account
the absence of any c-command relation between the NPs and pronouns
makes coreference possible either way:
(32) a. ``John's1 friends admire him1 .'' (coreference preferred but
optional)
b. ``His1 friends admire John1 .'' (coreference optional)
In short, we appear to need either some con®gurational constraints on
pronominal reference, or some analog to them in the form of semantic or
functional concepts. In the second approximation, we'll consider one kind
of solution to getting more structure into the account. An alternative,
pragmatic approach may be possible using the notions of theme and
rheme: essentially, the main reference of the theme may not be referred to
with a full form in the rheme (Bolinger 1979).
In what follows, I attempt to provide solutions of di¨erent kinds to
each of these problems. But before proceeding, the reader should bear in
mind that the moves taken in an A-®rst pragmatic analysis, which takes a
Condition A for granted and gets the B and C patterns partly by mirror
image from A, depend critically on the exact speci®cation of Condition A.
Thus the kinds of move that the pragmaticist should make here are very
much determined by the choice of a syntactic or semantic account. Let us
focus ®rst on an approach particularly intended to interact with a prag-
matic account. Later, however, we will return to consider some other
approaches drawn from some of the many di¨erent contemporary kinds
292 Chapter 4

of syntactic analysis of binding constraints, and consider how well these

might interact with the kind of pragmatic apparatus developed here.

4.3.2.2 A Second Approximation: Enlisting Reinhart's Account As just

noted, there appears to be more con®gurational structure involved in
constraining anaphoric linkages in English than can be predicted from
any simple pragmatic ordering rule for the introduction of pronouns.
Note though that on an A-®rst account, most of the con®gurational con-
straints in the Binding Theory are not lost. Rather the constraints in
Conditions B and C are now in part obtained by pragmatic re¯ection
from the syntactically speci®ed Condition A pattern, such that a pronoun
in any structural location where a re¯exive could have occurred, will be
implicated to be free (unbound) in that locus. But further speci®c con-
straints on the occurrence of full NPs and pronouns seem now to have no
account, as for example the just noted pattern where a c-commanded
pronoun may be coreferential even if fronted (as in Near him1 , John1 saw
a snake) whereas a pronoun c-commanding a lexical NP may not be, even
if fronted (*Near John1 , he1 saw a snake). We clearly need some addi-
tional machinery, either in the way of constraints on full coreferring NPs
in the rheme, or in terms of something like c-command. In this section, we
explore the possibilities of building in more con®gurational constraints
into the A-®rst account.
Reinhart (1983) develops an interesting account of the English binding
patterns which makes a fundamental distinction between grammatical
coindexing and mere coreference, an account since extended in various
ways (Reinhart 1986, Grodzinsky and Reinhart 1993). On this account,
only grammatical coindexing comes within the Binding Theory, and
grammatical coindexing is invariably semantically interpreted as vari-
able binding. Re¯exives get grammatically coindexed, and interpreted as
variables, as is usually assumed. Moreover, a re¯exive or nonre¯exive
bound by a quanti®er also of course has the variable interpretation. But
nonre¯exive pronouns also can get coindexed just in case they are c-
commanded by an NP outside their local domain, in which case they must
also be interpreted as variables. Thus there are parallels among (33a±c),
shown in the corresponding (informal) logical forms in (33g±i), respectively.
(33) a. ``Oscar admires himself.''
b. ``Everyone admires himself.''
c. ``Oscar thinks he is a genius.''
d. ``Everyone thinks he is a genius/they are geniuses.''
293 Grammar and Implicature

e. ``Oscar thinks he is a genius and so does Humphrey.''

f. *``Oscar1 's father thinks he1 is a genius and Harry2 's father does
too'' hsc. think Harry2 is a geniusi.
g. Oscar (lx (x admires x)).
h. Everyone (lx (x admires x)).
i. Oscar (lx (x thinks x is a genius)).
On this analysis, the he in (33c) can be read as a variable, just as the
pronoun he (or as some dialects prefer, they) undoubtedly is in (33d). This
is demonstrated by the optional sloppy-identity reading of the ellipsed
VP in (33e), where Humphrey thinks Humphrey, rather than Oscar, is a
genius. Reinhart points out that in most circumstances a referential versus
variable reading of (33c) amounts to an ``equivalence'' in interpretations
(Reinhart 1983; Grodzinsky and Reinhart 1993: 74). The structures that
allow these interpretations seem to be just the ones where a pronoun is
c-commanded by an NP, matching in person, gender, and number, and
where Condition B is met (the pronoun's antecedent is outside the mini-
mal governing category). Thus in (33f ), where Oscar does not c-command
the pronoun, the bound-variable reading for the pronoun appears to be
ruled out (although coreference is always possible in these structures).
Now on this analysis only bound-variable structures (i.e., re¯exives
generally and pronominals c-commanded by compatible NPs) come
under the Binding Conditions, and these alone can be syntactically co-
indexed. Any other kind of coreference is just that, reference to an entity
that happens to have been referred to before (such NPs will thus have
distinct syntactic indices, although they may be interpreted as coreferring
for other reasons). This leaves the following central ``ungrammatical''
patterns entirely unaccounted for at a syntactic level, although they would
be ruled out by the classical Binding Conditions B (for (34a)) and C (for
(34b±d)). To avoid confusion with grammatical coindexing, I use italics
to indicate the relevant coreferential readings:
(34) a. ??``John admires him.''
Avoided structure: John admires himself.
b. ??``He thinks John is a genius.''
Avoided structure: John thinks he is a genius.
c. ??``He admires John.''
Avoided structure: He admires himself.
d. ??``John admires the guy.''
Avoided structure: John admires himself.
294 Chapter 4

To explain these constraints, Reinhart appeals to pragmatics. She pro-

poses (Reinhart 1983: 167) a pragmatic principle, a Manner maxim, ``be
as explicit as conditions permit.'' When applied to the structures in con-
sideration, this amounts to the following addressee's strategy of inter-
pretation: If the speaker avoids the bound-anaphora options provided
by the structure he is using, then, unless he has reasons to avoid bound
anaphora, he did not intend his expressions to corefer. Thus we have a
speci®c explanation for why He thinks that Oscar is brilliant resists a co-
referential interpretation: the speaker has avoided the parallel structure
Oscar thinks that he is brilliant, which would have allowed the strong
semantic interpretation. But because the source of the ban is pragmatic,
defeasibility is permittedÐespecially where the bound-anaphora inter-
pretation would in fact imply something di¨erent (Grodzinsky and Rein-
hart 1993: 79). Thus in apparent violations of Binding Conditions B and
C, as in the following examples, these cases are not coindexed, merely
coreferential, and further, they are interpretable as coreferring just be-
cause alternatives with re¯exives (which would receive the bound-variable
interpretation) would not be what the speaker meant:
(35) a. ``His wife despises Oscar, his children despise him, even Oscar
despises him.''
b. ``Even Oscar despises Oscar.''
Reinhart's pragmatics may account for the above patterns but will leave
many other patterns unexplainedÐnamely, wherever there is a preferred
pattern of coreference or disjoint reference not associated with a c-
commanding antecedent:
(36) a. ``John came but he left quickly.'' (coreference preferred)
b. ``John came but the man left (disjoint reference preferred)
quickly.''
c. ``His father admires him.'' (coreference preferred)
d. ``His father admires the Ajax (disjoint reference preferred)
goalkeeper.''
If we replace Reinhart's ad hoc pragmatic maxim with our principles of
entirely general application, we obtain a more encompassing account
with no specially crafted pragmatic principles. We already have an
account of the normal unacceptability of John1 admires him1 in terms of
the Q-opposition between the stronger, more informative re¯exive and
the pronoun. In the same way, we can assimilate to the Q-principle the
295 Grammar and Implicature

opposition between bound-variable pronouns and other referring ex-

pressions: given that the structure NP c-commanding Pronoun allows
a bound-variable interpretation, if the speaker uses a structure Pronoun
c-commanding lexical-NP, which does not permit bound anaphora but
would allow other coreference, then no such coreference is intended.
Reinhart's account thus gives us two kinds of possibilities for syntactically
encoding coreference: either through re¯exives or through c-commanded
pronouns. As a result, we get two pragmatic re¯ections of syntactic con-
straints on interpretation: the ®rst, which we may call the ®rst mirror-
image rule, gives us a strong disjoint reading wherever a re¯exive could
have occurred (an approximation to Condition B as sketched above), and
a second inference, the second mirror-image rule, gives us a strong disjoint
reading for the pattern Pronoun c-commanding lexical-NP. The mirror
image is given to us in both cases by the Q-principle: if you have not used
an informative expression where you could have, then such an informa-
tive interpretation is not intended (italics in the example sentences indi-
cate once again readings where the italicized NPs are coreferential):
(37) a. ``John admires him.''
Q-implicates: disjoint reference (®rst mirror-image rule)
b. ``He thinks John admires Bill.''
Q-implicates: disjoint reference (second mirror-image rule)
c. ``Sue thinks she is adorable.''
d. ??``She thinks Sue is adorable.''
e. ``Everybody loves Sue. Her husband thinks Sue is adorable, her
parents think Sue is adorable, why even she thinks Sue is
adorable.''
f. ``He just couldn't believe that the FuÈhrer could lose the war.''
g. ``Her parents think Sue is adorable.''
h. John (lx (x admires x))
[logical form for ``John admires himself.'']
i. Sue (lx (x thinks x is adorable))
[logical form for ``Sue thinks she is adorable.'']
The strong resistance to reading (37d) as containing coreferential NPs is
now explained on two grounds. First, there is a (relatively weak) M-
implicature from the use of Sue where she might have been used. Second,
there is a Q-implicature that, because a bound-variable structure might
have been used but wasn't, leads one to suppose that coincidence of ref-
erence was not intended. Despite its robust character, this inference is
296 Chapter 4

defeasible, thus pragmatic, as shown by (37e, f ) (after Bolinger 1979: 307).

And the Q-implicature does not occur in structures where bound-variable
interpretations are excluded, as in (37g), thus allowing coreference here
despite the order of pronoun before lexical NP.
Reinhart phrases these two bound-anaphora rules as a Condition A
(governing re¯exives) and a highly restricted Condition B governing just
the NP c-commanding Pronoun case. But a novel way of thinking of
Reinhart's analysisÐnot one she promotes, by the wayÐis that it really
amounts to claiming that English has something a bit like long-distance
Anaphors (LDAs) disguised as ordinary pronouns: these are the pro-
nominals under the bound-anaphora reading. The semantics is essentially
the same for both re¯exives and bound pronominals (her informal logical
forms are shown in (37h,i)), but they must ®nd their antecedents either
within the local domain (for re¯exives) or outside it (for pronominals
acting as bound variables). LDAs indubitably exist in many languages,
where an analysis in terms of features both pronominal and Anaphoric
are sometimes proposed, as we shall see below. In any case, the central
fact captured by Chomsky's Condition B, namely the unacceptability of
interpretations like John1 admires him1 , is on Reinhart's account (like
ours) a pragmatic fact. Consequently, we can think about Reinhart's
approach as requiring a boosted Condition A, in line with A-®rst prag-
matic account proposed here, which now has two syntactic/semantic
components: a rule of interpretation for re¯exives ( just as in the classical
Condition A), and a rule of interpretation for pronouns c-commanded by
NPs.30
The skeptic will raise various doubts about both Reinhart's analysis
and the proposed interaction with our three GCI principles. Let us deal
with the latter ®rst. How exactly does the pronominal variable case (or
rather the pattern Pronoun c-commanding lexical-NP) come within our Q-
principle? Two doubts might be raised: (a) it cannot participate in a Horn
scale because the pattern is an opposition between structures, not between
expressions; and (b) the construction that would have to be held to be the
informationally strong one, is in fact rather weak because it is ambiguous
(between a variable and referential interpretation). In answer to the ®rst
doubt, we have already noted various systematic oppositions between
structures, capable of yielding Q-inferences. For example, we considered
in chapter 2 Gazdar's (1979) analysis of clausal implicatures, where it is
proposed that there is an opposition between such structures as Since p, q
and If p, q: the use of the less informative conditional, which does not
297 Grammar and Implicature

entail its embedded clauses, will implicate that the speaker is not in a
position to use the more informative Since-construction, which does so
entail its embedded clauses. In a similar way, a speaker who uses the
structure Pronoun c-commanding lexical-NP where he might have used
the alternative lexical-NP c-commanding Pronoun (thus permitting gram-
matical coindexing), appears to wish to avoid that stronger interpretation.
Let us turn now to the question of how an ambiguous construction or
expression could be held to be the stronger term in a Q-opposition of this
kind. In fact it seems quite generally to be the case that if the most precise,
concise way to say that p is U, even where U is semantically ambiguous or
general over p and q, and the speaker instead chooses to say W, which is
semantically or formally distinct from U, then he will suggest that he
is avoiding U and its implicatures. Take for example the M-opposition
between kill and cause to die: the lexical causative does not entail direct
causation but is instead semantically general over direct and indirect
causation. Nevertheless, kill comes to be pragmatically specialized by I-
implicature to direct causation, so that cause to die suggests indirect cau-
sation by M-implicature. More closely parallel, most is a quanti®er that
has a certain generality of meaningÐspeci®cally, it does not semantically
exclude `all'; otherwise Most, in fact all, of the boys have scarlet fever
would be a contradiction. Still if one says Many of the boys have scarlet
fever, one implicates `not most', thus showing that the semantic generality
of most is not inhibitive of Q-implicatures. Similarly, the quanti®er some
may be held to be actually ambiguous between the scalar quanti®er in
opposition with all versus a weaker inde®nite determiner (usually un-
stressed as in I want some beer; see Milsark 1974; Gazdar 1980: 64; Horn
1997), but this does not inhibit its role in a Horn scale.
Another kind of reservation may be directed squarely at Reinhart's
account. There are long-recognized di½culties with the c-command
analysis, so that Reinhart's predictions about the general unaccept-
ability of coreference when pronouns c-command lexical NPs are not
correct in all cases (see Kuno 1987: 53±55; Pollard and Sag 1994: 246¨ ),
and nor it is the case that all Anaphors must be c-commanded (Zribi-
Hertz 1989). These problems are exacerbated when one comes to look at
noncon®gurational languages, or languages with ¯at sentence structures,
where the c-command analysis will generally fail (see, e.g., Levinson
1987b). For these reasons, the pragmaticist would be well advised to
search for explanations for universal patterns that may be linked to ma-
chinery that appears less language-speci®c, and I shall discuss in section
298 Chapter 4

4.3.2.4 proposals that replace c-command with functional sentence per-

spective notions like theme/rheme (Bolinger 1979) or with more semantic
notions like thematic hierarchies (Jackendo¨ 1972, 1992a).
The pragmaticist may remain agnostic about this syntactic/semantic
analysis. The point is that it is the kind of analysis many syntacticians
think is reasonable. And if so, we can see that by adopting both a stan-
dard Binding Condition A plus an additional analysis of the pattern NP
c-commanding Pronoun as a bound-variable structure similar in interpre-
tation to re¯exives, we can obtain all the rest of the patterns by means of
our pragmatic rules. The substantial change from our ®rst approximation
to English, sketched above, is that now con®gurational structures play a
role not only in the interpretation of pronouns where re¯exives could have
occurred (by the ®rst mirror-image rule) but also (by the second mirror-
image rule) in the interpretations of pronouns where re¯exives could not
have occurred (e.g., as subjects of subordinate clauses). Con®gurational
patterns then play a role in the pragmatic interpretations, but they do so
only by re¯ection (i.e., by an inference from a structure that might have
been employed but wasn't). What is of interest to the pragmaticist is the
general strategy pursued here: build in to the syntax just as much struc-
ture as you need, so that when it interacts with the independently moti-
vated pragmatics the overall account has two desirable features: it is
hugely economical, and it predicts defeasibility just where it is allowable.
For the syntactician, it ought to be reassuring that the kind of pragmatic
apparatus that would be needed to ground Reinhart's proposals is already
independently motivated.

4.3.2.3 Could Condition C be Syntactic After All? The second approx-

imation given above goes some way to dealing with two of the three
problems we foresaw for the A-®rst account: Reinhart's account gets a
crucial element of further con®gurational structure into the account of
pronoun interpretation, thus accounting for some of the asymmetries of
interpretation that cannot be explained by the sheer order of pronoun and
antecedent. At the same time, it accounts for the robustness of the ban on
coreference in Pronoun c-commanding lexical-NP structures, now attrib-
uted to a Q-inference reinforced by an M-implicature. The third problem,
the occasional overlap in distribution between Anaphors and pronouns, is
still unaccounted for, but we will consider this shortly.
Despite the enormous volume of work that presupposes the three
Chomskyan Binding Conditions, there has been very little attempt to
299 Grammar and Implicature

defend them against the threat of pragmatic reductionism that goes back
at least to Dowty (1980) and Reinhart (1983).31 One exception is Lasnik
(1991), who explicitly defends Condition C against Reinhart's account
and pragmatic reductionism in general. Lasnik argues that Reinhart's
pragmatic rule, which enjoins a speaker to use the most explicit means at
his disposal, would predict that *John1 thinks that I admire John1 should
be preferred over John1 thinks that I admire him1 . And if one wishes to
include extra information in an epithet, that too should license the full
NP, as in John1 thinks that I admire the idiot1 . These complaints ought to
evaporate given our more explicit pragmatic apparatus, where the infer-
ence to conjointness is based on the I-principle and disjointness inferences
are produced by Q- and M-inferences. Even if the repetition of full NPs
were more explicit in some sense than a resort to anaphora, the systematic
semiotics of GCIs would give rise to marked interpretations, including
disjoint reference. GCIs constitute a market economy of forms engender-
ing default interpretations; one cannot freely innovate without semiotic
consequences (see the conclusions to this chapter). Nor is there anything
very vague about the notions employed: Q-inferences are based squarely
on semantic informativeness (aÁ la Bar-Hillel and Carnap), whereas M-
inferences based on markedness and prolixity give one the complement of
I-inferences from corresponding unmarked forms. Thus the account goes
schematically as follows (here italics indicate the NPs whose coreference is
in question):
(38) a. ``John thinks that I admire him.''
I ‡> coreference
b. ``John thinks that I admire John.''
M ‡> disjoint reference
c. ``Howard Lasnik admires the author of `Essays on anaphora.' ''
Q ‡> disjoint reference by virtue of the Q-opposition hhimself,
the authori
M ‡> disjoint reference by virtue of the M-opposition fhimself,
the authorg;
But both inferences will be overridden by world knowledge, if
one knows that Lasnik is indeed the author of Essays on
anaphora.
d. ``Howard Lasnik is the author of `Essays on anaphora.' ''
Q and M-implicatures to disjointness overridden by semantic
entailment, the stated identity
300 Chapter 4

e. ``Lasnik thinks that I admire the Lasnik of `Essays on anaphora',

while in fact I admire the Lasnik of `Lectures on binding and
empty categories.' ''
M-implicatures to disjointness overridden by the semantics/
syntax
f. ``The man who can pull the sword out of this stone is King
Arthur.''
Hence, where a pronoun could have occurred and would have I-
implicated conjoint reference (see (38a)), the use of a lexical NP (as in
(38b)) will M-implicate that disjoint reference is intended if that ®ts our
background knowledge. Where a re¯exive could have occurred (as in
(38c)), that would have restricted the referential interpretations further
and would, ceteris paribus, have been more informative; its avoidance
thus Q-implicates disjoint reference, reinforcing an M-implicature to the
same e¨ect. Note however that names rigidly refer and that descriptive
phrases may also uniquely refer to the same individual; then naturally the
referential semantics of Scott admired the author of Waverly overrides any
implicature to disjoint reference (and similarly in (38c). If Condition C of
the Binding Theory was really a rule of grammar, then (38d) would be
ungrammatical, as indeed would all identity statements.32 However, be-
cause the Condition C e¨ects are mere implicatures, they evaporate in the
face of the con¯icting semantics of asserted identity. Pronouns in English
(unlike in some languages) do not permit determiners, relative clauses,
modi®ers, or complements, especially in the accusativeÐone cannot
(alas) in modern English say things like Him of the lion heart, so addi-
tional commentary will require a prolix head noun (as in (38e)). In ex-
ample (38f ), we have another identity statement, which can be read either
referentially or attributively; but either way it is grammatically ®ne, so it
is not possible to explain away these Condition C violations in terms of
attributive versus predicative uses.
Lasnik's second line of attack is that a pragmatic account should be
universal, whereas a syntactic account can be parameterized. Given
which, the facts of a language like Thai should be problematic to the
pragmatic account, for in Thai one can say sentences that are glossed as
`John1 thinks that John1 is smart' or even `John1 likes John1 '. However, a
pragmatic account depends on the total local economy of NP types, for
full lexical nouns can only implicate, on our account, by opposition to
zeros, pronouns, and re¯exives, and such implicatures fall away in the
301 Grammar and Implicature

face of con¯icting semantic factors. Chinese is another language that

freely permits full NP repetition, and we are fortunate to have a careful
study of the whole system from a pragmatic point of view (Huang 1994).
In Chinese, for example, one must repeat lexical NPs to express a bound-
variable interpretation (as in `Every mother-in-law has mother-in-law
di½culties'), but where the choice is free, there are implicatures, but not
necessarily ones of disjoint reference (Huang 1994: 199¨ ). But Lasnik's
argument here is on soft ground anyway. English has registers that per-
mit, nay require, lexical-NP repetition instead of pronouns, notoriously
for example in legal documents: on Lasnik's line of argumentation we
would need a di¨erent parameterization of the Binding Conditions not
just for each language but for each discourse genre. Crosslinguistic and
cross-genre variation is not the stu¨ of syntactic universals. A pragmatic
theory of the kind developed here, on the other hand, uses general prin-
ciples interacting with language-speci®c conventions to predict both
unmarked crosslinguistic tendencies and language-speci®c e¨ects, while
allowing for marked types of genre without grammatical violations.
Lasnik claims that Condition C should be split into two subconditions:
(a) R-expressions (full lexical NPs) may not be bound by R-expressions,
this held to be parameterized crosslinguistically; and (b) R-expressions
may not be bound by pronouns, this holding universally. Both are sup-
posed to hold in English, but we have already glimpsed exceptions. At
the risk of repetition, let us lay out the scope of the many kinds of
counterexamples.

4.3.2.3.1 Where a Lexical NP C-Commands Another Lexical NP Here

counterexamples are easy to ®nd (Reinhart 1983):
(39) a. ``Only Felix voted for Felix.''
b. ``Margaret's husband is Dennis.''
c. ``Dennis and Margaret have one thing in common: Dennis
thinks Margaret terri®c and Margaret thinks Margaret terri®c.''
d. ``Clinton said the President was not informed.''
e. ``Frog went to Frog's house; Toad went to Toad's house.''
On our analysis, the reason why repetition of lexical NPs with intended
coreference is not generally found is always (at least) a matter of M-
implicature: if a shorter, pronominal, or re¯exive form can be used to
I-implicate coreference, it should be. In some cases, such as those where
a re¯exive could have been used, there will also potentially be a strong
302 Chapter 4

Q-implicature to disjoint reference (as in (39a,b)). The point to be made

about these counterexamples is that in each case we can see why a re¯ex-
ive or a pronoun has in fact not been used. In (39a) substitution of the
second Felix with himself would have di¨erent truth conditions; in (39b)
the substitution would make the identity statement functionless; in (39c)
what is being discussed is the property of thinking Margaret terri®c (not
oneself terri®c); in (39d) the choice of the lexical NP the President is
understood to indicate that Clinton was speaking in an o½cial capacity;
in (39e) (from a children's book) the marked genre of ``motherese'' clearly
embodies an assumption that pronominal reference tracking can be di½-
cult for children.33

4.3.2.3.2 Where a Pronoun C-Commands a Lexical NP This constraint

is much more robust crosslinguistically, although some languages seem to
exhibit it at most weakly.34 English counterexamples are perhaps a little
harder to ®nd (compared to the preceding pattern), as we'd expect if
Reinhart's account is right, but they certainly exist:35
(40) a. ``Napoleon was endowed with supreme self-con®dence. He
never doubted that Napoleon was what France needed.''
b. ``Clark Kent hurried. He realized that Superman was urgently
needed.''
c. ``She thinks the wife knows best; he thinks the husband knows
best.''
d. ``He thinks the only reliable teacher of linguistics is himself.''
e. ``She thinks it unfair that the Chair of the Department always
has to be her.''
f. ``She says the youngest member of the family, i.e., herself, should
choose the holiday.''
g. ``He's doing what John always does.''
h. ``He talks the way that only the author of the Pisa Lectures
could possibly do.''
i. ``She's called Madonna.''
j. ``He became known as Napoleon.''
k. ``He succeeded his father as James Ist of England.''
l. ``He managed to become the President of the USA.''
m. ``He is the very man you ®rst dated.''
n. ``She thought Oedipus' mother was someone else.''
k. ``She knows what Joan Bresnan currently believes, because she
is Joan Bresnan.''
303 Grammar and Implicature

It is thus quite easy to get the pronoun±c-commanding±NP coreferential

reading under certain semantic/pragmatic conditions36Ðfor example,
where the NP expresses an identity relation that is not obvious either to
the addressee or to the subject, or where the NP makes explicit a double
identity, or where what is relevant is the identity of the same referent
under a second description, and so on.37 Any attempt to hang on to a
syntactic version of Condition C runs straight into the morass of these
many kinds of exceptions. A pragmatic account on the other hand spe-
ci®cally allows for defeasibility under principled circumstances.

4.3.2.4 Noncon®gurational Approaches to Binding We have so far

taken for granted the view that binding consists in the coindexing of a
c-commanding antecedent with an Anaphor. C-command is de®ned over
phrase structure, and the Binding Theory relies on the hierarchical nature
of phrase structure as a crosslinguistic universal. However, it is well
known that this view is problematic with respect to languages that either
have ¯at-sentence structure (i.e., with no VP constituent; see, e.g., Kiss
1991), or worse, have little overt constituent structure altogether, as in
languages with not only free phrase order but free word order (see Hale
1983). Therefore it is worth looking for alternative accounts phrased in
terms of other grammatical or semantic notions that may have greater
crosslinguistic currency. In this section, let us ®rst consider a non-
con®gurational language and how it reinforces the need for a pragmatic
account, and then turn to alternative concepts that might be employed to
explain the patterns currently explained in terms of con®gurational
notions.

4.3.2.4.1 Binding in a Noncon®gurational Language Guugu Yimithirr,

a language spoken in northern Queensland (see Haviland 1979), may
serve to remind us of the di½culty of extending the standard Binding
Theory to noncon®gurational languages. Like many of the Australian
languages, it does not exhibit the structural properties that are the basis
for a Chomskyan account. For example, there seem to be only two rather
weak arguments for any kind of constituency at all in Guugu Yimithirr
(Levinson 1987b); this makes the application of notions like c-command
and structural domain for binding impossible to de®ne in a con®gura-
tional manner. Many other notions central to such accounts either fail to
apply or are counterexempli®ed (like the y-Criterion, which forbids two
arguments to have the same semantic role).38 Even the contrast between
304 Chapter 4

pronouns and lexical NPs in Guugu Yimithirr seems to be essentially

di¨erent from the contrast between pronominals and R-expressions in
GB theory (pronouns in Guugu Yimithirr being more like nominals in
various ways, like permitting modi®ers). In all these properties it is by no
means unique; Guugu Yimithirr may stand as one of a large class of lan-
guages where the application of such concepts is essentially problematical.
Guugu Yimithirr has no unequivocal re¯exive construction; re¯exivity
is expressed by use of what is often called the antipassive in Australian
languagesÐthat is, a verbal a½x that converts a transitive verb (with an
ergative subject) into an intransitive verb (with an absolutive subject). The
interpretations of this construction are potentially very wide, ranging
from inde®nite or accidental agency to reciprocal or re¯exive action of an
agent upon itself:
(41) Bama gudhiira gunda-dhi.
people-ABS two-ABS hit-ANTIPASSIVE
`Someone hit the two people accidentally.'
`The two people hit themselves.'
`The two people hit one another.'
Interestingly, the re¯exive or reciprocal interpretations are much
strengthened if the emphatic particle -:gu is added to the subject (a point
we will take up in connection with a B-®rst pragmatic account). This
construction therefore o¨ers a way of expressing re¯exive action of a
subject upon what would, without the antipassive a½x, be expressed as a
direct object:
(42) John-ngun nhangu gunday.
John-ERG him‡ACC hit‡past
`John hit him.'
No other kinds of arguments, such as adjuncts or oblique NPs, can thus
be re¯exivized, unlike in the English sentence John gave Mary the book
about herself. The issue of having to ®nd an antecedent within a speci®ed
domain (as with English Anaphors) is thus irrelevantÐthere is no Ana-
phor (only an indication that one argument is missing) and the ``anteced-
ent'' must be the subject.
Thus not only is the coding of re¯exivity not directly explicit (but rather
semantically general over re¯exivity and unknown agency, etc.), it is also
restricted to the relationship between what in a transitive sentence would
be core co-arguments, speci®cally subject and direct object. The prag-
matic account that we have developed therefore predicts less certainty of
305 Grammar and Implicature

disjoint reference between a subject and an overt object of a normal

(nonantipassivized) transitive verb and, further, no presumption of dis-
joint reference between NPs in other positions. This prediction seems to
be generally correct; sentences glossing as `John hit John', while having a
preferentially disjoint intepretation, would seem more freely to allow
coreferential interpretations than in English, and English John gave Mary
the book about herself would have to be translated as `John gave Mary the
book about Mary'. Further, in Guugu Yimithirr, just as in English, we
express `John1 asked his1 father' with a possessive pronoun, there being
no way to re¯exivize the possessive; but the possibilities of coreference
here, which on the Binding Theory would follow under Condition B from
the pronominal being free in the NP his father, are clearly due to the ab-
sence of a pragmatic contrast (as, on our account, they are in English).
Similarly, under Condition C we would expect sentences glossing `His1
father told him1 that John1 should make it' to be out; but in fact they
appear to be ®ne, which indicates that we have no strong inference here
against coreferring lexical NPs with pronominal antecedents.
What is striking therefore is the association of a preferred disjoint
interpretation of pronouns and NPs just where the antipassive might
have been used to code a missing argument (and thus, inter alia, coiden-
tity of the arguments) but was not. This can be neatly accounted for by
Q-inference. There is no evidence that the GB apparatus of binding
domains and disjoint reference conditions are in any way involved. Refer-
ential dependence is indicated by a construction that signals a necessarily
omitted argument (all arguments in Guugu Yimithirr in any position
whatsoever can be optionally omitted).
In fact Guugu Yimithirr o¨ers much further evidence for the operation
of principles of pragmatic cohesion and contrast. The general anaphora
pattern is nicely exempli®ed in a system where zero-anaphora is general,
and thus a three-way contrast in semantic generality is available (lexical
NP, pronoun, NP-gap). NP-gaps induce coreferential interpretations
where possible but inde®nite reference where there is no available appro-
priate antecedent, whereas the use of a pronoun where a gap may have
been expected is often su½cient to suggest (via M-implicature) locally
disjoint reference. But the reader is referred to Levinson 1987a and 1987b
for further discussion. The point here is that in languages of this sort, the
grammar of anaphora appears reduced to a minimum. Without a prag-
matic account of the anaphora patterns, we have no account of the clear
preferences in interpretation at all.
306 Chapter 4

4.3.2.4.2 Alternatives to C-Command Noncon®gurational concepts,

like hierarchies of grammatical relations or thematic relations, have been
explored from time to time as alternatives to c-command conditions on
binding. One candidate often proposed is a hierarchy of thematic rela-
tions (Jackendo¨ 1972, Wilkins 1988, Reinhart and Reuland 1991, Kiss
1991). Suppose we have a thematic hierarchy Agent > Goal > Theme,
and a constraint that the antecedent must outrank the Anaphor on this
hierarchy; then we can understand the following sort of pattern:
(43) a. John talked to Bill1 (Goal) about himself1 (Theme).
b. *John talked about Bill1 (Theme) to himself1 (Goal).
The predictions from a thematic hierarchy are in many cases the same as
from the grammatical relations hierarchy, but Pollard and Sag (1992:
297±299) argue that where they diverge a theory based on grammatical
relations does better. Pollard and Sag (1992, 1994) therefore advance an
account in a con®guration-based theory of grammar, HPSG, in which
nevertheless con®gurational constraints are (ironically enough) sup-
planted by a hierarchy of obliqueness across grammatical relations (see
also Hellan 1991). Clearly, for this to work, genuine Anaphors have to be
restricted to those whose antecedents are co-arguments, and thus ante-
cedents that also bear a grammatical relation to the same predicate. Their
theory simply requires that an Anaphor must be bound to an antecedent
that is a less oblique complement of the same head. Thus we obtain not
only the standard examples as in (44a), but also where a verb sub-
categorizes for both a primary and more oblique object, the former may
perform as antecedent for the latter.
(44) a. John1 likes himself1 .
b. Mary described John1 to himself1 .
c. Mary talked to John about himself.
d. Mary talked to John about herself.
In (44c), where both prepositions are nonpredicative and share their fea-
tures with their objects, the acceptability is predicted because to-phrases
are held to be less oblique than about-phrases (Pollard and Sag 1994:
264). Similarly (44d) is acceptable; the subject Mary is less oblique than
the Anaphor in a subcategorized PP that shares the features of its object.
This theory does not cover the full distribution of English re¯exives
(many of which are not co-arguments with their antecedents), holding
that the residue (and by implication the long-distance Anaphors of other
307 Grammar and Implicature

languages) are pseudo-Anaphors (see discussion in Section 4.3.2.5.3).

Because these pseudo-Anaphors overlap in distribution with pronouns on
a coreferential reading, Pollard and Sag (1994) can adopt a Binding
Condition B that requires a nonre¯exive pronominal to be free in just the
domain where an Anaphor must be bound, thereby predicting a comple-
mentary distribution between true Anaphors and pronouns.39 We could
therefore without problems apply an A-®rst pragmatic account to this
theory, and reduce their Condition B to implicature, along the lines
sketched in our ®rst approximation.
Such grammatical alternatives to con®gurational accounts are not the
only possibilities. At the outset, it was shown how linear order seems to
play a crucial role in cross-sentential anaphora, presuming some rule of
the kind ``®rst introduce your referents with full forms, thereafter use
reduced forms.'' Intrasententially, however, this con¯icts with many an
example, as notedÐthere are exceptions in both marked phrase orders
(Near him, John saw a snake) and in cases where the pronoun is embedded
in the subject (His children admire John). In early transformational
grammar, these observations gave rise to the ``precede and command''
analysis, which was observationally fairly good (see Kuno 1987: chap. 2
for review) and which preserved the ordering requirement, arguably an
essentially pragmatic or discourse element, later lost in the c-command
account, which makes the ordering facts epiphenomenal. It would seem
worthwhile to explore the possibility of reversing the argument, thus
making the con®gurational facts epiphenomenal from discourse princi-
ples. Something very near to the right principle is given by Bolinger (1979:
306), using the notions theme and rheme, and reidenti®cation (i.e., the use
of a full form after a referent has already been introduced in the dis-
course): don't reintroduce the topic of the theme in the rheme (or ``the
topic may be reidenti®ed easily in the theme, but in the rheme only if the
theme lacks a normally topical form''). As Reinhart (1983: 97) notes, this
gives a good account of many apparently con®gurational constraints, but
she notes that it would need extending to handle cases where both ante-
cedent and pronoun are in the rheme (e.g., Rosa locked him in Ben's room
where him cannot be identi®ed with Ben). My own suggestion (Levinson
1991: 114±115) was that there is also a gradient within the rheme, where
thematic role hierarchies would restrict coreferential possibilities (this ties
into Jackendo¨ 's (1972) observations). It is not at all clear that such an
account could not be made to work, but we should concede that it needs
further exploration (see also Van Valin 1990: 178±183).
308 Chapter 4

4.3.2.5 Where Complementary Distribution Fails: Long-Distance Ana-

phora and Related Phenomena Our pragmatic machinery nicely pro-
duces the very generally attested complementary distribution of pronoun
and Anaphor (by the Q-principle) and of pronoun and lexical NP (by the
M-principle). But our approximations to English appear to fail in a cru-
cial way: it is well known that there are marginal, but systematic, contexts
in which the pronoun and the Anaphor overlap in distribution. How can
the account be generalized to cover these cases? In the literature, there are
four main classes of context where this overlap between Anaphor A and
pronoun P has been noticed in English (Zribi-Hertz 1989: 698f ):
1. So-called picture-noun re¯exives:
a. A or P embedded within the subject of a subordinate clause:
They thought that pictures of themselves/them would be on sale.
b. A or P are embedded objects in an existential subordinate clause:
They said that there are pictures of themselves/them in the post o½ce.
2. Genitives (pronouns/reciprocals):
They love their/each other's wives.
3. Certain locative prepositional phrases:
They hid the money behind themselves/them.
4. So-called emphatic contexts, involving contrast, comparison and
restrictions of predication:
He thinks Mary is taller than himself/him.
He believes that men like himself/him are rare.
He believes that it was not sent to either him or her/either her or himself.
He thinks Mary loves him/himself, not Peter.
He thinks Mary hates even him/himself.
Enormous ingenuity has gone into seeking a way to reconcile these facts
with the Binding Conditions. Four main tacks have been taken:
1. One can try to ®nd ways in which what appear to be the same contexts
are in fact distinguished (e.g., there is an empty subject in the small
clause [e pictures of them] but not in the NP [ pictures of themselves], as
in Chomsky 1986: 173);
2. One may hold that emphatic or other pseudo-Anaphors fall outside the
theory (see Zribi-Hertz 1989 for discussion, also Bouchard 1983, Baker
1995);
3. One may hold that ``downstairs'' Anaphors are raised in LF, where the
Binding Conditions apply (see Harbert 1995: 203±208 for review);
309 Grammar and Implicature

4. One may tinker with the domains in the Binding Conditions A and B
so that just the right overlap is predicted (as in Chomsky 1986).
The last approach has been the most popular because it seems inde-
pendently motivated by the phenomenon of long-distance anaphora in
other languages. However, Chomsky's own response was directed
squarely at handling English examples like the following:
(45) a. They1 read [their1 books].
b. They1 read [each other's1 books].
c. The professors1 thought that [books about themselves1 ] would
be on sale.
d. The professors1 thought that [books about them1 ] would be on
sale.
To handle these, Chomsky (1986) suggested revising the Binding Con-
ditions utilizing the notion of a ``Binding Theory compatibility require-
ment,'' which e¨ectively extends the range of Anaphors slightly by
requiring that there is an actual binder or potential antecedent within the
domain D. Thus in the ®rst two sentences above, both an Anaphor and a
pronoun can occur in the same slot bound to the same antecedent. This is
because, on the 1986 account, each other simply cannot be coindexed
within the bracketed NP domain; therefore the actual domain is the ma-
trix sentence. The pronoun however must be free in its domain; therefore
the issue of an obligatory potential binder doesn't arise. The pronoun is
thus free in the NP which is its domain, and can be coreferential with the
subject of the matrix sentence just like the Anaphor. A similar analysis
can then be given of (45c,d): in (45c) the compatibility requirement for a
potential binder within the governing category (marked with square
brackets) is not met, so the Anaphor may be bound by the matrix subject,
but in (45d) the pronoun must still be free within the same domain.
A pragmatic account is not the slightest embarrassed by the failure of a
contrast between (45a) and (45b) in coreferentiality. This is because the
reciprocal and the possessive pronoun do not mean the sameÐthe pairing
of books and readers in (45a) is unconstrained, but in (45b) each reader
must read the books owned by the other. If one means the latter, one will
have to say it. There may be as consequence a (Q-based) assumption that
the plain possessive does not denote that particular reciprocal situation,
but in both cases the anaphoric link-up is the same. The reasons then are
entirely general: where two Q-contrasting expressions (or constructions)
310 Chapter 4

di¨er in meaning on two dimensions (here (i) necessary vs. optional ref-
erential dependency, (ii) reciprocal vs. unspeci®ed distribution), the use of
the informationally weaker expression is justi®ed if one is seeking to avoid
at least one of the stronger meanings. Many languages (e.g., Norwegian)
have possessive re¯exives that will contrast on just the one dimension with
the reciprocal, and on the other dimension with the possessive (non-
re¯exive) pronoun, which will then tend to be read as disjoint (Hellan
1991).
But the overlap in (45c,d) is more problematic for the simple A-®rst
pragmatic account developed so far. In fact, however, the pragmatic
explanation, which presumes some extension of the range at which
re¯exives can be bound, will turn out to be exactly the same in kind as
that given for the reciprocal versus possessive pronoun: the re¯exive here
contrasts with the pronoun on two dimensions, necessary referential
dependency and a subjective perspective called logophoricity in the litera-
ture. The general nature of this second dimension of contrast becomes
clearer when we look at other languages, which we shall shortly do.40
Meanwhile, note that Chomsky's (1986) emendation of the Binding Con-
ditions will not in fact achieve any generality of coverage. Reciprocals can
occur deeply embedded from their antecedents (as in They1 feared that the
prosecutor would reveal that Noriega had contributed to each other1 's
campaign funds; see Pollard and Sag 1994: 257), and even the nonre¯exive
pronoun in They1 put the books [beside them1 /themselves1 ] escapes the
account, because it ought to be free throughout the sentence (the brack-
eted PP not constituting a binding domain on this theory; Harbert 1995:
192). Within the GB framework there have been many attempts to ®nd
di¨erent kinds of solutions to the empirical inadequacy of the classical
Binding Theory (see Harbert 1995 for review). In fact, as we shall see,
English re¯exives are routinely long-distance bound, as are re¯exives in
many other languages. The patterns in these other languages, to which we
now turn, also give us a better idea of how to develop a pragmatic
account that can handle the kind of neutralization between disjoint/
conjoint reference that we ®nd in the above contexts in English.

4.3.2.5.1 Long-distance Anaphora and Logophoricity It has been clear

almost from the outset that the Binding Theory runs into di½culties with
so-called long-distance Anaphors in many languages (see, e.g., Maling
1984). Long-distance anaphora involves the occurrence of referentially
dependent anaphors which are just like Anaphors in that they must be
311 Grammar and Implicature

bound, and yet violate (the classical) Condition A of the Binding Con-
ditions by having antecedents that lie outside of their minimal governing
categories. Pretheoretically we may coarsely distinguish between di¨erent
kinds of referentially dependent Anaphors: those locally bound re¯exives
that require co-argument antecedents (like prototypical cases of English
re¯exives), those that require binding within the clause but can occur in
adjuncts (like English re¯exives embedded in PPs), and those (so-called
non-clause-bound re¯exives [NCBRs]) that may be bound outside the
immediate containing clause altogether (like Icelandic sig, Chinese ziji).
The last are of course particularly challenging to the theory, and a great
deal of work has gone into trying to understand the phenomena. The
distribution of such long-distance Anaphors may include the following,
as indicated by the English glosses with self standing for the Anaphor
(although the European languages tend to lack a nominative long-
distance Anaphor as in (46b)).
(46) a. `John hit self.'
b. `John said that self would come.'
c. `John said that Bill would hit self.'
Unfortunately, on close examination the patterns are fairly heterogenous,
even within the European languages (see the useful survey in Reuland and
Koster 1991). Thus in the Germanic languages (Icelandic, Norwegian,
Dutch, etc.) the long-distance Anaphors are simplex morphemes (like sig,
seg, zich) which may occur clause-bound, and prefer subject antecedents;
when bound outside the clause they are not necessarily in complementary
distribution with pronouns (i.e., pronouns can be conjoint with the same
antecedent). But despite the commonalities, even here the details vary in
intricate ways, according to such factors as the ®niteness or the subjunc-
tive nature of higher clauses, and according to the availability of other
(often compound) re¯exives in clause-bound contexts. Thus although
Icelandic, Danish, and Gothic allow structures in English gloss like (47a),
Icelandic exceptionally allows binding across a tensed-clause boundary as
in (47b) (Harbert 1995: 193±194).
(47) a. `John1 asked Peter to shave self1 .'
b. `John1 said that I had (subjunctive) betrayed self1 .'
Once we turn to languages further a®eld, the phenomena look rather di¨er-
ent. For example, Chinese has long-distance re¯exives that can be ``bound''
from an arbitrary distance by an antecedent that is not necessarily in a
312 Chapter 4

c-commanding position;41 such LDAs can be either simplex or compound

(both ziji and ta ziji), and if compound the antecedent need not be a
subject (Huang 1994: 76±78, 196±199). The Japanese LDA zibun also
exhibits a great freedom in the circumstances under which it must be
bound (Mazuka and Lust 1994).
Despite the substantial variation, LDAs do show certain similarities:
(a) when they are simplex, they lack (person, number) agreement fea-
tures;42 (b) when they have local antecedents, they contrast in reference
with pronouns,43 but when they have long-distance antecedents, they do
not necessarily contrast in reference with a pronoun (i.e., complementary
distribution with respect to coreference fails);44 (c) nevertheless, wherever
both a pronoun and a long-distance Anaphor could occur in a grammati-
cal slot with the same reference, the Anaphor always contrasts in mean-
ing with the ordinary pronoun, the associated meanings having something
to do with emphatic contrast, empathy, or protagonist's perspective, sub-
jective point of view, and so on.
The ®rst fact, the tendency to lack concord features, is presumably a
re¯ex of the fact that LDAs are (by de®nition) necessarily referentially
dependent on a discourse antecedent, unlike pronouns which may have
independent reference and with which LDAs contrast. But it is the second
and third facts that are crucial here: the complementary referential possi-
bilities between pronoun and re¯exive break down conclusively in these
long-distance environments, permitting (in gloss) both `John said self
would come' and `John said he would come' for the coreferential reading
(or for languages without nominative Anaphors, `John said Bill loves self '
and `John said Bill loves him'). This neutralization of the referential con-
trast between Anaphor and pronoun is of course problematic for the
classical Binding Theory, and it has become the focus of intensive re-
search and a substantial literature (see, e.g., Harbert 1995: 193¨; Huang
1994: chap. 4, in preparation; and Koster and Reuland 1991 for review
and references).
There are at least three major responses within the framework. First,
one can reclassify LDAs as something other than the Anaphors of Bind-
ing Theory (Bouchard 1983). Or, one can assume that in underlying LF
the Anaphor is actually locally bound just as presumed in the classical
theory (and is thus raised at LF; Pica 1987). Finally, one can parameterize
Anaphors, so that they have distinct binding domains in di¨erent lan-
guages, allowing that even di¨erent Anaphors within a language (e.g.,
313 Grammar and Implicature

complex vs. simplex ones) might have di¨erent binding domains (Wexler
and Manzini 1987).
All of these responses have their own problems (see, e.g., Huang 1994,
chap. 4). The ®rst option requires, for example, that one treats homo-
phonous expressions as distinct lexical items (e.g., Anaphors when locally
bound, referentially dependent pronominals elsewhere), and that one
minimizes the di¨erence between pronominals (with their exophoric and
deictic potential) and LDAs (which have only referential dependence and
distinct distribution). The second option, using a raising at LF analysis,
has been applied ingeniously to some of the Germanic languages (see
Harbert 1995: 203¨ ), but it runs into overwhelming problems with lan-
guages like Chinese (on which there has also been considerable work; see
Huang 1994: chap. 4 for review). The last option gives the best hope of
empirical coverage, but at the expense of an ad hoc variation of domains
which lacks the system-wide implications that parameters are supposed
to have in the Principles and Parameters framework. Some analysts are
optimistic that we may need just three kinds of domains: one for clause-
bound re¯exives (like the prototype English exemplars) as in classic
Binding Condition A, one for LDAs consisting of the ®rst ®nite clause to
be stated as a variant Condition A, and one for homophonous pseudo-
re¯exives that are subject to discourse conditions rather than the Binding
Conditions (Reuland and Koster 1991). The admission that many LDA
look-alikes escape the Binding Conditions is unavoidable because they
often seem to be ``bound'' by non-c-commanding antecedents, sometimes
outside the sentence altogether. But it is damaging because it makes the
claim that the Binding Conditions have, say, just two alternate domains
pretty unfalsi®able. Most commentators would concede that despite the
energetic exploration of these three di¨erent ways in which the Binding
Theory might be reconciled with the phenomena of long-distance ana-
phora, we have at present no convincing overall theory from a syntactic
point of view.
Now a pragmatic account of the (A-®rst) kind developed here would at
®rst sight appear to be in just as much trouble from the failure of the
complementarity of LDAs and pronouns as the Binding Theory accounts.
Clearly, this is the same kind of problem we found with the English
re¯exives, which turned out to overlap in distribution (on a coreferential
reading) with pronouns in a few special contexts; but here in the case of
LDAs the overlap is not marginal, but systematic, and for most of the
314 Chapter 4

relevant languages holds for all the long-distance loci. But the LDA facts
very clearly signal the direction in which the pragmatic theory can be
rescued: as mentioned, even where the Anaphor and the pronoun can
have the same reference, they always contrast in meaning in some other
dimension. And the way in which they contrast is similar to the contrast
in reference: the Anaphor is necessarily referentially dependent, whereas
the pronoun may or may not be; the Anaphor carries additional perspec-
tival information, which the pronoun is unmarked for. In short, in both
the referential and nonreferential contrasts, the Anaphor is information-
ally richer than the pronoun. This opens up the possibility that just as we
have a Q-inference contrast from the use of the pronoun to the absence of
conjoint reference that would have been signaled by an Anaphor in some
loci, so we have a Q-inference in other loci from the use of the pronoun to
the assumption that the perspective associated with the Anaphor is not
intended. Let us examine the meaning di¨erence in a little more detail.
The relevant meaning distinctions associated with LDAs were ®rst
noticed in connection with the logophoric pronouns and clitics of West
African languages (HageÁge 1974). Logophors are special expressions used
in subordinate clauses to indicate coreferentiality with a matrix subject,
but under speci®c conditions: usually the matrix verb must be a verb of
saying or a verb of mental state, and thus the prototype context is of the
sort `John1 said that logophor1 would come', where the logophor may in
some cases actually have evolved from a ®rst-person pronoun. HageÁge
noted that the logophor encodes the subjective perspective of the partici-
pant, and noted the parallel to long-distance anaphora in languages like
Japanese. In fact, recent work has shown that the similarities are far-
reaching (Sells 1987; Stirling 1993: chap. 6), and we will return brie¯y to
consider the distinctions between switch-reference systems, LDAs, and
logophors. Logophors are licensed by speci®c semantic contextsÐtypi-
cally a verb, or complementizer associated with a verb, of communica-
tion, epistemic or mental state, or perception and what is clearly signaled
is the perspective and subjective state of the protagonist, the so-called
logocentric NP or antecedent, not the current speaker. Sells (1987) ana-
lyzes these semantic contexts in terms of three features: whether or not
there is an internal, reported speaker or source of communication,
whether or not there is an internal self whose mental state is described,
and whether or not there is an internal pivot or relativized deictic center.
The strongest logophoric contexts are those in which there is a text-
internal source, self, and pivot, and the weakest are those in which there is
315 Grammar and Implicature

only an internal pivot; intermediate are contexts in which there is an ex-

ternal source, but internal self and pivot (i.e., the external speaker reports
the mental state and perspective of a protagonist). These contexts are
forced by, respectively, verbs of communication, verbs of psychological
attitude, and constructions (like verbs of perception) forcing the perspec-
tive of an internal protagonist. Logophoric languages appear to allow
logophors progressively down this implicational scale (Stirling 1993: 259±
260).
Just as in LDA, logophors and pronouns are not necessarily in com-
plementary distribution with regards to reference: whereas the logophor
must be coreferential with its antecedent (the so-called logocentric NP) a
plain pronoun may or may not be. As Stirling (1993: 266±267) puts it:
In all languages in which such a contrast is possible, it is associated with a mean-
ing distinction of a remarkably consistent kind: if the ordinary pronoun is used, it
indicates that the speaker has assimilated the proposition being reported into her
own scheme of things, and accepts its truth and/or approves of its content. If the
logophoric pronoun is chosen, it indicates that the speaker has not assimilated the
proposition into her knowledge base, and does not necessarily accept its truth or
approve of its content: in some sense, responsibility for its truth, content or lin-
guistic characterization is distanced, and left to the referent of the logophoric
pronoun.
Not surprisingly, then, there are collocational constraints with evidentials
(O'Connor 1993: 225f ) and with mood (Stirling 1993). The subjective
point of view associated with the logophor can be thought of as a complex
feature [‡logophoric] which is imposed by, or in concord with, such a
feature associated with logophoric verbs, the feature being in turn char-
acterized in terms of the atomic features source, self, and pivot mentioned
above (Sells 1987; for a simpler alternative, see Stirling 1993: 282¨ ).
Long-distance Anaphors and logophors have a great deal in common,
as was already noticed by HageÁge (1974). First, the semantic constraints
on LDAs and logophors appear to be very closely similar. For example,
in Icelandic, the LDA sig occurs especially in subjunctive clauses, where
the mood signals that the point of view is not the speaker's but that of the
subject of the relevant verb, as evidenced by the deniability of the truth of
the proposition (Stirling 1993: 268). Second, the perspectival distinctions
associated with logophors versus pronouns also correctly predict many
aspects of the distribution of LDAs versus pronominals (see, e.g., Kuno
1987, Sells 1987, Hellan 1991, and references therein for studies of Japa-
nese, Korean, and Germanic long-distance anaphora). For example, the
316 Chapter 4

Chinese LDA ziji appears to be very closely linked to logophoric contexts

(Huang 1994: 187¨ ), occurring especially under verbs of saying and
thinking or feeling, and just like the West African logophors, sometimes
being triggered by complementizers derived from verbs of saying. Indeed,
so close is the parallel that one may ask whether they are not in fact the
same basic phenomenon, the main di¨erences apparently being that
whereas LDAs may usually be locally bound, logophors are usually in
complementary distribution with locally bound re¯exives (Stirling 1993:
259). Scholars in the Binding Theory tradition have also noted the paral-
lels, but their reaction has been, for example, to distinguish between dif-
ferent uses of LDAs: (i) those uses of LDAs, subject to the Binding
Theory, that have antecedents within the ®rst ®nite clause, and (ii) those
so-called logophoric uses that occur beyond this maximal extent of
syntactic binding and are controlled by discourse factors (Reuland and
Koster 1991: 22±24).45 This retreat from a global structural theory is
motivated by the fact that both LDAs and logophors can have their
antecedents outside c-commanding positions, even outside the sentence
altogether, and thus elsewhere in the discourse (see examples in Stirling
1993: 262±263; Huang 1994: 195; Mazuka and Lust 1994).
Let us return then to the pragmatic account. The proposal (which
originates with O'Connor [1987]1993) should now be clear. Long-distance
Anaphors have two essential semantic properties: ®rst, they are always
referentially dependent; second, they are logophoric. That these are con-
junctive conditions is clear, from Kuno's (1987: 141) analysis of Icelandic
LDAs, for example:
A re¯exive pronoun can be used in a complement clause in Icelandic if it is
coreferential with a [‡logoÿ1] NP of the main clause and if the complement clause
is an expression of the logophoric NP's point of view rather than the speaker's
rendition of it

(where Kuno's [‡logoÿ1] feature marks the source or self in Sells's

terms). Both semantic features contrast with ordinary pronouns, which
are unmarked for referential dependency and unmarked for logophor-
icity, or schematically:
(48) LDAs Pronouns
[‡referential dependency] [Greferential dependency]
[‡logophoric] [Glogophoric]
We now have a privative opposition on two dimensions, and the use of
the pronoun will indicate that the more informative LDA is being
317 Grammar and Implicature

avoided, because one or the other of the features associated with the LDA
does not applyÐin other words, the conjunctive de®nition of the LDA
yields a disjunctive Q-inference from the use of the pronoun. We may thus
maintain the A-®rst analysis in all its essentials; we simply recognize that
a Q-based contrast should be a contrast in informativeness, but it need
not be a contrast in reference. Where the stronger item contrasts with the
weaker expression in more than one way, then the use of the weaker form
will be associated with avoidance of one or more of the contents of the
stronger expression. Schematically, using English glosses where self stands
for the LDA and him for the nonre¯exive pronoun:
(49) a. i. `John1 likes self1 .'
Coreference required by Condition A within an enlarged
binding domain.
ii. `John1 likes him2 .'
Q-contrast with the LDA in reference
b. i. `John1 says that self1 will come.'
Coreference required by Condition A within an enlarged
binding domain.
ii. `John1 says that he1=2 will come.'
Q-contrast with the LDA in logophorocity, plus or minus
contrast in reference.
Such an account will also hold for genuine logophors, where the pronoun
may implicate either disjoint reference or the adoption of the speaker's,
rather than the protagonist's, point of view, thus exhibiting the pattern in
(49b) (the pattern in (49a) is usually preempted by alternative short-range
re¯exives). It may also hold for switch-reference markers, which are again
closely related to both logophors and LDAs.46 Switch-reference systems
are systems of (normally) verbal su½xing that indicate whether a (usually
subordinate) clause has the same or di¨erent subject as the main clause;
they are found in areal clusters in North America, Africa, New Guinea,
Australia, and a few other locations (see Haiman and Munro 1983 for a
collection of case studies, and Stirling 1993 for analysis and survey). In
fact, a Gricean account of disjoint reference in switch-reference systems
was independently proposed by O'Connor (1993) for Northern Pomo,
which also has long-range re¯exives. O'Connor noted that the long-range
re¯exives overlap in distribution with coreferentially interpreted pro-
nouns, and she proposed the analysis we have just adopted, viz. that
318 Chapter 4

pronouns implicate a contrast either in reference or in logophoricity; but

she went on to extend the account to switch-reference systems.
The reason to suspect pragmatic factors in switch-reference systems
is that the di¨erent-subject marker (DS) often contrasts with the same-
subject marker (SS) in ways other than just reference, and in these cases
conjoint reference is possible with DS marking. Quite typically, the SS
marker is associated with a sequence of actions performed by the same
actor in proximate locations, so that it can be seen as one complex action
(as in the gloss `He went to the store and he bought-SS some whiskey'),
following a pattern we have already associated with the I-principleÐcor-
eference is associated with cohesive interpretations. In line with this ex-
pectation, SS markers tend to be the relatively unmarked morphological
option (sometimes zero) associated with sequentiality, whereas the more
marked (morphologically complex) DS markers are associated with
simultaneous actions by di¨erent actors (see Stirling 1993: 30±32). So far
we have the familiar I- versus M-inference pattern, discussed at length in
chapter 2, but apparently here fully grammaticalized. But there is also
evidence in some switch-reference systems with overt SS marking for a
privative opposition, with the DS marker actually unspeci®ed for dis-
jointness, but picking up that presumption by Q-inference. So we would
have an opposition of the following kind (where the ``single complex
action'' feature may be variously interpreted in di¨erent languages to
involve same degree of agency, or close spatio-temporal cohesion):
(50) SS marker DS marker
[‡referential dependence] [Greferential dependence]
[‡single complex action] [Gsingle complex action]
If this is correct, we would expect the disjointness of the DS marker to be
defeasible, and in many languages this seems to be the case. Further, we
would expect that there would be two main reasons to avoid the SS
marker: either the two subjects are disjoint in reference, or the two actions
do not form a single complex action, one action being less intentional or
separated in space or time from the other. DS markers then implicate a
disjunction, either disjoint reference or two distinct actions. Again, such
cases of DS markers unexpectedly allowing coreference coupled with a
contrast in action-coherence seem to be well attested, although the details
are complex, and depend of course on the exact speci®cation of the fea-
tures built into the SS marker in particular languages (see Stirling 1993:
33±34, 50, 98±114, 150±153). In short, for some switch-reference systems
319 Grammar and Implicature

with systematically opposed overt SS and DS markers, we may be able to

maintain just the same kind of analysis as for LDAs and logophorsÐ
coreference is stipulated by the SS marker (in line with the A-®rst
pragmatic analysis) and the DS-marker, which is semantically general,
picks up its disjoint reference implications by pragmatic opposition (see
O'Connor 1993 for a case, Nichols 1983 for discussion).
There are thus similarities between LDAs, logophors, and SS markers
in switch-reference systems. This reopens the fundamental question,
raised by Dowty (1980) and Reinhart (1983), but at ®rst sight closed by
the very existence of switch-reference systems, concerning whether we can
state an important linguistic universal that holds that disjoint reference
systems are never fully grammatically speci®ed whereas coreference sys-
tems may be so. It seems in fact unlikely that this is an interesting prop-
erty of Universal Grammar, unpredictable on other grounds. Rather, it is
probably a tendency that follows on Zip®an grounds of functional econ-
omy: given a grammatical speci®cation of the informationally richer pole,
namely coreference, disjoint reference can be signaled by avoiding that
grammatical construction.47 Just as the fact that some, by Q-contrast to
all, default implicates `not all' and thus e¨ectively blocks by redundancy
the lexicalization of a new form *nall, so the grammaticalization of Ana-
phors, logophors, and SS-markers may tend to make redundant the gram-
maticalization of systematic, specialized disjoint-reference mechanismsÐ
existing terms in the privative opposition (between conjoint-reference
marker and unspeci®ed default) will do the job.48 This is the logic of
the A-®rst pragmatic account. On the other hand, given that these disjoint
inferences from the weaker term of a privative opposition are general-
ized, this existing expression may well come to have its GCIs convention-
alized over time, as arguably in switch-reference systems or indeed even
in the Condition B behavior of English pronouns.

4.3.2.5.2 The Long-distance Uses of English Re¯exives Let us now re-

turn to the puzzles of English re¯exives. Can we ®nd meaning di¨erences,
similar to those associated with LDAs and logophors, between re¯exives
and pronouns when they can both occur in the same locus with a cor-
eferential meaning (as in John hid the book behind him/himself )? Kuno
(1987), marshalling a large number of earlier observations by Ross (1970),
Cantrall (1974), and others, has argued vigorously that we can. For ex-
ample, we noted above three classes of much examined cases where
re¯exives and pronouns have overlapping distribution (leaving aside the
320 Chapter 4

overlap of reciprocals and pronouns): locative PPs, picture-noun re¯ex-

ives, and various emphatic contexts.49 In each of these cases, Kuno points
out that there are reasonably clear if subtle meaning di¨erences between
the corresponding sentences with pronouns or re¯exives. For example, in
those locative PPs that allow this alternation, the re¯exives set up a pro-
tagonist's perspective on events, which is intuitively clear enough in (51a)
but even makes a di¨erence to truth conditions in (51b) (after Cantrall
1974, cited in Zribi-Hertz 1989: 704; see also Kuno 1987: 153).
(51) a. i. ``He1 pushed the brandy away from him1 .''
Gloss: observer ``camera-angle''
ii. ``He1 pushed the brandy away from himself1 .''
Gloss: protagonist's ``camera-angle''
b. i. ``The women1 were standing in the background with the
children behind them1 .''
Gloss: observer's perspective, with children beyond the
women
ii. ``The women1 were standing in the background with the
children behind themselves1 .''
Gloss: women's perspective, with children at their backs
Similarly, the picture-noun re¯exive cases also reveal a clear meaning
di¨erence between the re¯exive and the pronoun (Kuno 1987: 126, 174),
often rather subtle consequences of the subjective point of view associated
with the re¯exive, as in (52a). Consequently there are collocational con-
straints, where the sentential context makes it di½cult to take the per-
spective of the protagonist referred to with the re¯exive, as in (52b).
(52) a. i. ``John1 heard some strange gossip about him1 on the radio.''
Gloss: ``strange'' in the estimation of the speaker
ii. ``John1 heard some strange gossip about himself1 on the radio.''
Gloss: ``strange'' in the estimation of John
b. i. ``Mary told John that there was a picture of himself in the post
o½ce.''
ii. ??``Mary said about John that there was a picture of himself in
the post o½ce.''
The so-called emphatic re¯exives reveal a similar pattern (see Zribi-Hertz
1989: 698f for further examples and references), with the protagonist
perspective associated with the re¯exive often engendering presumptions
of restrictive exclusivity:
321 Grammar and Implicature

(53) a. ``John1 believes the letter was sent to him1 and Mary.''
Gloss: observer perspective: the letter was also perhaps sent to
others
b. ``John1 believes the letter was sent to himself1 and Mary.''
Gloss: participant perspective: letter not sent to others
To see that these rather subtle semantic factors are robust enough to
engender collocational constraints with other perspectival indicators,
consider the following pairs:
(54) a. ``John is happy that Anne has come so far to see himself.''
b. ??``John is happy that Anne has gone so far to see himself.''
Deixis not relativized to John's point of view.
(55) a. ``According to John, the paper was written by Anne and
himself.''
b. ??``Speaking of John, the paper was written by Anne and
himself.''
Antecedent embedded in observer-perspective phrase.
(56) a. ``John said frankly that Anne was infatuated with himself.''
b. ??``John said apparently that Anne was infatuated with himself.''
Observer-perspective associated with evidential apparently
(57) a. ``The Prime Minister1 was worried that The Sun would publish
that damn compromising memo about himself1 .''
Re¯exive collocates with positive psych-verb and expletive
determiners
b. ??``The Prime Minister1 was unaware that The Sun would publish
the compromising memo about himself1 .''
Negative psych-verb and absence of other perspectival markers
suggests observer perspective
Note too that, because a portion of discourse must maintain only one in-
ternal, subjective viewpoint, we have a natural account of why we cannot
easily have two Anaphors with distinct reference in the same sentence
(Zribi-Hertz 1989: 713¨; Pollard and Sag 1992: 275) as in ??John traded
Mary pictures of herself for pictures of himself.50
Zribi-Hertz (1989), utilizing a textual database, shows that there is little
reason to resist the conclusion that English re¯exives exhibit LDA be-
havior. The intuitions of clause-boundedness for English Anaphors come
about, she claims (1989: 722±723), because a clause is the minimal point-
of-view domain, and when we present an isolated clause we assume that
322 Chapter 4

this corresponds to such a domain. Hence although *Jane is speaking only

to himself is ungrammatical in isolation, it is ®ne when embedded with an
appropriate antecedent, as in John hopes that Jane is speaking only to
himself. Anaphors must be bound in this point-of-view domain and may
occur long-distance bound only if the antecedent is the closest ``subject of
consciousness''Ðthat is, the protagonist from whose subjective point of
view a¨airs are presented (only when locally bound can Anaphors com-
bine with an objective point of view, on this account). Zribi-Hertz seems
to presume that these semantic conditions exhaust the constraints on the
long-distance uses. She also points out that there is an intrinsic connec-
tion between the endophoricity, or obligatory referential dependency, of
re¯exives and logophoricity (1989: 724): ``re¯exive pronouns serve the
internal narrative point of view and they are bound within some structural
domain'' in which this point of view is systematically maintained.51
If we then assimilate the English cases of overlapping distribution of
pronoun and re¯exive to the same kind of phenomenon exhibited in LDA
generally, we have most of the basic ingredients of an implicatural ac-
count of the patterns.52 The A-®rst account gives us the complementary
distribution of pronouns and re¯exives by Q-implicature where it occurs,
essentially where antecedent and re¯exive are co-arguments, or at least
clausemates. To account for the long-range uses of English re¯exives,
Condition A of the Binding Theory would need to be extended; Kuno
(1987) o¨ers many rather speci®c structural conditions that could be
added to Condition A with a wider domain, whereas Zribi-Hertz, as just
mentioned, o¨ers rather simple semantic constraints on logophoricity
which can be set up to override the restrictions of Condition A. Either
way, a syntactico-semantic set of constraints could predict the long-range
occurrences of the re¯exive.53 Once again, then, wherever we can have a
re¯exive in these long-distance environments, the use of a pronoun will
contrast; but because the re¯exive is marked for both referential depen-
dency and logophoricity (or internal perspective), the contrast need not be
in terms of reference. And in the long-distance uses, the contrast is now
between a subjective versus objective point of view, or schematically:
(58) Q-contrast in coreference, rather than logophoricity
a. ``John was speaking to himself.''
Coreferential by Binding Condition A
b. ``John was speaking to him.''
Disjoint by Q-implicature
323 Grammar and Implicature

(59) Q-contrasts in logophoricity, rather than reference

a. ``John hoped that Anne would speak to himself.''
Coreferential and logophoric by an extended Condition A
b. ``John hoped that Anne would speak to him.''
Nonlogophoric by Q-contrast
Coreferential by presumptive I-inference
It is clear that there is one major residual problem for the pragmatic
account of LDA generally. It is to explain why, given that the re¯exive
encodes both coreference and logophoricity, the contrast in reference is
dominant where the pronoun could have a clausemate antecedent, and
the contrast in logophoricity is dominant elsewhere.54 If re¯exives with
LDA uses are in general marked [‡logophoric, ‡referential dependence],
then avoidance of the re¯exive with a pronoun should be interpreted as
either a contrast in reference, or logophoricity, or both. Yet in the clause-
bound cases, the contrast always seems to be one of reference sim-
pliciter, and with long-distance antecedents, the contrast generally seems
to be one of logophoricity (although it may be one of reference, too).
There seem to be essentially two lines of possible explanation here.
On the one hand, one may claim, in a line following Bouchard (1983)
through to Pollard and Sag (1994), that true re¯exives are restricted to
those with clausemate antecedents, whereas all long-distance bound re¯ex-
ives are essentially pronouns of a special contrastive sort that happen
to be homophonous with real Anaphors; we turn to this view in section
4.3.2.5.3. On the other hand, one may attempt to ®nd an additional
pragmatic factor that would explain the overriding concern with reference
within the clause. In the second half of this chapter, I will show that such
an additional factor (to be labeled the DRP) is independently motivated
by the existence of languages without conventionalized Anaphors at all:
such languages exhibit a preference for the reading of clausemate (co-
argument) NPs as disjoint. But because such languages have no gram-
matical Anaphors, this can only be a default presumption, and thus a
pragmatic principle and not a re¯ex of a universal Condition B. It turns
out that these languages are also relevant to understanding languages like
English that conform to the A-®rst patternÐEnglish indeed is diachroni-
cally derived from such a language. This additional pragmatic principle,
of presumed disjointness of arguments, ultimately involves superseding
the A-®rst approach with a more complex, diachronically laminated
account of the pragmatics of anaphora. Meanwhile, one can see that the
324 Chapter 4

existence of such a principle is exactly what is needed to complete the

above account: clausemate NPs will be presumed disjoint in reference
unless marked as re¯exive, and a pronoun will therefore be interpreted as
motivated in order to Q-contrast in reference; whereas outside the clause
of the antecedent, an LDA can more openly be interpreted as either con-
trasting in logophoricity or reference or both. Before pursuing this, how-
ever, we turn initially to the ®rst option given above.

4.3.2.5.3 LDAs as Homophonous with, but Distinct from, Real Anaphors

Condition A of the Binding Theory, in any of Chomsky's formulations,
requires that the Anaphor be bound by an antecedent at close range. It
follows that all sorts of re¯exives in English, like the following, are simply
not Anaphors on this approach:
(60) a. ``John said the paper was written by Ann and himself.''
b. ``John wanted the photo taken with Ann in front of himself.''
c. ``The college president thought that Bacon's picture of himself
was unattractive.''
d. ``The picture of himself in the commonroom bothered the
college president.''
Some go further and think that true Anaphors are more restricted still,
so that picture-noun re¯exives, as in John found a picture of himself, are
really pseudo-Anaphors (Pollard and Sag 1992, following, e.g., Postal
1974) and thus that Chomsky's (1986) e¨orts to extend slightly the bind-
ing domain for Anaphors were mistaken. True Anaphors, on this line, are
co-arguments with their antecedents, and all other uses of re¯exives are
homophonous pseudo-Anaphors (``exempt anaphors'' in Pollard and
Sag's terminology). The reason for abandoning the line, for example, that
picture NPs are real Anaphors is that acceptability appears to be gov-
erned by gradient pragmatic or processing factorsÐsubjects intervening
between the antecedent and the anaphor are acceptable the less animate
and referential they are:
(61) a. ??``Bill1 remembered that John saw a picture of himself1 in the
police station.''
b. ?``Bill1 remembered that The Times carried a picture of
himself1 .''
c. ``Bill1 remembered that nothing could make a picture of
himself1 look respectable.''
325 Grammar and Implicature

Pollard and Sag (1994: 269) attribute this to processing factors, but
equally we have an explanation in terms of the I-principle preferring the
most local plausible antecedent. A second factor governing the occurrence
of pseudo-Anaphors is the logophoricity requirement already reviewed in
detail. An increasing number of scholars seem prepared to consider LDAs
to be in general pseudo-Anaphors, subject to pragmatic constraints rather
than the Binding Conditions (see, e.g., Koster 1994, Mazuka and Lust
1994)Ðbut they do not develop the pragmatic analysis that would then
be required.
If this line is correct, we have a natural explanation of the way that the
Q-inferences from pronouns divide into disjointness-inferences for co-
arguments and a wider disjunctive inference, to either disjointness or
antilogophoricity, in long-range uses of re¯exives. The clausebound, co-
argument (nonre¯exive) pronouns are in Q-contrast with true Anaphors,
which must be bound by a coargument; pronouns elsewhere may be in Q-
contrast with quite di¨erent lexical items, namely pseudo-Anaphors, with
referential dependence and logophoric implications. Such a line would
also have some support (unnoted by Pollard and Sag) from languages
where re¯exivity is indicated by a marked verbal form that indicates that
two of its arguments are coreferential (see Faltz 1985)Ðhere re¯exivity
must be about co-arguments. The whole Pollard and Sag (1992, 1994)
approach, which is based on a noncon®gurational version of the Binding
Conditions within a di¨erent grammatical framework (HPSG), can be
made to dovetail with the pragmatic account just as well as those more
directly in the Chomskyan tradition.
However, there are some good reasons to resist this approach to
English. The ®rst is methodological: as Zribi-Hertz (1989: 721) puts it,
considering the same option, ``First-year students in linguistics already
know, however, that the homonymy solution should be regarded as a last
resortÐa bad solution per se, a `desperate' strategy.'' Why should there
be two sets of re¯exives, with exactly the same members, the same oblig-
atory referential dependency, but di¨erent binding constraints? The sec-
ond reason for skepticism is that long-distance-bound re¯exives occur
also in core argument positions, not just adjuncts, so that the distinction
between real and pseudo-Anaphors becomes more de®nitional than
observational. Indeed, Pollard and Sag (1992: 278±279) are forced to
suggest that many well-attested uses of such re¯exives are just ungram-
matical, but acceptable in prose.55 Clearly, if this line is to be pursued,
we need independent evidence for the distinction between real and
326 Chapter 4

pseudo-Anaphors (i.e., good tests that rest on other than binding-domain

factors). Meanwhile, we shall continue to explore the alternative line that
attempts to unify an account of re¯exives and seeks a further pragmatic
factor to explain the special concern with coreference versus disjointness
within the clause.

4.3.2.5.4 Conclusions to the A-First Pragmatic Analysis In the ®rst part

of this chapter we explored the idea that grammatical principles need only
state the conditions for the binding of Anaphors. Nearly all systematic
disjointness conditions, speci®cally the sort of constraint speci®ed in
Binding Conditions B and C, would follow by pragmatic ``mirror image''
from the kind of binding principle exempli®ed in the various versions of
Condition A. This would follow by principles of economy: unmarked
pronouns pick up the disjoint meaning by Q-contrast to re¯exives
wherever they might have been employed. Meanwhile, pronouns and full
lexical NPs tend to contrast in reference because of the I-inference to
coreference from minimal expressions associated with pronouns, and the
M-contrast to disjointness that then arises where a pronoun might have
been used but a full referring expression was used instead. Hence the
general contrastive behavior of re¯exives, pronouns, and lexical NPs with
respect to anaphoric linkage.
Such an approach has none of the inexactness of most functional
explanations: whatever syntactic and semantic conditioning there is on
the occurrence of Anaphors will get re¯ected into the pragmatic account
of the behavior of pronouns. If the distribution of Anaphors was
accounted for solely in terms of their lexical properties (and their neces-
sary referential dependence in particular), there would be little motivation
for maintaining an A-®rst account (for some languages, this may be the
case; see Huang 1994 on Chinese). But insofar as there are three terms
involvedÐantecedent, Anaphor, and binding domain de®ned in syntactic
or semantic termsÐan A-®rst account is well motivated (see Hellan 1991
for an interesting typology of di¨erent kinds of domain constraints). On
the other hand, while making precise predictions, the pragmatic account
does allow for principled defeasibility, which a syntactic account of the
Binding Conditions variety will never be able to explain.
The A-®rst style of account is not restricted to any speci®c kind of
grammatical framework. Modern approaches vary in various profound
ways: some theories aim for maximum coverage of re¯exive distributions
whereas others distinguish real from pseudo-Anaphors (which are then
327 Grammar and Implicature

disregarded); some theories presume con®gurational constraints whereas

others argue that they are based on hierarchies of thematic or grammati-
cal relations; some theories cut the pie di¨erently, into bound-variable
anaphora versus referential dependencies, for example, rather than into
Anaphors versus pronouns; some theories bring meaning di¨erences (like
logophoricity) into the account, others do not. Pronominalization and
anaphora were among the earliest subjects to be explored in generative
grammar, and after 40 years of intensive research there is still no adequate
explanation of all the phenomena nor any consensus about where it may
yet be found. I believe that the failure to take into account the systematic
contribution of a theory of generalized pragmatic inference is a substan-
tial reason why further progress has not been made. A theory of GCIs can
interdigitate with any of these frameworks, reapportioning some pre-
dictions from the syntax to the pragmatics and o¨ering some handle
on the unexplained residue of defeasible patterns. In each case, how-
ever, the corresponding pragmatic predictions will be di¨erent and worth
investigating as indirect evidence for or against the di¨erent syntactic
approaches. The overall power of the resulting theory, with the syntax
and the pragmatics working in tandem, is certain to be considerably bet-
ter than any of the purely syntactic accounts taken on their own. I have
sketched in passing how for each of the discussed frameworks the prag-
matics might be put to work.
The major challenge to an A-®rst account is undoubtedly the failure of
complementarity of pronouns and re¯exives (i.e., their overlap in certain
loci as potentially linked to the same antecedent). We have gone to some
e¨ort to establish that the pragmatic account does not necessarily break
down on account of this. Pragmatic contrasts can be on various dimen-
sions, and where an Anaphor encodes perspectival information as well as
referential dependency, the contrast with a pronoun can as well be in
terms of the logophoric dimension. However, unless we buy the distinc-
tion between Anaphors and pseudo-Anaphors, we clearly do need an
explanation of why the contrast is nearly always in terms of reference just
in the clausebound cases, whereas in the long-range binding cases the
contrast is either in terms of logophoricity or reference or both.

4.3.3 The B-First Account, with a Pragmatic Reduction of Binding Conditions A

and C56
Let us now explore a radical alternative to the preceding account which
has taken Condition A of the Binding Theory as a basic rule of grammar,
328 Chapter 4

and has developed pragmatic reductions of Conditions B and C as para-

sitic on Condition A. In this alternative, we shift the center of the account
and make the Condition B pattern basic and derive Conditions A and C
as pragmatic derivatives. Such an alternative account seems better suited
to certain languages, and because the account uses the same pragmatic
apparatus, there is no theoretical inconsistency between our A-®rst and
what we may call a B-®rst account. In either case, any one language (now
being thought of as a sociolinguistic entity) should be treated as having
stabilized a particular pattern of anaphoric interpretation as basic, as the
conceptual anchorage for the rest of the system. This stabilized core
(whether Condition A-like or Condition B-like) is then subject to our
familiar Q-, I-, and M-inferences, yielding ``for free'' the rest of the sys-
tem. What determines this kind of pragmatic parametrization is historical
process, and in the end I shall argue that A-®rst languages tend to exhibit
residues of a B-®rst stage, thus showing that there is no radical disconti-
nuity between these types.
Here is how a B-®rst account might be developed. The starting point is
a stabilized pattern of interpretation where clausemate NPs are preferen-
tially interpreted as distinct in reference. This could be thought of as a
matter of grammatical stipulationÐCondition B or C as part of Uni-
versal Grammar surfacing as an inferential presumption. Against that,
whereas Condition A is often thought to display parametric variation in
the size of the binding domain, there is little evidence that Condition B
behaves in the same way (speci®cally, the exclusion domain tends to
remain strictly local; see Harbert 1995: 194±195). Second, as we have
seen, the B-pattern is defeasible. Thus the core B-like pattern could itself
be pragmatically motivated. Farmer and Harnish (1987), for example,
advance a pragmatic presumption as the basic datum for pronominal
anaphora; they propose a pragmatic Disjoint Reference Presumption
(DRP), which holds that ``the arguments of a predicate are intended to
be disjoint, unless marked otherwise'' (1987: 557). The origin of such a
pragmatic presumption is left unclear on their account, but we may per-
haps be able to relate it to the GCI framework. Under a classi®cation of
actions of central human interest, agents normally act upon entities other
than themselves; the prototypical actionÐwhat is described by the pro-
totypical transitive clauseÐis one agent acting upon some entity distinct
from itself. If that is how the world stereotypically is, then an interpreta-
tion of an arbitrary transitive sentence as having referentially distinct
arguments is given to us by the I-principle, which encourages and
329 Grammar and Implicature

warrants an interpretation to the stereotype. Note that this is not some

kind of behaviorist presumption that the statistical preponderance of
nonre¯exive states of a¨airs, or even linguistic statements, is inductively
learned and then re¯ected unwittingly in pragmatic presumption. The
I-principle is a principle of default interpretation, which capitalizes on
the underspeci®cation of linguistic meaning together with mutual beliefs
about regularities in the world to yield rich interpretations as hypotheses
about utterance-meaning. It thrives on mutual assumptions of mutual
orientation to stereotypes (i.e., on complex mental constructs), not behav-
iorist response.
In language typology, crosslinguistic tendencies are used to suggest, for
example, a universal markedness hierarchy, so that Agents are prototypi-
cally animate and Patients inanimateÐthe Ur-act, as it were, is somebody
doing an action to somethingÐwith deviations from these expectations
progressively marked (Comrie 1981: 120¨ ). The same reasoning applied
to re¯exivity would set up a universal prototype or tendency where
Agents doing things to disjoint referents is the unmarked expectation,
any deviations being marked either by formal or pragmatic means (cf. a
similar presumption in Reinhart and Reuland 1993). Because this ten-
dency is not re¯ected in all languages by special morphology or syntax,
but in some languages may be merely re¯ected in patterns of use, it is
plausible that it is due to extralinguistic principles rather than innate ideas
or Universal Grammar.
There will be those who are skeptical. As Zribi-Hertz (1995) puts it,
criticizing the use of the DRP by Huang (1994), ``What criteria do we
have to decide what is `more `natural' than what? If we don't have any
independent test for `naturalness,' the DRP is nothing but a clone of
Principle B.'' But there are signs of naturalnessÐnamely, the fact that it is
re¯exive statements, not the nonre¯exive ones, that are the linguistically
marked alternative in all languages that mark re¯exivity at all. Second,
the DRP is only a defeasible tendency, so it is distinct in kind from a
grammatical stipulation like Condition B. Still, the critic may point to the
fact that some actions are stereotypically re¯exive, like brushing one's
teeth, bathing, and so on. But here, interestingly, languages often do rec-
ognize the special character of such stereotypically re¯exive actions. In
Frisian, for example, it is just these predicates that allow the use of the
normal pronoun for self-reference rather than the re¯exive or emphatic
form (thus one says Willem1 wasket him1 not *Willem1 wasket himsels1
for `William washes himself ', Reuland 1994: 239; the same was true in
330 Chapter 4

Middle English, see Faltz 1985: 242). In this case the unmarked pronoun
indicates the stereotypical situation, irrespective of whether one or two
agents are involved. However, because the DRP is the normal presump-
tion, languages more often indicate stereotypical re¯exive (or so-called
introverted) predicates in a way that signals both their departure from the
DRP and their stereotypical status: as a strong crosslinguistic tendency,
introverted predicates occur with a null or reduced re¯exive marking
(Haiman 1985a: 168±174), as in English ``He put a coat on '' (or as in
the ``weak'' re¯exives of the Germanic languages, discussed below). Here
the reduced form picks up the stereotype in line with the I-principle, and
the special marking indicates deviation from the DRP. This special treat-
ment of such exceptional predicates as shaving, bathing, dressing, adorn-
ing, and so onÐin accord with their stereotypical expectations of re¯exive
action but in contrast to the DRPÐis interesting indirect evidence in
favor of the pragmatic nature of the DRP, as based on stereotypical
expectations for disjoint arguments with regard to the majority of transi-
tive predicates.
Still, the critic may point out that re¯exivity in many languages is not
con®ned to core arguments of transitives, as in so-called raising contexts
like John revealed himself to be a Mason. The response here is that the
DRP is only a small part of an explanatory apparatus, which cannot
predict all the loci that a speci®c language will treat as requiring re¯exiv-
ization. In English, the DRP plays only a small part in explaining the
patterns (namely, it predicts where pronoun and re¯exive ®rmly contrast
in reference), which at least as far as Anaphors are concerned are ®rmly
grammaticalized.
Huang (1991, 1994: 130) notes that there are potential problems for
the neo-Gricean theory if the DRP is taken to be given to us by the I-
principle. Although there are good reasons, already reviewed in chapter 2,
for taking the priority of Q > M > I inferences to regulate implicature
projection, there are no worked out principles regulating, say, the relative
priority of subtypes of I-inference. But if the DRP follows from the I-
principle, and the preference for local coreference of reduced forms also
follows from the I-principle, the two will often be in con¯ict, unless we
recognize that the preference for coreference operates only beyond the
clause. Huang therefore entertains the conjecture that the DRP should be
thought of as part of our background presumptions about the world,
which would then, as always, override default implicatures. That, how-
ever, has problems of its own: as we shall see, the origin of many re¯ex-
331 Grammar and Implicature

ives lies in the use of emphatic particles, presumably because these are a
natural way of M-implicating that the DRP doesn't obtain. In that case,
the DRP itself must be cancelable by a higher-ranking inference type, and
all this ®ts with the assumption that the DRP is I-implicated, with M-
implicatures taking priority over I-implicatures in the usual way.57
But whatever the exact source of the DRP, once we presume it, we can
then derive the Condition A pattern this way. First note that re¯exives are
marked forms; if they are pronoun-like in grammatical category, they
tend to be longer, more morphologically complex than ordinary pronouns
(cf. himself with him); if re¯exivity is encoded on the verb, then the verbal
morphology is marked (Faltz 1985). Thus, what the marked, prolix forms
indicate is that the normal, stereotypical scenario associated with a tran-
sitive clause does not in fact obtainÐnotice is served by M-implicature
that the I-inference to the stereotype is not intended. If the stereotype is
disjoint reference for arguments, then the M-implicature is (by Horn's
1984 division of pragmatic labor) to the complement of that interpreta-
tion, namely to coreference. Hence (given that, by our projection rules for
implicatures, M-implicatures cancel any rival I-implicatures) we derive
the Condition A-like pattern for marked pronouns, so-called Anaphors.
It remains to account for the Condition C-like pattern. Here the ac-
count is familiar. Where a pronoun is in a lower clause and thus not a
clausemate with another NP (thus the pragmatic Condition B-like DRP is
not in play), as in He thinks that he should come, there will be nothing to
prevent the I-implicated tendency to coreference (assuming matched
agreement features, etc.). But if, just where such a pronoun would so im-
plicate coreference, a marked form (in this case a lexical referring expres-
sion) is employed instead, as in He thinks that John should come, then
again we get the complementary M-interpretation, this time to disjoint
reference.
Thus, to recapitulate: the sentence in (62a) below is interpreted with dis-
joint arguments in line with the (I-induced) stereotype ``clausemate
arguments are distinct''; the marked form of the pronoun in (62b) warns
(M-implicates) ``contrary to stereotype,'' and because M-implicatures
override I-implicatures, we obtain a conjoint interpretation. The pronoun in
(62c), on the other hand, does not fall under the presumption of clausemate
disjointness because the only potential antecedent is in another clause; the
pronoun is unmarked, so there is nothing to stop the I-preference for
coreferentiality of reduced forms going through. In (62d), by contrast, the
use of a full lexical NP suggests (M-implicates) that whatever might
332 Chapter 4

have been implicated by the use of the pronoun is being avoided and thus
that the complement of the interpretation in (62c) (now disjointness) is
intended.
(62) a. ``John likes him.'' (disjoint by DRP)
b. ``John likes himself.'' (conjoint by M)
c. ``John said he went.'' (conjoint by I)
d. ``John said the boy went.'' (disjoint by M)
There is considerable evidence of di¨erent kinds in favor of a B-®rst
pragmatic analysis of this kind for at least certain languages. One kind of
support comes from typological evidence for a widespread association of
re¯exives and emphaticsÐin language after language, the re¯exive seems
to be homophonous with an emphatic particle or a½x (Moravcsik 1972,
Langacker and Munro 1975).58 For example, just as in English we have
the emphatic use of the re¯exive pronoun in (63a), compared to the re-
¯exive use as in (63b), so in, for example, Tamil we have the emphatic
particle in (63c) homophonous with the root of the re¯exive pronoun in
(63d).
(63) a. ``She herself will come.''
b. ``She hit herself.''
c. Ava-taan varuva.
she-EMPH will come.
`she herself will come.'
d. Ava taan-e aicca.
she self-ACC hit.
`she hit herself.'
The B-®rst account predicts this of course; on that account, ``anaphors''
are just pronouns marked in such a way that they will trigger an M-
implicature to the complement of the normal interpretationÐif clausemate
co-arguments are normally disjoint, a marked pronoun will M-implicate
coreference. The contrastive, emphatic particle or a½x does the job
nicely. Note that many languages have two or more re¯exive AnaphorsÐ
for example, the Germanic languages typically have two re¯exive forms,
cognate with German weak re¯exive sich and strong re¯exive sich selbst
(see Reuland and Koster 1991). It is the strong re¯exives that tend to oc-
cur with clausemate antecedents, and the weak ones with long-distance
antecedents, except that they also occur locally bound with inherently
re¯exive (introverted) verbs. Note that this ®ts the DRP-based predic-
333 Grammar and Implicature

tion rather well: emphatic re¯exives occur intraclausally where a marked

alternative to the DRP-based assumption of disjoint reference is re-
quired; where in fact there is a stereotypical presumption of an agent
acting on himself (as in many inherently re¯exive verbs), the emphatic
re¯exive may not be required.59
What is being o¨ered here is, of course, a very simple explanation for
the oft-noted overlap between re¯exive and emphatic forms. That overlap
is su½ciently salient that there have been a number of attempts to try and
explain it. In the heyday of Generative Semantics there were attempts for
example to derive John himself came from John, he came (Moyne 1971) or
JohnÐJohn's own selfÐcame (Cantrall 1973; see Jayaseelan 1988 for a
revival). There have also been careful and interesting attempts to charac-
terize the exact semantic/pragmatic import of the intensi®er use of re¯ex-
ive forms; thus Moravcsik (1972) distinguishes a head-bound use (as in
John himself came) from an adverbial use (John came himself ), and
Edmondson and Plank (1978) suggest the ®rst has to do with a pragmatic
scale (in the sense of Fauconnier 1975) of unexpectedness, the second with
a scale of direct agency or experience.60 Philosophers, too, have specu-
lated on the meaning of phrases like He himself, Hintikka (1970) for ex-
ample claiming that the intensi®er rules out an intensional (or attributive)
interpretation, Geach (1972) that it translates a ®rst-person pronoun into
a form appropriate for indirect speech (cf. the discussion of logophoricity
above). All of this is usefully reviewed by Edmondson and Plank (1978)
and KoÈnig (1991), but summarizing the semantico-pragmatic e¨ects of
the intensi®er uses of re¯exive forms, we may note that (a) there is a
contrastive, contrary-to-expectation element; (b) there is a natural nega-
tive gloss, of the sort `and not anyone else', `and not the more expectable
persons', and so on; (c) the intensi®er often plays a role in reference, by
forcing a particular coreferential interpretation of a pronoun;61 (d) the
intensi®er could often be replaced with like e¨ect by stress; and (e) issues
of scope arise according to placement. All of these features are compatible
with, indeed explained by, an account in terms of M-implicature and
Horn's (1984) division of pragmatic labor (i.e., in my terms, the com-
plementarity of I-implicatures and M-implicatures). M-implicatures, by
picking out the complementary interpretation to the stereotypical I-based
interpretation that would have been generated by the use of simple
unmarked form, naturally give rise to contrary-to-stereotype, negative
interpretations. It is left open on this account that the contrast intended
may be either one of unexpectedness or point of view, on the one hand, or
334 Chapter 4

one of reference, on the other. That contrastive stress might do exactly the
same job is likewise explainedÐfor all the intensi®cation of the NP is
doing is warning of a marked interpretation. Thus simple though it is, the
B-®rst account of the overlap between re¯exive pronouns and emphatic
intensi®ers modifying pronouns has much to recommend it.
There is more general evidence for the use of marked forms to indicate
marked, re¯exive, interpretations. In those languages, like many Austra-
lian ones, where the re¯exive is signaled by a special verb form, that verb
form is a marked form. In languages like Guugu Yimithirr (Haviland
1979, Levinson 1987a,b) this special verb form does not directly encode
re¯exivity but only indicates directly that one argument is missing. There
is a pragmatic inference in many cases to the re¯exive. But this can be
further reinforced by the use of an emphatic a½x which then makes the
re¯exive or reciprocal reading central:
(64) Bama-gu gunda-dhi.
the people(ABS)-EMPH hit-ANTIPASSIVE
`The people hit each other.'
`The people hit themselves.'
`Someone hit the people.'
Thus in languages which currently lack, but which seem to be in the pro-
cess of acquiring, fully grammaticalized, unequivocal Anaphors, we can
®nd the same association between markedness and re¯exive readings.
A third closely related kind of evidence for the B-®rst account is the
diachronic fact that re¯exives seem in many cases to have derived from
emphatic or contrastive forms, even if there is no longer a homophonous
corresponding emphatic particle or a½x. This seems to be the case with
Tamil taan or Chinese ziji, for example, and I discuss the history of
English re¯exives in detail. But the most persuasive kind of evidence for
the B-®rst analysis comes from consideration of languages that lack
Anaphors altogether. As noted, such languages obviously cannot support
an A-®rst analysis. It is therefore crucial to see what patterns of anaphora
they exhibit.

4.3.3.1 Languages without Re¯exives There are scattered reports from

around the world of languages that do not directly code re¯exivity. Such
languages seem to have been largely ignored by grammatical theorists
even though they would appear to be crucial evidence against the cen-
trality of the binding apparatus.62 It is not yet clear just how common the
335 Grammar and Implicature

absence of re¯exives is. Faltz (1977: 210f ) cites the Low West Germanic
subgroup of languages (including Old English) and Biblical Hebrew;
Carden and Stewart (1986: 51) cite Fijian and other Austronesian lan-
guages, a Papuan language Harway (data from Comrie), and perhaps up
to half the pidgins and creoles (in many cases by independent innovation).
We can add a number of Australian languages, and there are no doubt
other groups of languages yet to be reported (e.g., Carden 1989: 3 notes
the same pattern in Isthmus Zapotec). In this section I provide a sketch of
the facts for some of these cases to aid and abet the rehabilitation of these
linguistic orphans.

4.3.3.1.1 Australian Languages We have already seen that there are

languages, like Guugu Yimithirr, that lack true re¯exives. In fact, the
majority of Australian languages would seem to lack Anaphors. Instead,
in many of these languages the re¯exive meaning is indicated by a
semantically general antipassive or detransitivized clause (Dixon 1980:
434, 448; see also Austin 1981: 151±155; Evans 1995: 278±279, 348±354).
The antipassive construction can be thought of as transforming a transi-
tive clause into an intransitive one with a single absolutive argument; it
thus simply indicates that, for one reason or another, one of the argu-
ments of the verb is missing, unknown, or irrelevant. Thus the re¯exive
meaning is just one natural interpretation of the antipassive.
For example, recollect the Guugu Yimithirr facts discussed previously:
There is no unequivocal re¯exive or reciprocal construction. To indicate
re¯exivity one uses a detransitivizing verbal a½x, which suggests the
irrelevance of the missing argument. That irrelevance may stem from the
speaker's not knowing (or not caring) who or what the agent was, or from
the accidental nature of the agency, or from the fact that the missing
argument is identical to the surviving one. However, the re¯exive reading
can be made prominent by marking the surviving NP with the emphatic
particle -: gu (as expected on the B-®rst account, which derives re¯exives
from marked pronouns, or more generally, marked re¯exive clauses from
unmarked nonre¯exive ones). The same use of the emphatic attached to a
pronoun can suggest coreferentiality where the antipassive construction
cannot be used, for example, in a sentence glossing `John1 helped his1 -
emphatic father'. Here we see an emphatic re¯exive in the very making.
In some languages though, the use of such an intransitive form to ex-
press re¯exive meaning is optionalÐfor example, Dixon (1983: 497) dis-
cusses Nyawaygi, where the re¯exive can be expressed by an antipassive,
336 Chapter 4

a regular transitive with two ordinary NPs (one in ergative and the other
absolutive case), or a transitive verb with two NPs in absolutive case. In
other languages one simply has two NPs in the absolutive with a normal
transitive verb, as if they were both the subjects (one the head, the other a
modifying adjunct) of an intransitive verb (Dixon 1981: 64±65). How-
ever, some of these languages do not employ any such device at all. For
example, in Gumbaynggir (New South Wales), the re¯exive is expressed
by a normal transitive sentence (Eades 1979: 312):
(65) Gua:du bu:rwang gula:na magayu.
3sg-ERG paint-PAST 3sg-ABS red paint-INSTR
`He painted himself/him with red paint.'
There would appear to be a su½x -w (gula:naw instead of gula:na in
(65) above) that forces a re¯exive reading, but this is de®nitely optional
(p. 313).63 Similarly, Austin (1987) reports that Jiwarli does not formally
express any distinction between an ordinary transitive sentence and a
re¯exive meaning: ``Jiwarli re¯exives are indistinguishable from simple
transitive clauses except that the subject and object nominals are corefer-
ential.'' Thus, in addition to the widespread antipassive or middle mark-
ing of re¯exives on the continent of Australia, there are clearly languages
where re¯exivity is not marked at all.

4.3.3.1.2 Austronesian A number of Austronesian languages are

reported to lack re¯exives and reciprocals, but Fijian is perhaps the best
documented. Thus in Dixon's (1988: 255±256) grammar we ®nd that
``For ®rst and second persons the re¯exive is simply shown by using the
appropriate pronoun in both subject and object slots,'' or if it is an indi-
rect object that is involved, the ordinary pronoun in the relevant case is
used. In the third person, a verb with the transitive marker -a and without
an explicit object is interpreted as having unmarked reference to a third-
singular object which is noncoreferential with the subject. If coreference
or re¯exivity is intended, a full object pronoun (e.g., 'ea, third-singular
object) is required, and although this might be interpreted disjointedly, it
encourages a coreferential reading:
(66) sa va'a-.dodonu-.ta'ini 'ea o Mika.
ASP correct 3sg‡Obj Art Mike
`Mike corrected himself ' or `Mike corrected him.'
The pattern looks like our hypothesized B-®rst type: the least marked
form of a transitive clause encourages the supposition that the arguments
337 Grammar and Implicature

are distinct in reference; any marked form raises some possible doubt
about this, and so it is the normal way to express coreference if intended.
Apparently, the same holds for the expression of the reciprocal relation-
ship, although there are also intransitive roots and so on that can also be
used to express this.
Kilivila, another Austronesian language, is a clear example of a lan-
guage without dedicated re¯exives; again a normal transitive verb is used,
but with one of a series of emphatic pronouns in addition to the subject
(Senft 1986: 54¨, personal communication):64
(67) m-to-na e-wai titole-la.
Dem-CPman-Dem 3-beat self±3.Poss.ProIV
`This man beats himself ' or `This man beats (someone).'
The emphatic pronoun titolela (or equally from another series alama-
guta), built on a possessive pronoun su½x -la (or possessive pre®x ala)
plus an emphatic particle, is construed as referring to the subject, but
either as an emphatic adjunct to the subject or as object of the verb. There
is no formal means to distinguish the readings. Yet another, distantly
related, Austronesian language, Tahitian, apparently follows the same
sort of pattern (Tryon 1970: 97). To express the re¯exive, one can only
use the pronoun:
(68) 'ua ha'opohe 'oia 'iana.
was kill he him
`He killed himself ' or `He killed him.'
But there is an emphatic a½x iho, which when added to the pronoun
seems to be the standard way to express the reciprocal meaning.
An Austronesian language that has been the subject of some detailed
GB work is Chamorro (Chung 1989). Chung notes (1989: 148¨ ) that
Anaphors and pronominals are morphologically identical just as in
Fijian, although the re¯exive interpretation can be forced by an optional
adverb maisa (glossed as `by oneself '). However, she goes on to argue that
two grammatical constraints are sensitive to the coindexing of NPs within
minimal governing categories, and therefore that Chamorro quasi-
pronominals are in e¨ect ambiguous (``morphologically syncretic''), dis-
tinguished into Anaphors and pronominals at the syntactic level. How-
ever, it is hard to see how this can be made an empirical proposition,
given that whenever a proform is locally bound, it can be claimed to be a
covert Anaphor, and where not, a pronominal. Additionally, a peculiar
338 Chapter 4

consequence of taking these constraints to be syntactic criteria for covert

Anaphors is that the same criteria then identify NP-trace and PRO as
non-Anaphors (p. 173¨ ). It seems to me then that it may be preferable to
treat the special constraints on Chamorro pronominals construed re¯ex-
ively as semantic. For example, there seems to be a syntactic ®lter ban-
ning any third-person-plural subject of a normal transitive clause, unless
subject and object are coreferential. But the exception is surely based on
the natural semantic conception of re¯exive clauses as quasi-intransitive,
as shown in the worldwide association of middle verb forms with re¯ex-
ivity both synchronically and diachronically (see, e.g., GeniusÏieneÇ, 1987).
The second constraint likewise gives no real support to a syntactic dis-
tinction between Anaphor and pronominal.65 In any case, Chamorro is
clearly a language where a single set of pronominals does double duty as a
set of re¯exive and nonre¯exive pronouns, as seems to be so often the case
in the Austronesian family.

4.3.3.1.3 Creoles66 A third and diverse group of languages that seem

to generally lack Anaphors are the pidgins and creoles of all origins.
These languages have recently been surveyed by Carden and Stewart
(1986, 1988, 1990) to see to what extent the Binding Conditions ade-
quately describe their anaphoric patterns. They point out that the issue is
rather interesting from the point of view of the Bioprogram hypothesis
advanced by Bickerton (1981); if the Binding Conditions are innate, and if
pidgins upon creolizing should reveal the essential stripped-down proper-
ties of human language, then we might expect to ®nd the Binding Con-
ditions rapidly instantiated upon creolization. This is not, however, at all
what was found in the survey. Instead, the evidence seems to show both
that pidgins and creoles tend to lack Anaphors and that they tend to lack
a grammaticalized Binding Condition B (pronouns obligatorily disjoint
with clausemate antecedents). I repeat a selection of their ®ndings.
Carden and Stewart (1986, 1988) report that in the Northern dialect of
Haitian Creole the ordinary pronouns may be used as re¯exives:
(69) a. Emile dwe ede li.
Emile1 should help him1=2 (can be coreferential)
`Emile should help him/himself.'
b. Emile dwe ede teÁt-a-li (or koÁ-a-li).
Emile should help himself (must be coreferential)
`Emile should help himself.'
339 Grammar and Implicature

They note that the literal meanings of teÁt (`head') and koÁ (`body') play a
role in the choice between these re¯exive forms and that teÁt is used as an
emphatic focusing particle as in teÁt-Emile (`it was Emile himself '). They
conclude from this that the Northern dialect of the creole can be treated
as an earlier, less grammaticalized stage of the creole. Historical data
seems to support thisÐthe earliest materials provide plenty of evidence
for the use of the plain pronouns as re¯exives but little or no evidence for
the specialized re¯exive forms. The implication is that the earliest forms
of Haitian Creole (pre-1800) had no re¯exives; the later re¯exives based
on body-part metaphors may be due either to ultimate (African) substrate
sources or just as likely to independent invention.
Other creoles also evidence no re¯exive forms, such as the Spanish-
based Palenquero (Carden and Stewart 1986: 24, 199), French-based
Guadeloupe, and Arabic-based KiNubi, not to mention various pidgins
including Chinese pidgin English and early Chinook Jargon (Carden and
Stewart 1990: 5). Moreover, perhaps half of all Creole lects allow the
re¯exive use of clausemate pronouns (1986: n. 58; Carden and Stewart
1987: 23¨ exemplify from, e.g., Martinique Creole, Chinook Jargon,
Bislama, KiNubi, Negerhollands). Corne (1988) shows that Mauritian
Creole uses the ordinary pronouns as re¯exives in the unmarked case,
even though it has re¯exive forms (although Corne believes this use of
pronouns to be a post-creole innovation, Carden and Stewart (1989)
reanalyze the historical material to argue that this use of the pronouns is
derived from the original pidginization process).
Comparison leads Carden and Stewart to hypothesize a diachronic
sequence in creole languages (I here generalize from their 1986 discussion
of Haitian Creole):
(a) Stage 1: no encoded re¯exives; plain pronouns used re¯exively
(b) Stage 2: gradual emergence of morphological re¯exives (e.g., based
on body-part metaphors) with a clausemate, subject-antecedent con-
dition, coexisting with but encroaching upon the use of ordinary pro-
nouns as re¯exives
(c) Stage 3: the loss of the re¯exive use of the ordinary pronouns
One point that they draw special attention to is that the tolerant treatment
of pronouns as possible re¯exives is not a feature that is by any means
lost abruptly (surviving some hundreds of years in some dialects of the
Haitian Creole), as might be expected if Condition B were part of the
Bioprogram that would assert itself immediately upon the acquisition of
340 Chapter 4

native speakers, in other words upon creolization of the original pidgin.

The gradual nature of the change towards a typologically unmarked sys-
tem (with re¯exives and Condition B-like patterns) argues either against
Bickerton's (1981) idea that creolization should reveal the Bioprogram in
the raw as it were, or against the idea that the Binding Conditions are
core elements of that Bioprogram.
Carden and Stewart also note the parallels in language acquisition,
pointing out that this supports the idea that creole patterns are tied closely
to patterns in ®rst-language acquisition. Studies by Solan (1987), Jaku-
bowicz (1984), and others (see Lust 1986 and Kaufman 1994 for review)
show that while the clausemate restriction on the interpretation of re¯ex-
ives (in languages that have them) is acquired early, the Condition B
nonre¯exive use of pronouns is much more uncertainly acquired, with
sometimes as many as 50% of four- to six-year-olds ®nding a clausemate
coreference between an NP and a nonre¯exive pronoun perfectly possible
(Solan 1987). As Carden and Stewart point out, this ®nding is surprising
in the light of the apparent rarity of languages that permit the re¯exive
use of ordinary clausemate pronouns.
In any case, the conclusion that we may draw here from the creole
studies is that there is distinct evidence in favor of a B-®rst analysis in
such languages, both synchronically in that many creoles just use a
marked or emphatic pronoun to indicate a re¯exive, and diachronically in
that creoles that do seem to have Anaphors also seem to have passed
through a B-®rst stage with no Anaphors in the cases for which we have
data.67

4.3.3.1.4 Old English As the diachronic sequence hypothesized for

creoles makes clear, a B-®rst analysis clearly projects a diachronic
sequence: Anaphors should develop out of marked pronouns that are
used to M-implicate conjointness in opposition to the I-based presump-
tion of disjoint clausemates. In that case, an A-®rst language, at least one
that codes re¯exivity through re¯exive pronouns, should develop out of a
B-®rst language (at least that should be one likely source). English is then
an interesting test case, because it was clearly at one stage, the earliest
recorded stage of Old English, a B-®rst language without encoded
Anaphors.
According to Visser's (1963: 420±439) compendious treatment of Old
English, self was not originally a re¯exive element:
341 Grammar and Implicature

in the earliest Old English texts self could be added to the personal pronouns in
the nominative whenever this was thought necessary for the sake of emphasis. . . .
It seems however to have taken some time before self was added to the re¯exive
pronouns [i.e., ordinary pronouns used re¯exively] . . . that already from the be-
ginning of the Old English period were widely used as objects without the addition
of self. (p. 420)

Thus self clearly began as an emphatic adjectival adjunct to nominative

pronouns ``with the purpose of emphasizing the identity of the person or
thing denoted by the object and the subject'' (p. 425), and at the time of
Beowulf a re¯exive object could only be expressed by the normal accusa-
tive pronoun, as in ic me cl0nsie (`I washed me').68 By the time of King
Alfred both that mode of expression and its counterpart with self (as in ic
me selfne cl0nsie) seem to have been current: thus both hie forseoD hie
selfe and hie forseoD hie (p. 421). Indeed the use of the simple pronoun as
a re¯exive continued right through the Middle English period, although
during that period (say 1200±1500):
there is a distinct change in the relative frequency of the two forms: the simple
pronouns, which in the beginning are numerically predominant, are gradually,
and to an ever increasing extent, encroached upon by the compound pronouns
(ˆ pronouns ‡ self ), with the result that by the second half of the ®fteenth century
they are more or less in the minority in most writings. (p. 432)
Throughout this period then ``such a phrase as he ofsticode hine might
mean either `he stabbed him' (someone else) or `he stabbed himself ' ''
(Visser 1963: 433, citing Sweet). Visser suggests that the emphatic was
increasingly used to avoid such possible ambiguities.69
It is tempting, of course, for modern commentators to make a sharp
distinction between the emphatic and re¯exive uses of self,
but to import it into OE is to follow a false native informant. As I see it Quirk and
Wrenn are on the right track when they say. . . : `For the most part, self was used
in OE simply to emphasize and was not, as in Mod. E., associated with being a
re¯exive sign or a pronoun-enclitic'.'' (Mitchell 1985: 189)
The essential fact about Old English is that both the plain pronoun (e.g.,
hine) and the corresponding emphatic pronoun (e.g., hine selfne) could be
used either re¯exively or nonre¯exively in, for example, object position
(op. cit. p. 115).70 Thus the re¯exive uses indubitably grew out of the
emphatic uses, and not the other way around. Nevertheless, text counts
show that ``potentially re¯exive uses are one typical context for use of an
emphatic'' (Faltz 1989: 328). Faltz (1985: 242) notes a further piece of
evidence for the evolution of re¯exives from the use of emphatics to mark
342 Chapter 4

nonstereotypical interpretations. The hypothesis suggests that where a

re¯exive interpretation might in fact be stereotypical, as with predicates
like `wash' and `dress', the emphatic pronouns would not in fact be used.
And this seems to have been the case in Middle English: one said he
cladde hym as a poure laborer rather than he cladde hym self. In these
stereotypically re¯exive actions one can see a tension between the global
stereotype re¯ected in the DRP (``presume no coarguments corefer'') and
the speci®c stereotype (`one washes, etc., oneself ') and consequent insta-
bilities in their treatment. Thus the progression from ic me cl0nsie > ic me
selfne cl0nsie > ic cl0nsieÐthat is, the loss of the re¯exive object with
verbs like wash, dress, shave, and so onÐhas been much commented on
(see, e.g., Jespersen 1927: 325f; Visser 1963: 145¨ ). The endpoint of the
development, the correlation of the omission of the object with the ste-
reotypical reading (as predicted by the I-principle), is of course striking.
The history of English thus seems to provide excellent evidence for an
evolution of re¯exives out of emphatics. The point at which the emphatic
became grammaticalized as a re¯exive can perhaps be equated with the
point at which it lost its in¯ection, at the transition of Old to Middle
English. But as just indicated, this long preceded the acquisition of
indefeasible (or grammaticalized) Condition B-like patterns outlawing
the re¯exive use of ordinary pronouns, a practice that survived well into
Shakespeare's time and beyond (Visser 1963: 435; Haiman 1995).
However, regardless of the diachronic progression, there is little doubt
that Old English was a classical B-®rst language, at least in the earliest
period: it lacked Anaphors entirely, operating with a presumption of
clausemate NP disjointness, a presumption that could be defeated by the
use of a pronoun with an emphatic adjunct. The history of Dutch (and no
doubt other West Germanic languages) might have been similar had it
not been for the relatively recent in¯uence of German: before then just as
in Old English there were no re¯exives, with pronouns doing double
duty.71 But, according to Everaert (1986: 3), ``in the fourteenth and ®f-
teenth centuries the eastern dialects of the Dutch language area started to
use a sich-re¯exive (which later developed into zich) under the in¯uence of
Middle Low German dialects involved in a process of sich adoption from
Middle High German. If we look at the present-day non-standard dialects
we can observe that the use of zich is still limited to the eastern dia-
lects (Ureland 1981).'' Everaert notes that West Flemish (1986: 43) and
Afrikaans (1986: 39±40) continue to operate without morphological
re¯exives.72
343 Grammar and Implicature

There is no doubt, then, that there are languages without Anaphors or

re¯exives of any kind. The most interesting thing about such languages
from the current point of view is that the opposition between coreference
and disjoint reference in such languages can only be achieved pragmati-
callyÐit can only be suggested, indeed implicated, by the choice of a
speci®c kind of referring expression. If there are general pragmatic pro-
cesses that can be reliably employed to make such referential distinctions
in such languages (and there must be), then there is every reason to think
that the same pragmatic processes operate in other languages to support
partially grammaticalized oppositions of the same kind. Thus the very
existence of languages of this sort is evidence in favor of the fundamental
role that we are attributing to generalized conversational implicature in
the interpretation of anaphora.

4.3.3.2 Interim Conclusions: The B-First Account There is, then, dis-
tinct evidence in favor of a B-®rst analysis. First, as we have seen, there is
evidence that there are languages that have no way of coding re¯exivity at
all. This makes an A-®rst analysis impossible, of course, because the
A-®rst analysis predicts all the patterns of anaphora as pragmatically
derived from a grammatical coreÐsomething like Condition A of the
Binding Theory. It goes without saying that such B-®rst languages are
even more serious challenges to theories like that of Chomsky (1981,
1986), where Condition A is claimed to have no basis in experience and is
held to govern NP-movement as well as Anaphor construal. We can then
ask whether in languages with an ill-developed or zero-coding of re¯ex-
ivity, we can nevertheless ®nd a grammaticalized Condition B. But the
question suggests its own answerÐif nonre¯exive core arguments of
transitive verbs were stipulated to be necessarily disjoint (as in Conditions
B and C), how could re¯exivity be expressed at all when there is no direct
coding of re¯exivity? Nonre¯exively marked NPs in such languages must
be construable, under the right circumstances, as coreferential with a co-
argument antecedent. So, in such languages, the assumption of disjoint
clausemate arguments can only be a tendencyÐin short, a pragmatic
inference to the stereotype. What we do see in such languages is the
development of some device to suggest the nonstereotypical coreferential
interpretation; the theory of GCIs predicts that such a device would be
merely a matter of marking the NP or the verb emphatically or con-
trastively, so suggesting a marked situationÐthat is, M-implicating con-
joint reference against an I-based assumption of disjoint clausemates.
344 Chapter 4

Thus we obtain Condition A-like patterns of interpretation of marked

pronouns as a pragmatic inference from a preestablished B-like pattern.
A B-®rst analysis therefore requires that Condition B-like patterns
of interpretation cannot have a deep grammatical or innate basis. Yet
Reinhart for one thinks that at least the disjoint pattern in bound-variable
anaphora as in Everybody likes him does indeed have an innate basis
(Grodzinsky and Reinhart 1993; see also Chien and Wexler 1990). But the
language-acquisition facts even in A-®rst languages like English, where
adult intuitions about locally disjoint pronouns are very strong, do not
support an innate Condition B. As mentioned, these studies seem to show
that children learning A-®rst languages learn the clausemate restriction on
re¯exives relatively early (as early as 3 years old) and relatively excep-
tionlessly (up to 95%); but the assumption that a nonre¯exive pronoun in
object position is disjoint from the clausemate subject is learned relatively
late (as late as 5 or 6 years old) and applied only some of the time (in
some experimental conditions as little as 36%Ðsee Solan 1987; also
Jakubowicz 1984, Read and Hare 1979). Surveying these studies, Kauf-
man (1994: 182) concluded that ``taken together, these studies demon-
strate that children's performance with pronouns is sometimes poorer and
often more variable than their performance with re¯exives,'' nor did she
®nd the recent tests of bound-variable disjointness to support Reinhart's
assumption that at least that aspect of Condition B was more robust (see
also C. Koster 1994). These particular investigations in A-®rst languages
support our A-®rst analysis, by suggesting that children must ®rst learn,
or learn to apply, Condition A, and only then would be able to slowly
appreciate the pragmatic implications of that pattern (i.e., Condition B
and C patterns), often not before middle childhood (C. Koster 1994: 201).
They also clearly indicate that Condition B is not as natural and robust as
it seems to the intuitions of adult speakers of A-®rst languages.
We seem to be left then with the possibility that there are two rather
radically di¨erent kinds of languages: those that have a grammaticalized
Condition A (or similar principle), and those that lack it but have a
pragmatically based Condition B-like assumption that clausemate argu-
ments are distinct unless otherwise indicated. The A-®rst languages are
numerically in the ascendant, and there is historical evidence that the B-
®rst languages tend to develop into A-®rst systems. In both cases, though,
our theory of GCIs gives some account of how the rest of the anaphora
patterns can be derived.
345 Grammar and Implicature

4.4 THE B-THEN-A ACCOUNT: SYNTHESIS OF THE A-FIRST AND

B-FIRST ACCOUNTS

Let us sum up so far. We have reviewed the earlier A-®rst account of

binding patterns, which makes the anaphoric patterns often attributed to
Conditions B and C of the Binding Theory pragmatically derivative from
grammaticalized Anaphors. We have shown how certain di½culties with
this may be circumvented. In particular, the most prominent problem, the
overlap in distribution between (especially long-range) re¯exives and
pronouns, was shown to not necessarily be an overwhelming di½culty at
all. Instead, by bringing the subtle logophoric aspects of long-distance
re¯exives into the picture, we can maintain the Q-based contrast between
pronoun and re¯exive. Thus we end up with a natural extension of the
A-®rst account into languages and phenomena (like switch reference)
beyond those for which it was originally developed. We can now say that
there is a large class of languages (those we have been calling the A-®rst
languages) that seem to fall within the scope of such a partial pragmatic
reduction of the Binding Conditions.
But the reader will recollect that one problem remains with the A-®rst
account, su½cient to suggest that we are still missing an essential ingre-
dientÐnamely, the fact that the re¯exive and the pronoun seem to be in a
di¨erent kind of opposition where there is a clausemate, co-argument
antecedent within the clause, compared to where the antecedent is more
remote. In the ®rst case, the opposition is nearly always one of reference;
in the second case, it is either one of reference or one of logophoricityÐ
hence the pronoun can often occur with the same reference but opposed in
logophoricity. On the face of it, this fact in many languages with long-
range re¯exives predisposes us once again to a grammatical account,
taking the re¯exive and the pronoun to have di¨erent binding domains.
But once the wider facts about distribution of such re¯exives are consid-
ered (like the fact that the antecedent can be far removed, even beyond
the sentence, or deictically implicit in the case of ®rst- and second-person
antecedents), the grammatical account of long-distance Anaphors no
longer looks so attractive (see, e.g., Mazuka and Lust 1994, J. Koster
1994). The basic factÐa semantic factÐis that the re¯exive is necessarily
referentially dependent, the pronoun only optionally so; the search for an
antecedent seems to be guided by matters of preferential interpretation,
not grammatical principles (for an extended justi®cation of this for
346 Chapter 4

Chinese, see Huang 1989). Besides, a grammatical account would leave

the logophoric contrasts unaccounted for.
But if we bring our B-®rst analysis back into the consideration of
A-®rst languages, we can now ®nd an answer to this rather central
puzzle. Suppose we assume that in A-®rst languages we can ®nd the same
presumption that is central to the anaphoric operations of B-®rst lan-
guagesÐnamely, the presumption that core arguments of a single clause
are disjoint (the DRP). Additionally, of course, we have the Q-implicated
contrast between re¯exive and pronoun, which in a language without
long-distance re¯exives is su½cient explanation of the tendency to read
the pronoun as disjoint with a clausemate antecedent. Thus in languages
without long-distance re¯exives, the presumption that core arguments are
disjoint is redundant, which is why it never ®gured in our original A-®rst
account. But in languages with long-distance re¯exives, although the two
pragmatic principles will be largely overlapping and reinforcing, in the
case of the long-distance use of the re¯exive there will be a Q-contrast
outside the scope of the clausemate-disjointness presumption. In those
long-distance locations, the Q-contrast need not be interpreted as a con-
trast in reference (although it may be); it may instead be interpreted as a
contrast in logophoricity, along the lines that O'Connor (1993) suggests.
To summarize the argument, then, we have two pragmatic principles at
work underlying the di¨erent patterns of interpretation for pronouns in
opposition to locally bound re¯exives on the one hand, versus pronouns
in opposition to long-range re¯exives on the other:
1. The presumption of clausemate co-argument disjointness, the DRP, a
stereotypical presumption that can be attributed to the I-principle.
2. A scalar Q-implicature contrast between re¯exive and pronoun, based
on the di¨erential semantic strength of the re¯exive and the pronoun,
the one being necessarily referentially dependent, the other only op-
tionally so; the one suggesting subjective perspective and emphasis, the
other lacking such suggestions.
Where the two presumptions overlap (i.e., where antecedent and ana-
phoric items are co-argument clausemates), the ®rst pragmatic presump-
tion will ensure there is always a contrast in reference. And the contrast in
reference will be su½cient motivation for the use of re¯exive versus pro-
noun, thus exhausting the pragmatic import. Outside these positions, a
long-range re¯exive is contrastive but not necessarily contrastive in refer-
347 Grammar and Implicature

ence. Such a synchronic account goes some of the way toward answering
our central puzzle. But it still leaves aspects of the puzzle unexplained. In
particular, it is still not clear why the re¯exive in long-range positions
should have a logophoric interpretation that is not prominent in the
clausemate use. Where does the deictic markedness come from?
The problem is only fully resolved, I believe, by taking a diachronic
perspective, by looking at the anaphoric patterns that languages exhibit as
accretions of pragmatic practices. We have already seen, in the review of
the facts about Creole languages, and more especially in the detailed facts
about the history of Old English, some reason to believe that B-®rst sys-
tems evolve into A-®rst systems. I would now like to propose that this is
the general evolutionary route by which the familiar A-®rst languages
have arisen. Restricting ourselves to the languages that mark re¯exivity
by NP type, on such a diachronic account, Anaphors arise as marked
pronominal NPs M-implicating contrast with the clausemate-disjointness
presumption. But the use in a B-®rst language of emphatics to encode
the marking, as is so often the case, is obviously consistent with a much
wider use of such forms. In particular, because a deictic perspective is the
perennial unmarked presumption, a marked or emphatic form may
always suggest a contrastive perspective, such as one where (as Stirling
1993 puts it) the ``validator'' is not the current speaker (see also Levin-
son 1988a). So here we have the explanation of the association of logo-
phoricity with re¯exives: both conjoint reference within the clause and a
marked deictic perspective are perennial possibilities of interpretation of
the use of a marked form in a B-®rst language, because it will suggest by
M-implicature an interpretation complementary to the stereotypical one
that would arise from the use of the corresponding unmarked form.
Putting these ingredients together, we may schematize the diachronic
progression from B-®rst languages to A-®rst ones as follows (where
English glosses sketch the various language patterns):73
(70) Stage 1: No re¯exives; core-arguments tend to be disjoint
Sentences of the form `John hit him' will I-implicate disjoint
reference, via the DRP;74 only ad hoc means such as the use of an
emphatic or marked intonation can be used to M-implicate the
complementary interpretation of exceptional coreference; the same
means are likely to be used to M-implicate a marked deictic
perspective.
348 Chapter 4

(71) Stage 2: Emphatic core-arguments may be conjoint

a. The exceptional coreferential interpretation of `John hit him' is
now regularly reinforced by the use of a particular focus or
emphatic particle: ``John hit him-EMPH'' M-implicates `John1
hit himself1 .'
b. An established usage of this sort reinforces the contrast with the
simple pronoun without the emphatic, so ``John hit him''
I-implicates more strongly `John1 hit him2 .'
c. But because the marking of pronouns is by a focus particle with
general uses, other contrasts (beyond reference) may be
intended; one such contrast being the point-of-view contrast
called logophoric, especially relevant outside the scope of the
DRP: ``John said he will come'' expresses the unmarked deictic
point of view, whereas ``John said he-EMPH will come''
M-implicates a marked point of view.
(72) Stage 3: Emphatics become re¯exives (A-®rst system)
An established system emerges: grammaticalized re¯exives encoding
necessary referential dependence within some domain (i.e.,
Anaphors) have evolved. So now:
a. Co-argument positions: the pronoun in ``John hit him'' is not
only presumed disjoint by the DRP, uncanceled by any
M-implicature, but also Q-implicates disjoint reference by the
scale hAnaphor, Pronouni. Hence the strong inference to
disjoint core arguments from the use of a pronoun.
b. Where antecedent and anaphoric expression are not clausemate
co-arguments: here we are outside the domain of the DRP, so
the inference to disjointness doesn't obtain; there is still a
Q-contrast between pronoun and Anaphor, but now a contrast in
reference is not the only possible one. The marked Anaphor may
still carry the marked-point-of-view meaning (from Stage 2):
hRe¯exive, Pronouni
‡coreference unmarked
‡logophoric unmarked
The consequence is that use of the pronoun Q-implicates that the
speaker is not in a position to use the Anaphor because either he
intends disjoint reference or he intends nonlogophoricity.
349 Grammar and Implicature

So far, for this hypothetical diachronic progression we have adduced

evidence only from Old English and the various mentioned Creoles. But
the hypothesis seems to be largely, if not entirely, compatible with the
diachronic hypotheses about the origins of re¯exives that Faltz (1985)
advances on the basis of typological reviews of re¯exivity encoding. His
®ndings are worth reviewing in detail. Faltz suggests (1985: 48¨ ) that the
main typological distinction is between languages that have true, un-
derived, pronominal re¯exives (like Russian sebja and German sich75)
and those with compound re¯exives. Compound re¯exives are of di¨erent
types: (a) nominal head re¯exives, usually modi®ed by a possessive pro-
noun, as in Turkish kendi-me (`self-mine');76 (b) adjunct re¯exives, based
on a pronoun modi®ed with a re¯exive adjunct as in Old English hine-
sylfne `3msg-ACC ‡ REFL' or Irish e feÂin `him + REFL'; and (c) fused
adjunct re¯exives, like modern English where himself can no longer be
analyzed simply as head ‡ modi®er.
In addition to all these types, there are verbal re¯exivizers, often
derived from cliticized pronominal or nominal re¯exives (e.g., French
se ‡ Verb from Latin se), but which end up as true verbal a½xes (like
Russian middle -sja from pronominal sebja). Diagrammatically:
(73)

Certain properties are said to be restricted to each of these types. For

example, P (pronominal) re¯exives require subject antecedents, strictly
require all coreferential clausemates to be encoded with the P form, but
may have long-distance uses; whereas C re¯exives tend to allow non-
subject antecedents, allow some oblique nonre¯exive coreferential clause-
mates, but don't permit long-range uses. (It's clear, in fact, on the basis of
more recent information that these could be at most systematic trends.)
Faltz (1985: chap. 6) then suggests that the following kinds of dia-
chronic progressions may be hypothesized:
1. All verbal re¯exives come from NP re¯exives via cliticization, thus:
P > V or C > V, together with an associated tendency for later semantic
350 Chapter 4

bleaching of the V-re¯exive verb form to a middle (or generalized intran-

sitivizing) verb form; reverse progressions do not occur,77 and it is
hypothesized that there are no progressions *P > C or *C > P (p. 233¨ ).
2. Perhaps all C re¯exives come from emphatic adjuncts on pronouns
Ðfor example, Old English self, French meÃme (from late Latin *met-
ipsimum),78 Dutch zelf, Papago hijil, and so on (1985: 243¨ ); some of
those based on forms glossing like `his head/body' also clearly derive from
emphatics (Biblical Hebrew), whereas others may possibly have been
directly coined to encode re¯exivization.
3. More speculatively, P pronouns derive from subordinate re¯exives
(long-range-only re¯exives)Ðthat is, nonre¯exive pronouns used corefer-
entially with nonclausemate antecedents, like the logophoric African
pronouns discussed in section 4.3.2.5.1.79 These subordinate re¯exives in
turn derive either from stressed pronouns or from a ®rst-person pronoun
used in direct reported speech.80
Thus, if in broad outline Faltz is right, nearly all re¯exives ultimately
arise from emphatic or stressed pronouns.81 As to why this might be the
case, he advances what is in e¨ect a brief and informal version of our B-
®rst account: ``an emphatic is added to an NP as a warning to the hearer
that the intended referent of that NP is unusual or unexpected'' (p. 240);
``in the absence of a re¯exive strategy, it is natural for an emphatic to be
used to signal that the more highly marked, less expected situation is
present, namely that the NP (argument) in question has the same referent
as another one in the same predication'' (p. 242). In addition to Old
English, Faltz cites Dutch, French and Spanish secondary re¯exives
(meÃme and mismo), Papago, Hebrew, among other languages, as provid-
ing clear evidence for re¯exives derived from emphatics.82 The German
secondary re¯exive selbst is also interesting here; cognate with English
self, it remains an intensi®er used to mark re¯exivity with nonsubject
antecedents (the primary re¯exive sich requiring a subject antecedent)
as in:
(74) Hans sprach mit Fritz uÈber ihn (selbst).
Hans spoke with Fritz about him (self )
`Hans spoke with Fritz about him (himself ).'
Here, without the addition of selbst, the reference of ihn is likely to be
taken to be a third party, but with the addition, the referent is taken to be
Fritz (Faltz 1985: 240). But selbst is also synchronically an intensi®er:
351 Grammar and Implicature

(75) Der PraÈsident selbst wird kommen.

the President self will come
`The President himself will come.'
and a focus particle:
(76) Selbst der PraÈsident wird kommen.
self the President will come
`Even the President will come.'
the di¨erent scopes perhaps being responsible for the meaning di¨erences
in these usages (Edmondson and Plank 1978; KoÈnig 1991: chap. 8). Dia-
chronically, the NP intensi®cation use seems to be prior to all the others
(KoÈnig, op. cit.).
Although Faltz's account antedates the great upsurge of work on
binding, it seems still to be the most comprehensive crosslinguistic survey
of re¯exivization available (although see also GeniusÏiene 1987; Sells,
Zaenen, and Zec 1987; Koster and Reuland 1991; Huang in prep.; and
references therein); and it suggests that the account we are advancing may
have some real crosslinguistic generality. Nevertheless, the diachronic
evidence is still relatively thin and still relatively biased by data from
Indo-European languages. Clearly we need much more work on the his-
tory of re¯exives in those languages, like Tamil, Chinese, and Japanese,
that provide long continuous records. In the history of all three of those
languages, we can ®nd evidence for at least a precondition for the
re¯exives-from-emphatics hypothesisÐnamely, evidence for overlap in
function of the same morpheme(s) for both emphasis and re¯exivity. Thus
Classical Chinese zi clearly had both re¯exive and emphatic uses like
modern Chinese ziji.83 Ancient Japanese had a range of re¯exives that
also functioned as subject intensi®ers or emphatics;84 modern zibun or
jibun (the ji- morpheme is written with the same Chinese character as
Classical zi and is obviously borrowed) clearly maintains the emphatic
adjunct function (Mazuka and Lust 1994: 152).85 Modern Tamil has the
same telltale overlap in function of the morpheme taan as an indeclinable
NP intensi®er and a declinable (long-distance) re¯exive; in Classical
Tamil, the intensi®er was a declinable postposition (Andronov 1969: 108).
The tendency to pronominalize is continuing in the more recent use of
taan as a nonre¯exive honori®c second- (or sometimes third-) person
pronoun. Even in the present language, the emphatic and the re¯exive are
more than mere accidental homonyms: ``with regard to the emphatic
352 Chapter 4

marker taan, one might wish to say that there is some semantic overlap
with the re¯exive pronoun taan,'' with another emphatic a½x, -ee, some-
times also being used to indicate re¯exivity (Asher 1985: 86). As Caldwell
(1961[1865]: 401) noted over a century ago, Dravidian taan ``must, I
think, have originated from some emphatic demonstrative base,'' a con-
clusion supported by later comparative Dravidian studies (e.g., Gondi
tanaa is ``only used in the third person and in the nominative'' (Burrow
and Emeneau 1961: 207); that is, I suspect, purely as an emphatic
adjunct).
Still, whatever the current empirical basis, the hypothesis seems plausi-
ble enough, and it is certainly open to empirical refutation. It clearly
makes interesting predictions about the evolution of anaphoric patterns
generally; it suggests, for example, that languages with long-distance
re¯exives may be midway between A-®rst and B-®rst patterns, and if so,
we might expect a greater ¯exibility in interpretation in such languages
than exists in English. Chinese is an interesting case in point, and it is
worth seeing how the facts might be incorporated into an account of this
sort.
In a thorough reexamination of Chinese anaphora, Huang (1991, 1994)
claims that no grammatical account of the Chomskyan type will be able
to handle the observed ¯exibility of interpretation associated, in particu-
lar, with the pronominal ta and the long-range re¯exive ziji. He thor-
oughly reviews the many recent attempts to bring Chinese within the
fold of the Binding Theory predictionsÐfor example by means of (a)
expanding the binding domain for ziji, (b) postulating movement at the
level of Logical Form, and (c) reclassifying ziji as a special kind of pro-
nominal. Each of these simply makes the wrong predictions. The essential
facts (Huang 1994: 76±78) are that Chinese has two candidate re¯exive
Anaphors, simple ziji and complex ta ziji (or pronominal ‡ ziji), which
distribute in mostly the same ways: they can be both short- and long-
distance bound (i.e., their antecedents can be local or inde®nitely distant),
and their antecedents need not be in c-commanding positions. They di¨er
in that ziji requires a subject antecedent and ta ziji does not; and ziji,
which is (unlike pronominal ‡ ziji) unmarked for person, assumes the
nearest possible antecedent is of the correct person (i.e., long-distance
linking is ruled out by a ``blocking rule'' if the distant antecedent is of a
di¨erent person than a local candidate). Rather than invoke the Binding
Theory apparatus, an accurate account of the behavior of these re¯exives
can be obtained, he claims, simply by the interaction of their lexical or
353 Grammar and Implicature

semantic content (referential dependency in particular) with pragmatic

principles of a neo-Gricean kind.
However, Huang (1991, 1994) argues that the A-®rst account (as in
Levinson 1987b) does not properly engage with Chinese, largely because
on that account there should be a strong tendency (via Q-implicature) for
the complementary interpretation of Anaphors and pronominals, whereas
in Chinese this is not the case. Instead, he argues that to account for the
patterns in Chinese one needs only M- and I-implicatures (with that order
of priority), together with a presumption of intraclause argument dis-
tinctness (the DRP). Note that this is closely related to the B-®rst account
already described (which was partly prompted by his work): on a B-®rst
account, the DRP is the foundation, and M-implicatures invoked by
emphatics are used to indicate coreference, whereas simple pronouns are
I-implicated to be closely coreferential beyond the clause. In a B-®rst
account, the Q-principle is otiose, because until a privative opposition
between pronoun and re¯exive has developed, there is no basis for scalar
(or Q-based) implicature. However, Huang (1994: 130) does not invoke
the M-principle to explain clausemate conjointness with ziji, which is
rather semantically explained: ziji is clearly lexically speci®ed as referen-
tially dependent and thus can override the pragmatic assumption of the
disjointness of co-arguments. The B-®rst account is tailored to a rather
di¨erent situation, where there is no such stabilized semantic encoding,
as in Old English. A schematic comparison between the accounts
may help to make the contrast clear (where > signi®es ``overrides'' and
where ``I-coreference'' indicates an inference to conjointness induced by
the I-principle, ``semantics'' indicates speci®ed referential dependency,
M‡> indicates M-implicates and so on):
(77) B-®rst account Huang's account of
Chinese
a. `John blames DRP>I-coreference86 DRP>I-coreference
him'
b. `John blames M>DRP Semantics>DRP
(him)self '
c. `John says that I-coreference I-coreference
he will come'
d. `John says that M‡>logophoricity M‡>logophoricity
self will come' I-coreference Semantics &
I-coreference
354 Chapter 4

Huang's (1994: 128±147) account of Chinese appears to work well to

explain both unacceptable and preferred interpretations of anaphoric ele-
ments. There is, in e¨ect, a global tendency to locally coreferential read-
ings of both the re¯exive ziji and the nonre¯exive pronoun ta, a tendency
defeated in the case of ta with a potential clausemate antecedent by the
presumption of co-argument distinctness. Contrasts between ta and ziji,
which may be expected by M-implicature, tend outside the domain of the
DRP to be interpreted as contrasts in emphaticness or logophoricity
rather than reference. Both ta and ziji contrast with lexical NPs, which
discourage coreferential interpretations by M. Because Chinese permits
zero-anaphora, this has to be brought into the picture too: except in spe-
cial circumstances, ta and zero tend to contrast in the expected way, with
the pronoun suggesting more distant coreference than the zero-anaphor
(p. 134¨ ).87
Chinese therefore di¨ers from a prototype B-®rst language like Old
English in having, in the form of ziji, a conventionalized pronominal that
is obligatorily referentially dependent but without a clear syntactic speci-
®cation of a binding domain. It is tempting to think of Chinese as transi-
tional between a B-®rst language and an A-®rst one, with ziji being in the
process of acquiring the Q-contrasts so typical of A-®rst languages.
Unfortunately, there seems to be no history of pronominals and re¯exives
in Chinese easily accessible to the nonspecialist. Perhaps some speculation
may be forgiven in the hope that specialists may be goaded into supplying
us with the facts. As already mentioned, ziji already in Classical Chinese had
both emphatic and re¯exive uses, and it continues to be used as an intensi®er
(e.g., in collocation with a subject; Li and Thompson 1981: 138f ). Thus, on
the hypothesis that ziji was in origin an emphatic, we can hypothesize an
earlier B-®rst state of the language with the following properties:
1. Intraclausal anaphoric patterns: use of the pronominal ta would
tend to I-implicate local coreference, except with clausemate antecedents
where the DRP would block coreference; this presumption being overruled
by the use of a marked, emphatic ta ziji or ziji that would M-implicate the
complementary interpretation, here coreference.
2. Cross-clause anaphoric patterns: here, where there is no presump-
tion of disjoint reference, ta would tend to pick up a local antecedent; the
use of (ta) ziji will M-implicate a contrastive interpretation but not nec-
essarily a contrast in reference (no longer especially salient, outside the
scope of the DRP)Ðthe contrast may be on the dimensions of unexpect-
edness, sole agency, or nondeictic (logophoric) point of view.88
355 Grammar and Implicature

Later, through intraclausal anaphoric usage, ziji must have acquired

speci®c semantic content, viz. necessary referential dependence, but it still
retains the unexpectedness or logophoric implications associated with its
original emphatic uses.
To what extent can we see the modern language as in diachronic pro-
gression from a B-®rst to an A-®rst language? First, let us see to what
extent Chinese can be seen as conforming to A-®rst predictions. We might
look at it this way: given that ta encompasses both deictic and anaphoric
uses, there is now an ordering of semantic informativeness over the two
expressions and thus the makings of a Horn scale hziji; tai, with the use
of the weaker ta suggesting that the use of ziji would be inappropriate.
But ziji, even in the present language, is not an Anaphor in the Binding
Theory sense, because it carries no speci®cation of the domain in which it
must be bound (Huang 1994: 88¨ ). The Q-opposition between ziji and ta
is thus rather weakÐziji is no more informative about where to ®nd an
antecedent, and, after all, most pronominal uses are anaphoric and thus
referentially dependent in any case. What strengthens the opposition is
the encoding of contrast or unexpectedness, emphasis, and logophoricity
associated with ziji. The use of ta therefore Q-implicates that either the
speaker wishes to avoid referential dependence or that he wishes to
avoid the logophoric or contrastive associations of ziji. Where reference
is the salient issue (especially, where there is a potential intraclause
antecedent such that the presumption of intraclausal argument distinctive-
ness is pertinent), the use of ta will strongly suggest disjoint reference,89
but elsewhere it is just as likely to suggest noncontrastiveness or non-
logophoricity. So we obtain the following patterns of preferred interpre-
tation (from Huang 1989: ch. 3, 1994: 125):
(78) a. Lao Wang piping le ziji.
Lao Wang criticize PFV self
`Wang1 has criticized self1 .'
b. Lao Wang piping le ta.
Lao Wang criticize PFV 3sg
`Wang1 has criticized him2 .'
(79) a. Lao Wang shuo ziji shi ge gongchengshi.
Lao Wang say self be CL engineer
`Wang1 says that self1 is an engineer.'
b. Lao Wang shuo ta shi ge gongchengshi.
Lao Wang say 3sg be CL engineer
`Wang1 says that he1 is an engineer.'
356 Chapter 4

Thus disjoint versus conjoint reference is the most salient opposition

between ta and ziji within the clause, because the DRP makes reference
the salient issue there, whereas the subjective, logophoric contrast is the
most salient opposition outside it. To this extent, an A-®rst account is
compatible with the facts.
As discussed above, Huang's (1991, 1994) account of all this is simpler:
in (79b) the presumption of clausemate argument distinctness overrides
the I-preference for local coreference, whereas the latter in turn is over-
ridden by the semantics of referential dependency in (78a). In (79a) the
preference for coreference goes through unchecked. But Huang's account
leaves questions unansweredÐfor example, why the ta/ziji opposition
fails to form a Q-implicating Horn scale. It also, perhaps, hides the
crosslinguistic generality of these patterns: if we bring the subjective
(logophoric) meaning contrasts back into the picture we can see why we
can have a pragmatic contrast without a change of reference. These
contrasts, subtle though they are, are surely a prominent part of these
systems, partially eclipsed by the focus on referential potential in gram-
matical theory but correctly emphasized in recent work by O'Connor
(1993), Kuno (1987), Sells (1987), Stirling (1993), and others. The contrast
between ta and ziji is therefore often interpreted nonreferentially, as
acknowledged by Huang (1991), where ziji is claimed to carry an ``unex-
pectedness message'' indicating either logophoric viewpoint or contras-
tiveness, when used in positions where a pronoun or zero-anaphor could
have the same reference. He has since gone on to document the logo-
phoric implications of ziji in detail (Huang 1994: 184±199), suggesting
that the logophoric implications of ziji are at least in part conven-
tionalized. One piece of evidence (op. cit. p. 196) in favor of this is that,
for example, a third-person antecedent cannot be searched for over a
more local ®rst-person antecedent, because the ®rst-person reference
de®nes a perspectival domain (hence, in gloss, *`John1 thinks that I trust
self1 too much'). In short, if we consider ta and ziji to constitute an
emerging privative opposition (with ziji necessarily referentially depen-
dent and logophoric, ta unmarked), we have an account of the Chinese
patterns that ®ts the crosslinguistic spectrum of systems, as intermediate
between A-®rst and B-®rst.
Such an intermediate position would also help to explain some rather
striking parallels between modern Chinese and the B-®rst Old English.
Consider, for example, the surprising possibility of the following ana-
phoric links:90
357 Grammar and Implicature

(80) a. Wang Xiaojie feichang xinshang ta ziji.

Wang Miss very admire 3sg self
`Miss Wang1 admires herself2 very much.'
b. Moyses, se De wñs Gode sua weorD Dñt he oft wiD hine
Moses he who was to-God so dear that he often with him
selfne sprñc.
self spoke
`Moses1 , who was so dear to God2 that he1 often spoke with
[God] Himself2 .'
The seeming paradox of these usages, where we ®nd apparent ``re¯exives''
associated with disjoint interpretations, is very hard to understand except
on a pragmatic account of a B-®rst kind. On a B-®rst account, an em-
phatic form may be used to implicate disjointness, overriding the DRP.
Double emphasis, though, may itelf override that interpretation in turn.91
As Huang (1994: 197±198) notes, both ziji and ta ziji continue to have
emphatic meanings: if ziji can implicate conjointness, then ta ziji can by
double contrast implicate the complementary interpretation (i.e., disjoint
reference). In some respects, then, Chinese appears to behave in the
manner of a B-®rst language like Old English, with ziji and ta ziji oper-
ating just like emphatics. This M-like behavior of ziji and ta ziji can also
be seen in the logophoric implications: example (81a) below (conforming
to the alleged ``blocking rule'' mentioned above) must be interpreted with
ziji linked to the ®rst-person antecedent, because this introduces a sub-
jective point of view, but the double form ta ziji in (81b) can override this
(Huang 1994: 196).
(81) a. Xiaoming1 renwei wo2 tai xiagxin ziji2 le.
Xiaoming1 thinks I2 too trust self2 CRS
`Xiaoming thinks that I trust myself too much.'
b. Xiaoming1 renwei wo2 tai xiagxin ta ziji2 le.
Xiaoming1 thinks I2 too trust 3sg self1 CRS
`Xiaoming thinks that I trust him(self ) too much.'
To this extent, then, Chinese looks rather like a B-®rst language, contrary
to the earlier discussed evidence for an emerging Q-opposition between ta
and ziji. These lines of evidence together converge to suggest an inter-
mediate, transitional state for Chinese. Note too that Old English with
its emphatic ``re¯exives'' could achieve long-distance anaphora just like
Chinese:
358 Chapter 4

(82) a. Xiaoming yiwei mama yao zeguai ziji le.

Xiaoming think mom will blame self CRS
`Xiaoming1 thinks that Mom2 will blame self1 .'
b. He [Gregorius] álmihtigne God b0d p0t
he Almighty God prayed that
`He1 prayed to Almighty God2 that
he hi mid his gife gescylde:
he to-them with his grace should-bestow
He2 should bestow His2 grace upon them:
7 p0t he him seolfum forgeafe p0t
and that he to-him self would-grant that
and that He2 to himself1 would grant that
he moste Done w0stm heora gewinnes
he might the fruits of-their labors
he1 might see the fruits of their labors
in heofona rices wuldre geseon.
in heaven's kingdom's glory see
in the glory of the Kingdom of Heaven.'
In the Chinese, the search for an antecedent for ziji very naturally leaps
over a clausemate subject, ®nding an antecedent in a matrix clause
(Huang 1989: 50, 91¨ ). Similarly, Mitchell (1985: 194) argues that him
seolfum in this extract from Bede ``refers to Gregorius, the subject of the
main clause, and not to God, the subject of the clause in which it stands.''
Thus the passage glossed `He1 prayed to Almighty God2 . . . that He2 to
himself1 would grant . . .' has the same pattern of anaphora as the Chinese
sentence. Mitchell (1985: 194) points out that often the -self form in Old
English ``is in e¨ect re¯exive, though not grammatically so'' in just this
way.
These are surely striking parallels across time and linguistic phylogeny.
In both Old English and Chinese, we seem to be dealing with a ¯exible
system of anaphoric interpretation, with only very partial coding of
Anaphoricity (i.e., referential dependency within a restricted domain),
where the referentially dependent pronominal clearly has primordial em-
phatic or intensi®cation functions that continue to a¨ect its use. In the
case of English, we know the further direction of development, toward
true Anaphors with restrictive binding domains; it may not be entirely
fanciful to imagine that Chinese is heading in the same direction, towards
359 Grammar and Implicature

an A-®rst system. But regardless of that, the categories A-®rst, B-®rst, and
transitional may provide a useful, if crude, ®rst step towards a cross-
linguistic typology of anaphoric systems, providing that they are taken
together with the pragmatic principles that generate the patterns and
severely restrict the possibilities of preferred interpretation.

4.5 CONCLUSIONS

4.5.1 Summary
This chapter has examined one crucial area, intrasentential anaphora,
where syntax and pragmatics seem to closely interact to generate patterns
of interpretation. I began by exploring the A-®rst account (Levinson
1987a,b), which attempts to reduce Binding Conditions B and C to
matters of generalized conversational implicature using an apparatus
independently motivated by much work in pragmatics. This analysis can
be extended to subsume Reinhart's theory that holds that only bound-
variable anaphora is grammatically speci®ed. I also noted that the prag-
matic theory could be made to engage with other approaches, such as those
with noncon®gurational theories of binding, but in each case precise pre-
dictions will be made which are a further constraint on theory construc-
tion. Now, the A-®rst account runs into a number of di½culties, notably
the failure of complementarity between pronouns and Anaphors. It was
shown, however, how this could be handled by the observation that pro-
nouns and re¯exives are in other kinds of contrast in these positions. In
developing this account, we also saw how it could be extended to cover
some of the facts about long-distance re¯exives, logophoric anaphoric
expressions, and even perhaps switch reference.
But there is one crucial piece of data that has been ignored by all the
work on AnaphorsÐnamely, the existence of languages that lack re¯ex-
ives altogether. Here an A-®rst account is clearly inapplicable. So for such
B-®rst languages (Levinson 1991) I have shown how the same Gricean
pragmatic principles can be used to predict a rather di¨erent strategy for
indicating coreference and disjoint reference. While the A-®rst account
explains how anaphoric patterns are projected by Gricean principles
applied to grammaticalized Anaphors, the B-®rst account shows how
grammaticalized Anaphors might arise in the ®rst place. The B-®rst
approach makes basic the DRP (i.e., the presumption of co-argument dis-
jointness) and explains proto-Anaphors in terms of the use of marked ex-
pressions to indicate a marked message, namely co-argument coreference.
360 Chapter 4

Thus it seems natural to suppose that all A-®rst languages have derived
from B-®rst ones. If we take this diachronic perspective, linking our A-
®rst and B-®rst accounts, we are then provided with a plausible explana-
tion of the persistent association of logophoricity (marked perspective)
and long-range re¯exives. The DRP, inherited by A-®rst systems from B-
®rst ones, explains why long-distance Anaphors are contrastive with pro-
nouns over reference inside the clause but may be contrastive over point
of view or logophoricity beyond the clause. Thus the third approach, a
diachronic A-from-B account, seems to o¨er a detailed explanation of
many of the facts of intrasentential anaphora across a broad spectrum of
languages. The fact that independently motivated Gricean principles, with
reasonably intricate projection rules, should happen to o¨er explanations
in the right directions over the details of, say, English or Chinese is quite
encouraging. Consider too that they throw light on such matters as
why English-speaking children should start learning Condition A before
learning Condition B, or why emphatics and re¯exives so often have the
same morphological form, or how English re¯exives evolved, or why the
typological facts about re¯exives seem to be the way they are.
All three of the pragmatic accounts are of course compatible: the A-
®rst and B-®rst accounts handle the synchronic patterns in di¨erent types
of languages, and the diachronic A-from-B account connects them his-
torically. But because the diachronic account gives a functionally moti-
vated account of Condition A, it further undermines the idea that even
this residue of the Binding Conditions is a direct re¯ection of innate pre-
dispositions, or Universal Grammar. Clearly, on this account, anaphoric
patterns are layered accretions of practices motivated by rational func-
tional principles, but which in becoming conventionalized make available
yet further pragmatic exploitations. The patterns are learnable on the
basis of evidence for the use of language taken together with the rational
principles underlying language use. The time scale for evolution of one
system into another (e.g., a B-®rst system into an A-®rst system), or of the
emergence of true Anaphors, may be measured in hundreds of years. All
this, of course, seems clearly at variance with a Chomskyan perspective,
where anaphoric patterns are instantiations of mental templates and are
considered to be in principle unlearnable, and where parameter switching
can only take place within the space of a single generation. Because the
conclusions are radical, it is worth brie¯y reviewing other evidence
against the view that the Binding Conditions are a deep part of our intel-
lectual makeup.
361 Grammar and Implicature

4.5.2 Pragmatics versus Parameters in Language Learning and Language Change

If the Binding Conditions were part of UG, the child's native endowment
for language acquisition, one might expect them to ®t well with the kind
of language-learning model associated with the Principles and Parameters
perspective, where the child's primary task is to select the correct settings
(for the local language) from a small, pregiven set of parameters. To re-
strict the possibility of unknowing overgeneralization (given the absence
of explicit information about the grammatical/ungrammatical boundary),
the child must, it seems, prefer parameter values that generate the
``smallest'' language consistent with the data, the so-called subset princi-
ple (Wexler and Manzini 1987). Applied to the Binding Conditions this
has the result that the child should always opt for the smallest domain D
for Anaphors consistent with the data, and the largest D for pronouns
(assuming that Anaphors and pronouns may have distinct Binding
domains, as with LDAs). In e¨ect, children should always opt for the
most restrictive language. But as we have already noted, this does not ®t
the data for child speakers of languages like English, at least, where chil-
dren allow nonre¯exives to have more unrestricted interpretations (Solan
1987, Kaufman 1994). Nor do children learning LDA systems seem to go
for a clause-bound interpretation initially (see, e.g., SigurjoÂnsdoÂttir,
Hyams, and Chien 1988 on Icelandic, and discussion in Lightfoot 1989:
370±371). And B-®rst languages like Old English would be predicted to
be highly marked, because the one set of pronouns can be interpreted
unrestrictively as either Anaphors or nonre¯exive pronouns.92 Yet the
evidence we reviewed showed that pidgins and creoles, which one would
expect to exhibit unmarked parameter settings, are also in initial stages
very likely to be B-®rst languages (Carden and Stewart 1988).
Compare now a pragmatic account of the learning of such systems. The
evidence for the Q- and I-principles abound in many aspects of language
data (see, e.g., Atlas and Levinson 1981; Levinson 1987a,b; Horn 1984,
1989). True, it is not clear exactly how the learner comes to formulate
such principles, some like Sperber and Wilson (1986) presuming that
something like these principles derives directly from innate predis-
positions to process information in certain ways, and others like Grice
and those who follow him more closely (like the present author) imagin-
ing that they arise from some complex interaction between native ratio-
nality and the demands of the communication setting (see Levinson 1989).
Be that as it may, such inferential principles (or something like them)
indubitably exist and are clearly utilized (to growing extents) by the lan-
362 Chapter 4

guage learner, as are many other, not speci®cally linguistic, inferential

procedures. Patterns of anaphora will then be inspected to see whether
they can be treated as instantiations of Q- or M-oppositions to the I-based
tendency to stereotypical and coreferential readings. If anaphoric patterns
are learned this way, then there are clear predictions for acquisition based
on the fact that the patterns arise from contrastive opposition. For ex-
ample, in an A-®rst language, where Condition B patterns are parasitic on
Condition A patterns, Condition A patterns would have to be learned
®rst. And so in fact they seem to be.93
The parameter-setting model also has clear implications for how one
would expect language change to occur. Lightfoot (1989: 325) suggests
such change should happen because a structural reanalysis by the lan-
guage learner of some syntactic patterns (a parameter switch) may force
certain data to be ignored as `impossible data', not part of the `trigger
(parameter setting) experience'. Vincent (1989a: 361) comments that one
would then expect a certain special kind of obsolescence, notable for ``the
relatively sudden disappearance from attested sources.'' For example,
once a himself form is analyzed as an Anaphor bound within a minimal
domain, one might expect rapid loss of homophonous forms bound fur-
ther a®eld; or once him is analyzed as a pronominal free in a domain, one
would expect rapid squeezing out of the use of him as a re¯exive. The loss
should take place in transmission across one generation, and although
one must allow for sociolinguistic di¨usion across a population, individ-
uals' usage should be consistent (Lightfoot 1989: 370).94 But again, none
of these predictions seem true of the history of anaphora in English, or
indeed other languages, where the establishment of, for example, Condi-
tion B±type patterns of usage began in Anglo-Saxon times but were not
complete in Shakespeare's. All this suggests that if language learning does
indeed generally proceed by the setting of parameters, the view that
Binding domains are candidate parameters does not have obvious em-
pirical support from the facts of acquisition or the indirect e¨ects that
modes of acquisition might have on language change.

4.5.3 Pragmatics and the Generative Program

There are striking crosslinguistic parallels in anaphoric patterns. There
are also some striking exceptions to what we have come to expect (e.g.,
the B-®rst languages). Finding a satisfactory explanation of both the deep
parallels and the anomalies is going to be di½cult. It will clearly involve a
universal theory of construal tendencies and the way this interacts with
363 Grammar and Implicature

language-speci®c details. One model for such an explanation is a para-

meterized Binding Theory interacting with other parameters in an innate
language-learning device. Another model, the one being suggested here, is
a universal pragmatic theory that covers the anaphoric tendencies, inter-
acting with language-speci®c details that are partly historical accretions of
practices.
The observations in this chapter, if correct, would seriously undermine
the view that the Binding Conditions are part of the native endowment of
human language. The A-®rst account suggests that all disjoint reference
conditions belong to pragmatics, with only Condition A as a plausible
candidate for a language universal; even here, the linguistic variation
suggests that perhaps we are dealing primarily with the semantics of nec-
essary referential dependency. The B-®rst account, though, suggests that
Condition A emerges from usage conditions. The nativist hypothesis, in
the absence of direct genetic evidence, can only be an argument from the
absence of any alternative plausible account, so the very possibility of a
pragmatic account of the core patterns of anaphora must raise doubts
about whether the Binding Conditions play as central a role in our innate
linguistic endowment as so much linguistic theorizing has presumed.
There is of course no general challenge here to some concept of a rich,
restrictive innate linguistic ability or Universal GrammarÐonly an argu-
ment that we need to look elsewhere. Equally, there is no suggestion here
that the current patterns of anaphora in, say, English, are entirely prag-
maticÐrather that the central facts can be seen to have functional moti-
vation, that diachronic explanations might refer more to pragmatic
principles than to principles of UG, and that defeasibilities of interpreta-
tion, at the margins as it were, are the stigmata of the functional origin of
currently grammaticalized patterns. Thus the presumed inability of a
pragmatic account to predict, for example, the details of re¯exives in
raising (or Exceptional Case Marking) contexts like John believes himself
to be ill does not undermine the thrust of the argumentÐonce a language
has grammaticalized an A-®rst system, such a system will have to treat
such weakened clause boundaries one way or the other, as permeable or
impermeable barriers to binding.
An important question is whether other well-researched areas of gram-
matical analysis might not turn out to allow pragmatic analyses. One such
possible target is the theory of Control, especially concerned with the
study of in®nitival complements, and the anaphoric linkage of the under-
stood subject (PRO) of that complement. On the classic analysis, the
364 Chapter 4

anaphoric linkage is determined by the subcategorization of the verb as

subject-controlling (as in (83a)) or object-controlling (as in (83b)).
(83) a. Max1 promised Mary PRO1 to come.
b. Max asked Mary1 PRO1 to come.
But as has long been noted, there are indications that the interpretation
depends also on pragmatic factors (e.g., the most probable scenario):
(84) a. Max asked Mary1 PRO1 to leave.
b. Max1 asked PRO1 to leave.
c. John appealed to Bill2 PRO2 to leave.
d. John1 appealed to Bill2 PRO1 to be allowed to leave.
In other languages, like German and Russian (Comrie 1984), Guugu
Yimithirr (Levinson 1987b), or Chinese (Huang 1994), the pragmatic
nature of these patterns is manifest, and the GCI apparatus applies with
considerable success to predict the correct outcomes (see especially Huang
1994). But even in English one can note that the patterns, which are at
least partially grammaticalized, conform to the pragmatic expectations.
Consider for example Chomsky's (1981: 65) Avoid Pronoun principle,
which speci®es a preference for PRO over an overt pronoun to indicate
coreference where both are allowable, as in:
(85) a. He1 'd prefer PRO1 going.
b. He1 'd prefer his2 going (disjoint reference preferred).
But this is of course nothing but Horn's division of pragmatic labor (as
Horn (1984: 23) points out), whereby an I-inference to local coreference
arises from a gap but is defeated by an M-inference from the more prolix
pronoun.95 In short, there are many other areas where a syntactic analysis
is generally presumed, but where a pragmatic analysis might predict the
crosslinguistic patterns from entirely general principles.
Sometimes, as in anaphora, rather far-reaching pragmatic reductionism
of allegedly syntactic principles may be feasible. But GCI theory in gen-
eral should be seen as complementary to syntactic theory, as part of the
explanatory theory for distributional patterns. The division of labor be-
tween syntax and pragmatics ought to be constantly in question in the
analysis of phenomena. If a pattern appears to be general but defeasible,
and explicable in terms of the general behavior of GCIs without any fur-
ther apparatus, the presumption must be that it does not belong to syntax.
If such a pattern is so explicable, but appears ®rm and not defeasible, we
can suspect frozen pragmaticsÐthat is, grammaticalized GCIsÐwith
365 Grammar and Implicature

consequent further pragmatic inferences made available (as in A-®rst

systems). Otherwise, we assume we are dealing with either general or
language-speci®c syntactic speci®cations. A pragmatic account working
in tandem with syntactic principles (as in the A-®rst account) can give us a
better overall explanation on three counts: (a) it is economical, because it
reduces what needs special, language-speci®c or arbitrary speci®cation;
(b) it accounts for restricted areas of defeasibility in interpretations just
where they occur; and (c) it further constrains the syntactic part of the
account by making predictions about further inferences from each syn-
tactic speci®cation of interpretation. That is the ultimate attraction of
bringing GCI theory into the overall explanatory apparatus.
However, even grammarians sympathetic to pragmatic explanations
have wondered whether pragmatic accounts, given their defeasible nature,
can ever play a signi®cant role in linguistic analysis and discovery: as
Zribi-Hertz (1995) puts it, ``The pragmatic approach thus leads to fuzzi-
ness and fuzziness is not heuristically productive.'' We have tried to show
that a system of interacting pragmatic principles, with well-de®ned inter-
actions between them, could in fact be heuristically valuable. Although
the default nature of such principles indeed raises methodological prob-
lems by the standards of exceptionless syntactic principles, major prag-
matic principles in turn have a signal methodological advantage over
syntactic principlesÐthey apply crosslinguistically across the board.96
Such pragmatic principles make us question, for example, any proposed
grammatical stipulation of disjoint reference, make us suspect any oppo-
sition between gaps and pronouns (or more generally simplex and com-
plex morphemes) as a potential case of Horn's division of pragmatic
labor, and make us scrutinize any privative opposition for its predicted Q-
implicatures. The I-principle suggests systematic links ought to be found
between clause-connectedness and coreferential interpretations, of the
kind represented by PRO-structures, but not su½ciently explored cross-
linguistically (see Foley and Van Valin 1983). Such principles also suggest
typological distinctions, like those between A-®rst and B-®rst systems of
anaphora (or systems with greater and lesser syntactic core; see Huang
1994: 259±260). They make rather detailed predictions about diachronic
trajectories, as in our B-®rst to A-®rst transition, from emphatics to
re¯exives, thus explaining the associations of re¯exives with logophor-
icity. It has to be conceded that these heuristics are still relatively broad
and crude, but then the detailed study of pragmatic principles has hardly
begun.
Chapter 5
Epilogue

In this book I have tried to champion the notion of a presumptive mean-

ing or preferred interpretation. The idea is simple but crucial to a theory
of communication: when we say something, we ®nd ourselves committed
to much more, just by virtue of choices between all the ways we could
have said it. There is nothing new in the idea, but the notion has never
been pushed to the forefront of the theory of meaning where it belongs,
together with the other, more familiar cast of charactersÐthe entail-
ments, presuppositions, metaphors, and so forth. One genus of such pre-
sumptive meanings is the category of GCIs, with the now-familiar species
of Q-, I- and M-inferences. Together with other principles governing idi-
omatic use of language, these constitute our third level of utterance-type
meaning, sitting between the coded meanings of linguistic expressions on
the one hand, and nonce-inferences to speaker-meaning on the other.
Because of the interstitial position of this third level of meaning,
reductionists will forever be tempted to assimilate it to either the level of
coded meaning or the level of nonce-inference. But this is a mistake, not
only because the interstitial nature of the phenomena with its defeasible
regularities will resist such a move, but because this third level interacts so
deeply with the structure of language that we will never build an adequate
linguistic theory without paying proper attention to it.
I hope to have been reasonably successful in making a case for GCIs
as an important kind of linguistic orphan. But why, one may ask, are
such pragmatic factors so relatively neglected in current linguistic theory?
One general source of resistance to incorporating pragmatic factors into
linguistic theory is the continued feeling that pragmatics is not well-
enough developed to yield any kinds of prediction.1 Second, the view that
pragmatics is concerned only with nonce-inference and with post hoc
accounts of speaker-meaning discourages the search for any systematic
368 Chapter 5

interrelation between pragmatics and core linguistic theory. Third, the

idea that pragmatics concerns itself with nonlinguistic reasoning that
happens to be applied to linguistic objects suggests that the essential
principles, insights, and goals lie far from those of linguistic theory.
Fourth, resistance to the invasion of pragmatic explanations into areas
erstwhile thought of as semantic or syntactic may be linked to a mistaken
view that such explanations go together with an extreme functionalism
that attempts to reduce all structure to use. Such reasons may add up to a
feeling that syntax, phonology, and semantics must go it alone.
GCI theory is very clearly not related to this perception of pragmatics.
Let me address the points above one by one. The theory makes clear
predictions, as I have tried to detail throughout this book (and I will
return to the point in a moment). It is not about nonce-inferences nor
about post hoc explanations about what a speaker may in full context
have meant: it is about what any reasonable interlocutor would under-
stand by a certain way of putting things. The theory is not about reason-
ing simpliciter: it is about special modes of inference that come into play
on the basis of our heuristics and their role in the special coordination
problems posed by meaningnn Ðindeed, as suggested in chapter 1, the
kind of reasoning involved may be the Ur-type of default reasoning. Nor
should it be misconstrued as a kind of radical reduction of structure to
function: GCIs are inferences from structure and meaning to further pre-
sumptive meanings. The role of the theory of GCIs in linguistic theory is
as an additional tool in the explanatory armory. It is not likely to replace
any of the alternative tools, but it o¨ers to shrink their domains of appli-
cation to more manageable proportions.
Let us pursue a few of these issues in a little more detail.

5.1 PREDICTIVE POWER OF THE THEORY OF GCIS

GCI theory in its current state is a set of generalizations about the

interrelations of linguistic expressions, the extra inferences that these in-
terrelations give rise to, and the conditions under which these default
inferences do not go through. These generalizations give us clear expec-
tations about inferences that ought to arise from particular expressions,
and in this sense the theory is predictive.2
GCI theory makes predictions about language structure at di¨erent
levels of generalization. At the lowest level, it makes predictions about the
369 Epilogue

presumptive meanings arising from individual morphemes and lexical

items in their respective semantic ®elds. For example, unmarked lexemes
will have specializations by I-inference to the stereotype, whereas their
marked doublets will pick up complementary interpretations; informa-
tionally asymmetric sets of items in the same ®eld will form scales,
inverted under negation, with regular scalar inferences; weak scalar items
will be in privative opposition with their stronger counterparts, with
compatible extensions excluded by implicature. GCIs do not of course
predict exactly what is lexically or grammatically codedÐthat is quite
largely a language-speci®c matter. But they do constrain what is lexi-
calized, and they do govern the likely relations between such codings.
Notice that the theory predicts multiple inferences even from quite small
semantic ®elds, as exempli®ed by the ``logical'' connectives with their Q-,
M-, and I-inferences (e.g., the scalar inference from or, the marked inter-
pretation of double negation, conditional perfection of if ) and the com-
plex interactions between them governed by the projection rule. Notice
too that the theory makes predictions not only about content but about
form/meaning relations: unmarked forms pick up I-inferences; marked
ones the complementary M-inferences.
At the next higher level, the theory makes general predictions about
constraints operating over the lexicalization patterns in such ®elds. Here,
for example, lies the crosslinguistic generalization that O-corners of the
square of opposition generally resist lexicalization, while blocking rules in
syntax and morphology (Arono¨ 1974, Williams 1997) are arguably
merely speci®c instantiations of Horn's (1984) division of labor between
I- and M-inferences. At a still higher level, it makes predictions about
the interaction of semantics and pragmaticsÐfor example, the intrusion
of pragmatic inferences into syntactic constructions where one constit-
uent is evaluated against the interpretation of another. Similarly, it sug-
gests economy in the syntactic component, with, for example, no need
for stipulation of disjoint-reference conditions. Putting these di¨erent
levels of abstraction together, we have clear predictions about pragmatic
constraints on language structureÐin e¨ect, predictions about possible
languages. These predictions arise because the Q-, M-, and I-heuristics
operate across the board, generating presumptive inferences that insert
themselves willy-nilly into the interpretations of utterances. They thus
play a decisive role in structuring lexical ®elds and syntactic construc-
tions, especially by redundancy constraints: what is implicated need not
be coded.
370 Chapter 5

In addition to predictions about synchronic structure at these di¨erent

levels of abstraction, GCI theory clearly ought to make predictions about
process. But here the predictions have not yet been worked out in any
detail. Take online utterance comprehension: the prediction clearly is that
GCIs should be available early, before full sentence-meaning is computed,
because of their association with linguistic expressions and their alter-
nates. There is very little psycholinguistic work directly addressed to
implicature, and still less of this concerns online processing, but one may
hope for rapid progress here.3 One might also expect Gricean inferences
to play a signi®cant role in hypothesis formation about the meaning of
linguistic expressions during language acquisition. The hypothesis would
be that children would have the principles early but apply them incor-
rectly because of incomplete control of the relevant semantic ®elds (and
this seems generally correct; see Jackson 1981; Jackson and Jacobs 1982;
Clark 1993, 1997). Finally, GCI theory should make interesting pre-
dictions about language change (as in the hypothesis in chapter 4 about
the evolution of A-®rst from B-®rst anaphoric systems). Because GCIs are
regularly associated with linguistic expressions, they may easily be mis-
analyzed by both theorist and language learner as part of the coded con-
tent. On the other hand, the general redundancy constraint (i.e., what is
implicated need not be coded) may be expected to put a brake on such
tendencies for GCIs to end up as part of the conventional meaning of
expressions. The present theory suggests that I-inferences might be pri-
mary engines of change, because inferences to the stereotype will re¯ect
changes in sociocultural environment. M-inferences to the complement of
I-inferences would then consequently be shifted also. But Q-inferences
might be less likely to become conventionalized on the grounds that they
arise from constrained sets of salient alternates, where change in the
meaning of one member would have immediate implications for the
understandings of the other members, the whole set acting as a constraint
on change within. These tendencies seem consistent with what is known
(see KoÈnig and Traugott 1988, Traugott 1998).
In sum, GCI theory makes widescale predictions about the association
of presumptive meanings with speci®c forms, and thus crosslinguistic
predictions about a myriad of such particulars as lexicalization con-
straints on quanti®ers, the kinds of predicates allowing Neg-raising, the
semantic generality of anaphoric expressions, the tendency for switch-
reference systems to lack strict coding of disjoint subjects, the rhetorical
import of periphrasis or double negations, and so on.
371 Epilogue

5.2 PRESUMPTIVE INFERENCE AND GENERAL REASONING

It may perhaps be useful to contrast the present approach with some rival
assumptions in Relevance theory. Wilson and Sperber (1986, 1991: 583)
have argued that pragmatics amounts to nothing more than central rea-
soning processes applied to linguistic stimuli: ``Grammar is a special-
purpose modular system; pragmatics is not a cognitive system at all. There
are no special-purpose pragmatic principles, maxims, strategies or rules;
pragmatics is simply the domain in which grammar, logic and memory
interact.'' They go on to suggest that ``many linguists have assumed
without question that speakers of English know a pragmatic code, anal-
ogous to a grammar, which enables them to recover the intended inter-
pretation. . . . They have assumed in other words that pragmatics is a
module, an extension of the grammar'' (1991: 584). There is little doubt
that they had Gricean theories of the present sort in mind as the target of
these remarks.
The present theory has indeed proposed ``special-purpose pragmatic
principles,'' but even here, just as Grice did with his maxims, we hope to
show that such principles follow from rational design characteristics (as
outlined in chapter 1). Regardless of our success in that, the other charges
certainly miss the target. GCI theory does not confuse inference under
heuristics with decoding (indeed the Relevance-theory claim that inter-
pretation involves the recovery of a single message under deduction seems
closer to a decoding view). Nor is there any assumption that pragmatics is
a module in the Fodorean sense of an encapsulated processing module.
However, GCI theory does suppose that there is a body of knowledge and
practice concerned with the use of language. This knowledge crucially
involves metalinguistic knowledge about the structure of the lexiconÐ
speci®cally, knowledge about the structuring of semantic ®elds, the avail-
ability of alternate expressions, subjective assessments of frequency and
markedness of speci®c expressions, knowledge about the stereotypical
associations of linguistic concepts in the speech community, mutual
assumption of principles for resolving con¯icts between inferences, and
so on. It also involves mutual awareness of the central heuristics and a
crucial understanding that recovering meaningnn involves solving coordi-
nation problems under those heuristics. These background assumptions
about the nature of communication radically change the nature of the
reasoning: if I see a man moving a hand in front of his face, the kind of
reasoning involved in trying to guess what he is trying to hit versus what
372 Chapter 5

he is trying to communicate seem very far apart, even though both infer-
ences are in the domain of intention-recovery (Levinson 1995b). The
essential di¨erence is that the mutual assumption of meaningnn and the
heuristics warrant the presumptive inferences that the theory tries to spell
out.
None of this amounts to anything like a module in the Fodorean sense.
But it does suggest that there are speci®c kinds of inferential analysis done
on the decoded content of utterances that are not applied to our analysis
of other things in the world, including any other noncommunicative
aspect of human behavior. Because (as I have argued) this special mode of
reasoning yielding default, defeasible inferences must interdigitate with
another kind of reasoning that yields monotonic inferences, pragmatic
reasoning of this sort must be organized in such a way that it can interact
with the results of semantic processing. We can, if we like, think about
pragmatic processing as speci®c kinds of subroutines that are called from
time to time. To that extent, pragmatics can be said to be a component in
an overall theory of grammar, without claiming that all the processes
involved have exclusively linguistic uses.
Sperber and Wilson also assume, along with many cognitive linguists,
that the semantic component outputs (possibly fragmentary) representa-
tions that are directly in the ``language of thought.'' This is improbable,
even assuming there is one such thing (most cognitive scientists assume
with good reason that there are multiple conceptual representational
systems). Semantic representations are going to be much too closely tied
to the properties of language in general (e.g., the linearity due to the
medium) and to the structure and content of speci®c languages (e.g., spe-
ci®c grammatical categories and constructions) to make this plausible.
Rather, it seems that semantic representations are likely to be intermedi-
ate formats, which interface with linguistic form at one end and general
conception (and indeed further specialized formats) at the other (Bierwisch
1996; Levinson 1997; see Jackendo¨ 1996a: 546±547 for discussion). If
pragmatic inferencing is working on semantic representations rather than
conceptual structures (which must, after all, include imagistic representa-
tions and the like), then that is another reason to think that pragmatic
reasoning is language-specialized. But, all of this is admittedly speculative.
Despite the close relationship between the inferences discussed in this
book and those that were the ultimate target of Grice's theory, many
researchers may feel that the notion of implicature has here been stretched
373 Epilogue

too far. First, there is the issue of how GCIs relate to the classical Gricean
scheme. In Grice's scheme, speakers implicate, but in my parlance,
utterance-types carry generalized implicatures (and the talk of heuristics
suggests a comprehension perspective): does this signal a category mistake?
Not at all: rational speakers meannn both what they say (except in non-
literal uses of language) and what that saying implicates; di¨erent levels
of meaning all come under the umbrella of meaningnn . Grice (1989: 88±
91, 117±137) himself made a number of distinctions within meaningnn Ð
for example, the distinctions between the applied timeless meaning of an
utterance-type (capturing disambiguations), the occasion-meaning of an
utterance-type (capturing what the speaker means by the use of the
words), and utterer's occasion-meaning (capturing what the speaker in-
tended). Although the distinctions are tortuous, I think that GCIs would
fall somewhere between the ®rst two distinctions (Grice indicates that
idioms would be part of applied timeless meaning, and what the words
are used to mean falls within occasion-meaning). Because GCIs are asso-
ciated with utterance-types, they can be accessed without considering the
utterer's occasion-meaning, and it is thus legitimate to view them neu-
trally, both from the speaker's and addressee's perspectives.
A second issue concerns whether GCIs are just too subliminal, as it
were, to count as implicatures. In chapter 3, I discussed possible termi-
nological distinctions (of the implicature, impliciture, and explicature
sort) in the area between what is said and what is not said at all but
implicated only. I concluded that boundaries are here hard to draw, pre-
cisely because of the way in which the very same inferences may or may
not ``intrude'' into the meanings of complex constructions, and can be
phrased in di¨erent ways by the analyst. All the authors there surveyed
presume that the inferential mechanisms would be essentially the same
in all cases, and I have made the same assumption, and argued there-
fore that there is no clear way to make such distinctions. This may be
wrong. It may be, for example, that the inferential procedures underlying
I-inferences are su½ciently distinct to belong to an entirely di¨erent
genusÐafter all, compared to the M- and Q-inferences they are much less
metalinguistic in character, and much more dependent on common pre-
sumptions about the world (i.e., stereotypes). We may expect that future
research will expose a much more detailed typology of pragmatic infer-
ence than we currently have at our disposal. But such a typology should
be based precisely on the underlying inferential procedures and corre-
sponding distinctions between the properties of inferences.
374 Chapter 5

5.3 ROLE OF GCIS IN LINGUISTIC THEORY

In chapter 4, I developed a pragmatic account of anaphoric patterns,

in partial competition with current syntactic accounts. The idea is that
anaphoric potential relies crucially on semantic generality, in line with the
I-principle, and that the intricate division of labor between NP-types
in di¨erent loci is explained by the complex interactions of Q-, M-, and
I-inferences, governed by their projection properties. Such moves have
been suggested before, but in general the attempts have been rejected by
syntacticians on the grounds that they simply don't get down to a su½-
cient level of detail to be interesting. I have tried to show that the detail
both intralinguistically and crosslinguistically, synchronically and dia-
chronically, ®ts the broad outlines of a pragmatic account. At the same
time, di¨erent possible analyses, with di¨erent degrees of pragmatic re-
duction, have been sketched to show that the general promise of such an
approach does not lie in a particular analysis but in the general way in
which it interdigitates with semantic and syntactic accounts.
Chapter 4 suggests a reanalysis of phenomena that have been thought
about as purely syntacticÐthat is, a shifting of the burden of explanation.
In the same way, many of the inferences discussed in chapter 2 have been
thought to be semanticÐthat is, part of the coded content of expressions;
here, an attempt is made to shift the burden o¨ the coded semantics into
pragmatics. These are issues about the division of labor between explan-
atory devices, and of course there will be plenty of room for further
argumentation. What I hope to have established is that GCIs have just
the right properties to assume part of this burden.
Chapter 3 introduces more radical issues about the relation of this
branch of pragmatics to linguistic theory. I have argued that GCIs force a
reconstrual of the semantics/pragmatics interface. Instead of a simple
feed-out, feed-in model of the interface, which continues to be assumed in
much work that is close to GCI theory (like Relevance theory), I have
argued that something more complex is required. But not much more
complex: what I have suggested is that we divorce the issue of represen-
tation from the issue of what processes or components contribute to that
representation.4 In that way we have a level of representation, say as in
DRT, to which the semantics contributes entailed conditions, whereas
default inferences contribute pragmatic content marked as subject to later
cancellation, and the whole structure can be semantically interpreted in
the normal way and can then be further enriched by additional pragmatic
375 Epilogue

inferences based on propositional content. Such a model seems to escape

with ease the chicken-and-egg problems (of pragmatic contributions to
semantic content) that assail the classical input/output model of the rela-
tion of semantics to pragmatics. At the same time, it keeps separate the
two crucially distinct kinds of informationÐthe defeasible and the inde-
feasible content.
There would be other ways to think about this model. One could
equally view it in the following way: GCI pragmatics is a theory that inter
alia maps (coded) semantic representations onto (enriched) semantic rep-
resentations. Semantics is now understood to do two things: map syn-
tactic structures onto (coded) semantic representations, and (enriched)
semantic representations onto propositions. Pragmatics likewise not only
maps initial semantic representations into enriched ones, but (if one likes)
propositions into additional propositions. And the two processes, each
masters of their own (indefeasible or defeasible) mode of inference, inter-
leave. Thus the general picture is simple enough, although we cannot
rule out additional interactionsÐfor example, the role of pragmatics in
extracting disambiguated (coded) semantic representations out of syntac-
tic structures in the ®rst place.
This picture though does change the way we think about (coded)
semantic content. The more power there is in the pragmatic theory, the
less rich such coded content has to be (a point well made by Sperber and
Wilson 1986). Other work in pragmatics and semantics is now converging
on the idea that such coded content may be abstract in the extreme and
may fail to determine anything like a logical form. This opens up intri-
guing questions about what exactly the shape of such representations is.
There are questions both about what we may call structural factors (the
syntax, as it were, of semantic representations), and content factors (rep-
resentations of lexemes in the object language). Many have assumed that
the structural properties of semantic representations are logic-like. But
there are reasons for doubt. Atlas (1989) argues persuasively that logical
negation is not part of our representation of meaning and that negation is
scope-indeterminate. Scope-free semantic representations have been pro-
posed by a number of authors (e.g., Bach 1982), and the motivations are
obvious. Hobbs (1996: 58) notes that a sentence like (1) has 120 possible
scopings:
(1) In most democratic countries most politicians can fool most of the
people on almost every issue most of the time.
376 Chapter 5

Yet we seem to understand it instantly, and it seems unlikely that speakers

or addressees actually compute even a small proportion of the possible
readings and then discard the irrelevant ones. Similar problems assail
other kinds of structural ambiguity (e.g., PP-attachment as in Look at the
man with a telescope), and rather than posit a combinatorial explosion of
readings, it makes good sense to consider the possibility that the structure
is indeterminate or underspeci®ed (see Matthews 1981, Van Deemter and
Peters 1995). The need for a pragmatically informed way of building
structure is particularly clear in the case of ellipsis (Kempson 1996).
Encoded lexical meanings may also be skeletal. Some authors have
arrived at this conclusion just by considering the apparently limitless
range of uses of lexemes, often crossing various syntactic categories (Ruhl
1989), or by documenting our ability to create what are clearly nonce uses
(Clark 1983, Clark and Clark 1979, Nunberg 1995). It has also been long
noticed that senses get specialized in verbal context, which argues against
both polysemy and classical conceptions of compositionality. Ross (1981),
for example, showed how the meanings of lexemes in construction with
one another alter each other's senses (consider the meanings of on in
barnacles on the hull, soap opera on TV, opinion on the matter, man on the
moon, cars on the move, and so forth). More recently, these issues have
come center stage in computational linguistics. Pustejovsky (1995), for
example, shows how ``co-compositional'' e¨ects can be achieved by
having semantically general speci®cations in one lexical item be ®lled in
by semantic speci®city in another, supplemented with pragmatic pre-
sumptions (so that John is enjoying the book suggests that the enjoyment
comes from reading it, whereas John is enjoying the ice cream suggests
another kind of pleasurable activity). As a consequence, we need seman-
tically general senses that will often be specialized by contextual or prag-
matic factors rather than intratextual ones. But not every combination of
words and background understandings can be forced into a meaningful
collocation (compare a sad day vs. *an afraid day). Consequently, Puste-
jovsky (1995) argues that we have to ®nd a middle ground between sense-
enumerative lexicons on the one hand, and fully unconstrained pragmatic
determinations of sense on the other.
Current work on semantic generality and underspeci®cation is gener-
ally conservative, modifying and relaxing older logical formalisms and
sense-representations in the direction of more open structures. But we
cannot rule out the possibility that much more radical rethinkings of the
concepts of semantic representation are in order (Atlas 1989). These are
377 Epilogue

surely among the most important current questions in the theory of

meaning. Here a developed theory of presumptive meanings may play a
crucial role: the more that can be said about default pragmatic inferences,
the more we can guess about the nature of the semantic representations
that must support them. Thus we can make signi®cant progress on these
issues only if we work simultaneously on both semantic and pragmatic
theory. The central contribution of GCI theory is to suggest speci®c ways
in which, given an underspeci®ed or indeterminate semantic representa-
tion of a sentence-meaning, the possible interpretations are not equi-
pollentÐthere is systematic biasing of interpretations, even before speci®c
contextual in¯uences have come into play. It is because of this predicted
biasing that GCI theory, unlike nonce-theories of inference, puts system-
atic constraints on a theory of underlying representations.
It was Grice's personal goal to try to build a synthesis between the
formalist and functionalist views of language (with Russell and Strawson,
respectively, as prototypes). He noted that the kind of phenomena sur-
veyed in this book are wrongly ignored by the formalists who fear weeds
in their tidy gardens, or as he more memorably put it, ``Like Cyrano de
Bergerac's nose, they are features which are prominent without being
dis®guring'' (1987: 373). I have tried to advance this same goal on more
thoroughly linguistic, and weedy, turf. The theoretical synthesis may or
may not be successful, but that is not ultimately the most important issue,
which is to keep those prominent pragmatic features in theoretical play.
For as Grice (1987: 379) remarked in a somewhat di¨erent context, ``The
continuation of progress depends to a large extent on the possibility of
`saving the phenomena'.''
Notes

Note to Students
1. The only careful collation of versions of Grice's texts that I know of is Arun-
dale 1997.

Introduction
1. Let me emphasize: this is not a general theory about the nature of communi-
cation. As far as I can see, such a theory will be a composite of many di¨erent
factors and considerations, including especially interactional factors almost totally
ignored in the present work. Most of the resistance to the theory of Generalized
Conversational Implicatures is based on a failure to understand its very limited
ambitions.
2. Parts of the manuscript have been in circulation for a long time and have their
own citation history (a version of chap. 3 in particular has gone the rounds since
1987, a version of chap. 1 since 1989, and a version of chap. 4 since 1991). In the
meantime, a number of younger scholars have written pertinent material, some-
times independently drawing attention to the same phenomena, and I have tried to
update all references correspondingly; apologies are due if I have overlooked some
recent work.
3. The problem is that, whereas the speaker reasoned from an intention to a ver-
bal means of achieving that intention, one can never reason back from conclusions
to the premises that produced themÐfor the simple reason that there are always
in®nite candidates. Because, as a matter of fact, we do seem to make correct
inferences of this kind, there has to be some heuristic limitation of the problem
space. I believe that these heuristics are of distinct kinds: the sort explored in GCI
theory, for example, as well as the di¨erent kind explored in the sequential anal-
ysis of discourse, which yields PCIs. All this is developed in chapter 1.
4. There are many suggestions about how to avoid this regress; in some version's
of Grice's theory of meaning it is avoided by a re¯exive (self-referring) intentionÐ
the speaker intending the recognition of that very intention. See, for example,
Bach 1987b.
380 Notes to Chapter 1

Chapter 1
1. Some authors think that this second version threatens an in®nite regress of
conditions (Schi¨er 1972), and thus some think it much better to stick to the
original account in terms of a single self-referring intention. See Harman 1974
(discussed in Avramides 1989: 56±58) and Bach 1987b, 1990: 391±393 on the
merits of accounts in terms of re¯exive versus iterated intentions.
2. An important departure from the classical Gricean account is sketched in Atlas
and Levinson 1981, where it is pointed out that it is the structure of the semantical
representation as well as the truth conditions that are the basis for the calculation
of implicatures. Manner implicatures can also generally be read o¨ the semantical
representation (e.g., p and ss p will have the same truth conditions, but ss p will
M-implicate that the stereotypical exemplication associated with p is not in-
tended), although in some lexical/morphological oppositions sheer markedness is
the operative factor.
3. The typographical emphasis is removed in the reprinting (Grice 1989: 37f ).
4. This construal of Relevance (quite unlike the Sperber and Wilson (1986)
informational version) correctly captures our pretheoretical intuitions that rele-
vance is about connectedness and collaborative activity, as I outlined in Levinson
1987a. It is also along the lines that Grice intended: the tying of relevance to par-
ticular plans and goals is very clear in Grice 1989 (p. 222), hence his objection
(p. 371±372) to the Sperber and Wilson conception. Plan-reconstruction accounts
of particularized implicature are now standard in the AI literature (see, e.g.,
Cohen, Morgan, and Pollack 1990, Green and Carberry 1993).
5. Compare for example these recent comments: ``His best known examples are
particularized implicatures; the discussion of generalized implicatures is restricted
to a few cases; and there is no evidence that he saw the distinction as theoretically
signi®cant'' (Sperber and Wilson 1987: 748) and ``The traditional distinction be-
tween generalized and particularized implicature is a false one, an artifact of the
inventiveness of analystsÐor lack thereof '' (Hirschberg 1985: 42). Yet Grice's
work on implicature opens with the issue of whether there is a way to account for
the systematic divergences between the logical connectives and natural-language
connectives (the issue which motivated his further development of the theory of
implicature; see Grice 1989: 374). In addition, there is every reason to think that he
was especially interested in the GCIs because they o¨ered some kind of evidence
for the way in which utterer's meaning might become associated with expression-
meaning (the theory of implicature was a preamble to that theory; see Grice 1989:
chap. 5 and 6).
6. This makes remarks like the following inapt: ``The distinction between `gener-
alized' and `particularized' conversational implicature is not represented in this
diagram [presented in Neale's paper] because it is theoretically inert (for Grice)''
(Neale 1992: 524, note 18).
7. For example (7), see Grice 1961 (1983: 251±252); 1978: 188; 1989: 44±46, 63f,
68. For example (8), see Grice 1981: 186; 1967: Lecture I, 10; 1989: 27, 68. For
example (9), see especially Grice 1989: chap. 4.
381 Notes to Chapter 1

8. The classical Gricean or Radical Pragmatics tradition took this semantically

broad meaning to be identical to the weaker or less speci®ed of the meanings
exhibited by the expression in question (e.g., some would have a lexical content
glossable roughly as `at least some', which is overlayed by the GCI `(and) not all'.
Atlas 1984a, and following him others, have argued that we should not presume
this uniformly: we need to entertain the possibility that the lexical content is not to
be equated with the weaker meaning but indi¨erent between the weaker and the
stronger. On his account, we need to distinguish, as candidates for the lexical
content, between a semantically broad but weakly speci®c meaning on the one
hand, and a semantically general meaning on the other.
9. Of course there are exceptions, notably neo-Griceans such as Bach (1984, 1995)
in philosophy of language or Horn (1989) or Huang (1994) in linguistics.
10. Speaker-meaning is here understood exclusivelyÐthat is, as speaker-meaning
minus sentence-meaning (of course in Grice's theory of communication, both
``what is said'' and ``what is implicated'' are ceteris paribus part of meaningnn , and
thus included in speaker-meaning). Similarly, I shall use utterance-type-meaning
as a shorthand for meaning that is not conventional meaning and not just speaker-
meaning (although following the Gricean scheme the speaker means it) but is
suggested by the utterance structure.
11. For Austin (1962), the perlocutionary act was a residual category that he was
keen to detach from the illocutionary act, the main focus of his interest, and it was
certainly broader and less well-de®ned than speaker-meaning. Nevertheless, this is
the Austinian category into which Gricean particularized implicatures would fall:
his prototype was an intended e¨ect on an audience which relates to what is said
at most obliquely (1962: 101), so that one can say ``by saying x, I did y'' (even the
terminology shows the kinship between Austinian and Gricean thought; see Grice
1986). Searle (1986: 210) for one holds that ``Grice argued that the intended e¨ects
of meaning were what Austin called perlocutionary e¨ects.''
12. Bach developed the notion speci®cally to account for explicit performatives
being primarily constatives while standardly (but indirectly) performing the action
they named. Later Bach and Harnish (1979) suggested various generalizations to,
for example, indirect speech acts.
13. Following a suggestion of John Lyons.
14. There is nothing easy, of course, about the notion of sentence-meaning. Most
formal semanticists operate with a notion of restricted utterance-meaning,
assuming deictic and, often, anaphoric resolution. See chapter 3.
15. I owe this observation to Rob van der Sandt.
16. Interestingly, it is not easy to ®nd all this spelled out in the textbooks. On the
special kind of processing of speech versus non-speech sounds, and other neuro-
physiological specializations, see Lieberman and Blumstein 1988: 148¨ . On the
perceptual match between auditory acuity and speech sounds see Lafon 1968: 81,
®g. 2.
382 Notes to Chapter 1

17. Uncontroversial it might seem, but Chomsky of course di¨ers. He has argued
(in, e.g., the 1981 Darwin lecture on ``Language and communication,'' or The
Minimalist Program, 1995: 18) that the mismatch between syntactic structures and
parsing strategies points to a more haphazard history of our communicational
abilities.
18. A range of measurements for di¨erent languages from di¨erent authors has
been given by Laver (1994: 541±544). Not all the estimates are in agreement. But
for English at least, the following facts look solid: in 20 hours of spontaneous
English, the fastest rate was 5.4 syllables/sec. or 13.4 segments/sec. Practiced
maximal rates for English peak, on optimal segments, at 8.2 syllables/sec., esti-
mated to be a physiologically conditioned maximum.
19. The basis for the calculation, which was kindly made for me by Bill Poser at
Stanford, is the following:
(a) the maximal normal speech rate is something like 7 syllables per second, with
each syllable consisting on average of 2.5 segments, making a rate of 17.5
segments per sec.;
(b) the information-theoretic content of each segment is taken to be log to base 2
multiplied by n (where n ˆ the number of the phonemes in the language); thus
for English (assuming that n ˆ 40) we get 5.5 bits per phoneme;
(c) so the data-transfer rate will be the number of segments encoded per second
multiplied by the information-theoretic value of each, thus 17:5  5:5 ˆ
96:3 bps.
Bill Poser tells me that there are specialists who wrangle over how exactly this
measure should in fact be computed for human speech. But the general point be-
ing made here is that such measures do not accurately calculate the rate of actual
human information transfer; only what the rate would be on the counterfactual
assumption that interpretive heuristics of various sorts were not in fact employed.
20. Calvert (1986: 178) notes a normal speaking rate of 10 sounds (segments) per
second, with an e¨ective maximum of 15 sounds/sec., and adds: ``While we typi-
cally produce speech at 10 sounds/second, we can understand it at as much as 30
sounds/second when paying careful attention'' (quoted in Laver 1994: 543).
21. This kind of argument from design was a favorite device of Grice's, most
evident in his seminars (see Grandy and Warner 1986: 31¨ ). The term pirot
may have been an oblique do¨ of the cap to Chomsky's (1957: 104) nonsense
example ``Pirots karulize etalically.''
22. Even from a recipient's perspective, inference seems cheapÐfor example,
there is no time advantage for resolving unambiguous pronominal anaphora over
``ambiguous'' zero-anaphors (Marslen-Wilson and Tyler, 1980). There may be
additional virtues of inference: inferred messages often seem to have greater im-
pact than coded ones, as long noted in traditional rhetoric. Green (1987) notes
that this runs counter to Relevance-theory metrics.
23. For example, information is informative relative to goals: we expect a route
description for a walker to be di¨erent than a route description for a driver, and
both to di¨er from a description of the same landscape for other purposes.
383 Notes to Chapter 1

24. The information-theoretic value of such a heuristic can easily be appreciated

by comparing it to the compression algorithm based on the idea ``only describe
what has changed'' with the decompression corrollary ``what isn't mentioned
hasn't changed.'' The heuristic may also be related to the closed-world assumption
used to reason in AI databases: if there's no information about whether x is a P,
assume that x is not a P (Reiter 1978).
25. These are not equivalent formulations of course, but here I am just interested
in sketching the range of possible domains for such a heuristic.
26. The information-theoretic value of this heuristic can be seen by comparison to
frames or data structures with inheritance by default, much used in AI (whereby,
e.g., if x is a bird, and birds generally ¯y, x can be assumed to ¯y). Such inferences
are potentially hazardous because they may be wrong, but where there is a tacit
communicational convention to speak in such a way that they can be assumed to
be true unless there is explicit warning to the contrary, the danger is removed.
Various formalizations of such default reasoning systems may be found in Gins-
berg 1987.
27. In the simplest case, ranked informativeness can be characterized in terms of
unilateral entailment. Thus in an arbitrary simplex sentence frame like the boys
came, All of the boys came entails Some of them came, provided the domain of
discourse contains at least one boy.
28. Con¯icting inferences would not alone justify the designation ``projection
problem,'' which has come to refer to how the inferences of complex expressions
may not just be the sum of the inferences due to the parts. But GCIs of complex
expressions involve just this kind of problem. See Levinson 1983: 142±144, 224±
225, and chap. 3 and 4 in this book.
29. Simpli®cations include: (i) Q inferences may be based on sets other than
informationally asymmetrical ones, (ii) I inferences may have other than stereo-
typical bases, (iii) the overiding or canceling of GCIs probably depends also on the
incremental order of clauses, and certainly Relevance PCIs may override all GCIs.
All of these issues are explored in chapter 2.
30. Some philosophers (e.g., Harman 1986: chap. 1), Kent Bach points out to me,
insist on a distinction between reasoning and argument (in the sense of proof in a
system), but those interested in capturing properties of human reasoning inevita-
bly think about reasoning in terms of systems of inference, which are the only
tools we have in which to model reasoning. Still, if one insists on this distinction,
defeasibility is a property of reasoning and nonmonotonicity is a property of
argument systems. Further, the exercise in this section may then be beside the
point: reasoning is an activity that may employ fragments of argument, jumps of
intuition, violent revisions of belief. Reasoning is an activity, not a system, and
di¨erent modes of reasoning cannot be captured in terms of distinct systems.
31. These properties of implicature are well known (see, e.g., Levinson 1983:
chap. 3), so those who want to hold onto a deductive account must, to preserve
monotonicity, argue that the apparent defeasibility of implicatures is not in fact a
matter of cancellation through the addition of a premise but rather through the
384 Notes to Chapter 1

replacement of the entire set of premises. Implicature cancellation would then be

exactly like having taken the wrong horn of an ambiguity, and being jolted by a
later context into backtracking and taking the other horn. But there is no evidence
for this and rather a lot against it, including the absence of garden-path phenom-
ena in implicature resolution (compare, e.g., the cancellation of the implicature in
``John stole some of the money, and possibly all of it,'' where there is no garden-
path phenomenon, to ``John didn't steal some of the money, he stole ALL of it,''
where there is). Another problem for the premise-swapping account is that there is
in fact no contradiction in an utterance like ``The facts are that John stole some of
the money, if not all of it. In fact, my own belief is that he didn't steal it all,''
where a scalar implicature is generated, then suspended by a clausal implicature,
but then the very same suspended inference is asserted. No serious attempt has
ever been made to explain all these well-known problems by the few (like Sperber
and Wilson 1986) who continue to maintain a deductive account (see Levinson
1989).
32. Some philosophers (e.g., Harman 1986: 4±5) think that the very notion of an
``inductive logic'' is a category mistake, arising from a confusion between reason-
ing and inference system.
33. Peirce (1932: 150±156) writes:
``The surprising fact, C, is observed;
But if A were true, C would be a matter of course;
Hence there is reason to suspect that A is true''
34. Thanks to Doug Edwards for pointing this out to me.
35. See, for example, Pople 1973, Reggia and Nau 1984, and Charniak and
McDermott 1987.
36. Actually, this account departs slightly from Gazdar's characterization. For
one thing, his model was concerned with a single speaker's commitment slate
rather than with a model of what is mutually taken for granted (Gazdar 1979:
130).
37. This tagging approach is in fact the one adopted in chapter 3, where a DRT-
type representation of utterance content is suggested, with both semantic and
pragmatic conditions represented, but the pragmatic ones kept demarcated so that
they are later erasable.
38. Perhaps this is what McCarthy (1980, 1986: 147) had in mind when he wrote
of his theory of nonmonotonic reasoning using circumscription: ``In puzzles, cir-
cumscription seems to be a rule of inference, while in life it is a rule of conjecture''
(where by ``puzzle'' he meant a linguistically posed problem).
39. Bach (1984) has, for example, suggested that default reasoning in general
(e.g., about the physical world) might be based on rather simple rules of thumbÐ
for example, `if it seems that p, infer p, unless a reason to the contrary occurs to
one'. Here the weight of keeping reasoning reasonable falls on some special, yet to
be elucidated, detection procedures for counterexamples (p. 45). But where tacit
coordination and the mutual presumption of heuristics are operative this gets
transformed: `if the speaker seems to mean p, (he knows that we'll mutually
385 Notes to Chapter 1

assume that) he does mean p, unless he signals otherwise'. The onus has been
transferred to the speaker. It seems to me that quite a few types of fallacy in
human reasoning can be attributed to confusing noncommunicational situations
with communicational ones (Levinson 1995b).
40. Some authors (e.g., Searle 1979) doubt that there is much value in a notion of
utterance-token meaning, on the grounds that an utterance-token has the same
semantic content as its type, and its import as a token is best captured in terms of
psychological notions (i.e., in terms of speaker-meaning).
41. GCIs are of course not untypical at all, if the phenomena surveyed in chapter
2 are anything to go by, and as I have argued they were certainly the ultimate
targets of Grice's investigations. Nevertheless, it should be emphasized that GCI
theory was never intended to replace the need for a theory of PCIs, although it
may give clues to how such a more general theory might be usefully developed (see
Hirschberg 1985). A great deal of misunderstanding has been generated by the
failure to see the need for the third level in the theory of communicationÐnamely,
a theory of utterance-type meaningÐand the failure to appreciate the very limited
ambitions of GCI theory.
42. The critiques of Relevance theory center on getting any kind of predictive
measure out of the theory (for a sympathetic attempt, which ends up adopting an
account Sperber and Wilson (1986: 130) reject, see Poznanski, 1992: 129±44). The
essence of the problem was given by Zipf 's (1949: 3±4) maxim of the ``singleness
of the superlative''Ðmaximizing the inferences while minimizing the e¨ort is a
double superlative that ``renders the problem completely meaningless.''
43. There are special controversies concerning the implicature analysis of cardinals;
see chapter 2. Those who prefer other scalar predicates can substitute examples of
the type John stole some of the books, if not all of them, and similar examples.
44. See Horn 1972, 1973; Gazdar 1979; Levinson 1983.
45. This holds except when the contradiction is read as a self-correction, as is
perhaps possible in (42b).
46. Thomason credits his student McCa¨erty with many of the central ideas here;
see McCa¨erty 1987.
47. Such accounts can be elaborated in various ways. For example, one can dis-
tinguish di¨erent kinds of accommodation (Thomason 1990: 352; see also Heim
1982: 381±384 and Soames 1989: 577±579), but these details are peripheral to the
present discussion.
48. However, I will not concede that their accounts are adequate even here; for
one thing, the concept of appropriateness or felicity on which the accounts rest is
problematic (see Levinson 1983: 24±27).
49. Heim (1982) did not accept the uniqueness condition; but see Kadmon (1987)
for a defense of it within a Kamp/Heim-type framework.
50. The entailments from A to I and E to O only go through, in modern standard
quanti®cational logic, if we assume existential import for the universals (i.e., if
there are some individuals quanti®ed over in the universe of discourse).
386 Notes to Chapter 2

51. Horn (1989: 260) discusses a principled exception: needn't is indeed inter-
preted as `not obligatory', but this is because need is a negative polarity item that
cannot be interpreted with internal negation as `obligatory not'. Thus there is not
the same pressure to interpret needn't in the same way that we interpret musn't, as
at the E corner of the square.
52. The prediction is thus that, for example, the She¨er Stroke (|, equivalent to
@(p & q)) will not appear as a surface lexical item in languages. For an extended
discussion of Gricean constraints on logical connectives, see Gazdar and Pullum
1976 and Gazdar 1979: 74±87.
53. Herb Clark has suggested to me that these lexicalization facts might be
accounted for without recourse to a theory of GCIs (which he is resistant to, being
himself a nonce-implicature theorist), along the following lines:
(a) what is infrequently used, fails to lexicalize;
(b) negative expressions are infrequently used;
(c) ergo, ``not all'' etc. fail to lexicalize.
The reasons this argument fails is precisely because of the asymmetry between the
lexicalization of the E corner (``none'', ``impossible'', etc.) and the absence of
lexicalization of the O corner.
54. The need for a theory of preferred interpretation in the theory of communi-
cation is perhaps just a re¯ex of a need for a new kind of theoretical object in the
social sciences generallyÐnamely, the maxim, the preferential mode of behavior.
Bourdieu (1977) has argued eloquently for this kind of construct (his ``habitus''),
distinct from the notion of norm or rule (with its strict complementarity between
permitted/prohibited or grammatical/ungrammatical), but capable of generating
regularities in behavior. It is the preferred ways of making bread, of cooking leeks,
of talking at dinner, or of marrying kin, passing on property, making a living, and
so on, that constitute a culture, just as much as the jural rules that have been the
major focus of anthropological research. I am chagrined to have to acknowledge
that John Haviland pointed this out me almost a decade ago before I could see
any sense at all in it.

Chapter 2
1. Thus Gazdar (1979: 46±47) argues that uttering ``p'' implicates in general that
the speaker knows that, or is committed to, pÐthus accounting for Moore's
Paradox (the oddity of asserting ``p but I don't believe that p''). Atlas 1993a
elaborates Grice's (1989: 42) objectionÐnamely, that this amounts to nothing more
than a restatement of the maxim. In Grice's scheme, Quality plays an important
part in the generation of PCIs like irony and sarcasm.
2. The reference (Grice 1989: 26) is to Kant's Critique of Pure Reason (1781: 70;
1933: 107). Irrespective of Kant, the four central Aristotelian categories of Quan-
tity, Quality, Relation, and Action (Manner or Modality) have long been central
in English philosophy of language. See, for example John Wilkins' 1641 Essay
towards a real character, and a philosophical language (Eco 1997: 240).
387 Notes to Chapter 2

3. I have phrased the principles in terms of a speaker's maxim and an addressee's

corollary for reasons of perspicuity. They could be phrased more neutrally as
heuristics to which both speaker and addressee are mutually oriented, as I
sketched in chapter 1. There is however a special onus on the speaker: he or she
will be understood ceteris paribus to have meant what (to employ the legal jargon)
``any reasonable man'' would have meant by the choice of expression that he or
she used under these heuristics.
4. Assuming the domain of discourse to be non-empty.
5. As already mentioned, ``U '' ‡> `p' designates that uttering U implicates p, or
more exactly an utterance of the type U engenders ceteris paribus the generalized
conversational implicature p. I shall however be loose for the sake of brevity and
reduce clauses to the relevant expressions; thus whereas one should say: `a sen-
tence of the form ``Some j 0 d'' ‡> `Not all j 0 d', I will sometimes reduce this to the
statement: ``Some'' ‡> `Not all'. This is indeed loose talk. Where important, I
distinguish between the implicated content alone (designated by ‡>) and the (rel-
evant) implicated content together with what is `said' (designated ‡‡>).
6. These utilize the epistemic logic formulated by Hintikka (1962). Readers may
®nd it helpful to think of this as essentially related to an S4 Modal Logic; thus
Kp 1@P @ p, and so on.
7. For example, whereas ``John has three children'' may have the weaker impli-
cature `the speaker doesn't know that John has four', ``I have three children''
clearly carries the stronger `the speaker knows he doesn't have four'. First-person
statements can then often be expected to carry the stronger inference.
8. There are however one or two distinct signs that the weak epistemic modi®ca-
tion is not correct for the scalar implicatures. For example, from the perhaps
rather special scale hknow, believei, an utterance of the form ``I believe that p''
will then generate an implicature `The speaker doesn't know that/whether the
speaker knows that p', which is incoherent or incorrect (whereas the stronger form
`The speaker knows that it is not the case that the speaker knows that p' is ®ne).
9. But Hirschberg (1985: 81) represents this as @BEL(S, `All came') in a three-
valued logic (so @BEL…S; p† ranges over cases where S believes p false and where
S does not know whether p is true).
10. Jay Atlas clari®ed the point to me thus: asserting a weak scalar ``A(W)''
implicates @A…S† because the addressee infers from ``A(W)'' that the speaker is
not in an epistemic position to assert ``A(S).'' The speaker intends that @A(S) is
recognized to be consistent with what the speaker knows, this to be the grounds
for adding plain @A(S) to the common ground.
11. The essence of the problem is that we need a wider range of distinct epistemic
commitments than are o¨ered by duals of K and P. The Horn-scale inferences are
in fact among the strongest commitments that can be implicated; much weaker
suggestions also arise, as I show later on.
12. Here the ceteris paribus clause allows of course for the defeasibility of impli-
catures. Gazdar (1979) made a distinction between potential implicatures and
388 Notes to Chapter 2

actual (uncanceled) implicatures, the output of the projection procedure, but I will
leave this implicit. It is in fact required for his projection solution, adopted in part
later on.
13. Thus K p can be read `The speaker believes p', and K p 1@P @ p, where
@P @ p is read `it is not consistent with what the speaker believes that not-p'. K
and P are thus related like r and t in, on Gazdar's 1979 formulation, an S4
modal logic.
14. See Gazdar (1979: 57±58) for more careful de®nition.
15. Van der Auwera (1995) attacks this lexicalization constraint because he wishes
to enlarge the scope of scalar implicature to account, for example, for conditional
perfection, (in the current framework, this will be given a completely di¨erent
account in terms of the I principle). In fact, the degree of lexicalization of scalar
items is crucial for understanding the complexities of scalar interpretations. The
point is made clearly by Horn (1989: 497±499), who notes that it is precisely be-
cause there is no lexicalization of, for example, `not everybody' that ``Everybody
didn't come'' has an ambiguity of scope. Matsumoto (1995: 45) also attacks the
lexicalization constraint in order to extend the scope of scalar implicature, even
though his own theory of the bounding of scalar implicatures by a ``conversational
condition'' (to the e¨ect that there should be no other reason why the speaker
might have avoided a stronger statement) is in fact more in line with the account
given here and the brevity condition of Levinson (1983: 135). (These remarks are
prompted by an unpublished note of Horn's.)
16. In dubbing this the ``aboutness'' constraint, Atlas and Levinson (1981: 42f )
gesture at the theory of ``aboutness'' developed by Putnam (1958); see also Atlas
1989: 100¨.
17. The phenomenon of scale reversal under negation was clear early on (see, e.g.,
Fauconnier 1975, 1979), but the fact that this solves the projection problem with-
out blocking under negation was apparently lost on later commentators.
18. Horn (1989: 169±170), reviewing the psychological literature, notes that the
latencies are best accounted for on the assumption that negatives are converted
into a½rmatives.
19. An early application of Gricean or proto-Gricean ideas to the analysis of the
Square is given by Fogelin (1967). For a review of the history of approaches here,
see Horn 1990. For interesting recent extensions in the analysis of modals, see Van
der Auwera 1996.
20. This is related to the restriction noted by Barwise and Cooper (1981) that
monotone-increasing and monotone-decreasing quanti®ers may not be conjoined,
thus *Few students and most faculty came, a constraint that can be described
alternatively as `Do not mix quanti®ers from positive and negative scales' (see
Horn 1989: 245¨ ).
21. The facts here are in fact underdescribed: many of the quanti®ers have speci®c
semantic constraints on their application. Thus most (as in Most of the men came)
requires at least three members in the domain of discourse, and many is not tol-
389 Notes to Chapter 2

erant in small domains (e.g., if many, say ®ve, of seven men came, then one could
not assert Many didn't come).
22. How do we determine this horizontal alignment? Partly in terms of which side
of the tolerance line they fall. But partly by Horn's stipulation that a quanti®er Q
and its internal negation Q. . .@ are linked by a horizontal line. Suppose then we
wish to know whether most has as a (sub)contrary few: we know that most is
intolerant, so can only have a contrary; so if few is to be a candidate, it must be
intolerant, which in fact it is (*Few of them came but few did not come). Most men
came has as contrary Most men did not come; so we can use the suspension tests to
see whether few is ranked higher or lower on the negative scale than most. . .not,
and what we ®nd is that we can say Most men didn't come, and in fact few came
but not ?? Few came and in fact most men didn't come, which suggests that few is
higher on the scale and thus not a true contrary for most. In fact, on Horn's
arithmetic model (sub)contraries Q and Q. . .@ should be so arranged that they
sum to zero (i.e., are linked by a horizontal line).
23. If a hint is required: the positive scale is (or at least includes) halways, almost
always, usually, often, sometimesi and its negative counterpart is hnever, hardly
ever, rarely/seldom, not alwaysi; see further hints given by Horn (1989: 238±239).
24. Such a paraphrase is not of course an analysis of the relevant semantic repre-
sentation, which might be formalized in various di¨erent ways.
25. Kempson also introduces the issue of examples like ``He doesn't have THREE
children, he has four'', rejecting Horn's account in terms of metalinguistic nega-
tion. I take up these issues in chapter 3.
26. Some scholars consider that there are a number of other fatal ¯aws in the
implicature account of numeral expressions. For example, Seuren (1998: 409)
thinks that exactly three would then incoherently mean `exactly at least three'; but
the solution is given by Kadmon (1987: chap. 4), who shows how an `at least' se-
mantics is compatible with modi®ers like exactly (and nonredundant with at least)
within a DRT framework. Essentially the solution works by treating three men
like an inde®nite phrase: three introduces an exact cardinality, but there is no
commitment that that the set of three men is exhaustive. She points out that such a
view is compatible with any further constraints on interpretation that may arise
from predicative uses or rhematic positioning (an observation that has led other
scholars, like Scharten (1997), to abandon the implicature account).
27. These data are from my own ®eld notes, but independent con®rmation can be
found in the glosses of Haviland (1979: 176).
28. Compare (i), where in B's utterance the cardinal has the normal GCI, which
thus contrasts rhetorically with A's utterance which lacks it; with (ii) where B's
utterance still has the GCI despite the fact that an `at least' interpretation would
contrast better rhetorically but is obscured by the GCI:
(i) A: I could earn at least ten dollars there.
B: Well, you will earn ten dollars here.
‡> no more than ten dollars
390 Notes to Chapter 2

(ii) A: I could earn only ten dollars there.

B: Well, you will earn ten dollars here.
‡> no more than ten dollars
A similar pattern holds for the classic scalar implicatures:
(iii) A: He stole at least some of the money there
B: Well, he stole some of the money here
‡> `not all'
(iv) A: He stole only some of the money there
B: Well, he stole some of the money here.
‡> `not all'
Thus although in these discourse contexts an implied contrast is required, the `at
least' interpretation fails to come to mind for both the cardinal and the quanti®er
case, where it would be most relevant. This suggests that in both cases there is a
GCI biasing interpretations in the same direction, regardless of the di¨erent dis-
course biasing (a pattern impossible to account for in a nonce-inference account
like Relevance theory).
29. This may be an over-simple analysis for English. Hawkins (1991: 414±416)
suggests the and that carry conventional implicatures that share the uniqueness
speci®cation but di¨er in the domains in which uniqueness is required, hence they
are not in a relationship of privative opposition.
30. The typology of tense systems is not well worked out. There are systems of
apparently six true tenses in Papuan languages, where the implicatural relations
can be expected to be rather di¨erent.
31. Alternatively, Larry Horn suggests to me, one could claim that in the stereo-
typical case one does not succeed without trying, and thus succeed I-implicates
`try'. But just how the generalized implicatures of expressions should play a prin-
cipled role in the construction of scales has never been worked out (indeed for
arguments against the transitivity of implicatures, see Atlas 1984b: 359, 372±373).
For remarks on the PCIs of context-dependent scales see Hirschberg 1985.
32. Another possible analysis, suggested by Kent Bach, is to hold on to the
entailment analysis by denying that (18c) is a literal use of language.
33. See Horn 1989: 268¨. The problem of the compatibility of distinct color
predicates is as old as the hills; see Frege 1884 (1953: 22) quoted in Harnish 1991:
316.
34. It is left unclear whether it is the linguistic expressions themselves that are
ranked or the entities they denote.
35. The classical theory of scalar implicature had also assumed partial orderings
(Gazdar 1979: 58, n. 22). Incidentally, to bring out the parallels with Horn scales,
I will display Hirschberg scales in the same (arbitrary) manner with the strongest
expressions to the left.
36. Hirschberg however indicates that on this theory, which attempts to reduce all
scalar implicatures to PCIs, some Horn scales will be reanalyzed. For example,
391 Notes to Chapter 2

Horn splits the temperature scale to give us two entailment scales hhot, warmi and
hcold, cooli; by contrast, Hirschberg (1985: 131) takes a single is-warmer-than
relation over both sets.
37. Another problem is that the uni®cation of phenomena that is one of the most
attractive parts of the theory is in fact partially illusory: we need in fact separate
rules for the generation of implicatures from unordered sets.
38. Burton-Roberts (1984) o¨ers a serious account of the pseudo-scale hnecessar-
ily p, pi. It was the need to rule out such absurdities that led Atlas and Levinson
(1981) to propose the lexicalization and aboutness constraints, which require salient
lexical alternates about the same relations.
39. An especially relevant cue may be a fall-rise intonational pattern indicating
focus within a given set. See Ladd 1980: 145±162; Horn 1989: 230±231.
40. At ®rst sight it seems that these clausal implicatures might be reduced to
scalar implicature, especially if (as argued, e.g., by Soames (1982)) one abandons
for scalar implicatures the strong epistemic modi®cation (K @ p) in favor of the
weaker modi®cation (@ K @ p) associated by Gazdar with clausal implicatures.
But intuitively, from the assertion of p or q more is (generally) implicated than
uncertainty about the applicability of the conjunctionÐnamely, uncertainty about
both individual propositions p and q. So alternatively (as Horn has suggested to
me) we could claim that it is the opposition between the assertion of p alone and
the disjunction that is responsible for that further inference. But this cannot be
achieved by a scalar inference because a scale of the sort h p, p or qi would fail the
lexicalization constraintÐessential to fend o¨ other pseudo-scales like hrp, pi.
What the concept of a clausal implicature captures is indeed this opposition be-
tween the assertion of p and the assertion of something that fails to entail p, but
the account relies on a constructional asymmetry between constructions that entail
p and those that fail to do so. Overall, I still ®nd Gazdar's concept essential.
41. This is because an overt violation of the second maxim of Quantity is also
likely to simultaneously infringe the maxim of Relevance or the maxim of
Manner, just as Grice foresaw.
42. The symbol ( : ) indicates a short pause.
43. M-intention is Grice's (1989: 105) shorthand for the complex re¯exive intention
involved in speaker's meaningÐnamely, the speaker's intention to cause an e¨ect in
the recipient just by getting the recipient to recognize that that was his/her intention.
44. For example, because of the stereotype that gold is yellow, white gold or blue
gold suggests a substance other than gold (cf. black gold for `oil'), although
jewellers in fact deal in both, and pure gold is white, uncontaminated with copper
(see also Putnam 1975: 250).
45. A formal machinery that might be adapted for characterizing such a notion
was explored originally in Carnap's (1947) concept of intensional isomorphism: to
capture the fact that say @…@ p† and p have the same intension but not the same
meaning, we need a notion of structured intensions in which they may be seen to
be nonisomorphic (see, e.g., Lewis 1972: 182±186).
392 Notes to Chapter 2

46. Those who believe in notions like explicature (Sperber and Wilson 1986) or
impliciture (Bach 1994) will ®nd in these characterizations of I-implicature further
reason for suspicions that many GCIs are really instances of these preimplicatural
enrichments of conventional meaning into something closer to what the speaker
intended. This is perhaps mostly a harmless terminological matter. However, if it
is thought that these notions will rescue us from serious dilemmas about the
interaction of semantics and pragmatics, then see chapter 3.
47. An inability based either on the speaker's knowledge that a more informative
proposition would be false, or on his or her lack of knowledge that it would be
true.
48. Horn (1984) dubs this the ``transderivational'' property, in reference to the
long-defunct rules in Generative Semantics that licensed a derivation only by
comparison to another parallel derivation.
49. Van der Auwera notes rightly that, in earlier formulations of the constraint,
Atlas and Levinson (1981) required equal lexicalization of expressions in a scale
and that negative scales (like hnone, not alli) are then counterexamples. The cor-
rect formulation of the constraint is that the stronger element must be equally or
more lexicalized than the weaker element. The properties of negative scales are
now correctly accommodated. But for the psychological reasons I discuss in sec-
tion 2.2.2 above, it may be better anyway to think of negative scales as introduced
by meta-rule from the positive scales. Either way, the lexicalization constraint now
escapes his strictures. However, in an interesting observation (1995: 5), he claims
that Chinese does lexicalize `i¨ ' and `if ' to approximately the same degree, and if
this is indeed so, the current theory predicts that the consequent Q-implicature
from the if-expression should defeat conditional perfection, because if one meant
the biconditional one should have said it.
50. Exactly how these utterances get read as conditionals is another matter; it has
much to do with the illocutionary force of advice/suggestion associated with this
use of the imperative, constituting a minor sentence-type (see Sadock and Zwicky
1985).
51. What I said was if p I-implicates q, then q must entail p, but that in both Q-
and I-induced implicatures this is only one of the conditions (Levinson 1987a: 74).
What I should have said is ``If p I-*implicates q, then q must entail p. However, in
neither the Q- nor the I-induced *implicatures is it [this necessary condition] any-
thing like a su½cient condition.'' I would have thought that the second sentence
made it su½ciently clear that the entailment condition between *implicatures and
what is said holds for both species of *implicature. The reason that Carston (1995)
attaches so much importance to this is that she hopes to make a distinction be-
tween explicature and implicature that hinges on implicatures not entailing expli-
catures. In fact, that distinction fails anyway, as I discuss in chapter 3.
52. This phenomenon was one of the ®rst to receive Gricean treatments by lin-
guists (see citations in Schmerling 1975).
53. An obvious problem with the generality of this account is that the phenome-
non of implicated temporal succession occurs with untensed verbs. Consider
393 Notes to Chapter 2

``Drinking a gallon of beer at the pub and driving home is wicked'', a type of
example made much of in the next chapter.
54. Most of Carston's (1995) supposed counterexamples to the proposed I-
implicature are therefore beside the point.
55. Guugu Yimithirr does not lacks complex clausal relators entirely; it is simply
that others are encoded. Guugu Yimithirr has three kinds of apprehensional
clause connectors that current English lacks entirely (Haviland 1979), for
example:
(i) Yuba-aygu dhadii bidha buli-igamu.
near-EMPH go ‡ IMP child fall-PRECAUTIONARY
`Go up close lest the child fall.'
56. Larry Horn points out to me that two indicatives can have a conditional in-
terpretation in English, too, as in You want it, you got it (especially with a rising
intonation on the ®rst clause). But the imperative plus indicative construction
clearly has (as he points out) more conventional conditional force, as indicated by
the negative polarity forms in, for example, Take any of that money and you're
dead.
57. This form of incremental processing is incorporated into formal semantic
models like DRT. Note too that presuppositions of the consequent can be asserted
in antecedents but not vice versa (as in If John has children, John's children must
have black hair too; cf. John has children. John's children must have black hair too.)
58. Why is this contrast between and and nothing (parataxis) not obvious in
English? It is obscured by a number of factors, such as prosodic marking, which
can turn a paratactic association into an unrelated list, and syntactic gapping and
reduction, which take the place of parataxis to suggest close association between
events.
59. However, verb omission or gapping proper in English actually inhibits a
causal reading (Levin and Prince 1986; Levin 1987): A man was shot and a woman
was arrested can be read as a causally related sequence, but A man was shot and a
woman arrested suggests a mere list. Clearly, gapping is restricted to sentences
with parallel structure and thus suggests parallel, perhaps simultaneous, events.
60. French windows in British English denote glass-®lled doors, or door-length
windows.
61. There are exceptions like hot, which participates both in the scale hhot, warmi
(and its negative corollary hnot-warm, not-hoti) and in the I-implicature from its
negation to its antonym cold. The prediction from my theory is that ``It is not hot''
should Q ‡> `It is warm', for this should defeat the inconsistent I-implicature `It
is cold'. This seems to be the case.
62. Assuming the subject is a rational thinker. Incidentally, I use the term ``read-
ing'' here for conciseness, without conceding that the construction is ambiguous.
Like Horn, I assume that there is in fact only one (semantically contradictory,
rather than contrary) reading and two understandings or interpretations.
394 Notes to Chapter 2

63. Although it does of course Q-implicate `It is possible that she'll come'.
64. For much crosslinguistic information, see Horn 1978, 1989.
65. In general, iconic factors enter into the present account in two ways: simple,
brief, unmarked expressions warrant maximally cohesive, stereotypical readings,
by virtue of mutual attention to the I-principle; and marked expressions indi-
cate the complementary interpretations by the M-principle that is about to be
explained. Iconicity itself is sui generis, but the exploitation of it by virtue of mu-
tual orientation to such licensing principles converts it into a matter of communi-
cative expectation and thus a pattern of generalized conversational implicature.
Hence it should occasion no surprise that many of the factors detailed in Haiman's
(1985a) notes on the iconicity of grammar surface again in a theory of GCIs.
66. The intoxicating I-implicatures of drink are now long conventionalized and
recognized in dictionaries as a separate sense, an autohyponym. In contrast, the
sexist I-implicatures of secretary or nurse are, argues Horn, matters of social
rather than linguistic convention. Certainly, I-implicatures seed the crystalization
of semantic change, and one can expect to see gradations of conventionalization.
But the distinction may not be so easy, because by changing social conventions
one can change linguistic ones: in British English vicar and vicaress meant, re-
spectively, male priest in the Church of England and his spouse, until the recent
innovation of female priests.
67. If this is so, notice that the child's task is greatly complicated by the tendency
for common vocabulary items to have irregular or suppletive forms instead of
regular derivations, as in the strong verbs in Germanic, with went instead of goed.
68. Like many observations useful to the GCI theorists, these derive from Gen-
erative Semantics, which attempted to ``syntacticize'' many acute observations
about meaning. The orphans of that theory, including the one discussed here or
Neg-raising discussed previously, can in many cases be happily adopted within
GCI theory. Newmeyer (1980), Harris (1993), and Seuren (1998: 502±526) o¨er
accounts of the Generative Semantics movement and reference to central works.
69. Note that the prior existence of sad, following the blocking principle men-
tioned earlier, requires that unhappyÐdespite being read as a contrary of happy
Ðsomehow means something di¨erent from sad; the solution with this (and other
cases like unintelligent) is to associate the derived form with a slightly less extreme
value on the scale (see Horn 1989: 279±280).
70. Schmerling (1975) discusses these cases and points to a number of signi®cant
syntactic di¨erences between the reduced and unreduced forms. For example, (i),
but not (ii) and certainly not (iii), seems to allow a violation of Ross's coordinate
structure constraint:
(i) ``Here's the whiskey that he went to the store and bought ''
(ii) ??*``Here's the whiskey that he went to the store and he bought ''
(iii) *``Here's the whiskey that he both went to the store and he bought ''
From this she concludes that the ®rst is not a real conjunction but more like a
complex VP of the He went and broke it kind. Note however that the implicature
395 Notes to Chapter 2

of (i) is not fully grammaticalized or conventionalized to the point of non-

defeasibility: ``He went to the store and bought some whiskey on the way home''
seems ®ne. An alternative possibility is that extractions of this kind are sensitive to
the full Gricean import of the conjoined VPs, the sort of heretical thought enter-
tained by Gordon and Lako¨ (1971).
71. This is not to deny that such patterns can become conventionalized in con-
structions of the kind studied in Construction Grammar, a candidate being There
are Ns and (there are) Ns, which suggests constrastive subsets of N.
72. A further wrinkle is the di¨erent possible phonetic realizations of the articles.
Thus the is normally reduced to [DP], but sometimes occurs with the close front
vowel as in [Di:]. As one might expect from the `division of pragmatic labour', [DP]
signals `business as usual' (by I-implicature), whereas [Di:] signals warning (by M-
implicature) of problems in production. See Fox Tree and Clark 1997. For other
examples of markedness signaling interactional problems, see Levinson 1987a.
73. Gazdar (1979) was careful to distinguish actual from potential implicaturesÐ
that is, inferences that may arise but be discarded as candidates for part of what
the speaker meant. I have been terminologically less exact, but I trust harmlessly:
a presumptive inference will have the status of a generalized implicature if there is
no reason to discard it.
74. For the reasons mentioned, it is relatively hard to ®nd cases where I-inferences
and Q-scalar inferences are in con¯ict. A puzzling case is discussed by Atlas and
Levinson (1981: 44±45), involving a negative scale and conjunction-buttressing:
i. ``It is not the case that John turned the switch and the motor started.''
Q-scale: h@( p or q), @( p & q)i
Q-inference: `p or q' i.e., `John turned the switch or the motor started'
I-inference: `It's not the case that John turned the switch and then as a result the
motor started'
Overall import of utterance:
Reading (a): (can be forced by stress on and ); Q defeats I:
`Either John turned the switch or the motor started.'
Reading (b): I-inference survives (where prosody indicates the conjunction is
one complex event): `It's not the case that John turned the switch thereby
causing the motor to start', i.e., the whole complex event did not occur.
The proper explanation of reading (i-b) is not clear. One possibility, explored in
chapter 3, is that the implicature here falls within the scope of the negation be-
cause it is, in a sense to be explored, incorporated within `what is said', or intrudes
into the truth conditions.
75. I have actually said little about ellipsis, but where, as here, the syntax does not
resolve the missing constituents, the I-interpretation to maximal cohesion and
coherence may supply them.
76. Another instance: in ``He was able to solve the problem and it's possible he
did so'', the ®rst clause carries an I-inference to `He did so'; but the Q (clausal)
inference from ``it's possible he did so'' to `for all the speaker knows he didn't do
so' defeats the I-inference.
396 Notes to Chapter 3

77. As mentioned, Soames (1982) suggests that both scalar and clausal implica-
tures should have the weak epistemic modi®cation, of the form sK p. Then
inconsistencies between scalars and clausal implicatures do not seem to arise and
one could dispense with the relative ordering of incrementation. But in fact this
proposal doesn't capture the intuitive di¨erence in epistemic content between the
two typesÐit is the epistemic uncertainty that is precisely the content of the
clausal inferences but not the scalars, which is why I have not followed him here.
78. This does remove one of the outstanding di¨erences between presupposition
projection and implicature projection, presuppositions by de®nition not being
blocked by negation. In the case of scalar implicatures, the reversal of the scale
comes to the same thing as blocking, but in the case of other implicatures, their
behavior can now be seen to be the same as presuppositions.
79. Gazdar (1979) added also presupposed content; in fact, he was especially
interested in the interactions between implicatures and presuppositions. But I
leave all this aside here.
80. Theoretically arbitrary distance has to be tempered by performance facts. In
fact, psycholinguists know well that conversationalists store only the gist of utter-
ances long term, not the form, and thus presumably not the distinction between
information coded and information implicated.
81. Such a notion has no relation to the Sperber-Wilson (1986) theory of Rele-
vance, which is in fact an informational principle like my I-principle, bounded by
e¨ort, or something like my M-principle.
82. Horn (1989: 145), having speci®ed antinomic Q- and R-principles of his own,
notes brie¯y in a column detailing the properties of conversational implicata
``Projection properties unclear, since conversational implicatures `may be indeter-
minate' (Grice); but cf. Gazdar 1979, Hirschberg 1985.'' Kent Bach points out to
me that one might hold that it is not the job of the GCI theorist to resolve the
projection problem anyway. The GCI theorist's job might be complete when the
default inferences are predicted, the rest of the story being left to the PCI theorist,
because projection properties will be a story about particular constructions and
usages. And as pointed out in chapter 3, some aspects of the projection problem
ought to be solved by a better understanding of the intrusion of pragmatic infer-
ences into the truth-conditional content of certain constructions.

Chapter 3
1. This chapter was, in an earlier form, widely circulated as a mimeo in 1987/
1988, and responses to that paper can be found in the literature; since then a
number of authors have taken up these themes and I have tried to incorporate
responses to them. The ideas were developed in lectures to the 1987 Linguistic
Institute of the LSA, and I am grateful to Larry Horn, who co-taught the course,
for much interaction over them. In the following year at Stanford, I bene®ted
from much discussion over these issues with Ivan Sag, and I am conscious of
having absorbed many ideas, references and ways of looking at issues from him,
which I have not acknowledged at every step. Also at Stanford, I gained much
397 Notes to Chapter 3

from discussions with Jerry Hobbs, Herb Clark, Stan Peters, Jon Barwise and
others; Craige Roberts put me in touch with then recent developments in DRT,
without which this chapter would have a lame ending. I thank Jack Hawkins,
Murvet EncË, Irene Heim, Ed Keenan, John Searle, Chuck Fillmore, and Paul Kay
for most useful comments on oral versions given at USC, UCLA and Berkeley.
The ideas have also bene®ted immensely from long discussion with Jay Atlas.
2. One such historical reason was the attrition of the ``wars of presupposition,''
between those who held that presuppositions were essentially pragmatic, and
possibly implicatural, in character, and those who held that they were essentially
semantic. The battles seem to have come to an end from mutual exhaustion
without ®nal resolution. Another such reason was the attempt to derail the
Gricean program by developing the theory of particularized implicatures at the
expense of the generalized, a program that arguably has little strictly linguistic
interest.
3. I consider this discovery of the di¨erent components of meaning to be the
fundamental advance in modern semantics. Of course, it may happen that (as I
myself have suggested) speci®c concepts like ``felicity condition'' and ``presuppo-
sition'' may in the end be reducible to others; but the point is that the observations
of apparently distinct processes constitute the advance. Those who, like Jacken-
do¨ or Sperber and Wilson, assume that there is no such variety of phenomena,
display, I think, a disregard for the ®ne-grained texture of linguistic meaning in
the hope of getting a simple macroscopic picture. Out of focus photographs can
have a certain charm, but out of focus X-rays are of no scienti®c use.
4. The notion of a component or level in linguistic theory has been thought about
both in terms of a distinct type of representation for each level and a distinct set of
operations over those representations. De®ning the relations between levels as
input/output relations requires that representations at the interface are compati-
ble, indeed essentially the same. Another way of thinking has it that there are spe-
cial mapping rules between levels, allowing much more signi®cant mismatches
between representations at the interface (Sadock 1991, Jackendo¨ 1997). Chomsky
(1995: 27) himself uses the term ``module'' in a special way of course, referring to
distinct sets of syntactic constraints (binding theory, Case theory, etc.) which in-
teract in more complex ways to yield, it is supposed, the observed and complex
patterns of grammatical well-formedness. Neither my use nor Chomsky's has
anything to do with Fodor's (1983) notion of module as a distinct mental faculty.
5. Any overall model of the theory of meaning could be construed in either an
abstract (``competence'') or a processual (``performance'') way. We shall be con-
cerned here primarily with an abstract perspective, essentially with what kinds of
information must be available to what processes, rather than the order in which
such processes might actually operate in speech production or comprehension. As
Bach (1994: 154±156) points out, despite noises to the contrary, we are in no
position to provide processing models at this stage.
6. The claim that semantics is not autonomous with respect to pragmatics could
easily amount to no more than an equivocation on the term ``semantics'' (Kent
398 Notes to Chapter 3

Bach points out to me). Indeed I assume that we need to distinguish one sense of
semantics, call it semantics1 , referring to a process that extracts an impoverished
level of semantic representation from surface structure and lexical content, from
another, call it semantics2 , which is a set of procedures that operate over a richer
level of semantic representation to compute sense relations, entailment relations
and truth conditions. The level of representation extracted by semantics1 may
perhaps be relatively autonomous, but even here we have to allow for pragmatic
resolution of ambiguities. However, semantics2 computes over a level of repre-
sentation to which pragmatics has contributed in fundamental ways. My claim is
that semantics2 is essentially linguistic and cannot be equated with conceptual
structure or thinking in general (as Sperber and Wilson 1986 or Jackendo¨ 1992a
suggest). It is the properties of semantics2 representations (not those of semantics1 )
that are described in semantic textbooks (under the rubrics of sense relations,
entailments, meaning-postulates, truth conditions), so that the claim that this level
is not autonomous with respect to pragmatics is not a vacuous claim.
7. The truth conditions of sentences or utterances? If utterances, so what? So asks
Kent Bach. But on the views developed below (and indeed in Bach 1982, 1987a,
1994), sentences don't have truth conditions, only utterances do. The interesting
question then becomes: what kinds of inferential process get us to the minimal
proposition-bearing entity, and my claim is that these are just the same kinds of
process, with the same kind of product, that elsewhere may take us from a prop-
osition expressed to something else the speaker meant, which is more obviously an
implicature.
8. Bach (1994: 143) points out that a slightly unfortunate consequence of this
de®nition is that `saying that p' then entails `meaning that p' (in Grice's sense of
meaningnn ), so that Grice is forced to say that a speaker who intends a nonliteral
interpretation of, for example, a trope has not `said that p' but merely `made as if
to say that p'. One cannot in pure Gricean terminology then talk about `saying
one thing but meaning another instead'. (Grice's concept of meaningnn has itself
been endlessly discussed; see, e.g., Schi¨er 1972, Avramides 1989.)
9. Compare, for example, Bach (1994: 142, citing Grice 1989: 87): ``Let me sug-
gest that . . . for Grice the contents of what is said are structured propositions
associated with syntactic forms,'' so that, for instance, active-passive pairs with
the same truth-conditional content will not count as `saying the same thing'.
Earlier, Bach and Harnish (1979: 29¨ ) had pointed out that the formula `U said
that p' must be taken opaquely; that is, what U said depends not only on the
extension but the intension, the sense, of the expressions used. Grice's own views
about how ®nely the contents of what is said are to be individuated beyond the
level of truth conditions are, it must be admitted, not clearly articulated.
10. It is a much discussed point whether anaphoric linkages should be thought of
as syntactic ambiguities (resolved by grammatical coindexing) or semantic ambi-
guities (resolved at the level of denotation assignment) or, as I argue in section
3.24 and in chapter 4, as semantic generality requiring pragmatic resolution (ex-
cept in the case of syntactically speci®ed coreference, as in re¯exives).
399 Notes to Chapter 3

11. Thus in example (4), if the utterance is interpreted with the bracketing ``He
likes [some [cats and dogs]]'', then the implicatures will include `He like some but
not all dogs'. If this is canceled by the preceding context, then we will have as
interpretation of the whole utterance `He likes some but not all cats and at least
some dogs', which does not seem to be what is conveyed.
12. Thus in the context of Montague's treatment of pragmatics as the study of
indexicality, Thomason (1974: 64) says ``the close similarity of pragmatic and
semantic theory raises the question of whether they are separate subjects at all.''
13. The two problems are obviously connected on some accounts; for example, on
the Russellian account of the uniqueness implications of de®nite descriptions, ref-
erential ``ambiguity'' amounts to falsity.
14. Principled exceptions are of course epithets, like the bastard, where a more
prolix form that would otherwise encourage a disjoint reading is clearly warranted
by the additional attitudinal information that can only be encoded in the fuller
nominal.
15. But because a tensed sentence usually follows another, it may pick up a ref-
erence time from the context; see discussion of conjoined or adjoined sentences in
chapter 2 and Partee's (1984) theory of implicit reference times involved in the use
of past tenses.
16. This is my appellation. Grice (1989: 138±139) himself identi®ed another circle,
possibly related, of this sort: can we give an account of the (timeless, literal)
meaning p of expression C in terms of people normally meaningnn (intending) p by
C, where what they intend or meannn by C will surely make reference to the
timeless, literal meaning of C? Incidentally, the phrase ``Grice's circle'' is not
meant to suggest that he had unwittingly tied himself in knots; Grice (1989: 375)
indicates that he was entirely cognizant of the problems treated here.
17. Other theorists of many di¨erent persuasions seem to have come to parallel
conclusions. Lewis (1979) supposed that accommodation (a kind of pragmatic
presumption) is involved in many uses of de®nite descriptions, in the narrowing
down of modal meanings to deontic versus epistemic, and so forth. Moore (1986)
in an AI context states ``the only way to hold fast to the position that the con-
struction of logical form precedes all pragmatic processing seems to be to put in
``dummy'' symbols for the unknown relation,'' and so concludes that ``a theoreti-
cally interesting level of logical form will have resolved contextually dependent
de®nite references, as well as the other `local' pragmatic indeterminacies'' (1981
(1986: 286)).
18. Whether that pragmatic process itself is peculiar to language interpretation is a
separate issue (I presume at least partially so; Wilson and Sperber 1986 presume not).
19. A convergent thought is Stalnaker's (1972) contention that semantics, qua the
theory of propositions, is essentially independent of linguistics: ``the subject has no
essential connection with languages at all, natural or arti®cial'' (1972: 382).
20. Recanati in that footnote reports a defense from Dan Sperber: the inference
from ``John has three children'' to `John has at most three children' should not be
400 Notes to Chapter 3

considered the relevant inference, as opposed to `John has exactly three children',
which entails what is said and is therefore an explicature. But this seems entirely
circular: we know the inference is an explicature, so we phrase it so it meets our
criteria!
21. Even though the notion of impliciture encompasses both the minimal propo-
sition expressed and the fuller enriched implicit proposition, Bach (1994: 157±
160), in contrast to Carston and Recanati, holds that the minimal proposition
plays a distinct and important functional role in getting to the expanded proposi-
tion. This is surely correct. But then why did he include the ``expansions'' beyond
the minimal proposition in the category of impliciture?
22. He adds that an implicature ``is a conceptually independent proposition, a
proposition with perhaps no constituents in common with what is said'', and gives
the example of an utterance of ``It's after 10'' implicating `the restaurant is closed'.
But we could have phrased the implicature as `The restaurant closes after 10', in
which case there is a constituent in common with what is said, and the inference
must be reclassi®ed as an impliciture.
23. Explicatures are derived just like implicatures by the monadic principle of
relevance (Sperber and Wilson 1986: 183±193), while ``both [completion and ex-
pansion aspects of impliciture] involve basically the same sort of pragmatic pro-
cess as in implicature proper'' (Bach 1994: 144).
24. This point is not, it seems, easy to get across. Thus Carston (1995: 239)
berates me for not supplying ``any coherent distinction between those implicatures
which contribute to what is said and those which function as independent
assumptions communicated by the utterance.'' But there is none. That is why I
don't subscribe to the terminological exercises surrounding ``explicature,'' ``impli-
cature,'' and so on.
25. This would be the line favored in some recent treatments of tense; see, for
example, Partee (1984: 257¨ ) who proposes ``if the main clause is an event-clause,
the last step in its processing is the resetting of the reference time to a time `just
after' the main clause event.'' Note though that not all sequences of simple past
tenses get interpreted this way:
(i) ``He read a book and made copious notes''
(ii) ``He played the pipes with one hand and he hit the drum with the other''
and that these other ways are various (in (i) the note-taking was not contempora-
neous with the reading, as the playing of pipes and hitting of drums was in (ii)).
These look more like a case of generality-plus-implicatures than a case of multiple
ambiguity (which would still involve resolution by implicatures).
26. It may be argued that there is some slippage here: some x are P can't be
intensionally equivalent to some and perhaps all x are P because only the latter
involves a modal notion; and in circumstances where `some but necessarily not all
x are P' is true, the ®rst will be true and the second false. However, I think the
objections are immaterial: the simpler nonmodal paraphrase as in (a 0 ) (with
stress on some and at least) would seem to give the same result. Despite the
401 Notes to Chapter 3

nondetachability of Q-implicatures, paraphrases with the same implicatures as

the target expression are hard to ®nd because periphrastic glosses generate M-
implicatures, thereby obscuring semantical equivalence. Granted that guarantee-
ing the truth-conditional equivalence of an implicature-lifting paraphrase is
di½cult, the fact that pretty much regardless of the exact paraphrase the sentence
is rescued from contradiction suggests that the only substantial di¨erence between
the A and B clauses in these `A is better than B' cases is that A has a scalar impli-
cature that B lacks.
27. Levinson (1983: 142±144) o¨ers an account of implicature cancellation in
cases like these.
28. Atlas (1993a: 171) discussing this argument of mine suggests a counter-
example: ``Eating some of the cake is better for my health than eating all of the
cake and drinking arsenic'', which has the structure A is better than B where B
entails A, and where the whole is clearly true without depending on implicatural
strengthening. Hence my additional constraint that the semantic structure of B
should not explicitly encompass A, as in privative opposites like horse versus
stallion. This is intended to capture the following kinds of cases, which seem to me
clearly anomalous without pragmatic strengthening:
(i) a. ``Eating some of the cake is better than eating all of it.''
b. Semantic gloss: `Eating at least some of the cake and perhaps all of it is
better than eating all of it'.
(ii) a. ``Owning a vehicle is better than owning a car.''
b. Semantic gloss: `Owning a vehicle, possibly a car, is better than owning a
vehicle which is a car'.
29. I take it in fact that the existential quanti®cation in all these accounts is too
weak.
30. Atlas (1993a: 171) complains that utterances of the form ``A is better than B''
show no anomaly when A is equivalent to B because these are intensional con-
texts. However, we can make all the same arguments with extensional predicates
like bigger than. For example, if the semantic content of ``A family with ®ve chil-
dren is bigger than a family with two children'' can be glossed `A family with at
least ®ve children is bigger than a family with at least two children', then what is
said would seem to be far from self-evidently true; the whole achieves its plausible
interpretation through upper-bounding implicatures on the cardinals.
31. For example, consider Atlas's (1984b) account of comparatives. The logical
form for `y is f-er than x' is as follows, where x* stands for the measures of f-ness
that x exhibits,
y* for measures of f-ness that y exhibits:
(by )(Ex )(x < y )
That is, there is some measure of f-ness that y exhibits such that for all measures
of f-ness that x exhibits, those measures are less than the y-measure. For example,
a stick y is longer than a stick x if there's some part of y that exceeds in length
every part of x.
402 Notes to Chapter 3

32. Note that in (23c) we need to make the auxiliary assumption that the domain
of discourse contains some exam that was cheated on, in order to get the entail-
ment relation between all and some. The logical form of such a sentence, Atlas
suggests, might be thought of as follows:
[ExEy (Student(y) & bu (exam(u) & cheats(y, u)) &
(Student(x) & Ev (exam(v) ! cheats(x, v))]
! better(y, x)
On this analysis, where there are no students who cheat on every exam the sen-
tence will be true (because the falsity of the antecedent will guarantee the truth of
the whole conditional); where there is a student who cheats on all exams, the sen-
tence will be false (because it is asserted that he's better than himself ).
33. Jay Atlas claims that there is no such entailment, and this follows from his
(1984b) account.
34. This kind of phrasing looks like a category mistake, Kent Bach points out to
me: traditionally, sentences entail and speakers implicate. But we are moving
steadily away from this picture. First, utterance-types carry with them presump-
tive meanings (GCIs, to which the speaker is committed of course); second, sen-
tences in the abstract are (at least in general) much too underspeci®ed to entail
anything. Thus in the new picture, it is utterances that both entail and implicate
propositions. The idea that it is only statements, not sentences, that entail or carry
truth, is of course not really new at all (see, e.g., Austin 1950).
35. Moore's paradox is of course the anomaly of the assertion ``p and I don't
believe that p.'' Grice (1978: 114) noted that this would follow from his maxim of
Quality but held that uttering ``p'' did not implicate `The speaker believes/knows
that p' on the grounds that such a presumption only expresses the maxim itself.
Others have been less cautious (see, e.g., Gazdar 1979: 46±48; Levinson 1983: 105)
but have noted that unlike most implicatures this inference is not deniable (hence
the paradox), indicating perhaps additional reinforcement through a sincerity
condition on assertion. See Atlas 1993a for further discussion.
36. In fact, Irene Heim points out to me, Lewis's account of because-clauses
relates them to counterfactuals; but there is no suggestion of such a move from
Barwise's camp.
37. An example of Ivan Sag's.
38. The argument was put forcefully by Greg Ward to Ivan Sag, who passed it on
to me, about the time that Bach (1987a) was putting it in print.
39. Bach argues that many uses of sentences are nonliteral: taken literally, they
express an absurdity or untruth and must be understood as expressing a related,
more explicit, more plausible proposition (Bach being in this regard, an incarna-
tion of our Obstinate Theorist). Thus, if I say ``I haven't eaten breakfast'', I mean
`not yet today' although the literal meaning of the sentence might entail `never'
(cf. ``I haven't eaten sheep's testicles''). On this account, we go around expressing
falsities, charitably interpreted as closely related propositions. Bach's account and
my account di¨er in that, using the mechanism of default interpretations, I think
403 Notes to Chapter 3

that many of these cases are pragmatically enriched before any truth conditions
are computed. I think my account ®ts better with intuitions about literality and
truth and falsity, but he thinks these can't be trusted.
40. In contrast, a PCI account of the numeral terms (as advocated by, e.g.,
Carston 1988, and reluctantly toyed with by Horn 1992b) under these circum-
stances of doubt should make no prediction at all.
41. A spontaneous textual example:
(i) ``I would've liked the two-cupboard cabinet; it's a three-cupboard one'' Rah:
P: 1: JMA p. 76.
A textual example of intrusion in the conditional is the following:
(ii) ``Imagine the di½culty of understanding this information if it were presented
one word at a time'' (from Ford and Thompson 1986: 359).
42. Of course, we here historicize: neither Russell (1905) nor Strawson (1950) yet
held the views of Grice (1967).
43. In some ways, the example is weakened by the presence of the second refer-
ring expression; given three visible children near the brother-in-law, he is (we
suppose) the only referent describable as the man with three children, so by elimi-
nation the speaker's brother must be the other man. But the second referring
expression is only there for expository reasons; the example stands without it.
44. The scale hsucceed, tryi is not quite prototypical; arguably, `he succeeded in
®nding the solution' doesn't entail that he tried; see chapter 2 and Horn 1972,
Hirschberg 1985, and Kay 1990 on Q-scales not based on entailment.
45. The hedge ``scale-like'' merely points to the fact that the opposition hlike, not
mindi, although it has all the logical properties of a Horn scale, would seem to fail
the lexicalization constraint posited by Atlas and Levinson (1981) and Levinson
(1987b). A less stringent constraint, posited by Horn, only requires that the
stronger item on a scale is more lexicalized than the weaker item (as in this case),
as discussed in chapter 2.
46. Note that The very tall man is even taller than the tall tall man should be
contradictory (or truth-valueless)Ðin short unusableÐon the treatment of tall tall
as very tall. But it seems to me that it could make good sense in the right context.
A related possibility is of course that the ®rst tall is a modi®er of the second tall, or
more exactly of the N 0 tall man, with a structure like:
[N 0 [ADJ tall] [N 0 [ ADJ tall] [N man]]]
and an interpretation that takes the set of tall men into a taller proper subset. But
we would then need an account of a red red sunset, a deep deep blue, an honest
honest man, the purest, purest honey, the tallest tallest building, and so on, where
intuitively the implicated content has less to do with objective qualities than it has
to do with speaker certainty. We'd also lose the intuitive distinction between a
[rich [rich]] man (rich among the rich) and a rich, rich man (rich, I assure you).
Note also that it is not only modi®ers that can be reduplicated; thus in American
English a dog dog or a salad salad implicates a prototypical dog and a real (not,
404 Notes to Chapter 3

e.g., fruit) salad, respectively. And there are contexts where one could say I gave
away the (ordinary) dog but kept the dog dog. But see the text for further exam-
ples.
47. Incidentally, Kripke's de®nition of speaker reference won't quite do even for
cases close to those he considers. His example:
i. ``Her husband is kind.''
said of a man who is in fact her kind lover (and distinct from her wicked husband).
But suppose the speaker says:
ii. ``Her lover is kind.''
where the speaker knows full well that the doting man in question is in fact her
husband. What the speaker says is true if the man (her husband) is her lover and
there are no other lovers in the wingsÐthere is surely no incompatibility between
being a husband and being a lover. But (ii) would be grossly misleading, of course,
to someone not apprised of the marital bond between that woman and that man;
example (ii) suggests that the man is not her husband by a Quantity implicature
(cf. Grice's (1975) discussion of his wife vs. a woman). Suppose further that the
addressee knows that the woman has a husband; then he will conclude that the
man being talked about is not her husband. Thus the proposition expressed will be
something like `That man is her lover rather than her husband and he is kind'.
In this second case, we intuitively want to say that the semantic reference results
in a true statement, and the speaker's reference results in a false or misleading one.
The point is that Kripke's de®nition of speaker-reference needs to be patched up
to allow for two further kinds of cases: (a) those (like the man with two children
discussed in the text) where the semantic conditions are knowingly not met but
successful speaker-reference takes place; and (b) those where the semantic con-
ditions are met but the pragmatic conditions of a nonmisleading referring expres-
sion are not met (as in our Her lover is kind said of her doting husband) and
speaker-reference is (arguably) not successful.
48. The passage is worth quoting:
If Donnellan had possessed a clear intuition that ``Her husband is kind to her'', uttered in
reference to the kind lover of a woman married to a cruel husband, expressed the literal
truth, then he would have adduced a phenomenon that conforms to the ambiguous D-
language [where referential and attributive references are truth-conditionally distinct] but is
incompatible with any Russell language [any language with the Russellian de®nite descrip-
tions]. (Square brackets embrace my explications.)
49. Note that what the semantics alone will provide, in the case of contradictions,
is singularly uninformative because every contradiction will have the same inten-
sion and extension. Not much semantic grist is thus provided for the pragmatic
mill.
50. Example from Jerry Hobbs.
51. For example, Katz speci®cally argues for a full semantic determination of the
modi®er relation, which we have taken to be given by I-inference: according to
him good knife can only mean `good for cutting with' (1987: 203¨ ), because
there's a `cutting' marker waiting as it were for modi®cation. This doesn't seem to
405 Notes to Chapter 3

be true: a good pocketknife may be good by virtue of its many accessories, and
indeed the Swiss Army knife company makes good business out of di¨erent knives
good for di¨erent purposes.
52. Consider for example the negative sentence:
i. ``I did not take some of your money''
and its positive counterpart:
ii. ``I took some of your money.''
The pronoun I is directly referential, and, given the same speaker, designates the
same entity in (i) and (ii)Ðit depends on no other element in the sentence and falls
outside the scope of negation or any other logical or modal operator. But the quan-
ti®er some falls of course under the scope of the negation in (i), and the scale upon
which any scalar implicature would depend is likewise sensitive to the negative scope.
Instead of the scale operative in (ii), viz. hall, somei, we have in (i) the scale hnot some,
not alli, and in fact therefore no scalar implicature at all (unless the utterance is
interpreted as a denial of a prior assertion). In short, implicatures have none of the
direct referential character that marks out indexicals. See EncË 1981: chap. 1.
53. In this discussion, I will argue that pragmatics should be considered a com-
ponent in a theory of meaning, not of course in the Fodorean sense (Fodor 1983)
of a specialized, encapsulated module, but in the sense (more normal in linguistics)
of a set of distinct principles (I continue to use the term module only because
component fails to yield an adjective like modular). Wilson and Sperber (1986)
argue explicitly against this (although there may be some confusing con¯ation
here between the notion of a linguistic level or component and a Fodorean module
proper). But their arguments are based on the assumption that Gricean mecha-
nisms are not inferential principles specialized to communication, and more par-
ticularly on the assumption that there are no GCIs. It should be clear that I make
neither of these assumptions: maxims like the I-, Q- and M-principles are, on my
account, speci®c heuristics adapted to the narrow bandwidth of human speech (or
other forms of human communication), and rely on metalinguistic reasoning. And
perhaps GCIs, by virtue of their resident default nature, have speci®c memory
requirements and thus need some mental home base, as it were. However, even if
pragmatic principles belong to no specialized linguistic component, but are free-
¯oating in a Fodorean central processor, none of the arguments below about the
interrelation of semantics and pragmatics seem to be weakened in any obvious way.
54. Heim (1983a: 171¨ ) does not map from syntactic structures directly to ®le
cards and their updating; rather she proposes an intermediate level of logical form
modeled on GB LF, at which level scope is disambiguated and anaphoric relations
assigned by coindexing.
55. Thus Heim notes in her conclusion (1982: 401):
We have derived considerable bene®t from permitting accommodation to be ``interspersed''
with interpretation. . . . This seems to me to add plausibility to the idea of a level of analysis
like our ®les, which are, so to speak, generated simultaneously by rules of diverse com-
ponents. The precise implications of this idea, and what it means for the relation between
semantics and pragmatics, remain to be explored.
406 Notes to Chapter 3

56. Kadmon notes (1987: 52f ) that the cancellation mechanism introduced by
Gazdar (1979) and generally assumed correct, viz. cancellation by inconsistent
prior background knowledge, cannot handle backwards cancellation as in I went
into a house; it was my house.
57. This is because within Kamp/Heim theories discourse referents are unbound
variables unless they occur within the scope of an explicit quanti®er or quanti®er-
inducing construction (like the conditional), and the rule of interpretation has the
e¨ect of an unselective existential quanti®cation over all unbound variables. This
peculiarity of the theory, whereby an assertion of the cardinality of x as n in the
DRS does not restrict the cardinality of x in the domain to exactly n, has the
bene®t that it is possible to get semantics for at least three distinct from that for
three (a problem that had always ba¿ed me). In e¨ect, at least three speci®es in
the DRS that the cardinality is equal or greater than three, thus not permitting an
upper-bounding implicature (Kadmon 1987: 80). Exactly three speci®es in the
DRS that any set in the domain that meets the conditions for that discourse ref-
erent is a member of the self-same threesome (p. 62).
58. An exception is made for numeral-determiner inde®nites in predicative posi-
tion that are claimed to have the semantic content `exactly three'; this is supposed
to be achieved automatically by Partee's type-shifting semantics (Kadmon 1987:
64).
59. However, let me record here one obvious inadequacy. The DRT treatment
looks as if it will work quite well for those implicatures we may call additive (i.e.,
GCIs and particularized implicatures that add pragmatic content to semantic
conditions). But there is another class of implicatures, which we may call sub-
tractive, which remove, coerce, or replace semantic conditions. (The terminology
comes via Sag from, I believe, Sadock.) Examples of subtractive implicatures are
ironies, and Lako¨ 's (1974) ``amalgams'' as in ``I saw you'll never guess how many
people at the party'' (where you'll never guess how many acts like a quanti®er).
Clearly these subtractive implicatures need to modify or replace semantic con-
ditions rather than merely add additional constraints. The problem is that this
process apparently requires semantic conditions to be defeasible. It may be that we
will need to insist that these apparent intrusions are indeed postsemantic, but I am
concerned by some evidence to the contrary (in part reviewed in section 3.5.3).
60. Copying operations are quite fundamental to DRTÐthat is what is involved
in anaphoric resolution. Various complex kinds of copying operation have also
been explored by, for example, Craige Roberts in unpublished work on the prag-
matic accommodation required to understand modal subordination.
61. Horn would no doubt respond that a generalization is being missed by this
accountÐwe now appear to have three kinds of negation: ordinary descriptive
negation, implicature-denying negation, and metalinguistic negation. In a sense,
this is correct: implicature-denying negations share one crucial property with the
descriptive cases (i.e., the meaning of negation as a truth function), but the retort
properties with the true metalinguistic cases. But in fact the DRT treatment sug-
gests that implicature-denying negation is just the intersection of descriptive
407 Notes to Chapter 4

negation with the echoic properties of retortsÐnothing special needs to be said,

unlike in the case of nonlogical metalinguistic negation. Geurts (1998) also holds
the view that metalinguistic negation is not in fact a unitary phenomenon.
62. This generalization just follows from the nature of implicature-inducing
scales. However, there are some reasons to think that negative scales should be
introduced by metarule over positive scales, as discussed in chapter 2.
63. If the process was entirely regular with the numerals, one might expect the
implicature `John has two children'. However, the numerals do seem weaker in
this regard. Yet Horn (1972) claims that it is quite general of scales that the middle
values are less strongly implicated.

Chapter 4
1. In this chapter I have drawn on material published in Levinson (1987a,b, 1991)
and the acknowledgements made therein are inherited here. However, I would
particularly like once again to thank Larry Horn for crucial initial ideas and Nigel
Vincent, Peter Matthews, and Yan Huang for early encouragement of heretical
views.
2. For discussion of so-called pragmatic agreement, see Pollard and Sag 1994.
Note that for some such cases one might hope for a semantic account. Consider:
(i) a. The ham sandwiches has left without paying.
‡> The man who ordered the ham sandwiches has gone without paying
b. His family are coming to dinner.
c. Five children in a car is the worst thing imaginable for any driver.
d. Two eggs every morning is bad for your health.
e. Vous eÃtes le professeur.
In (ia) we seem to have a syntactic re¯ex of pragmatic inferences associated with
an expression. In (ib) real-world number (optionally) overrides grammatical sin-
gularity, and in (ic) construal of ®ve children in a car as a singular situation per-
mits singular verb agreement, and in (id) two eggs functions as a singular menu.
One might hope to account for (ib±d) in terms of some general ambiguity of plural
NPs between, say, plural and singular collective readings. But this will not doÐwe
should then be able to say *The men is coming in and we can't. Inevitably, some
process of pragmatic construal seems to be involved. Implicature proper is more
clearly involved in other cases, for example, honori®c plurals may collocate with
singular predications, as in (ie), where the plural form originally merely implicated
respect, and in some languages continues to do so rather than being fully con-
ventionalized (Brown and Levinson 1987: 198¨ ).
3. I have borrowed the phrase, not entirely fairly, from Saleemi's (1997) review of
Huang's (1994) attempt to reduce anaphoric patterns in Chinese to pragmatic
principles.
4. Various syntactic alternates have been proposed to di¨er in just this way. For
example Steever (1977) argues that raised structures like John believes Sue to be
honest implicate subjective access or judgment (I-implicatures from reduced
408 Notes to Chapter 4

structures in our terms), compared to their unraised counterparts (M-implicating

the complement). Chomsky (1981: 65) points out that the Avoid Pronoun alter-
nations in John would prefer going (coreferential) versus John would prefer his
going (preferably disjoint) might be related to Gricean factors, a point taken up
by Horn (1984). Levinson (1987b: 417±420) argues that Control structures ®t the
same pattern and are not fully grammaticalizedÐsee Comrie 1984 on Russian
and Huang 1994: 149±158 on Chinese. I return to this point in section 4.5.3.
5. This and related issues are discussed by Horn (1989: 345±353). Indirect speech
acts are another class of cases where collocational constraints are associated with
standardized or short-circuited presumptive inferences. Thus preverbal please is
associated with indirect requests, but only those that have idiomatic status (Lako¨
1973; see Levinson 1983: 272¨ for references):
(i) a. Please open the window.
b. Would you mind please opening the window?
c. ?Wouldn't you mind please opening the window?
In a similar way, honori®c particles will not collocate happily with implicated
disrespect (Brown and Levinson 1987). Another interesting class of cases involve
sentential amalgams (Lako¨ 1974), where whole clauses can slot into modi®er
positions on the basis that those clauses would have implicated the relevant
modi®er:
(ii) a. ``John invited you'll never guess how many people to his party.''
b. ‡> `John invited a lot of people to his party'
c. ``You'll never guess how many people John invited to his party.''
d. ‡> `John invited a lot of people to his party' ''
e. ?John invited Bill will never guess how many people to his party.
Here the sentence in (iia) is interpreted to mean roughly (iib), by virtue of the fact
that uttering (iic) (but not, e.g., (iie)) would have implicated (iid). Thus the nonce
modi®er seems licensed by an implicature. The appropriate theoretical moves here
are not so clear, but the phenomena certainly support the idea broached in chapter
3 that implicature can play a role inside the process of semantic interpretation,
and that an incremental construction of a construal can play a role in licensing
grammatical processes like agreement, or treatment of a sentence fragment as an
NP as in (iia).
6. Perhaps the last point requires a further word. Defeasibility has to be, as far as
I can see, the litmus test for a grammatical versus pragmatic account of linguistic
patterns. Chomsky (1981: 227 n. 45) for one will not agree. He would rather sup-
pose that pragmatics may override grammatical stipulation. It is true, of course,
that under special conditions (e.g., a noisy environment), anything goes that
works. But the greater the range of interpretations, and the more normal defea-
sible interpretations seem, the less plausible the grammatical account becomes. Let
us concede though that the big picture counts too. If we have perfectly general
principles, which work across language and provide deep insights into the work-
ings of speci®c languages, and which would give us an account of the phenomenon
in question ``for free,'' then naturally we want to hang on to that explanation.
409 Notes to Chapter 4

However, both pragmatic principles and principles of Universal Grammar may be

of this kind, and in seeking to adjudicate between rival accounts we may then be
forced to consider the essential character of the phenomena to be explained. Are
they grammatical stipulations or merely preferred interpretations? It is this which
brings us back to defeasibility as the test.
7. The hypothesis is that any language with deictic pronouns will allow these to be
used both anaphorically and as variables. Certainly, there are also specialized
referentially dependent expressions, like long-distance anaphors to be discussed in
section 4.3.2.5.
8. Although semantic generality is clearly intrinsically connected to referential
dependence, the former does not of course automatically engender the latter. It is
quite possible to have a semantically general de®nite description like the man serve
as a referentially independent expression as well as (on some other occasion) a
referentially dependent one. Even though re¯exive pronouns encode referential
dependency in a certain domain (and are therefore in that respect not semantically
general), nevertheless in many languages they are interestingly stripped of num-
ber/gender features and thus conform to this general pattern (for an interesting
reanalysis of GB theory which makes this generality central, see Burzio 1991).
9. A possible exception is switch reference, in which a subordinate or coordinate
verb may indicate that its subject is identical to that of a prior matrix or coordi-
nate verb; in that case, perhaps the linkage is itself grammatically speci®ed (see
section 4.3.2.5.1). Another kind of exception would be local subject-oriented
Anaphors, those re¯exives or reciprocals that require an immediate subject as
antecedent.
10. Here and throughout, subscript indices merely serve to pick out the relevant
reading (same subscript indicating coreference) and have no theoretical status.
11. I use the term disjoint to mean local noncoreference or lack of anaphoric links
to immediately preceding NPs. Obviously, such locally disjoint NPs may be
coreferential with still earlier NPs (see Fox 1987 for empirical tendencies).
12. See Vonk, Hustinx, and Simons 1992 for experimental con®rmation of this
signaling function of choice between anaphoric expressions.
13. Kent Bach has objected to this kind of phrasing: addressees are not in the
business of inferring coreference or its complementÐrather, they are in the busi-
ness of determining reference. That I think is true. GCIs, on the account devel-
oped in this chapter, at most help to signal how the addressee is to ®nd the
reference; just as often (as in the case of I-inferences) the assumption of corefer-
ence will be a side product of a larger inference (e.g., ®nding a maximally cohesive
interpretation).
14. In a critique of my earlier papers (Levinson 1987a,b, 1991), Ariel (1994: 20)
suggests that I am deeply confused about informativeness: ``If anything, then,
nonstereotypic interpretations are more informative than stereotypic ones, in that
they eliminate from consideration a more deeply engrained belief. In other words,
the I-principle cannot both be informativity and stereotypy.'' In fact, however,
410 Notes to Chapter 4

she is herself confusing two very di¨erent notions: (i) the Shannon information-
theoretic measure, which de®nes informativeness as inversely related to the statis-
tical probability of signals; and (ii) the Bar-Hillel/Carnap measure, which relates
informativeness to the discrimination power of message contents. The former is
syntactic and probabilistic and plays absolutely no role in the theory here; the
latter is semantic and structural and is the relevant theoretical foundation to this
work. See Cherry 1966: 238¨ for elementary exposition.
15. To illustrate: given two predicates M and Y, and one entity a, M(a) will be
compatible with just two state descriptions or states of a¨airsÐnamely, (i) M(a)
and Y (a), and (ii) M(a) and @Y (a). But if we add another entity b, M(a) will now
be compatible with eight di¨erent state descriptions, as the reader may verify.
Enlarging the domain to three individuals will yield 16 state descriptions com-
patible with the assertion M(a), and so on.
16. Bar-Hillel and Carnap (1952) did not themselves explore the e¨ects of relative
size of domain of discourse on this concept of informativeness. (There appear to
be interesting theorems in this area, such as: the proportion of compatible to in-
compatible state descriptions for a given assertion stays constant whatever the size
of the domain.) This concept of informativeness thus makes the informativeness of
a statement relative to a domain of discourse a three-place relation: statement S is
more informative in domain A than in domain B. It is a concept of the informa-
tiveness of a statement relative to a world. I owe these clari®cations to Jay Atlas.
17. I am grateful to Jay Atlas for pointing out the force of the Popperian analysis
to me.
18. This argument I owe, once again, to Jay Atlas. He adds that if it is objected
that a particular addressee might prefer having a little information about two
referents to having a lot about one, then one must concede that of course that is
possible. But this is to invoke yet another notion of informativeness, namely
informativeness-for-a-person in a context, which is of course a matter entirely
relative to context. Such a notion of informativeness is of course not relevant to
the business at handЮnding a default preference for interpretation.
19. Incidentally, we shall make use of yet further arguments about the relative
informativeness of coreferential versus noncoreferential readings in section 4.3.2.
These are notions at the level of sense, rather than extension. I shall claim that a
sentence with a re¯exive or other Anaphor is more informative than the corre-
sponding sentence with a nonre¯exive or non-Anaphor, because the former entails
coreference where the latter at most permits it.
20. Sentence (ib), on a noncoreferential interpretation, requires an additional
existential quanti®er to (ia), as indicated by the associated logical forms.
(i) a. John likes himself.
bx [x ˆ j & L(x; x)Š
b. John1 likes him2 .
bxby [x ˆ j and x 0 y and L(x; y)Š
The re¯exive is thus more informative than an ordinary pronoun on a disjoint
reading.
411 Notes to Chapter 4

21. A similar line of argumentation goes back to Dowty (1980), Reinhart (1983),
Sadock (1983), and others. The only novel element here is the employment of the
speci®c pragmatic machinery developed in this book (this too pre®gured in Horn
1984: 24±25).
22. Further examples will be given immediately below. See also sections 4.3.2.4
and 4.3.3.1.
23. Some pragmatic inferences are undoubtedly stronger than othersÐthat is, they
are defeasible in fewer contexts, and alternative interpretations are intuitively less
salient, or even hard to conceive. It should be an aim of GCI theory to account for
at least some of these gradient e¨ects, but there will inevitably be many other factors
to do with functional sentence perspective, discourse structure, and so on. On the
current theory, we can account for gradient strengths only along the following lines:
(a) Some kinds of implicatures are intrinsically ``stronger'' than othersÐin our
theory Q-implicatures cancel both inconsistent I-implicatures and M-implicatures,
whereas M-implicatures only cancel inconsistent I-implicatures. That the inference
(if ungrammaticalized) from John hit him to the disjointness of the arguments is
due to a scalar Q-inference helps to explain why it might be stronger than the I-
inference to coreference of John and his in John hit his children, where coreference
is only a suggestion. The strength of Q-inferences is in turn due to their meta-
linguistic character, which unlike M-implicatures is not gradient in kind (mark-
edness of expression being upgradable in various ways).
(b) Implicatures from di¨erent sources may reinforce one anotherÐthis would
be one explanation of why John hit him is so strongly disjoint: not only is there a
Q-implicature as just noted, but on the theory sketched below that there is also an
I-inference to disjoint co-arguments, the two consistent implicatures would rein-
force each other.
(c) Finally, we may decide that some pattern like the disjoint reference in He
thought John could do it is an especially robust inference and handle it, as Reinhart
(1983) proposes, by building extra grammatical machinery that assigns to John
thought he could do it a strong variable interpretation, so that we have an addi-
tional Q-inference from the failure to use the strong structure.
24. I remind the reader that I am indicating the special Chomskyan sense of
Anaphor with a capital letter.
25. In fact, to handle preposed constituents, one needs a more complex formula-
tion: X c-commands Y i¨ either (i) the ®rst branching node a dominating X also
dominates Y, or (ii) the node a 0 immediately dominating a is of the same category
as a and a 0 also dominates Y. See Kuno 1987 for the history of these notions.
26. See, for example, work by Huang (1994) on Chinese, Cho and Hong (1988)
on Korean, Levinson (1987a) on Guugu Yimithirr, EncË (1986) on Turkish, and
Dimitriadis (1995) on Greek.
27. Ariel (1994) complains that such a scale from lexical NP to pronoun to zero is
far too crude, because there are many kinds of intermediate types. But this was
never intended as a full scale, and the account can be extended as required,
according to the NP types in the language (see remarks in chapter 2 with respect
to Gundel, Hedberg, and Zacharski 1989).
412 Notes to Chapter 4

28. Incidentally, as it has long been noted (e.g., Dowty 1980), there is no such
contrast in English between his own and his (as in He loves his/his own mother) as
one might at ®rst expect. This follows from the lexicalization constraint on Horn
scales, which requires the stronger expression (here his own) to be equally or more
lexicalized than the weak expression (his), a condition clearly not met in English.
In some languages, in contrast (like Norwegian) it is met, and then there is indeed
a Q-contrast of the predicted sort. See the discussion of example (45).
29. An obvious objection is that the mirror-image relation is not a perfect re¯ec-
tionÐthis issue of the noncomplementarity of pronouns and Anaphors is dis-
cussed at length in section 4.3.2.5. A less obvious objection is that Conditions A
and B appear to have independent existence, as shown by the fact that in certain
loci they overlap and exclude both pronoun and Anaphor. For example, both
They hate him (Condition B) and They hate himself (Condition A) seem ill-formed
when him or himself is understood as one of They (Sa®r 1992: 38, after Lasnik).
But in fact the Condition B±type case seems ®ne under the right contextual con-
ditions. Consider The Board must elect its own chairman unanimously and without
abstentionsÐThe bishop volunteered, so they elected him; or Renaldo was their best
playerÐthe whole team relied on him; or under cultural conditions where all must
drink to a toast, They toasted him.
30. Of course it does not matter at all whether one prefers, as Reinhart does, to
think about her account as absorbing Condition A, together with a fragment of
Condition B, with the rest of Conditions B and C obtained pragmatically. This is
terminology only, as long as we acknowledge that Condition B has not been
wholly reduced to pragmatics, because the NP±c-commanding±pronoun rule of
interpretation requires that the pronoun is not in the same domain as the NP.
31. Reuland and Koster (1991: 2) even suggest, counterfactually, that everyone
has given up on Condition C.
32. Larry Horn suggests to me that there has always been a tacit presumption that
the Binding Conditions or their predecessors hold only for presupposed, not
asserted or denied, coreference. But it is not clear why con®guration-based prin-
ciples should be sensitive to such semantic, or indeed, as it turns out, pragmatic
content (Sa®r 1992). An account in terms of defeasible pragmatic conditions is
attractive: asserted identity will automatically override pragmatic default infer-
ences to disjointness, and no special exceptions have to be stated.
33. More may be involved here, as Larry Horn points out to me: this kind of use
may depend on a parallelism of structure and thus a contrast in the frame X went
to X's house.
34. For example, in Guugu Yimithirr one can relatively easily break this con-
straint, as in:
(i) Nhangu-gu John-bi biiba nyulu gunday.
he-GEN ‡ ABS-EMPH John-GEN ‡ ABS father-ABS he-NOM hit
`He 1 hit John1 's father.'
See discussion in Levinson 1987b.
413 Notes to Chapter 4

35. Such usages are positively common in ®rst and second person, where mis-
communication is unlikely. Larry Horn kindly provides me with the following
textual examples:
(i) Dole is asked if he intends to stress the character issue in the campaign: ``I
don't think so,'' Dole said, ``My view is that I am going to talk about Bob
Dole.'' (ABC ``Nightline'' show, March 1996)
(ii) `` `I gave Pittsburgh every opportunity to sign Neil O'Donnell,' O'Donnell
said.'' Hartford Courant, March 1, 1996.
(iii) ``I 'm just doing what Neil O'Donnell can do,'' O'Donnell, quoted in New
York Times, November 3, 1997.
(iv) ``I just want to go to a place where Howard Johnson is going to put up some
big numbers,'' Howard Johnson in radio interview after he signed with the
Colorado Rookies following 1993 baseball season.
36. It is worth noting that Kuno (1987: 109) proposes an independent prag-
matic constraint that may help to explain the general rarity of coreference in the
pronoun±c-commanding±NP pattern. The constraint is partly a matter of linear
order but partly a matter of logophoricity, a concept reviewed in section 4.3.2.6.
37. To these examples it may be objected that the coreferential lexical NP is not
referentially dependent on the pronoun, even though it may be coreferential (thus
invoking Evans's (1980) distinction again). Thus if the Binding Conditions were
recast in terms of referential dependency rather than referential identity, these
examples, it might be claimed by, say, Higginbotham (1983), would not be coun-
terexamples. Even this though is not so clear in examples like:
(i) He is doing what John always does.
(ii) He it was who became King James Ist of England.
(iii) He who lifts the sword out of this stone is King Arthur.
For example, in (iii) the pronoun must be interpreted as a bound variable (as in
Whoever lifts . . .).
38. The T-Criterion speci®es that ``each T-role is assigned to one and only one
argument'' (Chomsky 1981: 36) or that ``every NP must be taken as the argument
of some predicate; furthermore it must be so taken once'' (van Riemsdijk and
Williams 1986: 243). On the face of it at least, the Guugu Yimithirr antipassive
construction (to be reviewed below) is a direct counterexample; it indicates that
single absolutive subject of the verb is to be taken twice, ®rst in agent role than in
patient role. And pleonastic pronouns in Guugu Yimithirr also violate this con-
dition (see Levinson 1987b: 391, n. 19).
39. In this theory, there is also a Condition C, two versions of which are consid-
ered, one of which makes use of a con®gurational notion and one of which does
not. Again, pragmatic reduction seems straightforward.
40. In English there is a phenomenon that indicates the essential ¯avor of the
pragmatic solution. Horn (1984: 24) notes that with ®rst- and second-person pro-
nouns, where reference is ®xed, a contrast is exploited between re¯exive and plain
pronoun:
414 Notes to Chapter 4

(i) Just go on home. I 'll worry about me.

In this textual example (from forthcoming work by Horn), the nonre¯exive is used
to avoid the logophoric or subjective associations of the re¯exive (cf. the introvert
I'll worry about myself with the extrovert or macho I'll worry about me; see also
Haiman 1995: 229). This indicates that the re¯exive in English does indeed carry
these logophoric associations, in addition to stipulated referential dependency, as
argued by Kuno (1987) and discussed further in this chapter.
41. I have put ``bound'' in scare quotes because technically, of course, within
Binding Theory an expression can by de®nition only be bound by a c-command-
ing expression, but I shall not always be so careful.
42. In the literature it is often stated that LDAs are invariably monomorphemic
(following Pica 1991); this is in fact not necessarily the case, as just noted for long-
range uses of Chinese ta ziji (Huang 1994). There are also the LDA usages of
English re¯exives like himself, as discussed by Zribi-Herz (1989) and Horn and
Lee (1995).
43. In nearly all cases, the LDA also has local, clause-bound uses, but the degree
to which these are marginal clearly di¨ers (see, e.g., Huang 1994: 111).
44. The complementarity of pronouns and LDAs does apparently persist system-
atically in a few languages, as in Gothic, and up to a point in Icelandic (see
Harbert 1995: 194). In general, the pattern in the European languages is for the
complementarity to evaporate once the re¯exive is outside the classical Binding
Theory domain (minimal governing category including a subject)Ðsee the tabular
presentation in Reuland and Koster 1991. For Icelandic the complementarity
persists when the LDA sig is bound within the ®rst containing ®nite clause but is
lost beyond that (ibid. 12±13).
45. The term ``logophoric'' is thus sometimes, confusingly, used to imply non-
Anaphoric.
46. The similarities and di¨erences between these systems may be tentatively
summed up thus (see Stirling 1993: 54±55, 259, Huang, in preparation):
LDAs Logophors Switch reference
Clause-bound uses ‡ ± ±
Logophoric contexts only ‡ ‡ ±
Referential overlap ± ‡ (limited) ‡
where ``referential overlap'' indicates that logophors may refer to a set of which
the antecedent is a member. Switch-reference systems allow not only this, but
the converse, and intersecting reference. It seems clear that there are family-
resemblance relations between these kinds of system, with intermediate cases. See
Stirling 1993, passim. Recent reports (e.g., Culy 1997) suggest that the African
logophoric systems in fact reveal considerably more diversity than previously
realized.
47. A tendency because, as about to be noted, grammaticalization processes may
partially ``freeze'' disjointness presumptions, and because there are isolated cases
415 Notes to Chapter 4

where disjoint-reference markers seem to exist in isolation, such as the famous

Dogrib anti-anaphor (see Harbert 1995: 214 for references).
48. DS markers remain the best systematic candidates for specialized disjoint-
reference devices. However, they have for the most part perspicuous diachronic
origins (see, e.g., Stirling 1993: 5, 144 for references).
49. There may perhaps also be cases of English LDA where there is no such
overlap of pronoun and re¯exive, especially where predicates of identity are
involved as in Frankenstein wants to produce a monster similar to himself/?him.
Sa®r (1992) presents a discussion in a framework that introduces pragmatic e¨ects
on Conditions A and B.
50. Interestingly, observes Larry Horn, this constraint is less clear with ®rst- and
second-person pronominals as in Will you trade me those pictures of yourself for
these pictures of myself, but perhaps that is explained by the fact that the protag-
onists share the same perspective in the current speech event.
51. Baker (1995), also utilizing textual data, casts some doubt on Zribi-Hertz's
(1989) logophoric analysis of English LDA. He suggests a systematic dialect dif-
ference between British and American English, with only British English system-
atically employing re¯exives as ``intensive'' NPs. An ``intensive'' NP is not only
contrastive but also refers to a central protagonist, a bit like Algonquian proxi-
mate (vs. obviative) pronouns. If he is correct about the dialects, this will certainly
help to explain the muddied intuitions. But the ``intensive'' content analysis can be
accommodated within the approach being developed here, although there are
reasons (given below) to think that the range of meanings associated with the re-
¯exive tend to be related to deictic markedness.
52. Zribi-Hertz (1989: 703) herself rejects a Gricean approach, apparently because
she sees the principles involved as belonging to a systematic realm of their own,
namely discourse grammar, from which sentence grammar is ultimately derivative
(p. 724).
53. As Zribi-Hertz (1989: 705, n. 12) points out, there has been an unjusti®ed
reluctance by workers in the Binding Condition framework to consider English
an LDA language. Consequently, such structural conditions as there are in long-
range uses of the re¯exive have been largely unexplored within this framework
(but see Kuno 1987 and review in Van Valin 1990; Parker, Riley, and Meyer
1990).
54. There is an interesting set of exceptions to this tendency in English. Horn
(1984: 24, and forthcoming work) notes that with ®rst- and second-person pro-
nouns, where reference is ®xed, nonre¯exive pronouns may be used with clause-
mate antecedents to signal the avoidance of the subjective, logophoric associations
of the re¯exive, as in (Horn's textual examples):
(i) I'm not looking at you, I 'm looking at me.
(ii) Take good care of yourself. You belong to you.
(iii) Well, I'm employed by them but I am working for me.
416 Notes to Chapter 4

These are the exceptions that prove the rule: in the clausemate case, pronouns
usually contrast with re¯exives primarily on the reference dimensionÐbut where
reference is ®xed by ®rst- or second-person reference, the referential contrast is
neutralized and the logophoric associations come to the fore. Burzio (1991) pre-
sents similar patterns in Romance languages.
55. Pollard and Sag (1992: 279) suggest that ``grammatical constraints can some-
times be relaxed by writers who exercise certain license with their language,'' bas-
ing their argument on the fact that in isolation native speakers ®nd literary
sentences like Its burden did not rest upon herself alone unacceptable. But as Zribi-
Hertz (1989) points out, this is exactly the wrong testÐisolated sentences are
analyzed as micro-discourses, and long-range Anaphors must be bound within the
discourse. See also Baker 1995 on dialect di¨erences.
56. These sections contain material published in the last part of Levinson 1991,
with modi®cations and bibliographical updates.
57. Huang is right that puzzles remain. If John hit him I-implicates disjoint refer-
ence by the DRP, why doesn't John hit the man M-implicate the complement,
given that R-expressions often M-contrast with pronouns? The observable fact is
that the DRP is strong and holds over all NPs unless the contrary is signaled by
grammar (re¯exivity, special marking of introverted predicates) or by marked or
emphatic forms (in languages without any re¯exive marking). Additionally, the
opposition between pronoun and lexical NP arises in loci where coreference is at
issue, and the pronoun is a minimal form suggesting coreference; but here disjoint
reference is at issue, and if the pronoun is disjoint, then certainly the less minimal
form will be too, unless specially marked by, say, emphatics in languages without
Anaphors.
58. The term ``emphatic'' is (as Faltz 1985: 239 remarks) regrettably vague.
Without attempting any careful analysis, we may note though that what we have
in mind are, for example, the various uses of the English ``re¯exive'' as illustrated
in:
(i) The Queen herself said so.
(ii) He cut the lawn himself.
(iii) He ¯ogged the Bishop himself.
The e¨ect in these cases may be glossed `and not anyone else', `not through the
agency of a third party', and so on. The e¨ect is thus to insist on the exact refer-
ence, to rebut a contrary suggestion or a stereotypical presumption, and so forth.
The scope of such particles can be complex, as indicated by the ambiguity of (iii)
(He himself ¯ogged the Bishop vs. The Bishop himself was ¯ogged by him). For a
most useful exploration of the semantics/pragmatics and scope of such particles,
see KoÈnig 1991. He notes (pp. 87±96) that a general phenomenon is the fact that
scalar additive particles often take the same form as the so-called emphatic
re¯exive pronouns or intensi®ers: German selbst, Dutch zelfs, Norwegian selv,
French meÃme, Irish feõÁn, and so on. Thus he relates scalar even to the emphatic
uses of himself (cf. French meÃme, German selbst cognate with English self )Ð
417 Notes to Chapter 4

both, to follow Edmondson and Plank (1978), ``associate a pragmatic scale with
propositions graduated in terms of the speaker's expectations of the involvement
of certain individuals in the relevant states, process and events.'' See also Kay
1990.
59. It is thus the weak re¯exives that (especially) participate in long-distance
binding and its associated logophoricity (see, e.g., Everaert, 1986: 212f; Hellan
1988: 87¨, 95). The strong re¯exives have not had the same intensive study as the
weak ones (Hellan 1988: 87), so perhaps the facts are not so clear. For example,
the strong re¯exives in Dutch, Norwegian, and Icelandic also have long-range uses
if they are stressed (Everaert 1986: 218f, 253 n. 4, 254 n. 12; Hellan 1986: 104), an
observation in line with the current theory. But clearly the main factor here is the
relative antiquity of the weak re¯exive (from proto-Indo-European *s(w)-, lost
from West Germanic and then reborrowed later in, for example, Dutch), which
may itself derive from an emphatic (see below). The strong re¯exive form in all
these languages seems to retain the pure emphatic functions (e.g., as a modi®er of
a subject) and to be in the process of extending its functions in the way predicted
by the argument here. For some generalizations about weak/marked versus
strong/unmarked re¯exives, see Faltz 1985 and Yang 1983. For an attempt to
predict their respective distributions, see Pica 1984, 1987; Everaert 1986: chap. 8;
among others. Rizzi (1989) treats the association of weak (morphologically sim-
ple) re¯exives and long-range binding (and conversely, compound re¯exives and
local binding) as so universally strong that it would su½ce for ``degree±0 learn-
ability''Ða child would not even need access to subordinate clauses to know what
the binding domain of a speci®c lexical Anaphor was! But Chinese, for example,
allows long re¯exivization with both simple ziji and compound ta ziji (Huang
1994: 77f ). The tendency for weak (morphologically simpler) re¯exives to have
long-distance uses is attributed by Faltz (1985: 256¨ ) to their diachronic origins as
stressed pronouns in subordinate clauses, as with the African logophoric pronouns
and perhaps proto-Indo-European *s(w)-. If the association between weak re¯ex-
ives and long-range uses does turn out to be a language-family wide generaliza-
tion, there is a direct potential explanation from our theory. The very fact that
clausemate re¯exivity is indicated by what was originally an emphatic form, which
in long-range uses may be indicating merely a contrast in perspective rather than
reference, may motivate double emphasis in the clausemate cases. In fact, there is
direct diachronic evidence for this process in the origin of many strong re¯exive
forms as double or triple emphatics. For example, Homeric Greek re¯exive he-
(from Proto-Indo-European *s(w)-) was later fused with autoÂs (a form used like
English -self to form compound re¯exives) to yield Classical heautoÂn in the third
person, this in turn becoming Modern eaftoÂs, now further modi®ed with a pos-
sessive pronoun indicating person, as in eafto tu, `self his'. Proto-Romance *met-
ipsimum is an even clearer case of triple emphatic derivation. See also Langacker
1977 on Amerindian languages.
60. Edmondson and Plank (178: 386) in fact suggest that there is a third himself
in English that indicates a reversal of semantic roles, as in Bill hit Sue and was
himself hit by Ann.
418 Notes to Chapter 4

61. For example, after Edmondson and Plank (1978: 380):

(i) John is a hippie. John's brother is a banker, happily married with three
children. He/He himself/HE is very much admired by people of his own type.
The tendency is to take the unadorned pronoun to refer locally to John's brother
(as predicted by the I-principle), but a marked pronoun (with the intensi®er him-
self or intonational/stress prominence) suggests coreference with John (by Horn's
1984 division of labor whereby a marked form M-implicates the complement of
the I-implicature that would have arisen from the unmarked form).
62. An anonymous referee, commenting on Levinson 1991, held that the existence
of languages without Anaphors is no more embarrassing to the Binding Theory
than the existence of languages without nasal consonants, although it might be
interesting to see whether such languages invariably fail to have NP movement
processes, as GB-theory might predict. Such a correlation is almost certainly
missing; for example, one can ®nd passives in most of the languages under con-
sideration. There is in fact one detailed GB-style analysis of a B-®rst language by
Chung (1989), which describes Chamorro as without overt morphological Ana-
phors but with canonical NP-traces in raising and passive constructions. Chung in
fact is driven to the reluctant conclusion that Chomsky's parallel between move-
ment and bound anaphora may have to be abandoned. See discussion in section
4.3.3.1.2. More recently, there have been moves to think of the Binding Theory as
simply failing to apply to languages that at ®rst sight seem to have good long-
distance Anaphors, like Japanese (see Mazuka and Lust 1994).
63. This could function like the optional English own in John admires his own
book (Larry Horn suggests to me), but much more likely it is just an emphatic.
Either way, it is clearly not an Anaphor.
64. The subject is composed of demonstrative a½xes ¯anking a classi®er for men
(marked `CPman'). This expression is semantically general with anaphoric uses, so
it could equally be translated as `he'.
65. The other syntactic re¯ex of the supposed Anaphors in Chamorro is a ``sub-
ject e¨ect'' whereby coreference is exceptionally ruled out where an antecedent
precedes a subject with an embedded pronominal (in general, Chamorro seems to
allow coreference wherever an antecedent precedes or c-commands a pronominal).
This is because, it is claimed, in these structures c-command is impossible, but real
Anaphors require it. It is unclear though, on this explanation, why a pronounÐ
which would look identicalÐshould not be able to be bound in this location. In
short, other factors seem involved. Chung (1989: 162±164) therefore considers an
alternative pragmatic explanation, that this disjointness is due to a pragmatic op-
position of the kind used by Reinhart (1983) or in our Q-based account of A-®rst
languages, preferring the use of the Anaphor (with a passive or active transform if
necessary). Overall, the attempt to distinguish pronominals from covert Anaphors
masquerading in the same form seems an almost-impossible task.
66. I am most grateful to David Perlmutter for putting me in touch with this work
on Creoles, not to mention useful discussions with him, conducted partly on the
jog.
419 Notes to Chapter 4

67. See also Carden 1989 on Guianese for diachronic evidence.

68. Mitchell (1985: 189) concurs with Visser's observation that:
emphatic self is found in the earliest texts in the nominative . . . its use with an accusative
re¯exive pronoun appears to come later. It is not found in Beowulf, Genesis, Exodus, Christ
and Satan, Deor or even Juliana, although it is evident in the prose of King Alfred's time.
69. Visser (1963: 439¨ ) documents the same ``ambiguity'' with the reciprocal
meaning: hi gecyston hi could mean either `they kissed them' or `they kissed each
other', this usage surviving into Middle English but being supplanted by various
means of reciprocal marking, such as by adverbs glossed as `mutually', and the
each. . .other forms already utilized in Old English.
70. Mitchell (1985: 115) gives the following kind of example of the nonre¯exive
use of the form hine selfne:
(i) Moyses, se De w0s Gode sua weorD D0t he oft wiD hine selfne spr0c
`Moses1 who was so dear to God2 that he1 often spoke with Him-self2 '
(from King Alfred's West Saxon version of Gregory's Pastoral Care 131.11)
The example is discussed further in section 4.4.
71. Thus, according to grammars of the seventeenth and eighteenth centuries, a
sentence like Hij beschuldight hem (`He accuses him/himself ') was ambiguous. See
Everaert 1986: 3¨ and references therein.
72. A re¯exive based on an emphatic, as in hem zelf, is also still current in stan-
dard Dutch, see Koster 1986: 344f, Koster 1994.
73. The evolution in the other direction (i.e., by the loss of Anaphors) is not of
course inconceivable; indeed this seems to be what exceptionally happened in the
West Germanic languages which are presumed to have lost the reconstructed
Proto-Germanic re¯exive *sik (see note 75). But the diachronic processes here
would be of a quite di¨erent kind, Nigel Vincent points out to me, perhaps owing
to phonetic change or language contact. The bidirectional possibilities should not
be confused with some natural cycle; for example, Vincent (1989b) insists, the
apparent synthesis-analysis-synthesis cycle in some Romance verbal in¯ections
disguises quite di¨erent underlying mechanisms. The historical change of B-®rst to
A-®rst systems would seem in accord with the kind of grammaticalization of
analytic into synthetic grammatical systems.
74. With introverted or inherently re¯exive predicates like `wash', `shave', `dress',
one may expect the pronoun to be ommitted (I-implicating coreference) or if
present to be understood in line with the stereotypical re¯exive action.
75. From Proto-European *s(w)-, proto-Germanic *sik, a form lost in the devel-
opment of English (Faltz 1985: 210).
76. This leaves re¯exives like Japanese zibun, a nominal head without modi®er,
somewhat stranded; Faltz (1985: 49) suggests that they must either be assimilated
to the true pronominal re¯exives or analyzed as head+zero modi®er.
77. Faltz rejects (1985: 278 n. 6) the idea that intransitivizing verbal a½xes might
specialize to become re¯exives; actually, though, this seems to be happening in the
420 Notes to Chapter 4

case of Guugu Yimithirr antipassive or middle (see Haviland 1979, Levinson

1987b), with re¯exive specialization of the intransitivized verb being currently
marked by an additional emphatic on the single NP. Of course, the Guugu Yimi-
thirr antipassive might already have passed through a re¯exive stage, with sub-
sequent semantic bleaching; but some evidence against this is that some verbs
(see Haviland 1979: 126f ) that obligatorily take antipassive in¯ection (by some
frozen, and thus presumably old, process) simply have an intransitive meaning
(e.g., `explode', `shine').
78. Nigel Vincent tells me that the Proto-Romance formation *met-ipsimum must
be a case of threefold emphasis: met was an emphatic (as in egomet `I myself '),
ipse was the standard Latin emphatic pronoun with case in¯ection, and -imum was
the superlative su½x as found in maximum. He adds that Italian stesso derives
from iste-ipse, a double pronoun combination. Langacker (1977) also discusses the
evolution of Uto-Aztecan re¯exives.
79. Faltz (1985) does not appear to have been aware of HageÁge's work and makes
no reference to the subjective point-of-view phenomena associated with long-
distance re¯exives, although he discusses some of the classical logophoric lan-
guages, like Ewe (p. 252).
80. Of course, the derivation of a long-distance pronoun from a ®rst-person pro-
noun in a language that has no system of indirect reported speech (i.e., from a
form glossing as ``John said I will come'' to a form ``John said self would come'')
would make immediate sense of the logophoric associations of such pronouns. But
Faltz (1985: 263¨ ) notes a number of problems for such a development, although
it seems eminently plausible for some of the African languages (like Ewe). How-
ever, it is quite clear that this cannot be the source of all such long-distance
re¯exives with associated logophoricity, because these are (as Faltz notes, p. 257¨ )
often (as in Igbo) clearly third-person in origin.
81. The exceptions will be those derived from ®rst-person long-range-only
``re¯exives'' and any C forms of the `body-his' kind that may have been directly
coined. For the record, Baker (1995: 94±95, n. 33) doubts this historical analysis
on semantic groundsÐhe thinks that the two functions (re¯exivity, emphasis or
discourse prominence) are ``conceptually independent'' and the diachronic relation
between them involves drastic reanalysis, but he owes us an account of why the
two conceptually independent functions so repeatedly get marked with the same
morphemes.
82. The evidence is of di¨erent kinds. The weakest is mere overlap in function
between emphatics and re¯exive morphemes. Stronger evidence, as in the case of
Papago, is where this is reinforced by comparison with related languages where
the morpheme only has a re¯exive function. Best of all, of course, as in the rest of
the cited languages, is where there are records showing increased grammaticali-
zation over time of the emphatic into a re¯exive.
83. Harbsmeier (1981: chap. 3) contains some interesting material on Classical
Chinese re¯exives. He claims that pronominal ji was a contrastive re¯exive pro-
nominal (related to, but distinct from the nominal ji referring to the Taoist `Self ';
421 Notes to Chapter 4

p. 178) with clausemate antecedents, but also a noncontrastive long-distance

re¯exive requiring coreference with the subject of the highest clause (p. 187);
whereas zi clearly had emphatic functions of many sorts, just like the adverbial
uses of modern Mandarin ziji (Li and Thompson 1981: 138).
84. For example, Syromiatnikov (1981: 81f ) notes both re¯exive and emphatic
adjunct uses for onoÈre and midukara in Ancient Japanese (cf. Modern Japanese
mizukara, an alternate to zibun; Hinds 1986: 116). These seem to be bleached head
nominals (e.g., kara means `body'), a common re¯exive source classi®ed by Faltz
(1985) as a kind of C pronoun.
85. Although rarely mentioned in the recent theoretical literature on jibun/zibun
(e.g., in neither Kuno 1973 nor Hinds 1986 nor Sells 1987), the emphatic functions
®gure prominently in traditional pedagogical grammars (e.g., Vaccari and Vaccari
1948: 358¨ list uses similar to English `He himself came' and `I bought it myself ';
they even note the double emphatic use of jibun jishin to translate `you did it
yourself ' [lit. ji-bun `self-part', ji-shin `self-body']). See also Mazuka and Lust
1994: 152.
86. It may be preferable, as previously mentioned, simply to assume that the
I-preference to coreference only operates outside the scope of the DRP.
87. Incidentally, Huang (1994: 144) wonders in fact whether his account might
not be generalized to languages like English, making an A-®rst account generally
unnecessary. But there are at least two basic di¨erences between the languages
which motivate holding on to an A-®rst account. Firstly, sentences of the form
`He1 hit him1 ' with the coreferential interpretation have a di¨erent status in the
two languages, being traditionally accepted by the grammarians as grammatical in
Chinese (Huang 1994: 130) and rejected in EnglishÐthus we need an explanation
(given by the Q-principle) of why the DRP is not so easily overruled in English.
Second, re¯exives in English (but not it seems in Chinese) have (it is argued)
grammatically speci®ed binding domains, and there are other signs of con-
ventionalization (as in the bound-anaphora patterns), which then in turn have
pragmatic implications, as modeled in the A-®rst account.
88. It is notable that in modern Chinese there is often a contrast in preferred ref-
erential interpretation of embedded pronominals that favors a local antecedent for
ta and a more distant one for ziji. For example, Huang (1989: 91) notes of the
following kind of example that the favored interpretation coidenti®es Wang
(rather than Li) and ziji:
(i) Lao Wang yiwei Lao Li bu zhidao ziji qu guo Meiguo.
Lao Wang think Lao Li not know self go EXP America
`Wang thinks that Li doesn't know that self has been to America.'
This tendency for ziji linkage to skip over a good candidate antecedent that would
have been the favored link with ta may therefore be partially explained by this
pragmatic theory (speci®cally, the M-contrast between ta and ziji). Huang (1989:
106 n. 6) points out that a similar pattern has been noted for Japanese. However,
our account cannot be the full explanation because even where there is no contrast
with ta, ziji still prefers a more distant antecedent. In a sentence with only one
422 Notes to Chapter 4

embedded clause glossing, say, as `Wang thinks Li has fallen in love with ta/ziji'
the nearest antecedent, Li, is a clausemate disfavoring identi®cation with ta, but
ziji would still naturally refer to Wang.
89. That this is only a suggestion was clear, Huang (1994: 130) points out, to
traditional grammarians of Chinese; he quotes Wang Li: ``As for ta ma ta (`He
rebukes him'), it is ambiguous: it can mean either `he rebukes himself ' or `he
rebukes another person'.''
90. The Chinese example is from Huang (1989: 96), and the source of the Old
English example is cited in note 70 (see Mitchell 1985: 115). I am grateful to April
McMahon for help in interpreting the Old English examples.
91. The logic is as follows: If an unmarked utterance U I-implicates p, and a
marked utterance M M-implicates the complement p, a doubly marked utterance
MM often implicates the complement of p, that is, p (most often with some addi-
tional restriction; see chapter 2).
92. For an attempt to construe the pronouns of a B-®rst language as systemati-
cally ambiguous between Anaphor and pronoun, see Chung 1989. More generally,
some theorists, like Koster (1986: 341¨, 1994), consider that pronouns are
ambiguously Anaphors and pronominals, while re¯exives may equally be Ana-
phors or pronominals (as in John himself came).
93. Apart from the delayed acquisition of Binding Condition B patterns found by
Solan and others mentioned above, Reinhart (1986: 140±142) notes that ®ndings
about children's use of backward anaphora favor treating Condition B patterns as
pragmatic. She argues that the data support her particular A-®rst analysis, with all
types of bound anaphora collectively distinguished from pragmatic anaphora (but
see C. Koster 1994). Her argument, like mine, is based on the presumption that
innate predispositions should show up earlier in acquisition than pragmatically
determined patterns.
94. Consider, for example, Platzack's (1987) consideration of the time scale for
the parameter switch from null subjects to realized subjects in Swedish. Platzack
argues that there are a number of syntactic re¯exes of this parameter (e.g., subject-
verb agreement, expletive pronouns vs. gaps, that-trace ®lter violations) and that
the historical record shows that ``these changes occurred together during a rela-
tively short space of timeÐmore or less within the seventeenth century'' and that
this supports ``the hypothesis that they are all re¯exes of a single parameter in the
grammar'' (p. 397f ). The statistics he produces for frequency of null subjects
might support a slightly more gradual account, in terms perhaps of a 200-year
process; and there were also special sociolinguistic factors during this period of
major state formation (as he points out). Nevertheless, there is a rather striking
contrast in timescale between this Swedish case, with relatively rapid linked
changes, and the history of English anaphora with the very slow acquisition of
Anaphors and the apparent lack of linkage between the acquisition of something
like Condition A, on the one hand, and Condition B±type patterns on the other.
Kroch (1989) takes a more skeptical view, maintaining that most linguistic change
is likely to be gradual and that language learners may happily operate with two
423 Notes to Epilogue

slightly inconsistent grammars, which makes rapid parametric change unlikely.

See also the much more gradual loss of pro-drop in Old/Middle French analyzed
by Adams (1987).
95. Chomsky (1981: 65) was perfectly aware that a pragmatic analysis might be in
order here, saying that the Principle ``. . . might be regarded as a subcase of a
conversational principle of not saying more than is required, . . . but there is some
reason to believe it functions as a principle of grammar.'' Although in recent work
Chomsky (1995) has stressed economy and lack of overdetermination within the
linguistic system itself which maps phonological form onto logical form, he now
concedes that ``we do expect to ®nd connections between the properties of the
language and the manner of its use'' (p. 168).
96. That is, we do not expect a Gricean principle to simply not apply to some
speci®c language, in the way that it has been supposed that syntactic principles
(especially outside the ``core'') might admit of marked exceptions. A claim to this
e¨ect has indeed been made by Elinor Ochs for Malagasy (then E. O. Keenan
[1976] 1998), but Brown and Levinson (1987: 9, 288±289 n. 27) point out that the
data do not substantiate this but rather point to speci®c social taboos overruling
direct speech in limited circumstances. Brown and Levinson (1987) develop a
theory in which the signaling of politeness works precisely by foregrounding
such departures against a Gricean background. In the same way, against such a
®rm background of Gricean expectations, exceptional practices will be found in
speci®c discourse genres. In these cases, Gricean principles are not even in limited
suspensionÐthe practices take their semiotic value from the departures from
Gricean expectations.

Epilogue
1. Gazdar (1979: 11±13) documented this sentiment 20 years ago, and in the in-
terim it has diminished perhaps only slightly. The complaint is often made that
pragmatics ``has found it di½cult to achieve the degree of formal precision con-
sidered desirable in the linguistic sciences today'' (Seuren 1998: 407). Note that
formalization and prediction are not necessarily related; most predictions in the
empirical branches of linguistics (e.g., psycholinguistics or language typology) are
not formalized but are clear enough. Because pragmatics is, more or less by de®-
nition, not concerned with rules or with monotonic inference, any formalization is
going to have an unfamiliar appearanceÐit is just not going to look like formal-
ization in syntax and semantics (good examples can be found in the work on AI or
natural language processing; see, e.g., Hobbs et al 1993). Large-scale formaliza-
tion is also likely to be premature and involve fundamental fudges because we
simply do not fully understand the underlying kinds of reasoning involved. This is
not to say that fragmentary formalizations (as in Gazdar 1979, Wainer 1991, or
Lascarides and Asher 1993) may not be extremely helpful. GCI theory, because of
its limited scope and because of the existence of models of default reasoning, is a
particularly promising area. Meanwhile, if the thesis of chapter 3 is roughly cor-
rect, much of what is currently done in formal semantics is the formalization of
the pragmatically enriched semantic content of utterances.
424 Notes to Epilogue

2. Admittedly, because the inferences arise under our heuristics from the under-
lying semantic representations as well as the form of expressions, precise pre-
dictions depend on the semantic analysis of the relevant expressions. Note that
because the conditions under which default inferences are defeasible are reason-
ably clear, defeasibility itself should not greatly weaken the predictions.
3. For o¨-line e¨ects of Gricean inferences, see, for example, Moxey and Sanford
1993, Newstead 1995 on scalar inference; Dulany and Hilton 1991, on scalar
inferences; Politzer and Noveck 1991, on Gricean e¨ects on the conjunction fal-
lacy. For on-line processing of anaphora, ambiguity, and so on, which may
involve Gricean e¨ects, see the reviews in Garrod and Sanford 1994 and Singer
1994. For a model of comprehension where pragmatic principles are involved
early in on-line processing and before general knowledge of the world, see Ni,
Crain, and Shankweiler 1996.
4. This is to depart from the otherwise attractive proposal of representa-
tional modularity by Jackendo¨ 1996b; see Bierwisch 1996: 41 for additional
motivations.
References

Adams, M. 1987. From Old French to the theory of pro-drop. Natural Language
& Linguistic Theory 5:1±32.
Allen, J. 1983. Recognizing intentions from natural language utterances. In M.
Brady and R. C. Berwick (eds.), Computational models of discourse (pp. 107±166).
Cambridge: Cambridge University Press.
Allen, S. 1994. Acquisition of some mechanisms of transitivity alternation in arctic
Quebec Inuktitut. Ph.D. dissertation, McGill University, Montreal.
Anderson, S., and Keenan, E. 1985. Deixis. In T. Shopen (ed.), Language typology
and syntactic description. Vol. 3, Grammatical categories and the lexicon (pp. 259±
307). Cambridge: Cambridge University Press.
Andronov, M. 1969. A standard grammar of modern and classical Tamil. Madras,
India: New Century Book House.
Ariel, M. 1994. Interpreting anaphoric expressions: A cognitive versus a prag-
matic approach. Journal of Linguistics 30:3±42.
Arono¨, M. 1974. Word formation in generative grammar. Cambridge, MA: MIT
Press.
Arundale, R. B. 1997. `Studies in the way of words': Grice's new directions in
conceptualizing meaning in conversational interaction. Mimeo, Department of
Communication, University of Alaska, Fairbanks. (Paper presented at the Inter-
national Communication Association, Chicago, May 1991.)
Asher, R. 1985. Tamil. London: Croom Helm.
Atlas, J. 1979. How linguistics matters to philosophy. In D. Dinneen and C.-K.
Oh (eds.), Syntax and semantics. Vol. 11, Presupposition (pp. 265±281). New
York: Academic Press.
Atlas, J. 1983. Comments on Horn (1985). Ms., Pomona College, Claremont, CA.
Atlas, J. 1984a. Grammatical non-speci®cation: The mistaken disjunction theory.
Linguistics and Philosophy 7:433±443.
Atlas, J. 1984b. Comparative adjectives and adverbials of degree: An introduction
to radically radical pragmatics. Linguistics and Philosophy 7:347±377.
426 References

Atlas, J. 1989. Philosophy without ambiguity. Oxford: Clarendon Press.

Atlas, J. 1993a. The implications of conversation: The 1990 Leuven lectures. Ms.,
Pomona College, Claremont, CA.
Atlas, J. 1993b. The importance of being `only': Testing the neo-Gricean versus
neo-entailment paradigms. Journal of Semantics 10:301±318.
Atlas, J. (in press). Pragmatic analysis: Logic, meaning, and conversational impli-
cature. Oxford: Oxford University Press.
Atlas, J., and Levinson, S. 1973. What is an implicature? Part 1: Kenny Logic.
MSSB mimeo.
Atlas, J. D., and Levinson, S. C. 1981. It-clefts, informativeness, and logical form:
Radical pragmatics (revised standard version). In Cole 1981: 1±61.
Austin, J. L. 1950. Truth. Proceedings of the Aristotelian Society. Supplementary
vol. 24. [Reprinted in J. L. Austin, 1970, Philosophical papers, pp. 117±133.
Oxford: Oxford University Press.]
Austin, J. L. 1962. How to do things with words. Oxford: Oxford University Press.
Austin, P. 1981. A grammar of Diyari, South Australia. Cambridge: Cambridge
University Press.
Austin, P. 1987. Cases and clauses in Jiwarli, Western Australia. Ms., La Trobe
University, Melbourne.
Avramides, A. 1989. Meaning and mind: An examination of a Gricean account of
language. Cambridge, MA: MIT Press.
Bach, K. 1975. Performatives are statements too. Philosophical Studies 28:229±
236.
Bach, K. 1982. Semantic nonspeci®city and mixed quanti®ers. Linguistics and
Philosophy 4:593±605.
Bach, K. 1984. Default reasoning: Jumping to conclusions and knowing when to
think twice. Paci®c Philosophical Quarterly 65:37±58.
Bach, K. 1987a. Thought and reference. Oxford: Oxford University Press.
Bach, K. 1987b. On communicative intentions: A reply to Recanati. Mind and
Language 2:141±154.
Bach, K. 1990. Communicative intentions, plan recognition, and pragmatics. In
Cohen, Morgan, and Pollack 1990: 389±400.
Bach, K. 1994. Conversational impliciture. Mind and Language 9:124±162.
Bach, K. 1995. Standardization and conventionalization. Linguistics and Philoso-
phy 18:677±686.
Bach, K., and Harnish, R. M. 1979. Linguistic communication and speech acts.
Cambridge, MA: MIT Press.
Bache, C., Basbùll, H., and Lindberg, C. E. (eds.) 1994. Tense, aspect, and action:
Empirical and theoretical contributions to language typology. Berlin: Mouton de
Gruyter.
427 References

Baker, C. L. 1995. Contrast, discourse prominence, and intensi®cation, with spe-

cial reference to locally free re¯exives in British English. Language 71:63±101.
Baker, G. P., and Hacker, P. 1984. Language, sense, and nonsense: A critical
investigation into modern theories of language. Oxford: Blackwell.
Bar-Hillel, Y. 1954. Indexical expressions. Mind 63:359±379.
Bar-Hillel, Y., and Carnap, R. 1952. An outline of a theory of semantic informa-
tion. MIT Technical Report 247. [Reprinted in Y. Bar-Hillel, 1964, Language and
information, pp. 221±274. Reading, MA: Addison-Wesley.]
Barton, E. 1988. Base-generated constituents: A theory of grammatical structure
and pragmatic interpretation. Pragmatics and Beyond Series, Amsterdam: John
Benjamins.
Barwise, J. 1986. Conditionals and conditional information. In E. Closs-Traugott,
A. ter Meulen, J. Reilly, and C. Ferguson (eds.), On conditionals (pp. 21±54).
Cambridge: Cambridge University Press.
Barwise, J., and Cooper, R. 1981. Generalized quanti®ers and natural language.
Linguistics and Philosophy 4:159±219.
Barwise J., and Perry, J. 1983. Situations and attitudes. Cambridge, MA: MIT
Press.
Berlin, B. 1963. Some semantic features of reduplication in Tzeltal. International
Journal of American Linguistics 29:211±218.
Bickerton, D. 1981. Roots of language. Ann Arbor, MI: Karoma.
Bierwisch, M. 1996. How much space gets into language? In Bloom, Peterson,
Nadel, and Garrett 1996: 31±76.
Blakemore, D. 1992. Understanding utterances. Oxford: Blackwell.
Blass, R. 1990. Relevance relations in discourse. Cambridge: Cambridge University
Press.
Bloom, P., Peterson, M., Nadel, L., and Garrett, M. (eds.) 1996. Language and
space. Cambridge, MA: MIT Press.
Bloom®eld, L. 1933. Language. New York: Holt, Rinehart, and Winston.
Bohnemeyer, J. 1998. Time relations in discourse: Evidence from a comparative
approach to Yukatek Maya. Ph.D. dissertation, Tilburg University, The Nether-
lands.
Bolinger, D. 1979. Pronouns in discourse. In T. GivoÂn (ed.), Syntax and seman-
tics. Vol. 12, Discourse and syntax (pp. 289±310). New York: Academic Press.
Botha, R. 1988. Form and meaning in word formation: A study of Afrikaans redu-
plication. Cambridge: Cambridge University Press.
Bouchard, D. 1983. On the content of empty categories. Dordrecht: Foris.
Bourdieu, P. 1977. Outline of a theory of practice. Cambridge: Cambridge Uni-
versity Press.
BreÂal, M. 1900. Semantics. [Trans. Mrs. H. Cust]. New York: Henry Holt.
428 References

Brown, P., and Levinson, S. C. 1978. Universals in language usage: Politeness

phenomena. In E. Goody (ed.), Questions and politeness: Strategies in social in-
teraction (pp. 56±289). Cambridge: Cambridge University Press.
Brown, P., and Levinson, S. C. 1987. Politeness. Cambridge: Cambridge Univer-
sity Press.
Burrow, T., and Emeneau, M. 1961. A Dravidian etymological dictionary. Oxford:
Clarendon Press.
Burton-Roberts, N. 1984. Modality and implicature. Linguistics and Philosophy
7:181±206.
Burzio, L. 1991. The morphological basis of anaphora. Journal of Linguistics
27:81±105.
Bybee, J., Pagliuca, W., and Perkins R. 1991. Back to the future. In Traugott and
Heine 1991: 17±58.
Calvert, D. R. 1986. Descriptive phonetics. 2d ed. Stuttgart: Thieme.
Caldwell, Bishop R. 1961 [1856]. A comparative grammar of the Dravidian or
South-Indian family of languages. Madras, India: Madras University Press.
[Originally London: Harrison.]
Cann, R. 1993. Formal semantics: An introduction. Cambridge: Cambridge Uni-
versity Press.
Cantrall, W. R. 1973. Why I would relate own, emphatic re¯exives, and intensive
pronouns, my own self. Proceedings of the regional meeting of the Chicago Lin-
guistics Society 9:57±67.
Cantrall, W. R. 1974. Viewpoint, re¯exives, and the nature of noun phrases. Janua
Linguarum, Series Practica 210. The Hague: Mouton.
Carden, G. 1989. Re¯exives constructions in Atipa and other early Guyanese
texts. In M. Fauquenoy (ed.), Special issue of Textes, Etudes, et Documents 7±8
(pp. 1±21). Paris: L'Harmattan.
Carden, G., and Stewart, W. 1986. Binding theory, bioprogram, and creolization:
Evidence from Haitian Creole. Paper presented at Linguistic Society of America
annual meeting. [Published as Carden and Stewart 1988.]
Carden, G., and Stewart, W. 1987. Mauritian Creole re¯exives: An alternative
historical scenario. Ms., University of British Columbia.
Carden, G., and Stewart, W. 1988. Binding theory, bioprogram, and creolization:
Evidence from Haitian Creole. Journal of Pidgin and Creole Languages 3:1±67.
Carden, G., and Stewart, W. 1989. Mauritian Creole re¯exives: A reply to Corne.
Journal of Pidgin and Creole Languages 4:65±101.
Carden, G., and Stewart, W. 1990. Clefts, topicalization, and coreference in Hai-
tian Creole: Implications for models of creolization based on L1 acquisition.
Paper presented at the Conference on Focus and Grammatical Relations in Creole
Languages, University of Chicago, May 1990.
Carnap, R. 1947. Meaning and necessity. Chicago: University of Chicago Press.
429 References

Carston, R. 1985. A reanalysis of some `quantity implicatures'. Ms., University

College London.
Carston, R. 1988. Implicature, explicature, and truth-theoretic semantics. In R.
Kempson (ed.), Mental representations: The interface between language and reality
(pp. 155±181). Cambridge: Cambridge University Press.
Carston, R. 1995. Quantity maxims and generalised implicature. Lingua 96:213±
244.
Carston, R. 1996. Metalinguistic negation and echoic use. Journal of Pragmatics
25:309±330.
Chafe, W., and Nichols, J. (eds.) 1986. Evidentiality: The linguistic coding of epis-
temology. Norwood, NJ: Ablex.
Changeux, P. 1994. Art and neuroscience. Leonardo 27:189±201.
Charniak, E., and McDermott, D. 1987. Introduction to arti®cial intelligence.
Reading, MA: Addison-Wesley.
Cherry, C. 1966. On human communication. Cambridge, MA: MIT Press.
Chien, Y.-C., and Wexler, K. 1990. Children's knowledge of locality conditions in
binding as evidence for the modularity of syntax and pragmatics. Language Ac-
quisition 1:225±295.
Cho, Y. Y., and Hong, K.-S. 1988. Evidence for the VP constituent from child
Korean. In E. Clark and Y. Matsumoto (eds.), Proceedings of the Annual Child
Language Forum: Papers and reports on child language development (pp. 31±38).
Stanford, CA: Stanford University.
Chomsky, N. 1956. Three models for the description of language. IRE Trans-
actions on Information Theory IT2: 113±114.
Chomsky, N. 1957. Syntactic structures. The Hague: Mouton.
Chomsky, N. 1959. On certain formal properties of grammars. Information and
Control 2:137±167.
Chomsky, N. 1965. Aspects of the theory of syntax. Cambridge, MA: MIT Press.
Chomsky, N. 1975. Re¯ections on language. New York: Pantheon.
Chomsky, N. 1981. Lectures on government and binding: The Pisa lectures. Dor-
drecht: Foris.
Chomsky, N. 1986. Knowledge of language. New York: Praeger.
Chomsky, N. 1995. The Minimalist Program. Cambridge, MA: MIT Press.
Chung, S. 1989. On the notion ``null anaphor'' in Chamorro. In O. Jaeggli and K.
Sa®r (eds.), The null subject parameter (pp. 143±184). Dordrecht: Kluwer.
Clark, E. 1993. The lexicon in acquisition. Cambridge: Cambridge University Press.
Clark, E. 1997. Conceptual perspective and lexical choice in acquisition. Cognition
64:1±37.
Clark, E., and Clark, H. 1979. Where nouns surface as verbs. Language 55:767±
811.
430 References

Clark, H. H. 1983. Making sense of nonce sense. In G. B. Flores D'Arcais and R.

Jarvella (eds.), The process of language understanding (pp. 297±332). New York:
Wiley.
Clark, H. H. 1996. Using language. Cambridge: Cambridge University Press.
Clark, H. H., and Clark, E. V. 1979. When nouns surface as verbs. Language
55:797±811.
Clark, H. H., and Haviland, J. 1977. Comprehension and the given-new contract.
In R. Freedle (ed.), Discourse production and comprehension (pp. 1±40). Hillsdale,
NJ: Lawrence Erlbaum.
Closs-Traugott, E., ter Meulen, A., Reilly, J., and Ferguson, C. (eds.) 1986. On
conditionals. Cambridge: Cambridge University Press.
Cohen, L. J. 1971. The logical particles of natural language. In Y. Bar-Hillel (ed.),
Pragmatics of natural language (pp. 50±68). Dordrecht: Reidel.
Cohen, P., Morgan, J., and Pollack, M. 1990. Intentions in communication. Cam-
bridge, MA: MIT Press.
Cole, P. (ed.) 1978. Syntax and semantics. Vol. 9, Pragmatics. New York: Aca-
demic Press.
Cole, P. (ed.) 1981. Radical pragmatics. New York: Academic Press.
Cole, P., and Morgan, J. (eds.) 1975. Syntax and semantics. Vol. 3, Speech acts.
New York: Academic Press.
Comrie, B. 1981. Language universals and linguistic typology. Oxford: Blackwell.
Comrie, B. 1984. Subject and object control: Syntax, semantics, pragmatics. Pro-
ceedings of the annual meeting of the Berkeley Linguistic Society 10:450±464.
Comrie, B. 1985. Tense. Cambridge: Cambridge University Press.
Corne, C. 1988. Mauritian Creole re¯exives. Journal of Pidgin and Creole Lin-
guistics 3:69±94.
Cruse, D. A. 1986. Lexical semantics. Cambridge: Cambridge University Press.
Culy, C. 1997. Logophoric pronouns and point of view. Linguistics 35:845±860.
Davis, S. (ed.) 1991. Pragmatics: A reader. Oxford: Oxford University Press.
Dimitriadis, A. 1995. When pro-drop languages don't: On overt pronominal sub-
jects in Greek. Penn Working Papers in Linguistics, 2:45±60.
Dixon, R. M. W. 1980. The languages of Australia. Cambridge: Cambridge Uni-
versity Press.
Dixon, R. M. W. 1981. Wargamay. In R. M. W. Dixon and B. Blake (eds.),
Handbook of Australian languages, Vol. 2 (pp. 1±143.). Canberra, Australia:
ANU Press.
Dixon, R. M. W. 1983. Nyawaygi. In R. M. W. Dixon and B. Blake (eds.), Handbook
of Australian languages, Vol. 3 (pp. 430±525). Canberra, Australia: ANU Press.
Dixon, R. M. W. 1988. A grammar of Boumaa Fijian. Chicago: University of
Chicago Press.
431 References

Dixon, R. M. W., and Blake, B. (eds.) 1983. Handbook of Australian languages.

Vol. 3. Canberra, Australia: ANU Press.
Donnellan, K. S. 1966. Reference and de®nite description. The Philosophical Re-
view 75:281±304.
Donnellan, K. S. 1978. Speaker references, description, and anaphora. In Cole
1978: 47±68.
Dowty, D. 1980. Comments on the paper by Bach and Partee. In J. Kreiman and
A. Ojeda (eds.), Proceedings of the Chicago Linguistics Society: Parasession on
pronouns and anaphora 16:29±40.
Dulany, D., and Hilton, D. 1991. Conversational implicature, conscious repre-
sentation, and the conjunction fallacy. Social Cognition 9:85±110.
Eades, D. 1979. Gumbaynggir. In Dixon and Blake 1983: 244±361.
Eco, U. 1997. The search for the perfect language. London: Fontana.
Edmondson, J. A., and Plank, F. 1978. Great expectations: An intensive self
analysis. Linguistics and Philosophy 2:373±413.
EncË, M. 1981. Tense without scope: An analysis of nouns as indexicals. Ph.D. dis-
sertation, University of Wisconsin, Madison.
EncË, M. 1986. Topic switching and pronominal subjects in Turkish. In D. Slobin
and K. Zimmer (eds.), Typological studies in Turkish linguistics (pp. 195±208).
Amsterdam: John Benjamins.
Evans, G. 1980. Pronouns. Linguistic Inquiry 11:337±362.
Evans, N. 1995. A grammar of Kayardild. Berlin: Mouton de Gruyter.
Everaert, M. 1986. The syntax of re¯exivization. Dordrecht: Foris.
Faltz, L. M. 1977. Re¯exivization: A study in universal syntax. Ph.D. dissertation,
University of California, Berkeley.
Faltz, L. M. 1985. Re¯exivization: A study in universal syntax. [Revised version of
Faltz 1977.] New York: Garland.
Faltz, L. M. 1989. A role for inference in meaning change. Studies in Language,
13:317±31.
Farmer, A., and Harnish, M. 1987. Communicative reference with pronouns. In
Verschueren and Bertuccelli-Papi 1987: 547±565.
Fauconnier, G. 1975. Pragmatic scales and logical structure. Linguistic Inquiry
6:353±375.
Fauconnier, G. 1979. Implication reversal in natural language. In F. Guenthner
and S. Schmidt (eds.), Formal semantics for natural language (pp. 289±301). Dor-
drecht: Reidel.
Fauconnier, G. 1985. Mental spaces: Aspects of meaning construction in natural
language. Cambridge, MA: MIT Press.
Fillmore, C. 1975. Santa Cruz lectures on deixis. Mimeo, Indiana University Lin-
guistics Club, Bloomington, IN. [Reprinted as Fillmore, C. 1997, Lectures on
deixis. Stanford, CA: CSLI Publications.]
432 References

Fodor, J. 1983. The modularity of mind. Cambridge, MA: MIT Press.

Foley, W., and Van Valin, R. D. 1983. Functional syntax and universal grammar.
Cambridge: Cambridge University Press.
Fogelin, R. 1967. Evidence and meaning. Atlantic Highlands, NJ: Humanities
Press.
Ford, C. E., and Thompson, S. A. 1986. Conditionals in discourse: A text-based
study from English. In Closs-Traugott, ter Meulen, Reilly, and Ferguson 1986:
353±372.
Fox, B. 1987. Discourse structure and anaphora. Cambridge: Cambridge Univer-
sity Press.
Fox, B. (ed.) 1996. Studies in anaphora. Amsterdam: John Benjamins.
Fox Tree, J., and Clark, H. 1997. Pronouncing ``the'' as ``thee'' to signal problems
in speaking. Cognition 62:151±167.
Frederking, R. 1996. Grice's maxims: ``Do the right thing.'' Paper delivered at the
AAAI Symposium on Computational Implicature. Stanford, CA, March 1996.
Frege, G. [1884] 1953. The foundations of arithmetic. New York: Harper and
Row.
Fretheim, T. 1992. The e¨ect of intonation on a type of scalar implicature. Journal
of Pragmatics 18:1±30.
Garnham, A. 1989. Integrating information in text comprehension: The interpre-
tation of anaphoric noun phrases. In G. Carlson and M. Tanenhaus (eds.), Lin-
guistic structure in language processing (pp. 359±400). Dordrecht: Kluwer.
Garrod, S., and Sanford, A. 1994. Resolving sentences in a discourse context. In
Gernsbacher 1994: 675±698.
Gazdar, G. 1979. Pragmatics: Implicature, presupposition, and logical form. New
York: Academic Press.
Gazdar, G. 1980. Pragmatic constraints on linguistic production. In B. Butter-
worth (ed.), Language production, Vol. 1 (pp. 49±68). New York: Academic Press.
Gazdar, G., and Pullum, G. 1976. Truth-functional connectives in natural lan-
guage. Proceedings of the Chicago Linguistic Society 12:220±234.
Geach, P. T. 1972. Logic matters. Oxford: Blackwell.
Geis, M., and Zwicky, A. 1971. On invited inferences. Linguistic Inquiry 2:561±565.
GeniusÏieneÇ, E. 1987. The typology of re¯exives. Berlin: Mouton de Gruyter.
Gernsbacher, M. (ed.) 1994. Handbook of psycholinguistics. New York: Academic
Press.
Geurts, B. 1998. The mechanisms of denial. Language 74:274±307.
Ginsberg, M. (ed.) 1987. Readings in non-monotonic reasoning. Los Altos, CA:
Morgan Kaufman.
Gordon, D., and Lako¨, G. 1971. Conversational postulates. Proceedings of the
Chicago Linguistic Society 7:63±84.
433 References

Grandy, R., and Warner, R. 1986. Philosophical grounds of rationality. Oxford:

Clarendon Press.
Green, G. 1987. Some remarks on why there is implicature. Studies in the Lin-
guistic Sciences, 17:77±92.
Green, N., and Carberry, S. 1993. A discourse-plan-based approach to a class of
particularized conversational implicature. ESCOL'93: 117±128.
Greenberg, J. 1966a. Universals of language. 2d ed. Cambridge, MA: MIT Press.
Greenberg, J. 1966b. Language universals, with special reference to feature hier-
archies. The Hague: Mouton.
Grice, H. P. 1957. Meaning. Philosophical Review, 67:377±388.
Grice, H. P. 1961. The causal theory of perception. The Aristotelian society: Pro-
ceedings, supplementary. Vol. 35 (pp. 121±153). [Reprinted in S. Davis (ed.) 1983,
Causal theories of mind (pp. 245±70).] Berlin: Mouton de Gruyter.
Grice, H. P. 1967. Logic and conversation. William James Lectures. Ms., Harvard
University.
Grice, H. P. 1969. Utterer's meaning and intentions. Philisophical Review 78:147±
177. [Reprinted in Grice 1989.]
Grice, H. P. 1975. Logic and conversation. In Cole and Morgan 1975: 41±58.
Grice, H. P. 1978. Further notes on logic and conversation. In Cole 1978: 113±
128.
Grice, H. P. 1981. Presupposition and conversational implicatures. In Cole 1981:
183±198.
Grice, H. P. 1986. Reply to Richards. In Grandy and Warner 1986: 45±106.
Grice, P. 1989. Studies in the way of words. Cambridge, MA: Harvard University
Press.
Grodzinsky, Y., and Reinhart, T. 1993. The innateness of binding and corefer-
ence. Linguistic Inquiry 24:69±101.
Gundel, J., Hedberg, N., and Zacharski, R. 1989. Givenness, implicature, and
demonstrative reference in English. Proceedings of the Chicago Linguistic Society.
Part 2: Parasession on language in context 25:89±103.
Gundel, J., Hedberg, N., and Zacharski, R. 1993. Cognitive status and the form
of referring expressions in discourse. Language 69:274±307.
HageÁge, C. 1974. Les pronoms logophoriques. Bulletin de la SocieÂte de Linguisti-
que de Paris 69:287±310.
Haiman, J. 1985a. Natural syntax. Cambridge: Cambridge University Press.
Haiman, J. (ed.) 1985b. Iconicity in syntax. Amsterdam: John Benjamins.
Haiman, J. 1985c. Symmetry. In Haiman 1985b: 73±95.
Haiman, J. 1995. Grammatical signs of the divided self. In W. Abraham, T.
GivoÂn, and S. Thompson (eds.), Discourse grammar and typology (pp. 213±234).
Amsterdam: John Benjamins.
434 References

Haiman, J., and Munro, P. (eds.) 1983. Switch reference and Universal Grammar.
Amsterdam: John Benjamins.
Hale, K. 1983. Warlpiri and the grammar of non-con®gurational languages. Nat-
ural Language & Linguistic Theory 1:5±47.
Hamblin, C. L. 1971. Mathematical models of dialogue. Theoria 37:130±155.
Harbert, W. 1995. Binding theory, control, and pro. In G. Webelhuth (ed.), Gov-
ernment and binding theory and the Minimalist Program (pp. 177±240). Oxford:
Blackwell.
Harbsmeier, C. 1981. Aspects of classical Chinese syntax. London: Curzon Press.
Harder, P. 1994. Verbal time reference in English: Structure and functions. In
Bache, Basbùll, and Lindberg 1994: 61±80.
Harman, G. 1974. Review of Stephen Schi¨er's Meaning. Journal of Philosophy
71:224±229.
Harman, G. 1986. Change in view: Principles of reasoning. Cambridge, MA: MIT
Press.
Harnish, R. 1976. Logical form and implicature. In T. Bever, J. J. Katz, and
T. Langendoen (eds.), An integrated theory of linguistic ability (pp. 313±392). New
York: Crowell.
Harnish, R. 1991. Logical form and implicature. In S. Davis (ed.), Pragmatics
(pp. 316±364). Oxford: Oxford University Press.
Harris, R. A. 1993. The linguistics wars. Oxford: Oxford University Press.
Haviland, J. 1979. Guugu Yimidhirr. In R. M. W. Dixon and B. Blake (eds.),
Handbook of Australian languages. Vol. 1 (pp. 27±182). Canberra, Australia:
ANU Press.
Hawkins, J. 1978. De®niteness and inde®niteness: A study in reference and gram-
maticality prediction. London: Croom Helm.
Hawkins, J. 1991. On (in)de®nite articles: Implicatures and (un)grammaticality
prediction. Journal of Linguistics 27:405±442.
Heim, I. 1982. The semantics of de®nite and inde®nite descriptions. Ph.D. disser-
tation, University of Massachusetts, Amherst.
Heim, I. 1983a. File change semantics and the familiarity theory of de®niteness. In
R. BaÈuerle, U. Egli, and A. von Stechow (eds.), Semantics from di¨erent points of
view (pp. 164±189). Berlin: Springer.
Heim, I. 1983b. On the projection problem for presuppositions. West Coast
Conference on Formal Linguistics 2:114±125. [Reprinted in Davis 1991: 397±
405.]
HeinaÈmaÈki, O. 1994. Aspect as boundedness in Finnish. In Bache, Basbùll, and
Lindberg 1994: 207±233.
Heine, B., Claudi, U., and HuÈnnemeyer, F. 1991. Grammaticalization: A concep-
tual framework. Chicago: Chicago University Press.
435 References

Hellan, L. 1986. On anaphora and predication in Norwegian. In L. Hellan and

K. K. Christensen (eds.), Topics in Scandinavian syntax (pp. 103±124). Dordrecht:
Reidel.
Hellan, L. 1988. Anaphora in Norwegian and the theory of grammar. Dordrecht:
Foris.
Hellan, L. 1991. Containment and connectedness anaphors. In Koster and Reu-
land 1991: 27±48.
Herskovits, A. 1986. Language and spatial cognition: An interdisciplinary study of
the prepositions in English. Cambridge: Cambridge University Press.
Higginbotham, J. 1983. Logical form, binding, and nominals. Linguistic Inquiry
14:395±420.
Hinds, J. 1986. Japanese. London: Croom Helm.
Hintikka, J. 1962. Knowledge and belief. Ithaca, NY: Cornell University Press.
Hintikka, J. 1970. On attributions of `self-knowledge'. The Journal of Philosophy
67:73±87.
Hirschberg, J. 1985. A theory of scalar implicature. Moore School of Electrical
Engineering, University of Pennsylvania, Technical Report MS-CIS±85±56.
[Reprinted as Hirschberg, J. 1991. A theory of scalar implicature. New York:
Garland.]
Hirst, G. 1987. Semantic interpretation and the resolution of ambiguity. Cam-
bridge: Cambridge University Press.
Hobbs, J. 1979. Coherence and coreference. Cognitive Science 3:67±90.
Hobbs, J. 1982. Towards understanding of coherence in discourse. In W. Lehnert
and M. H. Ringle (eds.), Strategies of natural language processing (pp. 223±244).
Hillsdale, NJ: Lawrence Erlbaum.
Hobbs, J. 1987. Implicature and de®nite reference. Stanford, CA: CSLI Report
87±99.
Hobbs, J. 1996. Monotone decreasing quanti®ers in a scope-free logical form. In
Van Deemter and Peters 1996: 55±76.
Hobbs, J., and Martin, P. 1987. Local pragmatics. Technical Memo, SRI Inter-
national, Menlo Park, CA.
Hobbs, J., Croft, W., Davies, T., Edwards, D., and Laws, K. 1987. Commonsense
metaphysics and lexical semantics. Computational Linguistics 13:241±250.
Hobbs, J., Stickel, M., Appelt, D., and Martin, P. 1990. Interpretation as abduc-
tion. Technical Note 499, SRI International, Menlo Park, CA.
Hobbs, J., Stickel, M., Appelt, D., and Martin, P. 1993. Interpretation as abduc-
tion. Arti®cial Intelligence 63:69±142.
Holdcroft, D. 1987. Conversational relevance. In Verschueren and Bertuccelli-
Papi 1987: 477±496.
Holland, J. J., Holyok, K. J., Nisbett, R., and Thagard, P. R. 1989. Induction:
Processes of inference, learning, and discovery. Cambridge, MA: MIT Press.
436 References

Hopper, P., and Traugott, E. 1993. Grammaticalization. Cambridge: Cambridge

University Press.
Horn, L. R. 1972. On the semantic properties of logical operators in English.
Mimeo, Indiana University Linguistics Club, Bloomington, IN.
Horn, L. R. 1973. Greek Grice: A brief survey of proto-conversational rules in the
history of logic. Proceedings of the Chicago Linguistic Society 9:205±214.
Horn, L. R. 1978. Remarks on Neg-Raising. In Cole 1978: 129±220.
Horn, L. R. 1984. Toward a new taxonomy for pragmatic inference: Q- and R-
based implicature. In D. Shi¨rin (ed.), Meaning, form, and use in context (pp. 11±
42). Washington, DC: Georgetown University Press.
Horn, L. R. 1985. Metalinguistic negation and pragmatic ambiguity. Language
61:121±174.
Horn, L. R. 1989. A natural history of negation. Chicago: University of Chicago
Press.
Horn, L. R. 1990. Hamburgers and the truth: Why Gricean inference is Gricean.
Proceedings of the Berkeley Linguistic Society 16:454±471.
Horn, L. R. 1991a. Duplex negatio a½rmat. . . : The economy of double negation.
Proceedings of the Chicago Linguistic Society: Parasession on Negation 27:80±106.
Horn, L. R. 1991b. Given as new: When redundant a½rmation isn't. Journal of
Pragmatics 15:305±328.
Horn, L. R. 1992a. Pragmatics, implicature, and presupposition. In W. Bright
(ed.), International Encyclopaedia of Linguistics, Vol. 1 (pp. 260±266). Oxford:
Oxford University Press.
Horn, L. R. 1992b. The said and the unsaid. SALT II: Proceedings of the second
Conference on Semantics and Linguistic Theory (pp. 163±192). Columbus, OH:
Ohio State University Linguistics Department.
Horn, L. R. 1993. Economy and redundancy in a dualistic model of natural lan-
guage. SKY 1993: 1993 Yearbook of the Linguistic Asssociation of Finland, pp. 33±
72.
Horn, L. R. 1996b. Presupposition and implicature. In Lappin 1996: 299±320.
Horn, L. R. 1997. All John's children are as bald as the King of France: Existen-
tial import and the geometry of opposition. Papers from the 33rd Meeting of the
Chicago Linguistic Society 33:155±179.
Horn, L. R., and Bayer, S. 1984. Short-circuited implicature: A negative contri-
bution. Linguistics and Philosophy 7:397±414. [Reprinted in A. Kasher (ed.), 1998,
Pragmatics, Vol. 4 (pp. 658±675). London: Routledge.]
Horn, L. R., and Lee, Y.-S. 1995. Progovac on polarity. Journal of Linguistics
31:401±424.
Hornstein, N. 1995. Logical form: From GB to Minimalism. Oxford: Blackwell.
Huang, Y. 1989. Anaphora in Chinese: Towards a pragmatic analysis. Ph.D. dis-
sertation, University of Cambridge.
437 References

Huang, Y. 1991. A neo-Gricean pragmatic theory of anaphora. Journal of Lin-

guistics 27:301±335.
Huang, Y. 1994. The syntax and pragmatics of anaphora: A study with special
reference to Chinese. Cambridge: Cambridge University Press.
Huang, Y. 2000. Anaphora: A cross-linguistic study (Oxford Studies in Typology
and Linguistic Theory). Oxford: Oxford University Press.
Jackendo¨, R. 1972. Semantic interpretation in generative grammar. Cambridge,
MA: MIT Press.
Jackendo¨, R. 1983. Semantics and cognition. Cambridge, MA: MIT Press.
Jackendo¨, R. 1990. Semantic structures. Cambridge, MA: MIT Press.
Jackendo¨, R. 1992a. Languages of the mind. Cambridge, MA: MIT Press.
Jackendo¨, R. 1992b. Mme. Tussaud meets the Binding Theory. Natural Lan-
guage & Linguistic Theory 10:1±31.
Jackendo¨, R. 1996a. Semantics and cognition. In Lappin 1996: 539±560.
Jackendo¨, R. 1996b. The architecture of the linguistic-spatial interface. In
Bloom, Peterson, Nadel, and Garrett 1996: 1±30.
Jackendo¨, R, 1997. The architecture of the language faculty. Cambridge, MA:
MIT Press.
Jackson, S. 1981. Conversational implicature in children's comprehension of ref-
erence. Communication Monographs 48:237±249.
Jackson, S., and Jacobs, S. 1982. Ambiguity and implicature in children's dis-
course interpretation. Journal of Child Language 9:209±216.
Jakobson, R. 1939. Signe zeÂro. In MeÂlanges de linguistique, o¨erts a Charles Bally
(pp. 143±152). Geneva: Georg.
Jakubowicz, C. 1984. On markedness and binding principles. Proceedings of the
North Eastern Linguistics Society 14:154±182.
Jayaseelan, K. A., 1988. Emphatic re¯exive X-self. Working Papers in Linguistics
5:1±20. Hyderabad, India: Central Institute of English and Foreign Languages.
Jespersen, O. 1917. Negation in English and other languages. Copenhagen: Hùst.
Jespersen, O. 1927. A Modern English grammar on historical principles, Part III.
Vol. 2, Syntax. Heidelberg: Carl Winters UniversitaÈtsbuchhandlung.
Johnson-Laird, P. 1983. Mental models: Towards a cognitive science of language,
inference, and consciousness. Cambridge: Cambridge University Press.
Josephson, J., and Josephson, S. 1994. Abductive inference. Cambridge: Cam-
bridge University Press.
Kadmon, N. 1984. Inde®nite noun phrases with cardinality indication. Ms., Uni-
versity of Massachusetts, Amherst.
Kadmon, N. 1987. On unique and non-unique reference and asymmatric quanti®-
cation. Ph.D. dissertation, University of Massachusetts, Amherst.
438 References

Kamp, H. 1981. A theory of truth and semantic representation. In J. Groenendijk,

T. Janssen, and M. Stokhof (eds.), Formal methods in the study of language, Vol. 1
(pp. 277±321). Amsterdam: Mathematisch Centrum.
Kamp, H., and Reyle, U. 1993. From discourse to logic. Dordrecht: Kluwer.
Kant, E. 1933 [1781]. Critique of pure reason. (Trans. N. K. Smith). London:
Macmillan. [Translation of 1781, Kritik der reinen Vernunft. Riga: Hartknoch.]
Karttunen, L. 1976. Discourse referents. In J. D. McCawley (ed.), Notes from the
linguistic underground (pp. 363±386). New York: Academic Press.
Kasher, A. (ed.) 1998. Pragmatics: Critical concepts. Vol. 4, Presupposition,
implicature, and indirect speech acts. London: Routledge.
Katz, J. J. 1987. Common sense in semantics. In Lepore 1987: 157±234.
Kaufman, D. 1994. Grammatical or pragmatic: Will the real Principle B please
stand? In B. Lust, G. Hermon, and J. Korn®lt (eds.), Syntactic theory and ®rst
language acquisition: Cross-linguistic perspectives. Vol. 2, Binding, dependencies,
and learnability (pp. 177±200). Hillsdale, NJ: Lawrence Erlbaum.
Kay, P. 1990. Even. Linguistics and Philosophy 15:59±112.
Kay, P. 1992. The inheritance of presuppositions. Linguistics and Philosophy
15:333±379.
Kay, P., and Zimmer, K. 1976. On the semantics of compounds and genitives in
English. Proceedings of the California Linguistic Association Conference 6:29±35.
Keenan, E. O. 1998 [1976]. The universality of conversational postulates. In
Kasher 1998: 215±229 [Reprinted from 1976, Language in Society 5:67±80.]
Kempson, R. 1975. Presupposition and the delimitation of semantics. Cambridge:
Cambridge University Press.
Kempson, R. 1980. Ambiguity and word meaning. In S. Greenbaum, G. Leech,
and J. Svartvik (eds.), Studies in English linguistics (pp. 7±16). London: Longman.
Kempson, R. 1986. Ambiguity and the semantics-pragmatics distinction. In C.
Travis (ed.), Meaning and interpretation (pp. 77±104). Oxford: Blackwell.
Kempson, R. 1996. Semantics, pragmatics, and deduction. In Lappin 1996: 561±
598.
Kempson, R., and Cormack, A. 1981. Ambiguity and quanti®cation. Linguistics
and Philosophy 4:259±309.
Kempson, R., and Cormack, A. 1982. Quanti®cation and pragmatics. Linguistics
and Philosophy 4:607±618.
Kenny, A. J. 1966. Practical inference. Analysis 26:65±75.
Kibrik, A. 1996. Anaphora in Russian narrative prose: A cognitive calculative
account. In B. Fox (ed.), Studies in anaphora (pp. 255±304). Amsterdam: John
Benjamins.
Kiparsky, P. 1982. Lexical morphology and phonology. In P. Kiparsky (ed.),
Linguistics in the morning calm (pp. 3±91). Seoul: Hanshin.
439 References

Kiss, K. EÂ. 1991. The primacy condition of anaphora and pronominal variable
binding. In Koster and Reuland 1991: 245±262.
Klein, E. 1980. A semantics for positive and comparative adjectives. Linguistics
and Philosophy 4:1±45.
KoÈnig, E. 1991. The meaning of focus particles: A comparative perspective. Lon-
don: Croom Helm.
KoÈnig, E., and Traugott, E. 1988. Pragmatic strengthening and semantic change:
The conventionalizing of conversational implicature. In W. HuÈllen and R.
Schultze (eds.), Understanding the lexicon (pp. 110±124). TuÈbingen: Niemeyer.
Koster, C. 1994. Problems with pronoun acquisition. In Lust, Hermon, and
Korn®lt 1994: 201±226.
Koster, J. 1986. Domains and dynasties: The radical autonomy of syntax. Dor-
drecht: Foris.
Koster, J. 1994. Towards a new theory of anaphoric binding. In Lust, Hermon,
and Korn®lt 1994: 41±70.
Koster, J., and May, R. 1982. On the constituency of in®nitives. Language
58:116±143.
Koster, J., and Reuland, E. (eds.) 1991. Long-distance anaphora. Cambridge:
Cambridge University Press.
Kripke, S. 1977. Identity and necessity. In S. P. Schwartz (ed.), Naming, necessity,
and natural kinds (pp. 66±101). Ithaca, NY: Cornell University Press.
Kroch, A. 1989. Language learning and language change: Comment on D.
Lightfoot, The child's trigger experience: Degree±0 learnability. Behavioral and
Brain Sciences 12:348±349.
Kuno, S. 1973. The structure of the Japanese language. Cambridge, MA: MIT
Press.
Kuno, S. 1987. Functional syntax: Anaphora, discourse, and empathy. Chicago:
Chicago University Press.
Ladd, R. 1980. The structure of intonational meaning. Bloomington, IN: Indiana
University Press.
Lafon, J.-C. 1968. Auditory basis of phonetics. In B. Malmberg (ed.), Manual of
phonetics (pp. 76±104). Amsterdam: North-Holland.
Lako¨, G. 1971. Presupposition and relative well-formedness. In D. Steinberg and
L. Jakobovits (eds.), Semantics: An interdisciplinary reader in philosophy, linguis-
tics, and psychology (pp. 329±340). Cambridge: Cambridge University Press.
Lako¨, G. 1972. Hedges: A study in meaning criteria and the logic of fuzzy con-
cepts. In Proceedings of the Chicago Linguistic Society 8:236±287.
Lako¨, G. 1973. Some thoughts on transderivational constraints. In B. B. Kachru
et al. (eds.), Issues in linguistics (pp. 442±452). Urbana: University of Illinois Press.
Lako¨, G. 1974. Syntactic amalgams. In M. W. Galy, R. A. Fox, and A. Bruck
(eds.), Proceedings of the Chicago Linguistic Society 10:321±344.
440 References

Lako¨, G. 1987. Women, ®re, and dangerous things. Chicago: University of

Chicago Press.
Landman, F. 1986. Con¯icting presuppositions and modal subordination. In
A. M. Farley, P. T. Farley, and K. E. McCullough (eds.), Proceedings of the
Chicago Linguistic Society: Parasession on pragmatics and grammatical theory
22:195±207.
Lang, E. 1984. The semantics of coordination. Amsterdam: John Benjamins.
Langacker, R. 1977. Syntactic reanalyis. In C. Li (ed.), Mechanisms of syntactic
change (pp. 57±140). Austin, TX: University of Texas Press.
Langacker, R. 1987. Foundations of cognitive grammar, Vol. 1. Stanford, CA:
Stanford University Press.
Langacker, R., and Munro, P. 1975. Passives and their meaning. Language
51:789±830.
Lappin, S. (ed.) 1996. Handbook of contemporary semantic theory. Oxford:
Blackwell.
Lascarides, A., and Asher, N. 1993. Temporal interpretation, discourse relations,
and commonsense entailment. Linguistics and Philosophy 16:437±493.
Lasnik, H. 1991. On the necessity of Binding Conditions. In R. Freidin (ed.), Princi-
ples and parameters in comparative grammar (pp. 7±28). Cambridge, MA: MIT Press.
Laver, J. 1994. Principles of phonetics. Cambridge: Cambridge University Press.
Leech, G. 1983. Principles of pragmatics. London: Longman.
Lehmann, C. 1988. Towards a typology of clause linkage. In J. Haiman and
S. Thompson (eds.), Clause combining in grammar and discourse (pp. 181±226).
Amsterdam: John Benjamins.
Lepore, E. (ed.) 1987. New directions in semantics. New York: Academic Press.
Levelt, W. J. M. 1989. Speaking: From intention to articulation. Cambridge, MA:
MIT Press.
Levesque, H. 1990. All I know: A study in autoepistemic logic. Arti®cial Intelli-
gence 42:263±310.
Levin, N. 1987. A pragmatic concomitant of gapping. ESCOL' 87:176±186.
Levin, N., and Prince, E. 1986. Gapping and causal implicature. Papers in Lin-
guistics 19:351±364.
Levine, J. E. 1990. Pragmatic implicatures and case: The Russian dative revisited.
Russian Language Journal 44 (147±149):1±27.
Levinson, S. C. 1979. Pragmatics and social deixis. Proceedings of the Annual
Meeting of the Berkeley Linguistic Society 5:206±223.
Levinson, S. C. 1983. Pragmatics. Cambridge: Cambridge University Press.
Levinson, S. C. 1987a. Minimization and conversational inference. In Ver-
schueren and Bertuccelli-Papi 1987: 61±129. [Reprinted in A. Kasher (ed.) 1998,
Pragmatics: Critical concepts, Vol. 4, pp. 545±612. London: Routledge.]
441 References

Levinson, S. C. 1987b. Pragmatics and the grammar of anaphora. Journal of

Linguistics 23:379±434.
Levinson, S. C. 1987c. Explicature explicated. Behavioral and Brain Sciences
10:722±723.
Levinson, S. C. 1988a. Putting linguistics on a proper footing: Explorations in
Go¨man's concepts of participation. In P. Drew and A. Wootton (eds.), Erving
Go¨man: Exploring the interaction order (pp. 161±227). Cambridge: Cambridge
University Press.
Levinson, S. C. 1988b. Generalized conversational implicature and the semantics/
pragmatics interface. Mimeo, Linguistics Department, Stanford University, Stan-
ford, CA.
Levinson, S. C. 1989. Relevance. Journal of Linguistics 21:455±472.
Levinson, S. C. 1991. Pragmatic reduction of the binding conditions revisited.
Journal of Linguistics 27:107±161.
Levinson, S. C. 1995a. Three levels of meaning. In F. Palmer (ed.), Grammar and
meaning (pp. 90±115). Cambridge: Cambridge University Press.
Levinson, S. C. 1995b. Interactional biases in human thinking. In E. Goody (ed.),
Social intelligence and interaction (pp. 221±260). Cambridge: Cambridge Univer-
sity Press.
Levinson, S. C. 1996. Frames of reference and Molyneux's question: Cross-
linguistic evidence. In Bloom, Peterson, Nadel, and Garrett 1996: 109±170.
Levinson, S. C. 1997. From outer to inner space: Linguistic categories and non-
linguistic thinking. In E. Pederson and J. Nuyts (eds.), With language in mind: The
relationship between linguistic and conceptual representation (pp. 13±45). Cam-
bridge: Cambridge University Press.
Lewis, D. 1969. Convention: A philosophical study. Cambridge, MA: Harvard
University Press.
Lewis, D. 1972. General semantics. In D. Davidson and G. Harman (eds.),
Semantics of natural language (pp. 169±218). Dordrecht: Reidel.
Lewis, D. 1979. Score-keeping in a language game. In R. Bauerle, U. Egli, and
A. von Stechow (eds.), Semantics from di¨erent points of view (pp. 172±187).
Berlin: Springer-Verlag.
Li, C., and Thompson, S. 1981. Mandarin Chinese: A functional reference gram-
mar. Berkeley and Los Angeles: University of California Press.
Lieberman, P., and Blumstein, S. 1988. Speech physiology, speech perception, and
acoustic phonetics. Cambridge: Cambridge University Press.
Lightfoot, D. 1989. The child's trigger experience: Degree±0 learnability. Behav-
ioral and Brain Sciences 12:321±375.
Linebarger, M. 1991. Negative polarity items and linguistic evidence. Proceedings
of the Chicago Linguistic Society: Parasession on Negation 27:165±188.
LoÈbner, S. 1987. Quanti®cation as a major module of natural language semantics.
In J. Groenendijk, D. de Jongh, and M. Stokhof (eds.), Studies in discourse rep-
442 References

resentation theory and the theory of generalized quanti®ers (pp. 53±85). Dordrecht:
Reidel.
Lucy, J. 1993. Re¯exive language and the human disciplines. In J. Lucy (ed.),
Re¯exive language: Reported speech and metapragmatics (pp. 1±32). Cambridge:
Cambridge University Press.
Lust, B. (ed.) 1986. Studies in the acquisition of anaphora. Dordrecht: Reidel.
Lust, B., Hermon, G., and Korn®lt, J. (eds.) 1994. Syntactic theory and ®rst lan-
guage acquisition: Cross-linguistic perspectives. Vol. 2, Binding, dependencies, and
learnability. Hillsdale, NJ: Lawrence Erlbaum.
Lyons, J. 1977. Semantics. Vols. 1 and 2. Cambridge: Cambridge University Press.
Lyons, J. 1981. Language and linguistics: An introduction. Cambridge: Cambridge
University Press.
Lyons, J. 1995. Linguistic smantics: An introduction. Cambridge: Cambridge
University Press.
Maling, J. 1984. Non-clause-bounded re¯exives in Modern Icelandic. Linguistics
and Philosophy 7:211±241.
Marslen-Wilson, W., and Tyler, L. 1980. Towards a psychological basis for a
theory of anaphora. Proceedings of the Chicago Linguistics Society: Parasession on
Pronouns and Anaphora 16:258±286.
Martin, R. M. 1987. The meaning of language. Cambridge, MA: MIT Press.
Matsumoto, Y. 1995. The conversational condition on Horn scales. Linguistics
and Philosophy 18:21±60.
Matthews, P. H. 1981. Syntax. Cambridge: Cambridge University Press.
May, R. 1987. Logical form as a level of linguistic representation. In Lepore 1987:
305±336.
Mazuka, R., and Lust, B. 1994. When is an anaphor not an anaphor? In Lust,
Hermon, and Korn®lt 1994: 145±176.
McCa¨erty, A. S. 1987. Reasoning about implicature: A plan-based approach.
Ph.D. dissertation, University of Pittsburgh, Pittsburgh, PA.
McCarthy, J. 1980. Circumscription: A form of non-monotonic reasoning. Arti®-
cial Intelligence 13:27±39, 171±172. [Reprinted in Ginsberg 1987: 145±152.]
McCarthy, J. 1986. Applications of circumscription to formalizing common sense
knowledge. Arti®cial Intelligence 28:89±116.
McCawley, J. 1978. Conversational implicature and the lexicon. In Cole 1978:
245±259.
Mehler, J., Sebastian, N., Altmann, G., Dupoux, E., Christophe, A., and Pallier,
C. 1993. Understanding compressed sentences: The role of rhythm and meaning.
In P. Tallal, A. M. Galaburda, R. Llinas, and C. von Euler (eds.), Annals of the
New York Academy of Sciences: Temporal information processing in the nervous
system 682:272±282.
443 References

Miller, G., and Johnson-Laird, P. N. 1976. Language and perception. Cambridge,

MA: Harvard University Press.
Milsark, G. 1974. Existential sentences in English. Ph.D. dissertation, Massachu-
setts Institute of Technology, Cambridge, MA.
Mitchell, B. 1985. Old English syntax. Vol. 1, Concord, the parts of speech, and the
sentence. Oxford: Clarendon Press.
Mitchell, J. E. 1986. The formal semantics of point of view. Ph.D. dissertation,
University of Massachusetts, Amherst.
Moore, R. C. 1985. Semantical considerations on nonmonotonic logics. Arti®cial
Intelligence 25:75±94. [Reprinted in Ginsberg 1987: 127±136.]
Moore, R. C. 1986. Problems in logical form. In B. Grosz, K. Jones, and
B. Webber (eds.), Readings in natural language processing (pp. 285±292). Los
Altos, CA: Morgan Kaufman. [Reprinted from 1981, Proceedings of the ninth
annual meeting of the Association for Computational Linguistics. Urbana±
Champaign, IN: ACL. pp. 117±124.]
Moravcsik, E. A. 1972. Some cross-linguistic generalizations about intensi®er
constructions. Proceedings of the Regional Meeting of the Chicago Linguistic
Society 8:271±277.
Morgan, J. 1978. Two types of convention in indirect speech acts. In Cole 1978:
261±280.
Moxey, L., and Sanford, A. 1993. Communicating quantities: A psychological
perspective. Hillsdale, NJ: Lawrence Erlbaum.
Moyne, J. A. 1971. Re¯exive and emphatic. Language 47:41±63.
Naylor, P. B. 1986. On the semantics of reduplication. In P. Geraghty, L. Car-
rington, and S. Wurm (eds.), FOCAL I: Papers from the Fourth International
Conference on Austronesian Linguistics: Paci®c Linguistics C±93:175±185.
Neale, S. 1992. Paul Grice and the philosophy of language. Linguistics and Phi-
losophy 15:509±559.
Newmeyer, F. J. 1980. Linguistic theory in America. New York: Academic Press.
[2nd ed. 1986.]
Newstead, S. 1995. Gricean implicatures and syllogistic reasoning. Journal of
Memory and Language 34:644±664.
Ni, W., Crain, S., and Shankweiler, D. 1996. Sidestepping garden paths: Assessing
the contributions of syntax, semantics, and plausibility in resolving ambiguities.
Language and Cognitive Processes 11:283±334.
Nichols, J. 1983. Switch-reference in the Northeast Caucasus. In Haiman and
Munro 1983: 245±265.
Nunberg, G. 1978. The pragmatics of reference. Mimeo, Indiana University Lin-
guistics Club, Bloomington, IN. [Reprint of Ph.D. dissertation, 1977, City Uni-
versity of New York.]
Nunberg, G. 1995. Transfers of meaning. Journal of Semantics 12:109±132.
444 References

O'Connor, K. 1987. Disjoint reference and pragmatic inference: Anaphora and

switch reference in Northern Pomo. Paper presented to Wenner-Gren Conference
on ``The role of theory in language description''. Jamaica, November 1987. [Pub-
lished as O'Connor 1993.]
O'Connor, K. 1993. Disjoint reference and pragmatic inference: Anaphora and
switch reference in Northern Pomo. In W. A. Foley (ed.), The role of theory in
language description (pp. 215±242). Berlin: Mouton de Gruyter.
Parker, F., Riley, K., and Meyer, C. 1990. Untriggered re¯exive pronouns in
English. American Speech 65:50±69.
Partee, B. 1984. Nominal and temporal anaphora. Linguistics and Philosophy
7:243±286.
Peirce, C. S. 1932. Collected papers of C. S. Peirce, Vol. 2. Cambridge, MA:
Harvard University Press.
Pica, P. 1984. On the distinction between argumental and non-argumental ana-
phors. In W. de Geest and Y. Putseys (eds.), Sentential complementation (pp. 185±
194). Dordrecht: Foris.
Pica, P. 1987. On the nature of the re¯exivization cycle. In J. McDonough and
B. Plunkett (eds.), Proceedings of the North Eastern Linguistic Society 17:483±500.
Pica, P. 1991. On the interaction between antecedent-government and binding:
The case of long re¯exivization. In Koster and Reuland 1991: 119±136.
Platzack, C. 1987. The Scandinavian languages and the null-subject parameter.
Natural Language & Linguistic Theory 5:377±401.
Pollard, C., and Sag, I. A. 1992. Anaphors in English and the scope of binding
theory. Linguistic Inquiry 23:261±303.
Pollard, C., and Sag, I. A. 1994. Head-driven phrase structure grammar. Stanford,
CA: CSLI and Chicago University Press.
Politzer, G., and Noveck, I. 1991. Are conjunction rule violations the results of
conversational rule violations? Journal of Psycholinguistic Research 20:83±103.
Pople, H. E. 1973. On the mechanization of abductive logic. International Joint
Conference on Arti®cial Intelligence 73:147±152.
Popper, K. 1959. The logic of scienti®c discovery. London: Hutchinson.
Poznanski, V. 1992. A relevance-based utterance processing system. Technical
Report No. 246, University of Cambridge Computer Laboratory, Cambridge.
Postal, P. 1974. On raising. Cambridge, MA: MIT Press.
Prince, E. 1981. Toward a taxonomy of Given-New information. In Cole 1981:
223±255.
Pustejovsky, J. 1995. The generative lexicon. Cambridge, MA: MIT Press.
Putnam, H. 1958. Formalization of the concept ``about.'' Philosophy of Science
25:125±130.
Putnam, H. 1975. Mind, language, and reality: Philosophical papers. Vol. 2.
Cambridge: Cambridge University Press.
445 References

Quirk, R., Greenbaum, S., Leech, G., and Svartvik, J. 1972. A grammar of con-
temporary English. London: Longman.
Radford, A. 1981. Transformational syntax. Cambridge: Cambridge University
Press.
Read, C., and Hare, V. C. 1979. Children's interpretation of re¯exive pronouns in
English. In F. R. Ekman and A. J. Hastings (eds.), Studies in ®rst and second
language acquisition (pp. 98±116). Rowley, MA: Newbury House.
Recanati, F. 1989. The pragmatics of what is said. Mind and Language 4:295±329.
Recanati, F. 1993. Direct reference: From language to thought. Oxford: Blackwell.
Reggia, J., and Nau, D. 1984. An abductive non-monotonic logic. In Proceedings
of the 1984 Non-monotonic Reasoning Workshop (pp. 385±395). New Paltz, NY:
American Association for Arti®cial Intelligence.
Reinhart, T. 1983. Anaphora and semantic interpretation. London: Croom Helm.
Reinhart, T. 1986. Center and periphery in the grammar of anaphora. In Lust
1986: 123±150.
Reinhart, T., and Reuland, E. 1991. Anaphors and logophors: An argument
structure perspective. In Koster and Reuland 1991: 283±322.
Reinhart, T., and Reuland, E. 1993. Re¯exivity. Linguistic Inquiry 24:657±720.
Reiter, R. 1978. On reasoning by default. In D. L. Waltz (ed.), TINLAP 2:
Theoretical issues in natural language processing 2 (pp. 210±218). New York:
ACM.
Reiter, R. 1980. A logic for default reasoning. Arti®cial Intelligence 13:81±132.
[Reprinted in Ginsberg 1987: 68±93.]
Reuland, E. 1994. Commentary: The non-homogeneity of Condition B and
related issues. In Lust, Hermon, and Korn®lt 1994: 227±246.
Reuland, E., and Koster, J. 1991. Long-distance anaphora: An overview. In
Koster and Reuland 1991: 1±26.
Rizzi, L. 1989. On the format for parameters: Comment on D. Lightfoot, The
child's trigger experience: Degree±0 learnability. Behavioral and Brain Sciences
12:355±356.
Roberts, C. 1987. Modal subordination, anaphora, and distributivity. Ph.D. disser-
tation, University of Massachusetts, Amherst.
Rosch, E. 1977. Human categorization. In N. Warren (ed.), Advances in cross-
cultural psychology, Vol. 1 (pp. 1±49). London: Academic Press.
Ross, J. F. 1981. Portraying analogy. Cambridge: Cambridge University Press.
Ross, J. R. 1970. On declarative sentences. In R. A. Jacobs and P. S. Rosenbaum
(eds.), Readings in English transformational grammar (pp. 222±272.) Waltham,
MA: Ginn.
Ruhl, C. 1989. On monosemy: A study in linguistic semantics. Albany, NY: SUNY
Press.
446 References

Russell, B. 1905. On denoting. Mind 14:479±493. [Reprinted in R. Marsh (ed.),

1956, B. Russell: Logic and knowledge. London: George Allen and Unwin.]
Sacks, H., and Scheglo¨, E. 1979. Two preferences in the organization of refer-
ence to persons in conversation and their interaction. In G. Psathas (ed.), Every-
day language: Studies in ethnomethodology (pp. 15±21). New York: Irvington.
Sadock, J. 1978. On testing for conversational implicature. In Cole 1978: 281±297.
Sadock, J. M. 1981. Almost. In Cole 1981: 257±272.
Sadock, J. M. 1983. The necessary overlapping of grammatical components. In
Proceedings of the Chicago Linguistic Society: Parasession on the interplay of
phonology, morphology, and syntax 19:198±221.
Sadock, J. M. 1984. Whither radical pragmatics? In D. Schi¨rin (ed.), Meaning,
form, and use in context (pp. 139±149). Washington, DC: Georgetown University
Press.
Sadock, J. M. 1991. Autolexical syntax: A theory of parallel grammatical repre-
sentations. Chicago: University of Chicago Press.
Sadock, J., and Zwicky, A. 1985. Speech act distinctions in syntax. In T. Shopen
(ed.), Language typology and syntactic description. Vol. 1, Clause structure (pp.
155±196). Cambridge: Cambridge University Press.
Sa®r, K. 1992. Implied non-coreference and the pattern of anaphora. Linguistics
and Philosophy 15:1±52.
Sag, I. 1981. Formal semantics and extralinguistic context. In Cole 1981: 273±294.
Saleemi, A. P. 1997. Review of Y. Huang, The syntax and pragmatics of ana-
phora. Journal of Pragmatics 28:103±128.
Sanford, A., and Garrod, S. 1981. Understanding written language. Chichester,
UK: Wiley.
Sapir, E. 1921. Language. An introduction to the study of speech. New York:
Harcourt, Brace.
Sapir, E. 1930. Totality. Language Monograph Number 6. Washington, DC:
Linguistic Society of America.
Saussure, F. de 1916. Cours de linguistique geÂneÂrale. (Ed. C. Bally and A. Seche-
haye). Paris: Payot. [Republished as Course in general linguistics, Trans. W.
Baskin, 1966. New York: McGraw-Hill.]
Scharten, R. (1997). Exhaustive interpretation: A discourse-semantic account.
Ph.D. dissertation, Catholic University Nijmegen.
Scheglo¨, E. A. 1972. Sequencing in conversational openings. In J. J. Gumperz
and D. Hymes (eds.), Directions in sociolinguistics (pp. 346±380). New York:
Holt, Rinehart, and Winston.
Scheglo¨, E. 1979. Identi®cation and recognition in telephone conversation
openings. In G. Pasthas (ed.), Everyday language (pp. 23±78). New York: Irvington.
Scheglo¨, E. 1996. Some practices for referring to persons in talk-in-interaction.
In B. Fox (ed.), Studies in anaphora (pp. 437±486). Amsterdam: John Benjamins.
447 References

Schelling, T. 1960. The strategy of con¯ict. Cambridge, MA: MIT Press.

Schi¨er, S. R. 1972. Meaning. Oxford: Blackwell.
Schmerling, S. 1975. Asymmetric conjunction and rules of conversation. In Cole
and Morgan 1975: 211±231.
Searle, J. 1969. Speech acts. Cambridge: Cambridge University Press.
Searle 1975. Indirect speech acts. In Cole and Morgan 1975: 59±82.
Searle, J. 1979. Expression and meaning. Cambridge: Cambridge University Press.
Searle, J. 1986. Meaning, communication, and representation. In Grandy and
Warner 1986: 209±228.
Sells, P. 1987. Aspects of logophoricity. Linguistic Inquiry 18:445±479.
Sells, P., Zaenen, A., and Zec, D. 1987. Re¯exivization variation: Relations
between syntax, semantics, and lexical structure. In M. Ida, S. Wechsler, and
D. Zec (eds.), Working papers in grammatical theory and discourse structure (pp.
169±243). Stanford, CA: CSLI. Publications.
Senft, G. 1986. Kilivila, the language of the Trobriand Islands. Berlin: Mouton de
Gruyter.
Seuren, P. (ed.) 1974. Semantic syntax. Oxford: Oxford University Press.
Seuren, P. 1998. Western linguistics: An historical introduction. Oxford: Blackwell.
Shibatani, M. 1972. Three reasons for not deriving ``kill'' from ``cause to die''. In
J. P. Kimball (ed.), Syntax and semantics (pp. 125±138). New York: Seminar
Press.
SigurjoÂnsdoÂttir, S., Hyams, N., and Chien, Y.-C. 1988. The acquisition of re¯ex-
ives and pronouns by Icelandic children. In E. Clark and Y. Matsumoto (eds.),
Proceedings of the Annual Child Language Forum: Papers and reports on child
language development (pp. 97±106). Stanford, CA: Stanford University Press.
Singer, M. 1994. Discourse inference processes. In Gernsbacher 1994: 479±516.
Soames, S. 1982. How presuppositions are inherited: A solution to the projection
problem. Linguistic Inquiry 13:483±545. [Reprinted in Kasher (ed.) 1998: 69±140.]
Soames, S. 1989. Presupposition. In D. Gabbay and F. Guenthner (eds.), Hand-
book of philosophical logic IV (pp. 553±616). Dordrecht: Reidel.
Soames, S. 1998. How presuppositions are inherited: A solution to the projection
problem. In A. Kasher (ed.), Presupposition, implicature, and direct speech acts
(pp. 69±148). London: Routledge.
Solan, L. 1987. Parameter setting and the development of pronouns and re¯exives.
In T. Roeper and E. Williams (eds.), Parameter setting (pp. 189±210). Dordrecht:
Reidel.
Sperber, D., and Wilson, D. 1986. Relevance. Oxford: Blackwell.
Sperber, D., and Wilson, D. 1987. PreÂcis of Relevance. Behavioral and Brain
Sciences, 10:697±754.
Sperber, D., and Wilson, D. 1995. Relevance. 2nd ed. Oxford: Blackwell.
448 References

Stalnaker, R. 1972. Pragmatics. In D. Davidson and G. Harman (eds.), Semantics

of natural language (pp. 380±397). Dordrecht: Reidel.
Steever, S. 1977. Raising, meaning, and conversational implicature. Proceedings of
the Chicago Linguistic Society 13:590±602.
Stirling, L. 1993. Switch-reference and discourse representation. Cambridge: Cam-
bridge University Press.
Strawson, P. 1950. On referring. Mind 59:320±344.
Strawson, P. 1986. `If ' and `I'. In Grandy and Warner 1986: 229±242.
Syromiatnikov, A. 1981. The ancient Japanese language. Moscow: Nanka.
Thagard, P. 1978. The best explanation: Criteria for theory choice. Journal of
Philosophy 74:76±92.
Thomason, R. H. 1974. Introduction. In R. Thomason (ed.), Formal philosophy:
Selected papers of Richard Montague (pp. 1±70). New Haven, CT: Yale Univer-
sity Press.
Thomason, R. H. 1987. Accommodation, meaning, and implicature: Interdisci-
plinary foundations for pragmatics. Ms., University of Pittsburgh. [Printed as
Thomason 1990.]
Thomason, R. H. 1990. Accommodation, meaning, and implicature: Inter-
disciplinary foundations for pragmatics. In P. R. Cohen, J. Morgan, and M. E.
Pollack (eds.), Intentions in communication (pp. 325±364). Cambridge, MA:
MIT Press.
Toney, A. 1963. 150 masterpieces of drawing. New York: Dover Publications.
Traugott, E. 1998. The role of pragmatics in semantic change. Paper presented to
the Conference of the International Pragmatics Association, Rheims, July 1998.
Traugott, E., and Heine, B. (eds.) 1991. Approaches to grammaticalization, Vol. 2.
Amsterdam: John Benjamins.
Traugott, E., and KoÈnig, E. 1991. The semantics-pragmatics of grammaticaliza-
tion revisited. In Traugott and Heine 1991: 189±218.
Tryon, D. T. 1970. Conversational Tahitian. Canberra, Australla: ANU Press.
Ureland, S. 1981. The development of Dutch and West Frisian re¯exives between
1879 and 1979. In M. Gerritsen (ed.), Taalverandering in Nederlandse dialekten
(pp. 250±263). Muiderberg, The Netherlands: Coutinho.
Vaccari, O., and Vaccari, E. 1948. Japanese conversation-grammar. Tokyo:
Vaccari.
Van Deemter, K., and Peters, S. (eds.) 1996. Semantic ambiguity and under-
speci®cation. Stanford, CA: CSLI publications.
Van der Auwera, J. 1995. Conditional perfection. In A. Athanasiadou and R.
Dirven (eds.), On conditionals again (pp. 169±190). Amsterdam: John Benjamins.
Van der Auwera, J. 1996. Modality: The three-layered Scalar Square. Journal of
Semantics 13:181±195.
449 References

Van der Auwera, J. 1997. Pragmatics in the last quarter century: The case of
conditional perfection. Journal of Pragmatics 27:261±274.
Van der Sandt, R. 1988. Context and presupposition. London: Croom Helm.
Van Kuppevelt, J. 1996. Inferring from topics: Scalar implicatures as topic-
dependent inferences. Linguistics and Philosophy 19:393±443.
Van Riemsdijk, H., and Williams, E. 1986. Introduction to the theory of grammar.
Cambridge, MA: MIT Press.
Van Valin, R. 1990. Functionalism, anaphora, and syntax. Studies in Language
14:169±219.
Verschueren, J., and Bertuccelli-Papi, M. (eds.) 1987. The pragmatic perspective.
Amsterdam: John Benjamins.
Vincent, N. 1989a. Observing obsolescence: Comment on D. Lightfoot, The
child's trigger experience: Degree±0 learnability. Behavioral and Brain Sciences
12:360±361.
Vincent, N. 1989b. Constituentization. Paper delivered at the Ninth International
Conference on Historical Linguistics. Rutgers University, New Brunswick, NJ,
August 1989.
Visser, F. T. 1963. An historical syntax of the English language, Part 1. Leiden:
Brill.
Vonk, W., Hustinx, L., Simons, W. 1992. The use of referential expressions in
structuring discourse. Language and Discourse Processes 11:301±334.
Von Wright, G. H. 1971. Explanation and understanding. London: Routledge and
Kegan Paul.
Wainer, J. 1991. Uses of nonmonotonic logic in natural language understanding:
Generalized implicatures. Ph.D. dissertation, Pennsylvania State University. [Pub-
lished by UMI Ann Arbor].
Ward, G., and Hirschberg, J. 1985. Implicating uncertainty: The pragmatics of
fall-rise intonation. Language 61:747±776.
Webber, B. 1979. A formal approach to discourse anaphora. New York: Garland.
Welker, K. 1994. Plans in the common ground: Toward a generative account of
conversational implicature. Ph.D. dissertation, Ohio State University, Columbus.
Wexler, K., and Manzini, M. R. 1987. Parameters and learnability in Binding
Theory. In T. Roeper and E. Williams (eds.), Parameter setting (pp. 41±76).
Dordrecht: Reidel.
Wheeldon, L. R., and Levelt, W. J. M. 1995. Monitoring the time course of pho-
nological encoding. Journal of Memory and Language 34:311±334.
Wilkins, D., and Hill, D. 1995. When `GO' means `COME': Questioning the
basicness of basic motion verbs. Cognitive Linguistics 6:209±259.
Wilkins, W. 1988. Thematic structure and re¯exivization. In W. Wilkins (ed.),
Syntax and semantics. Vol. 21, Thematic relations (pp. 191±214). New York:
Academic Press.
450 References

Williams, E. 1997. Blocking and anaphora. Linguistic Inquiry 28:577±628.

Wilson, D. 1975. Presuppositional and non-truth-conditional semantics. New York:
Academic Press.
Wilson, D., and Sperber, D. 1986. Pragmatics and modularity. Proceedings of the
Chicago Linguistics Society: Parasession on pragmatics and grammatical theory
22:67±84. [Reprinted as Wilson and Sperber 1991.]
Wilson, D., and Sperber, D. 1991. Pragmatics and modularity. In S. Davis (ed.),
Pragmatics (pp. 583±595). Oxford: Oxford University Press.
Winograd, T. 1972. Understanding natural language. New York: Academic Press.
Yang, D.-W. 1983. The extended binding theory of anaphors. Language Research
19:169±192.
Zi¨, P. 1975. On H. P. Grice's account of meaning. In D. Steinberg and L. Jako-
bovitz (eds.), Semantics (pp. 60±65). Cambridge: Cambridge University Press.
Zipf, G. K. 1949. Human behavior and the principle of least e¨ort: An introduction
to human ecology. New York: Hafner.
Zribi-Hertz, A. 1989. Anaphor binding and narrative point of view: English
re¯exive pronouns in sentence and discourse. Language 65:695±727.
Zribi-Hertz, A. 1995. Review of Y. Huang, The syntax and pragmatics of
anaphora. Lingua 96:179±189.
Name Index

Adams, M., 423 Blumstein, S., 381

Allen, S., 45, 52, 142 Bohnemeyer, J., 96
Ameka, F., xxi Bolinger, D., 291, 296, 298, 307
Anderson, S., 177 Botha, R., 152
Andronov, M., 351 Bouchard, D., 308, 312, 323
Appelt, D., xvii Bourdieu, P., xiii, 386
Ariel, M., xvii, 273, 409, 411 BreÂal, M., 137
Aristotle, 30, 64, 74 Brown, P., xxi, 13, 24, 45, 46, 89, 264, 407,
Arono¨, M., 369 408, 423
Arundale, R. B., xxi, 5, 379 Burrow, T., 352
Asher, N., 123, 161, 352, 423 Burton-Roberts, N., 80, 391
Atlas, J., xv, xxi, 1, 4, 22, 41, 43, 45, 46, 64, Burzio, L., 409, 416
74, 76, 77, 78, 79, 80, 82, 88, 89, 90, 91, Bybee, J., 263
109, 111, 114, 115, 116, 117, 156, 159,
171, 177, 185, 189, 203, 207, 222, 241, Caldwell, R., 352
256, 257, 258, 361, 375, 376, 380, 381, Calvert, D. R., 28, 382
386, 387, 388, 390, 391, 392, 395, 397, Cann, R., 95
401, 402, 403, 410 Cantrall, W. R., 319, 320, 333
Austin, P., 23, 335, 336, 381, 402 Carberry, S., 380
Avramides, A., 13, 380, 398 Carden, G., 335, 338, 339, 340, 361, 419
Carnap, R., 31, 34, 115, 273, 278, 299, 391,
Bach, K., xxi, 23, 24, 165, 169, 170, 188, 410
194, 195, 196, 197, 198, 214, 217, 231, Carston, R., xvii, 48, 89, 91, 114, 120, 121,
257, 258, 262, 375, 379, 380, 381, 383, 123, 134, 194, 195, 196, 199, 238, 254,
384, 392, 396, 397, 398, 400, 402, 409 257, 258, 392, 393, 400, 403
Bacon, F., 43 Chafe, W., 237
Baker, G. P., 192, 308, 415, 416, 420 Changeux, J. P., 2
Bar-Hillel, Y., 31, 34, 115, 163, 273, 278, Charniak, E., 45, 384
299, 410 Cherry, C., 410
Barton, E., 183, 184 Chien, Y-C., 344, 361
Barwise, J., xvii, 167, 177, 191, 208, 209, Cho, Y. Y., 411
210, 226, 228, 229, 388, 397, 402 Chomsky, N., xiii, 5, 8, 29, 163, 166, 190,
Bayer, S., 24, 134, 262, 264 210, 248, 261, 265, 266, 280, 281, 282,
Berlin, B., 152 283, 296, 308, 309, 310, 324, 343, 364,
Bickerton, D., 338, 340 382, 397, 408, 413, 418, 423
Bierwisch, M., 372, 424 Chung, S., 337, 418, 422
Blakemore, D., xvi Clark, E., xxi, 237, 370, 376
Blass, R., 148 Clark, H. H., xix, xx, 29, 30, 117, 126, 139,
Bloom®eld, L., 125, 137 140, 237, 258, 376, 386, 395, 397
452 Name Index

Claudi, U., 263 Geach, P. T., 333

Cohen, L. J., 31, 44, 46, 198, 199, 206, 214, Geiss, M., 117
380 Geniusiene, E., 338, 351
Cole, P., xv, 256 Geurts, B., 407
Comrie, B., 95, 96, 239, 329, 335, 364, 408 Ginsberg, M., 43, 45, 46, 275, 383
Cooper, R., 388 Goodman, N., 208
Cormack, A., 211 Gordon, D., 395
Corne, C., 339 Grandy, R., xv, 382
Crain, S., 424 Green, G., 380, 382
Cruse, D. A., 90, 101, 103 Greenbaum, S., 147, 150
Culy, C., 414 Greenberg, J., 126, 128, 137
Grice, H. P., xv, xix, 6, 11, 12, 13, 14, 15,
Davis, S., xv 16, 17, 18, 19, 20, 21, 22, 29, 35, 37, 38,
De Morgan, A., 68 41, 46, 48, 53, 55, 63, 72, 73, 74, 75, 78,
Dimitriadis, A., 411 91, 92, 108, 111, 112, 113, 135, 136, 165,
Dixon, R. M. W., 90, 152, 335, 336 166, 169, 170, 171, 172, 181, 186, 187,
Donnellan, K. S., 217, 226, 227, 228, 229, 189, 194, 195, 198, 199, 225, 236, 258,
230, 404 260, 361, 371, 372, 373, 377, 379, 380,
Dowty, D., 290, 299, 319, 411, 412 381, 382, 386, 391, 396, 398, 399, 402,
Dulany, D., 424 403, 404
Grodzinsky, Y., 292, 293, 294, 344
Eades, D., 336 Gundel, J., 94, 95, 411
Eco, U., 386
Edmondson, J. A., 333, 351, 417, 418 Haas, W., 101
Edwards, D., xx, 384 Hacker, P., 192
Emeneau, M., 352 HageÁge, C., 314, 315, 420
EncË, M., 177, 192, 242, 243, 246, 397, 405, Haiman, J., xvi, 75, 112, 113, 123, 126, 135,
411 136, 142, 146, 148, 276, 317, 330, 342,
Erasmus, D., 145 394, 414
Evans, G., 278, 284, 335, 413 Hale, K., 303
Everaert, M., 342, 417, 419 Hamblin, C. L., 49, 163
Hamilton, Sir W., 68
Faltz, L. M., 263, 325, 330, 331, 335, 341, Harbert, W., 265, 282, 308, 310, 311, 312,
349, 350, 416, 417, 419, 420, 421 313, 328, 414, 415
Farmer, A., xx, 328 Harbsmeier, C., 420
Fauconnier, G., 104, 236, 237, 245, 333, Harder, P., 96
388 Hare, V. C., 344
Fillmore, C., 99, 177, 397 Harman, G., 54, 380, 383, 384
Fodor, J., 187, 190, 257, 397, 405 Harnish, R., xv, xx, 1, 13, 23, 95, 98, 100,
Fogelin, R., 388 117, 159, 170, 184, 207, 262, 328, 381,
Foley, W., 276, 365 390, 398
Ford, C. E., 403 Harris, R. A., 394
Fox, B., 271, 273, 291, 409 Haviland, J., 117, 126, 303, 334, 386, 389,
Fox Tree, J., 395 393, 420
Frederking, R., 111 Hawkins, J., xvi, 92, 93, 127, 155, 390, 397
Frege, G., 390 Hedberg, N., 94, 95, 411
Fretheim, T., 90 Heim, I., 60, 62, 63, 93, 167, 182, 193, 248,
249, 250, 252, 253, 255, 385, 397, 402,
Garnham, A., 269 406
Garrod, S., 269, 424 HeinaÈmaÈki, O., 97
Gazdar, G., xv, xix, 1, 20, 35, 48, 49, 64, Heine, B., 263
75, 76, 77, 78, 79, 80, 81, 98, 108, 109, Hellan, L., 306, 310, 315, 326, 417
110, 111, 125, 158, 161, 162, 163, 198, Herskovits, A., 118, 207, 223, 237
201, 212, 214, 252, 255, 296, 297, 384, Higginbotham, J., 413
385, 386, 387, 388, 390, 391, 395, 396, Hill, D., 94
402, 406, 423 Hilton, D., 424
453 Name Index

Hinds, J., 421 Kamp, H., 25, 62, 88, 122, 163, 167, 193,
Hintikka, J., 77, 333, 387 205, 247, 248, 249, 250, 251, 252, 385, 406
Hirschberg, J., xvi, xvii, 64, 78, 80, 82, 99, Kant, E., 75, 386
101, 102, 104, 105, 107, 108, 120, 162, Kaplan, D., 191, 243, 245, 246, 247
212, 255, 380, 385, 387, 390, 391, 396, Karttunen, L., 180
403 Kasher, A., xv
Hirst, G., 174 Katz, J. J., 190, 240, 241, 404
Hobbs, J., xvii, xx, 12, 45, 46, 60, 61, 62, Kaufman, D., 340, 344, 361
63, 117, 125, 126, 165, 167, 237, 238, Kay, P., 77, 161, 207, 397, 403, 417
276, 375, 397, 404, 423 Keenan, E., 177, 397, 423
Holdcroft, D., 17, 52 Kempson, R., xvii, xix, xx, 25, 88, 89, 90,
Holland, J. J., 44 103, 182, 190, 210, 211, 212, 240, 254,
Holyoke, K. J., 44 255, 256, 257, 376, 389
Hong, K. -S., 411 Kenny, A. J., 43, 45
Hopper, P., 263 Kibrik, A., 273
Horn, L. R., xv, xvi, xix, xx, xxi, 1, 4, 5, 6, Kiparsky, P., 139
12, 15, 24, 38, 40, 41, 47, 52, 64, 68, 69, Kiss, K. EÂ., 303, 306
74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, Klein, E., 203
85, 86, 87, 88, 89, 90, 91, 98, 102, 103, KoÈnig, E., 263, 333, 351, 370, 416
105, 107, 109, 110, 111, 115, 117, 120, Koster, C., 325, 344, 351, 422
122, 124, 127, 128, 129, 130, 131, 132, Koster, J., 265, 282, 286, 311, 312, 313,
133, 134, 135, 136, 137, 139, 140, 141, 316, 332, 345, 412, 414, 419, 422
142, 143, 145, 146, 150, 155, 156, 157, Kripke, S., 188, 226, 228, 229, 404
162, 169, 181, 185, 190, 194, 197, 210, Kroch, A., 422
211, 212, 213, 222, 224, 237, 239, 240, Kuno, S., 237, 297, 307, 315, 316, 319,
254, 255, 262, 264, 274, 277, 278, 286, 320, 322, 356, 411, 413, 414, 415, 421
287, 290, 296, 297, 331, 333, 355, 356,
361, 364, 365, 369, 381, 385, 386, 387, Ladd, R., 102, 391
388, 389, 390, 391, 392, 393, 394, 396, Lafon, J. -C., 381
403, 407, 408, 411, 412, 413, 414, 415, Lako¨, G., xix, 192, 243, 395, 406, 408
418 Landman, F., 193, 253
Hornstein, N., 265 Lang, E., 122, 125
Householder, F. 140 Langacker, R., 243, 263, 332, 417, 420
Huang, Y., xv, xvi, xx, xxi, 210, 301, 312, Lascarides, A., 123, 161, 423
313, 316, 326, 329, 330, 346, 351, 352, Lasnik, H., 299, 300, 301, 412
353, 354, 355, 356, 357, 358, 364, 365, Laver, J., 382
381, 407, 408, 411, 414, 416, 417, 421, Lee, Y. -S., 414
422 Leech, G., 14, 74, 134, 147, 15
HuÈnnemeyer, F., 263 Lehmann, C., 276
Hustinx, L., 409 Lehrer, A., 101
Hyams, N., 361 Levelt, W. J. M., 28, 46
Levesque, H., 48
Jackendo¨, R., 8, 192, 243, 245, 282, 298, Levin, N., 393
306, 307, 372, 397, 398, 424 Levine, J. E., 97
Jackson, S., 370 Levinson, S. C., xii, xiii, xvi, xvii, xx, 4, 7,
Jacobs, S., 370 12, 14, 17, 23, 24, 30, 38, 40, 41, 43, 45,
Jakobson, R., 115, 137 46, 51, 54, 55, 57, 59, 64, 74, 75, 76, 79,
Jakubowicz, C., 340, 344 80, 89, 91, 96, 111, 112, 114, 115, 116,
Jayaseelan, K. A., 333 117, 120, 121, 156, 157, 158, 159, 171,
Jespersen, O., 68, 84, 342 177, 178, 181, 195, 197, 201, 207, 214,
Johnson-Laird, P. N., 33, 42, 54 217, 222, 224, 242, 252, 257, 258, 263,
Josephson, J., 42 264, 272, 277, 283, 284, 297, 303, 305,
307, 334, 347, 353, 359, 361, 364, 372,
Kadmon, N., xx, 64, 88, 89, 90, 162, 179, 380, 383, 383, 384, 385, 388, 391, 392,
182, 193, 219, 248, 249, 250, 385, 389, 392, 395, 401, 402, 403, 407, 408, 409,
406 411, 412, 413, 416, 418, 420, 423
454 Name Index

Lewis, D., 12, 29, 60, 177, 233, 391, 399, Pagliuca, W., 263
402 Parker, F., 415
Li, C., 354, 421 Partee, B., 122, 151, 208, 399, 400
Lieberman, P., 381 Pederson, E., xxi
Lightfoot, D., 361, 362 Peirce, C. S., 44, 384
Linebarger, M., 134, 262 Perlmutter, D., xx, 418
LoÈbner, S., 86, 131 Perkins R., 263
Lucy, J., 237 Perrault, R., xx
Lust, B., 312, 316, 325, 340, 345, 351, 418, Perry, J., xvii, 167, 177, 191, 226, 228,
421 229
Lyons, J., xx, 5, 93, 101, 102, 127, 128, 237, Peters, S., xx, 25, 240, 376, 397
381 Pica, P., 312, 414, 417
Plank, F., 333, 351, 417, 418
Maling, J., 310 Platzack, C., 422
Manzini, M. R., 313, 361 Politzer, G., 424
Marslen-Wilson, W., 382 Pollack, M., 46, 380
Martin, P., xvii, 46, 218 Pollard, C., 282, 297, 306, 307, 310, 321,
Matsumoto, Y., 80, 102, 388 323, 324, 325, 407, 416
Matthews, P., xx, 8, 125, 376, 407 Pople, H. E., 42, 384
May, R., 240, 286 Popper, K., 31, 115, 274
Mazuka, R., 312, 316, 325, 345, 351, 418, Poser, W., 382
421 Postal, P., 324
McCa¨erty, A. S., xvii, 46, 52, 385 Poznanski, V., 385
McCarthy, J., 48, 276, 384 Prince, E., 94, 393
McCawley, J., 137, 140, 141, 142 Pullum, G., 386
McDermott, D., 45, 384 Pustejovsky, J., 240, 376
McMahon, A., 422 Putnam, H., 115, 217, 388, 391
Mehler, J., 28
Meyer, C., 415 Quirk, R., 147, 150, 151
Mill, J. S., 68
Miller, G., 33 Radford, A., 281, 286
Milsark, G., 297 Read, C., 344
Mitchell, B., 177, 341, 358, 419, 422 Recanati, F., 170, 192, 194, 195, 196, 197,
Mitchell, J. E., 237 238, 258, 399, 400
Mohanan, K., xx Reggia, J., 384
Moore, R. C., 48, 209, 386, 399, 402 Reinhart, T., 181, 182, 248, 265, 278, 280,
Moravcsik, E. A., 332, 333 283, 290, 292, 293, 294, 295, 296, 297,
Morgan, J., 23, 24, 46, 134, 262, 380 298, 299, 301, 302, 306, 307, 319, 329,
Moxey, L., 424 344, 359, 411, 412, 418, 422
Moyne, J. A, 333 Reiter, R., 47, 383
Munro, P., 317, 332 Reuland, E., 265, 280, 282, 306, 311, 313,
316, 329, 332, 351, 412, 414
Nau, D., 384 Reyle, U., 122, 163, 247, 248
Naylor, P. B., 150, 151 Riley, K., 415
Neale, S., 380 Rizzi, L., 417
Newmeyer, F. J., 394 Roberts, C., xx, 193, 397, 406
Newstead, S., 424 Rosch, E., 101
Ni, W., 424 Ross, J. R., 319, 376, 394
Nichols, J., 237, 319 Ruhl, C., 165, 184, 240, 376
Nisbett, R., 44 Russell, B., 180, 219, 231, 377, 403,
Noveck, I., 424 404
Nunberg, G., 217, 236, 245, 376
Sacks, H., 113, 149
Ochs, E., 423 Sadock, J. M., xix, 8, 15, 88, 184, 290, 392,
O'Connor, K., 237, 315, 316, 317, 319, 346, 397, 406, 411
356 Sa®r, K., 412, 415
455 Name Index

Sag, I., xx, 191, 233,236, 237, 242, 245, 246, Van Kuppevelt, J., xvii, 25, 90
282, 297, 306, 307, 310, 321, 323, 324, Van Riemsdijk, H., 413
325, 402, 406, 407, 416 Van Valin, R. D., 276, 307, 365, 415
Saleemi, A. P., 407 Vincent, N., xx, 362, 407, 419, 420
Sanford, A., 269, 424, 424 Visser, F. T., 340, 341, 342, 419
Sapir, E., 84, 150 Vonk, W., 409
Saussure, F. de, 137
Scharten, R., xvii, 90, 389 Von Wright, G. H., 30, 45
Scheglo¨, E., 113, 114, 149, 179 Wainer, J., xvii, 48, 423
Schelling, T., 6, 29, 30, 53, 258, 259 Ward, G., 102, 402
Schi¨er, S. R., xix, 29, 380, 398 Warner, R., xv, 382
Schmerling, S., 123, 124, 392, 394 Webber, B., 267
Searle, J., xix, 23, 134, 217, 381, 385, 397 Welker, K., xvii
Sells, P., 237, 314, 315, 316, 351, 356, 421 Wexler, K., 313, 344, 361
Senft, G, 337 Wheeldon, L. R., 28
Seuren, P., 203, 389, 394, 423 Wilkins, D., xxi, 94, 306
Shankweiler, D., 424 Williams, E., 369, 413
Shibatani, M., 141, 142 Wilson, D., xii, xvi, xix, xx, 4, 5, 12, 22, 24,
Sigurjonsdottir, S., 361 25, 30, 42, 52, 55, 56, 57, 59, 74, 118,
Simons, W., 409 135, 164, 167, 174, 184, 186, 187, 190,
Singer, M., 424 194, 195, 196, 199, 207, 209, 210, 214,
Soames, S., 77, 78, 161, 163, 385, 391, 396 236, 238, 239, 240, 257, 258, 361, 371,
Solan, L., 340, 344, 361 372, 375, 380, 384, 385, 392, 396, 397,
Sperber, D., xii, xvi, xix, xx, 4, 5, 12, 22, 398, 399, 400, 405
24, 25, 30, 42, 52, 55, 56, 57, 59, 74, 118, Winograd, T., 31, 174, 271
135, 164, 167, 174, 184, 186, 187, 190,
194, 195, 196, 207, 236, 238, 239, 240, Yang, D-W., 417
257, 258, 361, 371, 372, 375, 380, 384,
385, 392, 396, 397, 398, 399, 400, 405 Zacharski, R., 94, 95, 411
Stalnaker, R., 49, 249, 399 Zaenen, A., 351
Steever, S., 407 Zec, D., 351
Stewart, W., 335, 338, 339, 340, 361 Zi¨, P., 29
Stickel, M., xvii Zimmer, K., 207
Stirling, L., 314, 315, 316, 317, 318, 347, Zipf, G. K., 6, 26, 112, 115
356, 414, 415 Zribi-Hertz, A., 282, 297, 308, 320, 321,
Strawson, P., 20, 180, 217, 219, 231, 377, 322, 325, 329, 365, 414, 415, 416
403 Zwicky, A., 117, 392
Svartvik, J., 147, 150
Syromiatnikov, A., 421

Thagard, P., 44
Thomason, R. H., xvii, 12, 17, 46, 52, 60,
61, 63, 385, 399
Thompson, S., 354, 403, 421
Toney, A., xxi, 2
Traugott, E., 263, 370
Tryon, D. T., 337
Tyler, L., 382

Ureland, S., 342

Vaccari, E., 421

Vaccari, O., 421
Van der Auwera, J., 119, 388, 392
Van Deemter, K., 240, 376
van der Sandt, R., xx, 77, 381
Subject Index

`A-®rst account' of Binding patterns, 283, and implicature, 264, 280, 344, 361±362,
285±327 370
characterized, 283, 286±289 of re¯exives (see Anaphors, Binding
diachronic relation to B-®rst, 347±348 Conditions, re¯exives)
disjoint reference pragmatic, 318±319, 414 addressee, vs. speaker. See speaker
n.47 adjectives, epistemic, 131±132
®rst approximation, 285±291 gradable, semantic generality of, 184±185
second approximation, 292±298 I-implicatures of, 127±129, 404 n.51
problems for: scalar implicatures of, 86±87
c-command, 297±298, 303±305 adverbs, quanti®cational, 86
constructional oppositions, 296±297 a½rming the consequent, 44
contrasts over reference vs. logophoricity, a½xal negation. See negation
316±317, 322±323, 327, 345±347 African languages, Western, 314, 350, 414
order of antecedent/anaphor, 290±291 n.46, 416 n.59
language variation, 300±301, 310±311, Afrikaans, 152±153, 342
313±317 agreement, pragmatic vs. syntactic, 262,
overlap of pronouns and re¯exives, 290, 407 n.2
308, 313±317, 319±324, 345±347 AI. See arti®cial intelligence
pseudo-Anaphors, 324±326 Aktionsart, 124
See also Binding Conditions alethic modality, 84
A-vertex of square of oppositions, 65, 82, all. See quanti®ers
83±84 `all over' inference, 100. See also color terms
abnormal expressions. See markedness Algonquian, 415 n.51
abductive inference, 42±44, 384 n.33, n.35 alternates, marked vs. unmarked, 136±137
and implicature, 45, 46, 60 constructional, 108±111
of laws, 44 lexical, saliency and, 32, 36, 80, 100
in local pragmatics, 46 scales and, 75±111
nonmonotonicity of, 43 referring expressions, 149
from nonpropositions, 257±258 See also contrast, Q-implicature
`aboutness' ambiguity
constraint on scales, 80, 121 of Anaphors with pronouns, 337±338, 418
informativeness and, 275 n.65, 422 n.92
accommodation, xvii, 12, 60±63, 233, 249, avoidance of, 14, 135, 290
251, 385 n.47, n.48, 399 n.17, 405 n.55 and informational strength, 297
de®ned, 60 lexical, 176±177
and failed reduction of GCIs, 60±63 of negation, 129±130
vs. gratuitous nature of GCIs, 61, 63 of pronouns, 269±270, 337
acquisition, of language, xii, 23, 140, 264, of referring expressions, 226±230, 233, 398
340, 344, 361±362, 416 n.59, 422 n.93 n.10
458 Subject Index

ambiguity (cont.) and epithets, 277, 399 n.14

vs. semantic generality, 20±21, 88±89, 135, ¯exible interpretation of, 358
165, 184, 228±229, 376 (see also and gaps, 270, 285, 305 (see zero-, and
monosemy) pronouns)
of scope (see scope) `general anaphora pattern', 267±268, 272,
structural, 174±176, 376 285, 305
Amerindian languages, 416 n.59 in generative grammar, 265 (see also
Anaphors, Chomskyan, 267, 270, 281±283, Binding)
286±290, 324±326 grammatical relations and, 306±307
acquisition of (see re¯exives) grammatical speci®cation of, coreference
ambiguous, 337±338, 418 n.65, 422 n.92 only, 305
anti-, 414 n.47 domains only, 270
covert, 337±338, 418 n.65 implicature and, 272, 326
`exempt', 324 (see pseudo-) I-Principle and, 272±276, 285, 299, 305,
and language change, 263±264, 338±342, 318, 325, 329±331
344, 347±352, 355±358, 417 n.59, 419 M-Principle and, 272, 277, 285, 288, 290,
n.68±73, n.75 301
languages without, 334±343, 418 n.62 Q-Principle and, 277±280, 286±287, 290,
long-distance (LDA), 282, 283, 296, 310± 294±297, 302, 305, 313±317, 319, 322±
317, 324±326, 414 n.42, n.43, n.44 323
characteristics of, 312, 413 n.42, n.44 restricted role of, 271±272
non-clause-bound (NCBR), 311 linkage to antecedents, pragmatic, 270±
`one only' rule, 321, 415 n.50 271, 409 n.9
origin as emphatics, 332±334, 341±342, and logophoricity, 314±317
347±350, 358, 416 n.58, 420 n.81, n.83, long-range, 282, 283, 296, 310±317, 325
421 n.84±85 order of antecedent/anaphor, 290±292, 307
parameters and, 312±313, 360±363 pragmatic nature of, 272, 343
proto-, 358±359 and pragmatic resolution, 179, 181±183,
pseudo-, 282, 307, 313, 316, 323, 324±326 217, 224±225, 249±250, 271, 324
raising of, 308, 312±313, 330, 363 `precede and command' analysis of, 307
simplex vs. compound morphemes, 311± and referential dependence, 268, 270, 353
312, 332±333, 416 n.59 and semantic generality, 268±269
weak vs. strong, 332, 413 n.38, 416 n.59 and semantics, 248
See also anaphora, Binding Conditions, sentential, vs. discourse, 251, 267±273
reciprocals, re¯exives temporal, 122±124, 151
anaphora thematic roles and, 306±307
`anaphoric potential', 269±270 VP, 251, 267
bound, 292±298 zero-, 146±149, 285, 305, 382 n.22
and Binding (see Binding Conditions) See also Anaphors, Binding Conditions,
in Chinese, 352±358 coreference, pronouns, reference,
clause-bound vs. long-distance, 311±317 re¯exives
co-arguments, and, 304, 306±307, 325, anaphoric expressions, 269±270
328±332, 353±354 scalar analysis of (see scales)
complementary distribution of pronouns semantic generality of, 269±270
and re¯exives, 283±284, 288, 290, 308 signals of switched reference, 95
breakdown of, 308±327, 414 n.44 anatomical adaptations for language, 28
contrasts in more than reference, 316± and, 19, 80, 91, 122±127, 155, 199, 206,
317, 320±323 242. See also conjunction, connectives
con®gurational constraints on, 291, 292± anomaly, pragmatic, 199, 209±210
303, 306 (see also Binding Conditions) semantic, 202±204, 209±210, 226 (see also
coerces implicature, 179, 181, 250±251 contradiction)
de®ned, 267 antecedents, blocked, 290
and deixis, 179 linkage is pragmatic, 270±273, 324±325
and determiners, 91±95 no grammatical speci®cation of, 270
in discourse, 271±273, 316 order with respect to anaphor, 290±291,
and ellipsis, 183 307
459 Subject Index

uniqueness condition, 250 given by I-Principle, 329±331

See also anaphora over-ridden by M-Principle, 331±333,
antipassive, 304±305, 335, 419 n.77 353
antonymy, 127±129. See also sense emphatics and re¯exivity, 332±334, 337,
relations 347±348, 350, 420 n.81
arbitrariness vs. motivation, in pragmatics, M-implicated conjoint reference, 331±333,
15. See also iconicity 336±337
argument from design, for GCIs, 27±35, Old English and, 340±342
382 n.21 See also Binding Conditions
Aristotelian categories, 386 n.2 bias, in interpretation, 90, 389 n.28
articles, 17±18, 155, 156 biconditional, 37. See also conditional
accommodation of, 60±63 perfection
de®nite (see de®nite article) Binding Conditions, 265, 280±344
Grice's theory of, 17, 91 acquisition of, 340, 344, 361, 416 n.59
inde®nite, 17, 62±63, 89, 91±93 not in accord with nativist views, 344,
phonetic subtypes of, 394 n.72 361±363, 422 n.93
scalar account of, 63, 91±93 c-command and, 281, 290, 292±298, 301±
vs. zero-determiner, 146±147, 156 303, 306
arti®cial intelligence, 25, 45, 167, 170, 275, `compatibility requirement', 309
423 n.1 counterexamples to, 282, 290, 301±303,
aspect and implicature, 96, 124 310±317, 413 n.35, 413 n.40
attributive uses of referring expressions. See cross-linguistic variation in, 301, 303±
referential 305
Australian languages, 90, 125, 152, 269, diachronic account of (see `B then A'
270, 283, 303±305, 335±336 account)
Austronesian languages, 335, 336±338 de®ned, 281±283
autoepistemic logic, 48 domains (see domain)
autohyponyms, 102±103, 394 n.66 and language-change (see parameters)
autonomy of semantics from pragmatics. and LF, 265, 281, 308, 312
See semantics nativism and, 338, 340, 343±344, 360,
auxiliary, 263 361±363
avoidance of expressions, 294. See also noncon®gurational constraints on (see
M-implicature, Q-implicature noncon®gurational, theme)
Avoid Pronoun principle, 364, 407 n.4, 423 parameters and (see parameters)
n.95 pidgins and `bioprogram', 338±340, 361
pragmatic reduction of
bandwidth of human speech, 28, 35, `A-®rst account', 285±327 (see `A-®rst
169 account')
`basic level' terms, 101, 102 `B-®rst account', 327±344 (see `B-®rst
because-clauses, 80, 210, 402 n.36 account')
belief vs. knowledge, 78±79, 387 n.9 `B then A' account, 345±360 (see `B then
believe vs. know, 87, 129±133 A')
best interpretation, inference to. See complimentarity of pronouns vs.
inference re¯exives in, 288, 290, 295, 313±317,
`B-®rst account' of Binding Patterns, 284, 319±327, 346
327±344 Condition A reduced, 284, 327±344
Australian languages and, 335±336 Condition B reduced, 283±284, 285±
Austronesian languages and, 336±338 327
characterized, 328±332 Condition C reduced, 283, 286±289,
Chinese and, 353±358 298±301, 305, 412 n.31
Creoles and, 338±340 Y. Huang's account, 352±358, 421 n.87
derivation of Condition A, 331 Lasnik's rejection of, 299±303
derivation of Condition C, 331 logophoricity and, 314±317, 322±323
diachronically prior, 328 Reinhart's account, 292±298
disjoint reference presumption (DRP), style and dialect and, 301, 415 n.51, 416
328±331, 342, 347, 353 n.55
460 Subject Index

Binding Conditions (cont.) ceteris paribus assumptions, 45, 79, 208±

variant versions of, 265, 286, 290, 292± 209
295, 303, 310, 312±313, 324±325, 326± Chamorro, 337±338
327 character vs. content (Kaplan), 245±247
See also Anaphors, anaphora, pronouns child language. See acquisition
binding of variables. See variables Chinese, 125, 285, 301, 311±312, 313, 316,
bioprogram hypothesis (Bickerton), 338, 334, 346, 352±358, 364, 420 n.83, 421
339±340, 361 n.88±89
blocking, in morphology, 139, 161, 369, 394 clausal implicatures, 35±37, 50, 76, 82,
n.69 108±111, 161, 296
in default reasoning, 161 conditions for, 111
blocks world, 31 constructional basis of, 108
bottleneck, on speech encoding, 6±7, 28, and epistemic uncertainty, 78, 108±109
381 n.16, 382 n.18, n.19 role in presupposition cancellation, 111
asymmetry with comprehension, 28, 382 vs. scalar, 161, 391 n.40
n.20 clausemate NPs, and anaphora, 306, 323±
boundedness of events, 96, 97±98 324
brevity, 14, 38, 114±115, 126 clauses, inferred connections between, 276.
equal, 108, 111 See also cohesion, I-implicature
and semantic generality, 115, 135, 181, closed-class morphemes, 90±98, 155
267, 277 circumscription, 48, 275±276, 384 n.38
See also I-Principle, Least E¨ort principle, See also nonmonotonic reasoning
minimal expressions, minimization coded meaning, vs. inference, 14, 15, 20, 21,
bridging inferences, 37±38, 61, 117, 126± 31, 34, 54, 128, 130, 142, 162, 169,
127, 182, 232±233, 249 257, 273, 287, 334, 343, 374±376
with inde®nites, 62±63, 127, 182 See also logical form, `said', semantic
`B then A' account of Binding Patterns, representation, semantics
284, 340, 345±360 coercion, 246. See also intrusion
for Chinese, 352±358 cognition, and pragmatics, 5, 21
as diachronic pattern, 347±348, 360 and semantics, 7, 241±242
for English, 340±342, 347±349, 357±358 See also psychology, psycholinguistics
Long-range re¯exives and DRP, 346 cognitive, e¨ort, 57, 135
linguistics, 21, 243, 372
calculability of implicatures. See cohesion, presumption of, 114, 116, 126,
conversational implicature 148, 276, 318, 409 n.13
cancellation coindexing, 282, 284, 292±293, 409 n.10.
of implicatures, 15, 49±54, 56 See also coreference, reference
as ambiguity, 384 n.31 collocational constraints, 261±262
at arbitrary distance, 51, 56±57, 163, 247, color terms, 36, 100, 105, 140, 152, 184, 390
253, 384 n.31, 396 n.80, 406 n.56 n.33
by common ground, 49±52, 162±163, 299 come, and deictic verbs of motion, 94
by entailments, 50, 56±57, 155, 299±300 commitment, of speaker. See speaker
by relevance, 88 common ground, augmentation of, 47, 48,
of presupposition, 252 49±54, 60, 193, 249, 384 n.36
See also defeasibility, projection communication, apparent certainty of, 53±
cardinal numbers. See number words 54
case, grammatical, 97, 186 theory of, 11¨, 21±35, 187
causal inferences, 122±123, 148, 182 standard vs. three-leveled theory of, 22±
direct vs. indirect causation, 141±142, 224, 27, 54±63
297 See also meaning, meaningnn
causatives, 141±142, 208 communicative content, of utterances
causatives, lexical vs. periphrastic, 39, 141± communicative load, of entailments and
142 implicatures, 70
c-command, 281, 290±298, 303, 411 n.24 communicative intention. See intention
alternatives to, 306 comparatives, 87, 199±205
See also Binding Conditions degree adjectives as, 184±185, 242
461 Subject Index

intrusion of implicatures in, 199±205 asyndetic, 125±126

semantic analysis of, 201±204, 401 n.31, buttressing, 37±38, 117, 121, 122±127,
n.32 138, 158, 199
See also intrusion Grice's analysis of, 19, 199
compatible alternates, 100, 102, 103 paratactic, 124±126, 147±148
competence, vs. performance, 172, 397 n.5 reduction, 148±149, 394 n.70
pragmatic, 22 See also connectives, and, or, but
complementary distribution connectedness, see cohesion
of pronouns and re¯exives, 283, 288, 290, connectives, sentential, 18±20, 84, 91, 125,
308 153±154, 386 n.53. See also
breakdown of, 283±284, 308±326, 414 conjunction, disjunction, conditionals
n.44 consistency, epistemic, 76±77
complementary interpretations, 34, 137, 150 of inferences, 77, 155±164
completions and expansions. See and common ground, 49±52, 162±163,
impliciture 299
complex sentences, and implicature. See not the only constraint on implicatures,
projection problem 52, 163
componential analysis, of meaning. See constituency, 303. See also c-command,
lexical decomposition noncon®gurational languages
components of linguistic theory, 166±167, Construction Grammar, 395 n.71
171, 187, 243±244, 365, 368, 405 n.53. constructions. See intrusion
See also levels, modules context
compositionality, bottom-up processing of, and anaphoric resolution, 271
245±247 bias in and default interpretations, 25±26
vs. noncompositional fusion, 376 change, and dialogue, 49±50
pragmatic intrusion in, 189, 199, 205±206, and defeasibility (see defeasibility)
213±214, 218, 233±234, 245±246 dependence, and indexicality, 191±192,
See also intrusion 230, 242±243
compound nominals, 61, 117, 147, 245 incremental model of, 50±54, 162±164, 193
comprehension, 259 normal vs. special, 16
asymmetry with production, 28, 382 n.20 types, 26
conceptual structure. See cognition, See also common ground
language of thought contextuals, 191
concessives, 82 contraction, 69±70
conditional perfection, 37, 58, 117, 119± contradiction (self-contradictory utterances),
120, 159, 388 n.15, 392 n.49 57, 199±200, 202, 203±204, 208, 209,
conditionals 214±216, 233, 253, 404 n.49
Grice's analysis of, 19±20 contradictories, vs. contraries, 63±65, 83±
implicated, 125±126, 393 n.56 86, 127±129, 130±132, 142±145, 222.
implicatures of, 19, 20, 37, 50, 58, 108± See also negation
109, 158, 296±297 contrast, of expressions in salient
incremental processing of, 126, 163, 393 oppositions, 32, 36, 76±111
n.57 of forms, 135±142
inferred vs. coded, 125±126 vs. gaps, 285 (see also anaphora, articles,
intrusion of pragmatic inference in, 205± pronouns)
210, 251±253 in morphological derivation, 139
Kamp's analysis of, 163, 205 principle of, 140, 150
and material conditional, 20 in child language, 140
order of antecedent vs. consequent, 126 of reference (see reference)
and presupposition projection, 252 in subjective meanings, 314±317, 319±321,
semantics of, 20, 205, 208±209 327, 407 n.4, 415 n.54
suspension of inferences by, 81, 158, 252 See also M-implicature, Q-implicature,
See also connectives, if, projection, scales
suspension control, grammatical theory of, 363±364,
conjunction 408 n.4
asymmetric, 122±127, 154 ¯ow of, in processing, 243±245
462 Subject Index

convention, and meaning, xix vs. bound variables, 267, 292±293, 301
Lewis's theory of, 29 and cohesion, 276, 318
of use, vs. convention of language, 23 and I-implicature, 181, 272, 273, 277,
See also standardization, short-circuited 287±288, 331
inference and I-Principle, 273±276
conventional implicature, 13±14, 166, 170, inferring, 273±276
213 and logophoricity, 312, 314±317, 346, 356
conventionalization and M-implicature, 331±332, 347, 354
of implicatures, 90, 134, 159, 237, 263± preference for, 273±276
265 See anaphora, Binding Conditions,
of uses, 23 disjoint reference, variables
See also grammaticalization, short- creoles, 338±340. See also pidgin and creole
circuited inference languages
conversation, 14, 23, 113±114 Chinook Jargon, 339
analysis of, 113 creolization process, 338±340
conversational implicatures Guianese, 419 n.67
actual vs. potential, 387 n.12, 395 n.73 Haitian, 338±339
calculability of, 15, 137, 141, 214 KiNubi, 339
cancellability of, 15 (see also cancellation) Mauritian, 339
clausal (see clausal implicatures) Palenquero, 339
and communication, 21 cross-linguistic generalizations, 69, 118,
conventionalization of (see 123, 134, 142, 146, 365, 370
conventionalization) cultural variation, xii
vs. conventional implicature, 13±14
de®nition of, 15, 171, 194 deductive inference, 42±43, 54
and ®gures of speech (see ®gures of and implicatures, 42, 55±57, 257, 383 n.31
speech) and sub-propositional representations, 257
generalized vs. particularized, 16±21 (see See Relevance (Sperber-Wilson theory of )
also those terms and GCIs, PCIs) deep or underlying structure, 129
language change (see language change, default
grammaticalization) inferences and grammar, 169
and logical connectives (see connectives) interpretation, xi, 11±12, 15, 21, 22¨
and logical form (see logical form) and metalinguistic reasoning, 6, 40±41,
and metaphor (see metaphor, ®gures of 104, 371
speech) logics, 45, 46±49, 52, 53±54
nondetachability of, 15, 99 modeling GCIs, 46±48
particularized (see particularized reasoning, 12, 42±54, 259
conversational implicature, PCI) basis in communication, 53±54, 372, 383
and presupposition, 13, 111, 252 n.26, 384 n.38, n.39
projection of (see projection) vs. default contexts, 25±26
typology of, 13±14, 16±17, 35±42, 373, vs. default logics, 54
406 n.59 derivative from implicature, 53±54
universality of, 15, 423 n.96 vs. ¯outs, 216
See also generalized conversational GCIs as prototype, 53±54, 372
implicature, I-implicature, rules, 47, 51
M-implicature, Q-implicature, speci®c rules block general ones, 161
Relevance, etc. from sub-propositions, 259
cooperative principle, 112, 214 types of, 42±49
de®ned, 14 defeasibility, 5, 42±54, 88, 123, 155±164, 192
and nonverbal activity, 112 and binding, 295±296, 299±303, 326
coordination problems, 29±30, 53±54, 258± de®nition of, 42
259, 371 di¨erential, 78
coreference, 39, 115, 117, 149, 177, 181, and extralinguistic context, see
225, 250, 267±276, 288, 292±295 cancellation (by common ground)
`accidental', 293 as ®ltering by consistency, 163
asserted vs. presupposed, 412 n.32 and incremental context, 49±54, 162±164
463 Subject Index

as litmus test of grammar vs. pragmatics, grammaticalization, language change,

268, 278±279, 364±365, 408 n.6 semantic change
models of, 45 diagnostics, for kinds of GCI, 40±41
nonmonotonicity and, 42 for scalar implicatures, 81±82, 84±86
and semantics vs. pragmatics, 192 disambiguation, 172, 232
and syntax vs. pragmatics, 265, 268, 278± implicatures and, 174±177
280, 294, 295±296, 300±303, 318, 327, postponed, 189, 232
363±365, 408 n.6 prosody and, 174
See also cancellability, context, projection and `what is said', 171
de®nite article, 60, 62±63, 92, 182±183, discourse, and anaphora, 267±273. See also
217±235, 249±250 conversation
anaphoric use, 93 discourse referent, 63, 179, 180, 248±249.
in bridging inference, 62±63, 127±127, 182 See also anaphora
pragmatic resolution of, 180±183, 217±230 discourse representation theory (DRT), 26,
scalar implicature and, 220 122, 163, 167, 172, 193, 218, 247, 248±
uniqueness condition, 92, 219, 250, 385 256, 406 n.57
n.49, 390 n.29, 399 n.13 implicatures in, 248±256, 406 n.59
See also articles, bridging pragmatic vs. semantic conditions in, 253±
de®nite descriptions, 60±63, 182±183, 217± 256
235 disjoint reference, 224±225, 267, 277±280,
de®nite expressions, 61±63 281±282, 295±303, 305, 310
as variables, 182 characterized, 277, 281, 409 n.11
de®niteness, 63, 248±250. See also articles, defeasibility and, 182, 277, 295±296, 299±
de®nite article 303
deictic determiners, 93±94, 390 n.29. See disjoint reference presumption (DRP),
also deixis 280, 328±331, 329±331, 342, 346±347,
deixis, 8, 93±94, 177±180 353
and anaphora, 179, 347 de®ned, 328±330
`exotic' parameters, 177 I-Principle and, 330±331, 416 n.57
and de®nite descriptions, 217 I-implicature and, 280, 328±331
and indices, 246 M-implicature and, 277, 288±289, 301,
and noncancellability, 242, 247 305, 331
and point of view, 347 pragmatic nature of, 319, 326, 343
pragmatic resolution of, 178±180 Q-implicature and, 277±280, 302, 305,
preemption over nondeictic expressions, 317, 318±319
99, 180 switch-reference and, 318±319
and the `said', 171 universally not grammatically stipulated,
scalar implicatures of, 94 319, 326, 365, 414 n.47, n.48
and truth-conditions See also Binding Conditions, coreference,
tests for deictic expressions, 242 reference
and visibility-conditions, 177 disjoint reference presumption (DRP), 323,
See also indexicals 329±331, 346±347, 353
demonstratives. See deictic determiners disjunction, 19, 67, 81, 84, 108±109, 125,
denial. See negation, metalinguistic 154, 158, 209, 252
denotation, 137, 144, 245±246 clausal implicatures of, 108±109, 154
shifted, 246 exclusive vs. inclusive, 108, 154
See also intension vs. extension, reference Grice's analysis, 19
deontic modality, 84 scalar implicatures of, 108±109, 154
derivation. See morphology and suspension of implicatures, 81, 159
design perspective on human distributional constraints, 261±262
communication, 27±35 division of pragmatic labor (Horn), 38,
detachability, 15, 99 137±155, 157, 277, 285, 333, 364, 365,
determiners, 91±95. See also articles 418 n.61
diachrony, and pragmatics, 69±70, 103, Dogrib, 415 n.47
139, 185, 262±264, 268, 279±280, 338± domain
342, 344, 347±352, 419 n.73. See also for anaphoric linkage, 271
464 Subject Index

domain (cont.) subjective meaning of, 320, 413 n.40, 415

in discourse, 271, 321±322 n.54
grammatical, 270±271 enrichment rule, 114. See also I-Principle
binding domains, 281, 282, 286, 292±293, entailment
309, 312±313, 326, 354 and cancellation of implicatures, 50, 56±
in acquisition, 361, 422 n.93 57, 155, 299
and long-distance anaphors, 310±317, 345 and de®nition of scales (see scales)
circumscription of, 48, 275±276 and di¨erential informativeness, 36, 76,
of discourse, 274±275, 383 n.27, 385 n.50 79, 108±109, 273±274, 383 n.27
minimizing, 274±276, 410 n.16 and implicature, 196, 253
for point of view, 321±323, 354, 357, 414 nonentailing constructions, 108±109, 296±
n.47 297
double negation, 38, 142±145. See also and partial ordering, 105, 390 n.35
negation scales (see Horn-scales, scales)
double processing. See psycholinguistics and semantic representations, 191, 194,
Dravidian, 352 203, 241, 243
DRS (discourse representation structure), and square of oppositions, 65, 385 n.50
122 in theory of meaning, 168
DRT. See discourse representation theory of what is said by what is implicated, 121,
duals, logical, 67, 84 392 n.51
Dutch, 311, 342, 350, 416 n.59, 419 n.71± enthymematic argument, 56, 283 n.31
72 epistemic modality, 84, 131±132
Dyari, 152 epistemic modi®cation
Dyirbal, 152 of implicatures, 76±77, 387 n.6, n.7, n.10,
n.11, 391 n.40
E-vertex of square of oppositions, 65, 82, and particularized strengthening, 78
83±84 variable strength of, 77±78, 108±109, 387
economy, speaker's, 6, 112±113 n.11, 396 n.77, 411 n.23
redundancy avoided in lexicon, 69±70, See also knowledge, belief
319, 326, 369±370 ergative/absolutive, 304, 335±336
See also least-e¨ort principle, Zipf ethnography of speaking, 23
ellipsis, 183±184, 189, 293, 376, 395 n.75 euphemism, 119, 128, 134, 138, 239, 264
and `what is said', 172, 183±184 even
implicatures of elided material, 184 events, 122±126, 148. See also conjunction
elsewhere condition. See blocking evidentials, 237, 315
empathy, 312 evolution of language, 28, 381 n.16, 382
emphatic markers, 304, 312, 331, 332±334, n.17
348, 350, 416 n.58, 420 n.81 Ewe, 420 n.79±80
di¨erent types of emphasis, 333, 351 exclusive vs. inclusive disjunction. See
intensi®cation, 333, 350±351, 415 n.51, disjunction
416 n.58 existential
encyclopedic knowledge import, 385 n.50
and cancellation of implicatures, 49±52, quanti®cation, 92, 200, 274±275
162±163, 299 statements, 93, 274
and disambiguation, 174, 176, 271 See also some
and implicatures (see also common explanation in linguistic theory. See levels,
ground, context) of linguistic theory
English explicature, 24±25, 167, 190, 194±198, 211,
binding patterns in, 285±298, 319±324 238±239, 254, 392 n.46, 392 n.51, 400
dialects in, 415 n.51 n.22
historical change in, 323, 340±342, 362 de®nitions of, 194±197, 399 n.20
Old, 335, 340±342, 349±350, 419 n.68± explicitness, pragmatic maxim of, 294,
70 299
Middle, 330, 341 exploitation, of the maxims. See maxims,
re¯exives, long-distance, 319±324, 415 exploitation of
n.53 extension. See intension, reference
465 Subject Index

factives. See verbs as constraints on lexicalization, 64±71,

fallacy, of a½rming the consequent, 44 143, 319, 369
in human reasoning, 384 n.39 cross-linguistic generalizations and, 69,
falsi®ability, 274 118, 123, 134, 142, 351, 365, 370 (see
felicity conditions, 23, 60±61, 166, 249±251, also cross-linguistic)
385 n.48 defeasibility of (see defeasibility)
®gures of speech, or tropes, 144±145, 163, default nature of, 16±17, 45±49, 162
210, 234±235, 238, 390 n.32. See also de®ned, 16, 162
exploitation, ¯outing, irony, metaphor, diachronic relation to PCIs, 263
metonymy diagnostics for types, 40±42, 81±82, 83±84
Fijian, 335, 336 existence doubted, 380 n.5, n.6
®le-change semantics, 167, 193, 248±249 as explicatures, 196±197, 373, 385 n.41,
®lters. See projection 392 n.46
Finnish, 97 generation and ®ltering of, 162
¯outing, of the maxims of conversation. See as gratuitous inferences, 59, 61
®gures of speech, maxims and idiomatic uses of language, 22±24,
focus, 94, 339, 341, 348, 351 134, 141, 367
force, illocutionary. See illocutionary force inconsistent, 39±40, 50±51, 77, 122 (see
form, linguistic also projection)
marked vs. unmarked (see markedness) inside scope of logical operators, 172
as trigger of inferences, 136 interactions between, 39±40, 153±164 (see
formalization, of pragmatic theory, 1, 49, also projection)
52, 54, 78, 377, 423 n.96 intrusion into semantic content (see
formulae, ritual, 23 intrusion)
frequency, of expressions, 6 joint e¨ect of, 153±155
and markedness, 136 and linguistic theory, 374±377
See also markedness mistaken for coded content, 18, 20, 68±69,
Frisian, 329 374
functionalism, 266, 363, 368 (attempted) reduction of, to nonce-
functional load, 132±133 inference, 54±71, 104±108, 123, 134,
function words, 90±98 367, 386 n.54
predictive power of, 59, 69±70, 289
game theory, 6, 29 presumptive quality, 123, 127
gapping, 124, 126, 393 n.59. See also projection problem (see projection)
conjunction psycholinguistic plausibility of, 162, 259,
`Gazdar's bucket', 49±50, 162 370
generality, semantic, 37 regularity of, 68, 134
narrowing of, 37, 184±186, 232 relative priorities between, 39±40, 50±54
GCIs (see also projection problem)
introduced, 11±21 rule-like, 47±48, 55
vs. PCI, 16±21 scalar, 58 (see also Q-implicature, scales)
See generalized conversational implicature from semantic representations, not truth-
generalized conversational implicature conditions, 380 n.2
(GCI), 11±21, 372±373 and semantics (see semantics/pragmatics
borderline with PCIs, 102, 104±108, 396 interface)
n.82 speaker's vs. addressee's perspective (see
calculation of, 141 speaker)
cancellation of, 49±54, 56±57 and square of opposition, 68±71
by background assumptions, 50 stable patterns of, 64±71
by entailments, 50, 56±57, 155, 299 strength of, 411 n.23
by inconsistent higher-ranked subliminal nature of, 17±18, 20, 69, 373
implicatures, 50±51, 77, 122, 155±164 tagging of, for cancellation, 163
by relevance, 51±52, 88, 163±164 `tra½c rules' between (see projection
as (supposed) category mistake, 55, 372±373 problem)
close relation to linguistic structure, 17, and truth-conditions, 167
22, 368±370 typology of, 35±42, 170
466 Subject Index

generalized conversational implicature role in recovering speaker's intentions, 30±

(cont.) 35, 53±54, 257±258, 371
See also conversational implicature, `the marked message', 33
GCI, I-implicature, M-implicature, `the not said', 31
Q-inference `the simply described', 32
generative grammar, 265. See also Binding See also I-Principle, M-Principle,
Theory, Government and Binding, Q-Principle
grammar, Minimalist program, syntax Hindi, 125
generative semantics, 237, 243, 333, 394 n.68 Horn's scales. See scales
German, 342, 349±351, 364 maxims, 40±41, 76, 136±137
Germanic languages, 311, 315, 330, 332, honori®cs, 264, 408 n.5
335, 342, 416 n.59, 419 n.73, n.75 and agreement, 407 n.2
gesture, 179 hyponyms, 101±103. See also
givenness hierarchy, 94 autohyponyms
goals, of interlocutors
and relevance, 17, 51±52, 163±164 (see I-vertex of Square of Oppositions, 65, 82,
also plans) 83±86
Gothic, 311 pragmatic relation to O-vertex, 68±71
governing domain, see domain Icelandic, 311, 315, 316, 414 n.44, 416 n.59
Government and Binding Theory (GB), iconicity, xvi, xix, 6, 75, 135±136, 142, 146,
265, 270, 281±284, 304 150, 394 n.65
grammar identity statements, 299±300, 412 n.32
and conventionalized implicature, 264, 266 idioms
overall models of, see linguistic theory generative theory of, 24
and pragmatics, 261±265, 362±365 and three-leveled theory of meaning, 24±
as `frozen pragmatics', 268 27, 54, 134, 367
See also syntax if, 19±20, 36±7, 91
grammaticalization, 27, 69±70, 74, 124, and suspension of implicatures, 81, 252
263±264, 268, 279, 339, 341±342, 347± (see also conditionals, connectives)
350, 362, 364 I-implicatures, 112±134
Greek, 285, 416 n.59 anaphora and, 115, 181±182, 217, 269±
Grice's 272, 287±290
circle, 172±198, 225, 236±243, 399 n.16 bridging (see bridging)
de®ned, 186 case, and, 186
and two pragmatics, 187 (see also of causatives, 141±142, 297
semantics/pragmatic interface) cohesion and, 276, 318 (see also bridging,
maxims (see maxims of conversation) cohesion, conjunction, coreference)
program, 12±21 compounds, 146±147, 186 (see also
theory of meaning, 12±13 (see also compounds, nominal)
intention, meaning nm ) conditional perfection (see conditional
Gumbaynggir, 336 perfection)
Guugu Yimithirr, 90, 125, 285, 303±305, conjunction buttressing (see conjunction)
334, 364, 393 n.55, 412 n.34, 413 n.38, from contradictories to contraries, 127±
419 n.770 134
contrast with Q-implicatures, 116, 119
Haitian creole, 338±339 coreference, 149, 177, 181 (see also
Head-driven Phrase Structure Grammar coreference)
(HPSG), 306±307, 325±326 and default logics, 48±49, 53
Hebrew, 335, 350 diagnostics for, 40±41, 119, 133±134
heuristics, 4±6, 31±35 and ellipsis, 183±184, 395 n.72
and bottleneck on speech encoding, 28±35, and gaps (see zero-)
382 n.18 grammaticalization of (see
multiplication of message content, 31, 34 grammaticalization)
mutual assumption of, 34, 53±54, 73, 258± heterogeneity of, 118
259, 371 intrusion of (see intrusion)
convert probabilities into certainties in and language change, 347±348, 355±359,
communication, 53±54, 258 370
467 Subject Index

and narrowing of meaning, 139, 185 vs. implicatures, 191±192, 242±243, 405
negative strengthening (see negative n.52
strengthening, negation) intrusion and, 192, 242
introduce semantic relations, 116 nondefeasible nature of, 242, 247
and M-implicatures, 137±155 (see also require pragmatic resolution, 177±180
division of pragmatic labor) and rigid designation, 242, 246, 405 n.52
and morphology, 143, 146±149 and scope of negation, 405 n.52
and Neg-Raising, 131±134 (see also Neg- and semantic interpretation, 177±180, 399
Raising) n.12
not independently glossable, 116, 120± standard view challenged, 177±180
121 See also deixis
and parataxis, 124±126, 147±148, 393 n.58 indirect speech acts, 23, 141, 237, 408 n.5
(see also parataxis) indirection, 239. See also euphemism,
possessives, 118, 146, 186, 207, 223 (see politeness
also possessives) Indo-European, proto, 417 n.59, 419 n.73,
and stereotypes, 114±115, 119, 138, 147, n.75
152, 223, 328±329, 333, 342, 370 (see induction. See inductive inference
also stereotypes) inductive inference, 42±43, 257, 384 n.32
tension with Q-implicatures, 116, 121, nonmonotonicity of, 43
128±129, 133±134, 156 not basis of implicatures, 45±46
and zero-morphs or gaps, 146±149, inference
285, 336±337, 364, 382 n.22, 411 abductive (see abduction)
n.26 ampliative, 258
See also anaphora, articles, pronouns and ampli®cation of message content, 30±
illocutionary force, 23, 49, 381 n.11, 392 35
n.50 to best interpretation, 37, 44, 177, 257±
vs. perlocutionary force, 23 259
implicature and circumscription, 48, 275±276, 384
vs. explicature and impliciture, 167, 194± n.36
198, 257±258, 392 n.46, 400 n.21, n.22, and default logics, 42±49
n.23, n.24 defeasible, 42±54, 162±163
kinds of, 13±14, 40±41 deductive (see deductive inference)
additive vs. subtractive, 406 n.59 inconsistency between inferences, 39±40,
phrasing of, 120±121, 196, 198 48, 50±54, 56±57, 77, 122
See also conventional implicature, inductive (see induction)
conversational implicature, generalized kinds of, 13±14, 42±54, 257±259
implicature, particularized implicature nonmonotonic (see nonmonotonic
implicitness in communication. See inference)
I-Principle, impliciture and processing e¨ort, 55, 57±58
impliciture, 24±25, 167, 194±198, 392 n.46, relative cheapness of, 29, 55, 382 n.22
400 n.21 and rhetoric, 382 n.22
de®ned, 194±195, 197, 400 n.22 and semantic representations, 241, 257±
incorporation of negation. See negation 259
inde®nite article, 91±93 from sub-propositions, 257
and nonuniqueness implicature, 92 information
See also de®nite article asymmetry of
inde®niteness. See de®niteness balance of, 112
indeterminacy, of semantic content, 89, increase by heuristics, 30±35
375±376. See also semantic measures of, 31, 34
representation, underspeci®cation of semantic, 31±35
of syntax, 376 theory of, 5±6, 383 n.24, n.26
indexicals, 171±172, 242±243, 246 informational strength, 76, 102
covert `contextuals', 177, 185, 191±192, informativeness, 14, 273±276, 287, 409n.14
242 Bar-Hillel and Carnap's theory, 115, 273±
diagnostics for, 242 275, 278, 299, 382 n.23, 410 n.15
enlarged theory of, 191±192, 230, 242± de®ned, 115, 273±274
243, 245±247 di¨erential, 36, 76, 79, 100
468 Subject Index

informativeness (cont.) default inference predicts, 259

and entailment (see entailment) de®ned, 168
and existential quanti®cation, 274 ®gures of speech and, 210, 238±239
information-theoretic, 383 n.24, n.26, 409 GCIs vs. PCIs and, 169, 236, 385 n.41
n.14 Grice's views on, 260
Popperian, 273±275, 278 I-inferences and, 199, 201, 206±207, 210,
Principle of, 38, 114 (see also I-principle) 221±223, 251
semantic vs. information theoretic, 409± as indexicality, 192
410 n.14 kinds of, 239
types of, 273±275, 409 n.14, 410 n.16, M-inferences and, 201, 207±208, 223±224
n.18, n.19 Q-inferences and, 200±201, 205±206, 210±
See also information, Q-implicature, 211, 213, 219±221, 249±251
Q-principle, scales and theory of meaning (see meaning)
innate knowledge, of grammar. See See also discourse representation theory,
nativism, universals pragmatics, situation semantics
input/output relations of components. See intrusive constructions. See intrusion
theory of meaning, architecture of Inuktitut, 142
intensi®cation I-Principle, 112±134
and M-implicature, 333, 348, 351. See also and coreference, 115, 274±276 (see also
emphatic markers coreference)
and reduplication, 150±153 and ellipsis, 183
intension vs. extension, 217±218, 225, 245± de®nition, 114±115
246, 391 n.45. See also character, and disjoint reference, 328±331 (see also
reference disjoint reference)
intensional vs. extensional contexts, 401 as heuristic, 32, 37, 48
n.30 licenses inferences, 118
intention, communicative, 4, 12±13 speaker's maxim of minimization, 114, 126
recognition of, 13, 29±30, 52±53, 258±259, recipient's enrichment rule, 114
372 and referential parsimony, 115
logical problem of, 30, 379 n.3 relation to Grice's maxims, 37, 74±75, 112
re¯exivity of, 379 n.4 irony, 163, 216, 234±235, 239±240
regress of, 379 n.4, 380 n.1 isomorphism
See also meaning-nn, m-intention of I-implicatures and semantic
interaction, social, 21 representations, 115, 116, 391 n.45
interface, semantics/pragmatics. See Italian, 420 n.78
semantic/pragmatic interface
of syntax and semantics (see syntax) Japanese, 102, 142, 312, 314, 315, 351, 419
See also meaning, semantics n.76, 421 n.84±85
intolerant vs. tolerant quanti®ers. See Jiwarli, 336
quanti®ers
intonation, 212, 391 n.39. See also prosody Kilivila, 337
`introverted' predicates, 329±330, 342, 419 kinship terms, 102
n.74 know, 80, 133, 208
intrusion, of pragmatics into semantic repre- knowledge
sentations, 122, 164, 189±190, 192, speaker's knowledge and implicatures, 76±
198±234, 236±237, 238, 248±253, 373 77
anaphora, and, 251, 272 of use of language, 371
as coercions, 246±247 See also encyclopedic, epistemic, mutual
constructions, intrusive, 198±217, 232, knowledge
239, 251±256, 259 Korean, 285, 315
de®ned, 198 Kwakiutl (Kwakwala), 177
comparatives, 199±205
conditionals, 205±210, 252±253 language change, 139, 185, 262±264, 339±
negation, 210±213, 253±256 342, 362
referring expressions, 217±236, 238±239, I-inferences as main engine of, 370, 394
249±251 n.66
469 Subject Index

implicature and, 263±264, 370 vs. implicated content, 214±215, 240 (see
See also grammaticalization, semantic also intrusion)
change vs. indirect speech acts (see indirect speech
language learning. See acquisition acts)
language of thought, 7, 242, 372 vs. nonliteral uses, 23, 130, 132, 187, 197,
Latin, 186, 350, 420 n.78 204, 214±217, 226, 229, 232, 238, 240,
learnability, of pragmatics vs. universal 402 n.38 (see also maxims, exploitation
grammar, 265, 279±280, 360, 361±363 of; reference, speaker-)
subset principle and, 361 litotes, 142±145. See also negation
least-e¨ort principle (Zipf ), 6, 112±113, `local pragmatics', 25, 167, 399 n.17
135, 277, 319 logic
law of abbreviation, 115 autoepistemic, 48
law of economic versatility, 115 deductive, see deductive inference
levels, of linguistic theory, 8, 243±244, 367± inductive, 43, 44
368. See also components of linguistic modal, 48
theory See also inference, reasoning
and the burden of explanation, 374±377 logical form (LF), 8, 190±191, 195, 211,
of meaning (see three-leveled theory of 231, 238, 240±241, 249, 257, 292±293,
meaning) 295, 375±376
lexical decomposition, 81, 140 Chomskyan LF, 190, 240, 265, 281, 405
lexical items n.54
constraints on (see lexicalization) and implicature, 171, 238
contrast sets of (see contrast) underspeci®cation and, 241, 249, 257±258,
doublets, 138±139 375±376, 399 n.17
meaning underspeci®ed, 376 See also semantic representation
vs. periphrasis, 140±142, 400 n.26 logical priority, of components in linguistic
and register, 138 theory, 166, 172
in semantic ®elds, 101±104 logophoric pronoun. See pronoun
taxonomies, 101±103 logophoricity, 284, 310±317, 356
compatibles, 102, 103 characteristics of, 314±315 (see also
incompatibles, 105 pronoun)
See also monosemy London school, 190, 256±257. See also
lexicalization Relevance (Sperber-Wilson theory of )
constraints on scales (see scales)
of negation (see negation) M-implicatures, 38±39, 89, 135±155
pragmatic constraints on, 12, 69±71, 91, and alternates, 136±137
143, 319, 369, 386 n.53, n.54 and anaphora, 181, 272, 277, 288±289,
lexicon 301±302, 305, 331±332, 333, 343±344
metalinguistic knowledge about, 6±7, 40± and cases vs. postpositions, 97
41, 119, 371 and compounds, 147
pragmatic constraints on (see and conjunction vs. parataxis, 126, 147±
lexicalization) 148, 393 n.55
structure of, 79±108 and disjoint reference (see disjoint
adjectives, 86±87, 100 reference)
closed-class items, 90±98 and double marking, 146, 357, 416 n.59,
logical operators, 82±86 420 n.78, 422 n.91
marked/unmarked pairs, 138±140 and double negation (see negation)
numerals (see number words) and derivation, 139±140
taxonomies, 101±103 diagnostics for, 40±41
verbs, 87, 98 (see also verbs) form-basis of, 34, 38, 40, 126, 136
See lexicalization, markedness, scales intrusion into semantics (see intrusion)
linguistic theory. See components, levels, and lexical doublets, 138±140
meaning markedness and, 38, 136±155, 277, 288,
literal meaning 326, 333±334, 337, 347, 369 (see also
vs. ®gurative use, 214±215, 402 n.38 (see markedness)
also ®gures of speech) and negation, 142±145
470 Subject Index

M-implicatures (cont.) meaning

and periphrasis, 140±146 composite nature of (see hybrid theory of
and possessives, 146, 147 below)
and prepositions, 96, 147 conventional (coded), 170
and reduplication, 150, 152±153 theory of, 7±9, 166, 374±377
and re¯exives, 329, 331±334, 337, 341, 347 architecture of, 7±9, 166±168, 188, 243±
and repetition, 149±153 244
take priority over I-implicatures, 39±40, with double pragmatics, 187, 232±236
137±138 with double semantics, 232±234, 236
and zero-morphs, 146±149 input/output structure, 166, 171±172,
See also division of pragmatic labor, 187, 242, 243±244, 374±375
Manner, markedness received view of, 171±173
M-Principle, 38, 91, 135±155 two vs. three levels in, 22±27, 54±63, 73,
as heuristic, 33, 38, 136 367
related to Grice's maxims, 38, 75 hybrid theory of, 7¨, 13±14, 21, 166, 397
relation to Relevance, 135±136 n.3
See also Manner literal (see literal meaning)
Malagasy, 423 n.96 nonnatural, 12 (see also meaningnn)
Manner, maxim of picture theory of, 4
Grice's submaxims, 14, 19, 38, 135±136 sketchy nature of encoded-, 2±8, 375±
reduction of, 135±136 377
Reinhart's, 294 speaker-, 22, 170 (see also speaker-
See also markedness, M-Principle meaning)
Manner implicatures, 135±136. See also M- timeless, 170
implicatures See also Discourse Representation Theory,
markedness pragmatics, semantic representation,
concept of, 115, 137 semantics, Situation semantics,
formal vs. semantic, 137 utterance-meaning
heuristic based on, 33±34 meaning nn , 12, 379 n.4, 398 n.7. See also
and Horn's division of pragmatic labor, 6, M-intention
26, 38, 136±139, 144, 277, 285, 364 metalinguistic
and iconicity, 136, 142 character of I- and M-implicatures, 40±41,
and Manner, 91 119
in morphology, 139±140, 152±153 reasoning and heuristics, 6±7, 371
and negation, 70, 142±144 negation (see negation)
and periphrasis, 39, 140±142 metaphor, and intrusion in semantics, 239
and phonetics and prosody, 332±333, 395 and implicature, 163
n.72 metarules, 81
and prolixity, 6, 14, 38, 79±80, 114, 135± metonyms, 236±237, 238, 245±246
136 minimal expressions, 37, 62, 113±115, 124,
and reference (see reference) 126, 268
as trigger of M-implicature, 136 de®nition of, 115
See also M-Principle, privative opposition See also I-principle, maxim of
maxims of conversation minimization, semantic generality
clash between (see tension between below) minimal governing category. See domains,
exploitation (¯outing) of, 112, 172, 214± Binding Conditions
216, 231±232, 234±235, 238±239 Minimalist program, 265, 281
Grice's, 14¨, 35, 37, 38, 386 n.2 minimization, maxim of, 113±115, 277. See
as heuristics, 31±39, 136 also I-principle
number of, 55, 74±75, 164 m-intention, 13, 29, 114, 391 n.43. See also
observation vs. ¯outing of, 15, 112 meaning nn
reduction of, 55, 74±75 mirror-image inference, 290, 295, 298, 326,
tension between, 55, 116, 121, 133±134, 412 n.29. See complementary
156±158 (see also projection) distribution
See also I-Principle, Manner and modality, 83±86, 140, 154, 391 n.38
Relevance, M-Principle, Quality, modal logics, 48, 387 n.6, n.9, 388 n.13
Quantity, Q-principle model-theoretic semantics. See semantics
471 Subject Index

modularity properties of, 254

Fodorean, 187, 190, 244, 372 scope over Q- and M-implicatures, 41,
vs. linguistic component, 244, 372, 397 197, 210±213, 253±256
n.4, 405 n.53 psycholinguistic complexity of, 70, 81,
in grammar, 21, 166, 261, 397 n.4 145, 388 n.19, 392 n.49
in meaning, 21, 166, 171, 187, 192±193, Q-implicatures and (see scales, square of
192, 243±244 oppositions)
and pragmatics, 371±372, 405 n.53 raising (see Neg-Raising)
representational, 424 n.4 and scale reversal, 64, 388 n.17
monosemy, and implicature, 18, 20, 165, scope of, and lexicalization, 70
184 strengthening by I-implicature, 127±134,
monotonicity 143
and inferences, 42±55, 56±57, 192, 243. sub-contraries, 65, 83, 132
See also nonmonotonicity See also square of oppositions
mood, 315 negative
Moore's paradox, 209, 386 n.1, 402 n.35 character of Q- and M-inferences, 40, 119
morphology, derivational, 139±140 expressions and M-implicatures, 142±145
causative, 142 polarity items, 134, 212, 237, 262
negative, 142±143 scales and Q-implicature, 36, 64±68, 80±
vs. periphrasis, 146 82, 128±129, 254±255
reduplication, 152±153 strengthening by I-implicature, 117, 127±
mutual knowledge. See common ground, 134, 143, 222
context, heuristics, salience See also Q-implicatures, scales, square of
oppositions
narrowing, of meaning Neg-Raising (NR), 129±134, 159, 237, 262
Q-based vs. I-based, 185±186 neo-Gricean pragmatics, xv, 353, 381 n.9.
nativism, 338, 344, 361, 363 See also radical pragmatics
necessity vs. possibility. See modality neurocognition and language, 28
negation nonce interpretations, 12, 24, 25, 54±63,
a½xal, 69, 134, 142±143 164
and blocking of implicatures, 64, 80±82, inability to explain stable inference
162, 252±256 patterns, 71, 134, 386 n.54
contraction of, 69±70 problem of inconsistent, 164
of contradictories, 144±145 noncon®gurational
of contraries, 145 constraints on binding, 303±307, 325
contraries vs. contradictories, 64±65, 83± languages, 297, 303±305
86, 127±129, 130±132, 142±145, 222 nondetachability, 99
and contrast sets, 103 de®ned, 15
conventionalization of implicatures of, 134 nonmonotonic reasoning, 42±54, 57, 192,
denials (see metalinguistic) 275±276, 384 n.38, n.39
double, 38±39, 127, 142±145 types of, 42±49
I-implicatures of, 127±134, 142, 222 nonnatural meaning, 12. See also meaning,
and implicatures vs. presuppositions, 252± meaningnn, m-intention
256 normal expectations, 16, 32±33. See also
implicit, 213 stereotypes
incorporation of, 69±70, 82±83, 130±131, Norwegian, 310, 311, 412 n.28, 416 n.59
142±143, 212 noun phrases (NP)
lexicalization and, 69±70 lexical (R-expressions), 270, 281, 285,
litotes, 127, 142±145 286±287, 292±296, 298±303
M-implicatures of, 127, 142±145 scale of reduction of, 270, 285, 288, 305,
metalinguistic, 210±213, 253±256 411 n.27
as contradictions, 214 types of, 277±286, 292, 300±301, 304, 326,
de®nitions of, 211±212 409 n.12
denials and retorts, 212, 253±256 See also Binding Conditions, pronouns,
kinds of, 406 n.61 R-expressions
as logical negation, 211±212, 253±254 NP. See noun phrases
nonlogical type of, 212, 254 NR. See Neg-Raising
472 Subject Index

number words, 25, 58, 87±90, 178, 249± See also PCIs, conversational implicature,
250, 389 n.22 nonce interpretations, Relevance,
ambiguity analysis of, 88±89 speaker-meaning
`at least' interpretation of, 88, 250±251, performance. See competence vs.
406n.57, 389 n.26, 406 n.57 performance
vs. cardinal numbers, 88 periphrasis, 37, 140±146
`exactly' interpretation of, 87±90, 175, perlocutionary acts/e¨ects. See
204, 206, 219, 406 n.58 illocutionary acts
PCI analysis of, 403 n.40 permissions, 49
semantic content of, 88, 250, 389 n.22, person, ®rst/second vs. third, in anaphora,
406n.57 387 n.7, 413 n.35, 413 n.40, 415 n.54
special properties of, 89±90 perspective, 312, 314, 320±321, 346±347,
Nyawaygi, 335 354
perspicuity, 14, 38, 135
oblique arguments, 282 phonetics and implicature. See markedness,
Obstinate Theorist, 187, 215, 217, 225, 227, prosody
230±236, 240, 402 n.39 phonology, 8
Occam's razor, 20 philosophy of language, 12±20, 23, 24, 64,
on, 33, 34 68, 84, 87, 217, 226±231
Ono, 148 pidgin and creole languages, 338±340
oppositions. See contrast pirots, 29, 382 n.21
or, 19, 81, 91, 153±154. See also disjunction planning, in speech production, 46. See also
order, of reported events, 14, 19, 75, 122± goals, inference, plans, relevance
127, 135, 199±201. See also conjunc- plans, 45, 46
tion, buttressing reconstruction of, 46, 380 n.4
of antecedents and anaphors, 290±291, recognition of, 60
307 and relevance, 17, 46, 51±52, 380 n.4
of antecedents vs. consequents, 126, 163 See also goals, inference, relevance
of incrementation of context, 50, 162 plurality, 262
of lexemes in semantic ®elds, 102 point of view, 237, 284, 320±322
underlying scales, 105±108 and logophoricity, 284, 314±316, 354, 356,
O-vertex of square of oppositions, 65, 82, 414 n.47, 415 n.54
83±86, 143 See also perspective, re¯exives
relation to I-vertex, 68±71 politeness, 14, 128, 145, 264, 423 n.96
O-to-E drift, 70, 143 polysemy, and implicature, 103
See also square of oppositions Pomo, Northern, 317
poset, 105. See also order
Papago, 350 possessives
Papuan languages, 335 interpretations of, 118, 146, 192, 207
parameters alienable vs. inalienable, 146
pragmatic, 328 and anaphora, 291, 305, 308, 309±310,
syntactic, 300±301, 312±313, 360, 361± 412 n.28
363 possibility. See modality
vs. pragmatic explanations, 360±362 PP-attachment, 174, 176
and language change, 362, 422 n.94 practical reasoning, 30, 43, 45, 52±53
parataxis, 124, 147±148 and intention, 45, 53
partial ordering, 105, 390 n.35 and implicature, 45, 46, 52, 60
particularized conversational implicatures nonmonotonic, 45
(PCIs), xvi±xvii, 13±21, 78, 103±108, and di½culties of reconstruction, 30,
163, 166, 169, 190, 236, 238, 263, 379 45
n.3, 380 n.5±n.6, 385 n.41 `pragmantics', 192±193, 243, 244
de®nition, 16 pragmatic intrusion, into semantic
vs. generalized, 16±21, 74, 89 representations. See intrusion
and inconsistencies with GCIs, 161 pragmatic resolution of underspeci®cation.
(attempted) reduction of GCIs to, 104± See semantic representation
108, 164 (underspeci®cation)
473 Subject Index

pragmatics detachability of, 99, 111

as component of linguistic theory, 166, and factives, 61, 252
243, 365, 368, 371±372 and `®lters', 252, 256
de®nitions of, 7±9, 168, 187, 243 projection of, 111, 249, 252±253, 255±256
and explanation in linguistics, 266, 364± triggers for, 60, 111
365, 374±377 principles of language use. See I-, M-, and
formal theory and, 377, 423 n.1 Q-Principles, maxims
local, 25, 167, 237 Principles and Parameters. See Government
generative grammar and, 237, 261±265, and Binding
362±365 privative opposition, 202, 277, 286, 316,
heuristic potential of, 365 319, 356, 365. See also scales
innate bases, xii, 361 PRO, 363±364
interface, with semantics, see semantics/ processing, and pragmatics. See
pragmatics interface psycholinguistics
nonmonotonic basis, 166, 192, 243 production. See bottleneck, comprehension,
predictive power of, 365, 367, 368±370 psycholinguistics
as process rather than representation, 168, projection
243 of implicatures, 39±40, 80±82, 134, 155±
and psychology, 28, 40, 78, 81, 90, 162, 164, 212, 252±253, 255±256, 289, 383
258±259, 405 n.53 n.28, 396 n.82
and psycholinguistics (see psycho- consistency, with incremented context, 49,
linguistic)s 162
vs. semantics, 7±9, 23, 25, 192, 374±375, with common ground, 49±52, 162±163,
399 n.12 299, 330
and syntax, 261±365, 374, 407 n.4 (see `®lters' and `holes', 252, 255±256
also syntax) M vs. I-implicatures, 137±155, 156±158,
border between, 264±266, 268, 362±365 160, 277 (see also division of pragmatic
universals of, xii±xiii, 15, 69, 266, 300± labor)
301, 361, 370, 423 n.96 of presuppositions, 158, 163, 252±253, 255
See also meaning, semantics and negation, 162, 212±213, 254±255
Prague school, 115, 137 Q-clausal vs. Q-scalar, 161±162, 396 n.77
preemption, 99±100 Q vs. I-implicatures, 133±134, 155±159,
and anaphora, 99±100 395 n.74, n.76
of deictic words, 99 Q vs. M-implicatures, 157, 160±161
and synonymy (see blocking) and Relevance, 52, 163±164
preferred interpretation, xi±xiii, 1, 4, 9, resolution schema, 157±158, 161±162, 289,
11¨, 21, 22±27, 54, 59, 355, 367, 386 330
n.55 rationale for, 161, 411 n.23
and explanation of stable inference uni®ed account for implicatures and
patterns, 71, 386 n.54 presuppositions, 255±256, 396 n.78
See also default interpretation, generalized See also consistency, defeasibility,
conversational implicature, intrusion
presumptive meanings prolixity, 38, 114, 267. See markedness
prepositions, 96, 222 pronouns, 39, 181, 217, 277, 280
presemantic pragmatics. See semantics/ agreement features of, 277, 312
pragmatics interface ambiguity vs. general purpose nature of,
presumptive meanings, 1, 22¨, 127, 169, 269±270, 409 n.7, 422 n.92
258±259, 367, 368. See also default c-commanding NPs, 290, 292±298, 302,
interpretation, generalized 413 n.35, n.36, n.37
conversational implicature clausemate, as coreferential, 278±280
presupposition, 13, 23, 61, 166, 249, 252, as disjoint, 280, 288, 331
397 n.2 complementary distribution of pronouns
accommodation of, 61, 249, 251 vs. re¯exives, 283, 288, 308±326
cancellation of, 111 failure of, 307, 308±326, 412 n.29, 414
and conditionals, 256 n.44
and conversational implicature, 111 and control, 363±364
474 Subject Index

pronouns (cont.) and negation (see negation)

and determiners, 300, 304 and pragmatics, 5, 28, 162, 259, 370, 382
emphatic, 263, 331, 333, 337, 348, 350 n.22, 423 n.1, 424 n.3
vs. gaps, 148±149, 285, 305, 336±337, and processing of implicatures, 5, 55, 162,
364±365, 382 n.22, 411 n.26 245, 361, 370, 382 n.22, 424 n.3, 396
logophoric, 314, 350, 414 n.45, 420 n.79± n.80, 397 n.5, 409 n.12
80 double-processing in tropes, 216, 234±
contrast with ordinary pronouns, 317± 235, 239±240
318, 346±347, 356 psychology, and pragmatics. See
vs. LDA and Switch-Reference, 317, 414 pragmatics
n.46
nonre¯exive, 287±290 Q-implicatures, 75±111
used re¯exively, 336±343 anaphora and, 181±182, 271±272, 286±
obviative, 415 n.51 289, 294±298, 301±302, 305, 317±318
possessive, 291, 305, 308, 309±310, 412 cancellation of, 50±51, 81±82, 161
n.28 clausal (see clausal implicatures)
referential dependence, and, 268, 270, 284, constraints on, 156±157 (see also scales)
287, 310, 312, 353±355 and constructional opposition, 294±297,
vs. referring expressions, 39, 62, 181, 267, 305 (see also clausal implicatures)
269±270 and contraries vs. sub-contraries (see
vs. re¯exives, 181, 237, 263, 267, 270, 271, negation)
277±280, 306±327, 410 n.19, n.20 (see and contrast sets, 32, 36, 64, 79±80
also Anaphors, anaphora, Binding contrast with I-implicatures, 119
Conditions, re¯exives) diagnostics for, 40, 84±85
semantic generality of, 181, 269±270, 409 and di¨erential informativeness, 36, 76,
n.8 79, 100, 102, 287
as variables, 267, 279, 292 inconsistency between, 50±51
vs. zeros (see gaps) and lexical alternates (see scales,
See also coreference, noun phrases, alternates)
reference and lexicalization constraints, 69±70, 91,
proposition, 172, 186, 190, 194, 198, 236, 143, 369
239, 256±259, 273 and narrowing, 185±186
inferred from nonpropositions, 257 negative character of, 40±41
informativeness of (see informativeness) and numbers, see number words
minimal expressed, 195±197, 258 and opposition between structures, 295±
nonlinguistic nature of, 190, 399 n.19 297
pragmatic resolution of, 256±259 scalar (see scalar implicatures, scales)
sentences fail to express, 256 and square of oppositions (see square of
vs. subpropositional content (see oppositions)
propositional content) and switch-reference systems, 317±319
and what is `said', 194±198, 398 n.7 take priority over others, 39±40, 50±51
See also truth-conditions tension with I-implicatures, 116, 121±122,
propositional attitudes, 110 156±158
propositional content, 169, 193 See also clausal implicature, contrast,
determination by implicature, 174, 180, scalar implicature, scales
194 Q-Principle, 75±111
vs. sub-propositional, 194, 240±241, 257± de®nition of, 76
258 as heuristic, 31, 35
See also semantic representation, truth- inferences due to (see Q-implicatures)
conditional content recipient's maxim, 76
prosody, 123, 174, 418 n.61. See also relation to Grice's maxims, 35, 74±75
intonation, stress speaker's maxim, 76
proto-Indo-European, 417 n.59, 419 n.75 See also Q-implicatures
psycholinguistics Quality, maxim of, 14, 19, 20, 386 n.1. See
and asymmetry of production vs. also Moore's paradox, ®gures of
comprehension, 28, 382 n.20, n.22 speech
475 Subject Index

quanti®ers and coreference (see coreference)

domain of (see domain) disjoint (see disjoint reference)
English, 85, 389 n.22 implicatural determination of, 180±183,
existential (see existential) 189, 217±236
implicatures of, 68±71, 76, 82±86 overlapping, 412 n.29, 414 n.46
monotone-increasing, 388 n.20 parsimony in, 115 (see also I-Principle)
negative, 83±86 (see also negation) to persons, 113±114, 149
and pragmatic resolution, 182±183 pragmatic approaches to, 217, 267±273
scope of, 165, 237, 257 and presupposition
and square of oppositions, 65 and referential dependence, 268, 270, 284,
tolerant vs. intolerant, 86, 131±133, 388 287, 305, 409 n.7
n.21, 389 n.22 necessary, 270, 310, 312
universal (see universal) resolution of the `said', 171
Quantity implicatures. See clausal-, Q-, rigid designation and, 242, 246, 300
scalar-implicatures speaker-, vs. semantic- (Kripke), 228±229,
Quantity, maxims of, 14, 19, 35, 37, 74±76, 404 n.47, n.48
112 Strawsonian vs. Russellian approaches,
®rst (see Q-Principle) 218±219, 226±227, 228, 231, 232
second (see I-Principle) -tracking, 273 (see also anaphora)
See also I-implicatures, I-Principle, Q- See also anaphora, extension, pronouns,
implicatures, Q-Principle noun phrases
referential, vs. attributive uses of referring
radical pragmatics, xv, 184, 185, 188, 242, expressions (Donnellan), 226±230,
256, 257, 381 n.8 300
raising as implicaturally determined reference,
of Anaphors (see Anaphors) 226±227
of negation (see Neg-raising) referring expressions
of subjects, 363, 407 n.4 alternative, 149
rather-construction, 213 scale of, 92±94, 149
rationality, 14, 48, 361, 371 See also reference
Realism, 7±8, 167 re¯exives, 181, 237, 263, 267, 270, 271,
reasoning, types of, 42±54 277±280, 306±327
with nonpropositions, 257 acquisition of, 340, 344, 361±362, 416
vs. rule-based inference, 54, 383 n.30 n.59, 422 n.93
See also inference Anaphors vs. pseudo-Anaphors, 306±307,
receiver, of communication. See speaker 308, 323, 324±326
reciprocals, 267, 304, 309±310, 336±337. antecedents of, 271, 345±346, 409 n.9
See also Anaphora, pronouns, binding domain of (see Anaphors, Binding
re¯exivity Conditions, domain)
recognition, of speakers, 113 clause-bound, 311, 323, 324
reduction vs. long-range (see long-range below)
of linguistic expressions (see minimal complementarity with pronouns, 283±284,
expressions) 288, 290, 308, 319±321, 345±347
of generalized implicature to nonce breakdown of, 308±327, 412 n.29, 414
inference, 25, 54±63, 73, 386 n.54 n.44
of grammar to pragmatics, 266 (see also in reference vs. logophoricity, 312, 314±
syntax) 317, 319±326, 345±347
of pragmatic inference to semantics, 25, diachronic origins of, 347±352
73, 169 (see also semantics) discourse-bound, 321±322
redundancy constraints, on lexicalization, disjoint interpretations of proto-re¯exives,
70, 80, 143, 319, 369, 370 357
reduplication, 150, 152±153, 208. See also emphatic-, 308, 320±321, 331±334, 341,
M-implicature 348, 416 n.58
reference origin from emphatics, 332±334, 341±
attributive vs. referential (see referential) 342, 347±348, 416 n.58, 420 n.81, n.83,
and anaphora, 94, 222, 225 421 n.84±85
476 Subject Index

re¯exives (cont.) Sperber-Wilson theory of, xvi, 12, 22, 52,

languages without, 304±305, 323, 334±343 55±60, 135, 167, 197, 236, 238±240,
and logophoricity, 284, 310±317, 320, 322, 257, 371, 380 n.4, 382 n.22, 385 n.42,
345±348 390 n.28, 396 n.81, 399 n.18
characteristics of, 314±315 as balance of I- vs. M-principles, 55
long-range, 282, 283, 296, 310±317, 319± critique of, 55±59
327, 345, 357±358 repetition, 149±153
acquisition of, 361, 416 n.59 respect. See honori®cs, politeness
contrast with pronouns, 319±327, 345± retraction in conversation, 163
347 R-expression, 281, 301. See also noun
more informative than pronouns, 278, phrase
287, 316±317, 410 n.19, n.20 rhetoric, 23, 145, 382 n.22
nonuniversal, 334±343 rigid designators, 242, 246. See also
in oblique phrases, 290±291, 305, 308, reference
319±321 Romance, 416 n.59, 419 n.73, 420 n.78
and perspective or point of view, 314±316, routinization, of inferences, 24. See also
320±323, 346±347, 356 grammaticalization, short-circuited
`picture nouns' and, 308, 320, 324 inference, standardization
pragmatic resolution, of antecedents, 271, R-principle (Horn), 40±41, 135±136, 396
409 n.9 n.82
and referential dependence (see reference) Russian, 97, 349, 364
semantic content of, 277±278, 312, 314±
317, 319±324, 345 `said' vs. implicated, 13±14, 116, 194±198
and stereotypical actions, 329, 332±333, contradiction and, 215
342 Grice's conception, 170±171, 398 n.7, n.8
simplex vs. compound, 311±312, 352, 413 implicatural contributions to `said', 172±
n.38 186, 217±225 (see also intrusion)
singularity of, 321 intuitions about, 197, 199±200, 205±206,
types of, 349±350, 359 216 (see also truth-conditions)
See also Anaphor, anaphora, Binding reference and, 171
Conditions, reciprocals, re¯exivity relation to logical form, 171
re¯exivity, 280, 304, 329±330, 334 relation to truth-conditions, 170±171
without Anaphors, 280, 304±305, 334± indexicals, role of, 171
343 variant views of, 194±198
grammatical relations and, 304 salience
implicated only, in some languages, 343, of antecedents, 273
334±343 and contrast, 32, 36, 80, 100, 107
and introverted predicates (see introverted and games of pure coordination, 29¨,
predicates) 258±259
languages without coding of, 334±343 of readings, 90
marked nature of, 329, 331±334 satisfaction conditions, 171
and `middle' verbs, 338, 349±350 scalar implicatures, 35±37, 50, 58, 76±108,
and stereotypical transitive actions, 329± 161, 383 n.29
330, 332±333, 342 of adjectives, 86±87, 128±129
verbal, 304±305, 325, 334, 349, 419 n.77 blocking of by negation (see negation)
register, 138 cancellation of, 50±51
reinforceability of implicatures, 15 by clausal implicatures, 81±82
relation, maxim of. See Relevance as default rules, 47
relevance de®nites vs. inde®nites, 63, 91±95, 220
and ellipsis, 183 diagnostics for, 81±82, 84±85
generates only PCIs, 74 entailment and (see scales)
goal- or plan-driven theory of, 17, 46, 51± epistemic commitment and, 78±79, 387
52, 55, 163, 380 n.4 n.8, n.8, n.11, 391 n.40
Grice's maxim of, 14, 74 in general vocabulary, 86±98
interaction with GCIs, 51±52, 163±164 intrusion into semantics, 200±201, 205±
and other maxims, 112, 135 206, 210±213, 217±221, 234, 250±251
477 Subject Index

lower vs. upper bound, 88, 178 reversal of (see numbers, negation)
modals, 82±86, 154 in semantic ®elds, 101±104
negation and, 36, 64±68, 80±82, 128±129, strong vs. weak items, 76, 81
211±213, 254±255 summary of properties, 82
numerals (see number words) and square of oppositions, 65, 83±86, 388
quanti®ers, 36, 47, 76, 82±86 n.19
positive vs. negative scales (see scales) tenses, 95±96, 128±129
processing of, 245 verbs, 87, 98±99, 110
and re¯exives (see scales) See also Q-implicature, scalar-implicature
and semantic ®elds (Type I and II), 101± scienti®c reasoning, 44
103 scope, of logical operators
and switch reference, 317±319 ambiguities of, 165, 375, 388 n.15
See also Q-implicature, scales implicatures inside, 41, 172, 195, 196±197,
scales 253±256
anaphoric expressions, 94, 181±182, 286± raising and, 129±134, 165
289, 310, 317±318, 348, 355 and scope-free semantic representation,
cases, 97 241, 257, 375
closed class morphemes and, 90±98, 155 See also ambiguity, negation
constraints on, 79±80, 82, 98, 122, 156, self, 340±342
391 n.38 self-identi®cation, 114
lexicalization constraint, 79±81, 120, 388 semantic change, 185, 262±264, 339, 341±
n.15, 392 n.49, 412 n.28 342, 347±350, 362. See also diachrony,
`aboutness' constraint, 80, 121, 388 n.16 grammaticalization
salience, 107±108 semantic entailment. See entailment
and contrast sets, 32, 36, 79±80 semantic ®elds
de®nition, 79 structure of, and implicatures, 101±104
determiners, 91±95 (see also lexical items, scales)
entailment and, 79±82 semantic generality, 115, 135, 184±186,
nonentailment scales, 98±103 189, 274, 376
function words and, 90±98 and anaphora, 225, 267±270, 287, 304
general vocabulary and, 86±104 vs. indexicality, 191±192
Horn-, 36, 47, 52, 79±108, 274, 278, 286, resolution of, 184±186, 257±258, 376
296, 297 and vacuity of representations, 184, 189
conditions on (see constraints above) weak univocality vs. nonspeci®c (Atlas),
Hirschberg-, 104±108, 120, 390 n.36, n.37 185, 189, 242, 257, 381 n.8
inversion, 88 semantic interpretation, 172
and lexicalization patterns, 69±71, 80±81, role of indexicals in, 177±180
143 semantic representation, 8, 168, 171, 240±
metascales, 81, 182 242, 247, 256±259, 372, 375±377
mid-scalar items, 77, 83±86 vs. conceptual representations, 241±242,
negation and, 64, 80±82, 83±86, 128±129, 372
162, 212±213, 255±256 constraints on, from pragmatics, 377
nonentailment-, 36, 98±108 decoded from syntactic structures, 193,
non-GCI scales, 82 194, 240, 248, 375
nonce- (see Hirschberg-scales) and entailment, 191, 194
number words, 87±90, 178, 219 enrichment by pragmatics, 194±198
order relations underlying, 105 and features (see `markerese' below)
particularized, 103±108 fragmentary, 168, 191, 238, 256±259, 399
placement on square, 84±86 n.17
positive vs. negative scales, 64, 254±255, generality of, 135, 184±186, 240±242
388 n.17 and implicature, 171, 238, 377, 380 n.2
pragmatic vs. semantic nature of, 88, 104± hybrid semantico-pragmatic nature of,
108 168, 193, 247, 248±256, 259±260, 374±
prepositions, 96 375, 405 n.55
pseudo-scales, 79, 80, 99, 391 n.38, n.40 pragmatic conditions `in red ink' and
re¯exives and, 278, 286, 353, 316±317 erasable, 249, 253, 259, 384 n.37
478 Subject Index

semantic representation (cont.) interleaving of processes, 168, 187, 236,

interpretation, 7±8, 168, 191 242, 245, 374, 398 n.6, 405 n.55
levels of, 168 intermediate phenomena (see intrusion,
vs. logical form, 375±376, 380 n.2 explicature, impliciture)
`markerese', 190, 240 modularity and, 187, 243±245, 371±372,
nature of, 375±377 397 n.4
and scope, 165 presemantic pragmatics, 187±190, 215±
scope-free, 241, 257, 375±376 217, 218, 226, 230, 231, 232±234, 238,
semantic generality and (see semantic 239±240, 245, 247
generality) recent trends, 167
sub-propositional nature, 194, 257 and scope of operators, 165
and time, 122 terminology, 194±198
and underspeci®cation, 8, 89, 168, 171, semiotics, 13, 26, 273
240±241, 256±259, 375±377 sense. See intension, monosemy
and `what is said', 171, 194±198 sense-relations, 168, 191, 194, 398 n.6
See also logical form, intrusion sentence vs. discourse, 267, 321±322, 415
`semantic retreat', 190, 240±242 n.52
semantics vs. elliptical fragments (see ellipsis)
author's assumptions, 7±9 -frame, 76, 108, 136
autonomy of, 166, 226, 232, 235, 244±245, vs. utterance (see utterance)
397 n. (see also Grice's circle, sentence-meanings. See utterance
presemantic pragmatics, semantics/ sentence-types, minor, 392 n.50
pragmatics interface) She¨er stroke, 386 n.53
controls pragmatic processes, 245 short-circuited implicature, 24±25, 134,
de®nitional problems, 168 262±263, 408 n.5. See also
division of labor with pragmatics, 165, conventionalization, standardization
168, 243±245, 398 n.6 simultaneity of events, 124
feature-theories of, 190 since, 36, 263
interface with pragmatics (see semantics/ Sissala, 148
pragmatics interface) situation semantics, xvii, 26, 27, 167, 172,
and `language of thought', 7, 242, 372 191, 208±210, 218, 229±230, 242±243,
model-theoretic, 6±7, 172, 193 256
vs. pragmatics, 7±9, 170±172, 186±193, described situation vs. utterance situation,
233±234 191, 229±230
as distinct processes over same indexicals in, 191±192, 230, 242±243
representations, 168 utterance situation vs. resource situation,
terminological di½culties, 194±195 191
restricted theories of, 190 sloppy-identity, 293
scope of, 186, 190±191 some, 16±17, 36, 47, 65, 68, 77, 175, 200,
and sense relations, 191, 241 297
truth-conditional, 7±8, 241, 248±253 space, deixis of, 94
See also, DRT, meaning, Situation linguistic expressions for, 94, 96
Semantics, truth-conditions prepositions of, 96, 118, 207, 222, 237
semantics/pragmatics interface, 165±260, Spanish, 350
374±375 speaker
vs. abandoning the distinction, 192±193, vs. addressee's perspective, 6, 28, 76, 114,
243 136, 382 n.20, n.22, 373, 387 n.3
architecture of (see meaning) commitment of, 76±79, 163, 384 n.36
common representation with distinct intentions of (see intention)
contributions, 168, 193, 244, 374, 424 knowledge of, 40
n.3 speaker-meaning, 4, 22±27, 54±63, 381 n.10
cyclical processing model of, 245 distinctions in, 27
and `Grice's circle', 172±173, 186±187, vs. linguistic(sentence)-meaning, 228, 234
236±243 (see also utterance)
input/output models of, 166±167, 187, and reduction of GCIs to, 25, 54±63, 367
217, 236, 242, 243±244, 374±375, 397 See also nonce meaning, utterance-token-
n.4 meaning
479 Subject Index

speaker-reference. See reference switch-reference, 317±319, 409 n.9, 414

speci®city n.46
de®ned, 115 syllogisms, of di¨erence kinds of inference,
vs. generality, 135, 185 42±43
and I-principle, 114±116 symbols and notations, ix±x, 387 n.5, n.6,
speech, bottleneck in production of, 28, 382 388 n.13
n.18, n.20 synonymy
neuroanatomical specializations for, 27± avoidance of, 139
28, 381 n.16 denotational, and the M-Principle, 40±41,
transmission rates in bps, 28, 382 n.19 137
speech acts, and satisfaction conditions, 171 and sense relations, 191, 241
and utterance-types, 23 near synonymy and Neg-Raising
square of oppositions, 64±71, 82±86, 104, predicates, 134
131±132, 388 n.15 syntactic amalgams, 408 n.5
arithmetic, 82, 85±86, 131±132, 389 n.22 syntax
and connectives, 66±67 and anaphora, 261±365
contraries vs. contradictories, 64±65 restricted role in, 270±271
and implicature (see scales) binding (see binding)
and logical operators, 66±67 indefeasible patterns of, 265, 342, 364, 365
and modals, 66±67 interface with semantics, 8±9, 172, 190±
and quanti®ers, 65 191, 240±241, 243±244, 248, 281, 362±
sub-contraries as pragmatic, 65±71, 132 365, 374
standardization of use, 23±24, 262, 381 and interpretation, 262, 407 n.4
n.12 and negation, 129
state descriptions, 115, 273±274. See also and pragmatics, 261±365
informativeness constraints on rules, 262, 407 n.4, 408 n.5
statements vs. sentences, 256, 402 n.34. See constructions with pragmatic origins,
also utterance 264±265
stereotypes, 6, 115, 391 n.40 as rival explanations, 261±265, 362±365
de®nition, 115 tests for pragmatic vs. syntactic
and heuristics, 32±33, 37, 38 character, 268, 278±279, 364±365, 408
and I-implicature, 114±115, 222±223, n.6
328±329 universals of, 265, 297, 303, 319, 329
inference in absence of, 127
inference to, 117, 136±137, 142, 147, 148 tacit coordination, 29, 258
nonstereotypes and M-implicature, 136± Tahitian, 337
137 Tamil, 97, 153, 285, 332, 334, 351±352
strength tautology, 184
di¨erential, of GCIs, 41, 77±78, 411 n.23 taxonomies, 101±104
of scalar items (see scales) temporal inference, 122±127, 158, 180, 199.
semantic (see informativeness) See also order, tense, conjunction
stress, contrastive, 333, 334, 418 n.61. See buttressing
also intonation, prosody tense
strong vs. weak and implicature, 95±96, 122±124, 180,
constructions (see clausal implicatures) 263, 390 n.30, 400 n.25
expressions (see scales, semantic past, 95, 400 n.25
generality) perfect, 96
structuralism, in semantics, 101 and reference points, 122, 151, 199, 399
subcontraries, 65±71, 82±86. See also n.15, 400 n.25
square of opposition semantic analysis of, 95
subjectivity, and anaphora, 312, 315, 320± Thai, 300
321, 413 n.40, 415 n.54. See also the, 60±62, 147, 217±230. See also de®nite
logophoricity article
`suspenders', 81, 84 thematic roles, 282, 298, 303, 306
suspension of implicatures, 81±82, 131, 252. theme vs. rheme, 291, 298, 307. See also
See also cancellation topic
Swedish, 422 n.94 theta criterion, 303, 413 n.38
480 Subject Index

theta roles. See thematic roles universe of discourse. See domain

three-leveled theory of meaning, 23±27, 54± univocality of meaning. See monosemy
63, 71, 367. See also meaning unmarked utterances. See markedness,
together-implications, 37±38, 117, 207 I-principle
tolerant vs. intolerant operators, 86, 131± Uto-Aztecan, 420 n.78
133 utterance
topic vs. comment, 279 anaphoric reference to, 253±254
transderivational constraints, 392 n.48 de®nition of, 26±27, 163
transitivity, of implicatures, 390 n.31 form, 26
transmission rates, of human speech, 28 vs. sentence, as bearer of truth-conditions,
`triggers' of inferences, 60, 111 256, 402 n.34
tropes. See ®gures of speech, irony, maxims types vs. tokens, 25
(exploitation of ), metaphor, metonymy -type meaning (see utterance-type)
truth-conditional content, of utterances, utterance-meaning, 22
166, 241, 248±249, 259. See also truth vs. sentence-meaning, 381 n.14
conditions utterance-type vs. utterance-token, 1, 22¨,
truth-conditional semantics. See semantics 25, 54±63, 71±72, 373, 381 n.10, 385
truth-conditions n.40
intuitions about, 186, 189, 197, 199, 204, See also meaning
205±206, 208, 209±210, 216, 219, 227,
229, 231, 232±233, 234±235 variables, 182, 267, 279
and nonliteral uses of language, 402 n.39 indexical, 191
(see also literal meaning, maxims, lexical NPs as, 301
exploitation of ) pronouns as, 279, 292±296
not the input to pragmatics, 380 n.2, 398 verb phrase (VP), 293, 303. See also
n.7 noncon®gurational languages
pragmatic contributions to (see intrusion) verbs
properties of utterances not sentences, 256, acquisition of, 140
398 n.7, 402 n.34 Aktionsarten of, 124
and semantic representations (see semantic causative, 141±142, 208, 297
representations) and clausal implicatures, 109±111
and nondeclaratives (see speech acts) denominal, 140
truth-functional connectives. See factives, 61, 252
connectives introverted (inherently re¯exive), 329, 332,
try, 98, 219, 390 n.31, 403 n.44 342
Turkish, 148, 285, 349 modal, 140
type/token distinction, 25 of motion (see come)
typology, markedness and, 329 Neg-Raising predicates (see Neg-Raising)
of anaphoric systems, 334, 344, 349, 359 of propositional attitude, 110, 129±131
of implicatures (see conversational of saying, 110, 314
implicature) and scalar implicatures, 87, 110
of inference types (see inference) subject-controlling vs. object-controlling,
Tzeltal, 152 364
Tzotzil, 152 Vietnamese, 125±126
visual interpretation, 2±4, 7
underspeci®cation of semantic vocabulary. See lexicon, semantics ®eld
representations. See semantic
representation Yucatec, 96
understatement, 145
universals Zapotec, 335
Greenberg's, 126, 128 zero anaphora. See anaphora
of pragmatics, xii, 15, 69±70, 280, 300, zero conjunction. See parataxis
319, 329, 365, 370, 408 n.6, 423 n.96
of semantics
and universal grammar (UG), 265, 279±
280, 297, 319, 329, 360, 408 n.6