10.

9 Semantics

We have already touched upon semantics at the lexical level when we discussed
semantic prosody/preference and pattern meanings in unit 10.2. But corpora are also
more generally important in semantics in that they provide objective criteria for
assigning meanings to linguistic items and establish more firmly the notions of fuzzy
categories and gradience (see McEnery and Wilson 2001: 112-113), as demonstrated
by Mindt (1991). This section considers semantics in more general terms, with
reference to the two functions of corpus data as identified above by McEnery and
Wilson (ibid).
Corpora have been used to detect subtle semantic distinctions in near synonyms.
Tognini-Bonelli (2001: 35-39), for example, finds that largely can be used to
introduce cause and reason and co-occur with morphological and semantic negatives,
but broadly cannot; yet while broadly can be used as a discourse disjunct for
argumentation and to express similarity or agreement, largely cannot. Gilquin (2003)
seeks to combine the corpus-based approach with the cognitive theory of frame
semantics in her study of the causative verbs GET and HAVE. The study shows that the
two verbs have a number of features in common but also exhibit important differences.
For example, both verbs are used predominantly with an animate causer. Yet while
with GET the causee is most often animate, the frequencies of animate and inanimate
causees are very similar with HAVE. Nevertheless, when causees are expressed as an
object (i.e. not demoted), the proportion of animates and inanimates is reversed, with
a majority of animates with GET and a predominance of inanimates with HAVE. While
Tognini-Bonelli (2001) and Gilquin (2003) can be considered as examples of
assigning meanings to linguistic items, Kaltenböck (2003) further exemplifies the role
of corpus data in providing evidence for fuzzy categories and gradience in his study of
the syntactic and semantic status of anticipatory it. Kaltenböck finds that both the
approach which takes anticipatory it to have an inherent cataphoric function (i.e.
referring it) and the view that it is a meaningless, semantically empty dummy element
(i.e. prop it) as have been proposed previously are problematic as they fail to take into
account the actual use of anticipatory it in context. The analysis of instances actually
occurring in ICE-GB showed very clearly that delimiting the class of it-extraposition
(and hence anticipatory it) is by no means a matter of ‘either-or’ but has to allow for
fuzzy boundaries (Kaltenböck 2003: 236): ‘anticipatory it takes an intermediate
position between prop it and referring it, all of which are linked by a scale of
gradience specifying their scope of reference (wide vs. narrow)’ (Kaltenböck 2003:
235). The functionalist approach taken by Kaltenböck (2003) is in sharp contrast to
the purely formalist approach which, relying exclusively on conceptual evidence,
identifies anticipatory it as meaningless. Kaltenböck argues that:
the two approaches operate with different concepts of meaning: a formalist will be
interested in the meaning of a particular form as represented in the speaker’s
competence, while for the view expressed here ‘meaning’ not only resides in
isolated items but is also the result of their interaction with contextual factors.
(Kaltenböck 2003: 253)
Let us now turn to a core area of semantics – aspect. According to Smith (1997:
1), ‘aspect is the semantic domain of the temporal structure of situations and their
presentation.’ Aspect has traditionally been approached without recourse to corpus
data. More recently, however, corpus data has been exploited to inform aspect theory.
Xiao and McEnery (2004a), for example, have developed a corpus-based two-level
model of situation aspect, in which situation aspect is modelled as verb classes at the
lexical level and as situation types at the sentential level. Situation types are the
composite result of the rule-based interaction between verb classes and complements,
arguments, peripheral adjuncts and viewpoint aspect at the nucleus, core and clause
levels. With a framework consisting of a lexicon, a layered clause structure and a set
of rules mapping verb classes onto situation types, the model was developed and
tested using an English corpus and a Chinese corpus. The model has not only
provided a more refined aspectual classification and given a more systematic account
of the compositional nature of situation aspect than previous models, but it has also
shown that intuitions are not always reliable (e.g. the incorrect postulation of the
effect of external arguments). We will return to discuss aspect in unit 15.3 of Section
B and case study 6 of Section C of this book.
The examples cited above demonstrate that corpora do have a role to play in the
study of meaning, not only at the lexical level but in other core areas of semantics as
well. Corpus-based semantic studies are often labour-intensive and time-consuming
because many semantic features cannot be annotated automatically (consider e.g.
causer vs. causee and animate vs. inanimate in Gilquin’s (2003) study of causative
GET/HAVE). Yet the interesting findings from such studies certainly make the time and
effort worthwhile. In the next section we will review the use of corpora in pragmatics.

10.10 Pragmatics

As noted in unit 4.4.6, pragmatics is strongly – though not exclusively – associated

with spoken discourse. This is hardly surprising considering that written registers tend
to be referentially explicit whereas spoken registers typically ‘permit extensive
reference to the physical and temporal situation of discourse’ (Biber 1988: 144). In
Kennedy’s (1998: 174) words, ‘What we say and how we say it is influenced by who
we are talking to and where the interaction is taking place.’ Until the mid-1990s
corpus-based pragmatic studies were severely constrained because there was only one
reasonably large, publicly available corpus which was sufficiently marked up for
prosodic and discourse features, the London-Lund Corpus (i.e. LLC, see unit 7.5) (cf.
Kennedy 1998: 174). It is, therefore, unsurprising that earlier corpus-based work on
pragmatics was based fairly exclusively on the LLC. For example, Svartvik (1980),
on the basis of a sample of 45,000 words from the LLC, found that the discourse
marker well is an important device which allows the speaker time to think online
while keeping a turn in conversation. The pragmatic functions of well include polite
disagreement, qualified refusal, reinforcement, modification, indirect and partial
answers, and delaying tactics. Aijmer’s (1987) study of oh and ah in a 170,000-word
sample from the LLC provides a full account of the major pragmatic functions of the
two ‘disjunct markers’ (Jefferson 1978: 221). Tottie (1991), on the basis of a
comparison of the LLC and the Santa Barbara Corpus of Spoken American English
(i.e. SBCSAE, see unit 7.5), finds that American speakers use backchannel agreement
markers (e.g. yeah, sure and right) three times as frequently as British speakers.
Aijmer (1987: 63) notes that one of the pragmatic functions of oh and ah is to
signal ‘a shift or development to something not foreseen by the speaker’, thus
construing what comes afterwards as ‘topically not coherent’ (Jefferson 1978: 221).
Discourse markers such as anyway, however and still help to establish coherence in
spoken discourse (see Lenk 1995, 1998a, 1998b). Lenk (1998b), for example, uses the
LLC and SBCSAE corpora to investigate how however and still are involved in the
process of achieving conversational coherence. It was found that ‘the function of both
of these discourse markers is to connect parts of the discourse that are not
immediately adjacent, or that are not topically related’ (Lenk 1998b: 256).
Nevertheless, while ‘however closes digressions that are relevant to the development
of the main topic, or that bear interactional significance’, ‘still closes off subjective
comments within a quasi-objective narration or presentation of facts’ (Lenk 1998b:
256). It is also interesting to note that however is used as a discourse marker only in
British English (Lenk 1998b: 251).
Spoken language, and face-to-face conversation in particular, takes place on the
basis of a shared context, avoids elaboration or specification of reference, and reflects
the needs for real-time processing (Leech 2000). It is, therefore, hardly surprising that
conversation is more vague than most written genres. Vagueness is pervasive in
conversation where it plays an important role. The most obvious reason for using
vague expressions is uncertainty at the time of speaking. In this case, vagueness
allows speakers to maintain fluency even though they lack information about a given
quantity, quality or identity, or, when such information is potentially available, they
cannot access or process it in time. However, speakers may still choose to be vague
even when they could in principle be more precise. This is because vague language
can serve a number of pragmatic functions. Jucker, Smith and Lüdge (2003), for
example, analyze the vague additives (i.e. approximators, downtoners, vague category
identifiers and shields) and instances of lexical vagueness (i.e. vague quantifying
expressions, vague adverbs of frequency, vague adverbs of likelihood, and
placeholder words) in a corpus of semi-controlled spoken interactions between
students in California. They find that vagueness is an interactional strategy which
plays an important role in managing conversational implicature. First, vague
expressions may serve as focusing devices, directing the hearer’s attention to the most
relevant information. Second, vague expressions of quantities provide information
about the significance of the quantity and may provide a reference point in terms of a
scale. Third, vague expressions may also convey several aspects of propositional
attitude (e.g. conveying different levels of certainty regarding the propositional
content, conveying the newsworthiness or expectedness of a statement, and conveying
evaluative meaning). Finally, vague expressions may serve various social functions
(serving as politeness strategies, softening implicit complaints and criticisms, and
providing a way of establishing a social bond). As such, vague language helps to
‘guide the hearer towards the best interpretation of the speaker’s intention’ (Jucker,
Smith and Lüdge 2003: 1766).
Similarly, Drave (2002) studies vague language (VL) in intercultural
conversations. The corpus he used was the Hong Kong Corpus of Conversational
English (HKCCE), a corpus consisting of 98,310 words of native speaker English
(NSE) and 84,208 words of English produced by native speakers of Cantonese (NSC).
Drave (2002: 27) finds that vague language can be used in naturally occurring
conversations strategically for promoting politeness and intersubjectivity and for
managing asymmetries of knowledge, particularly in intercultural interaction. It was
found that while quantitatively NSE seems to be ‘vaguer’ than NSC, the two groups
do not differ qualitatively, ‘with very few VL items being used exclusively by one
group or the other and the rank orders of VL items of the most frequent items virtually
identical’ (Drave 2002: 29).
McEnery, Baker and Cheepen (2002) explored the relationship between
directness and lexical markers of politeness with reference to operator requests to
‘hold the line’, using a corpus of telephone-based transactional dialogues. They found
that of the various types of request strategies (bare imperative, deletion, conditional if,
prediction and question), only the bare imperatives were unambiguously direct while
all of the other types were to some extent indirect imperatives. It is also interesting to
note that while bare imperatives are the most common request strategy, they are
typically softened by mitigators such as please and just (McEnery, Baker and
Cheepen 2002: 64-65).
While politeness strategies are particularly important in transactional dialogues
as explored by McEnery, Baker and Cheepen (2002), conversation is not always
polite. Complaining is unavoidable. Laforest (2002) presents an interesting study
which characterizes ‘the complaint/complaint-response sequence in everyday
conversations between people who are on intimate terms’ (Laforest 2002: 1596), in
this case, peer family members (i.e. husbands/wives and brothers/sisters). The
complaints exchanged between people who are not peers (i.e. parents vs. children)
were excluded in order to neutralize the variation introduced by a difference in
hierarchical position between interactants. The data used in this study was taken from
a corpus of about 50 hours of family conversations recorded in Montréal. The
complaints analyzed in this study illustrated the numerous ways in which speakers
expressed dissatisfaction with the behaviour of people close to them. They had
preferential realization patterns that could be linked in part to the intimacy of the
relationship between the interactants: in many ways, they were uttered without the
special precautions generally associated with face-threatening acts (FTAs) outside the
private sphere (Laforest 2002: 1617-1618). Laforest found that while the complainees
most often reject the blame levelled at them, well characterized arguments are
virtually absent from the corpus; the entry into the argument is negotiated in the
speech turns that follow the complaint/response sequence, and the argument only
breaks out if the complainer questions the value of the complainee’s response. The
study also shows that both complainer and complainee use various strategies for
avoiding an argument and, more often than not, succeed in doing so (Laforest 2002:
1596).
Nowadays, pragmatic studies are more varied than before. One area of increasing
interest is historical pragmatics which, like general diachronic studies, depends
heavily upon corpus data. For example, Arnovick (2000) examines the speech event
of parting, focusing on the development of Goodbye, which was originally an explicit
blessing God be with you. She finds that the formal development from God be with
you to Goodbye is linked to functional shifts. In the English Drama section of the
Chadwyck-Healey corpus, the original form, which appeared in closing sections of
dialogue in Early Modern English, was used as a blessing as well as a greeting at
parting while the contracted form became stronger in the force of the polite closing
greeting. Arnovick’s study shows that the end of the 17th century and the beginning of
the 18th century marked a crucial period during which the blessing declined and the
closing form Goodbye increased in frequency. Jucker and Taavitsainen (2000)
undertake a diachronic analysis of one particular speech act, i.e. insults, through the
history of English on the basis of a corpus composed of both literary and non-literary
data. Their analysis of written materials of the past periods indicates an evident bias
towards the conventionalized insults. Most early examples are found in literary texts,
which reflect generic conventions of the time and the culture that gave rise to these
literary forms. Jacobsson (2002) used a pilot version of the Corpus of English
Dialogues (CED, see unit 7.7) to study gratitude expressions such as Thank you and
Thanks in Early Modern English. The author found that while these expressions
themselves were probably the same in the Early Modern period as they are today, they
‘had not developed the discourse-marking features of today’s British English; nor is it
possible to see the complex patterns of thanking in different turn-positions in the CED
material’ (Jacobsson 2002: 78). Biber (2004) explores, on the basis of the ARCHER
corpus (see unit 7.7), the patterns of historical change in the preferred devices used to
mark stance across the past three centuries. He finds that of the grammatical
categories marking stance, modal verbs have undergone a decrease in use whereas
other devices such as semi-modals, stance adverbials, and stance complement clause
constructions have all increased in use across the historical periods in his study (Biber
2004: 129).
Pragmatics is an area in which more and more corpus data is being used.
However, meanings dependent upon pragmatics cannot easily be detected
automatically. As in semantics, the automatic extraction is not likely unless the
corpora used for such studies have been annotated manually with the required
analyses.

10.11 Sociolinguistics

While sociolinguistics has traditionally been based upon empirical data, the use of
standard corpora in this field has been limited. The expansion of corpus work in
sociolinguistics appears to have been hampered by three problems: the
operationalization of sociolinguistic theory into measurable categories suitable for
corpus research, the lack of sociolinguistic metadata encoded in currently available
corpora, and the lack of sociolinguistically rigorous sampling in corpus construction
(cf. McEnery and Wilson 2001: 116).
Corpus-based sociolinguistic studies have so far largely been restricted to the
area of gender studies at the lexical level. For example, Kjellmer (1986) compared the
frequencies of masculine and feminine pronouns and lexical items man/men and
woman/women in the Brown and LOB corpora. It was found that female items are
considerably less frequent than male items in both corpora, though female items were
more frequent in British English. It is also interesting to note that female items were
more frequent in imaginative (especially romantic fiction) than informative genres.
Sigley (1997) found some significant differences in the distribution patterns of
relative clauses used by male and female speakers/writers at different educational
levels in New Zealand English. Caldas-Coulthard and Moon (1999) found on the
basis of a newspaper corpus that women were frequently modified by adjectives
indicating physical appearance (e.g. beautiful, pretty and lovely) whereas men were
frequently modified by adjectives indicating importance (e.g. key, big, great and
main). Similarly, Hunston (1999b) noted that while right is used to modify both men
and women, the typical meaning of right co-occurring with men is work-related (‘the
right man for the job’) whereas the typical meaning of right co-occurring with women
is man-related (‘the right woman for this man’). Hunston (2002: 121) provided two
alternative explanations for this: that women are perceived to be ‘less significant in
the world of paid work’, or that ‘men are construed as less emotionally competent
because they more frequently need “the right woman” to make their lives complete.’
In either case, women are not treated identically (at least in linguistic terms) in society.
Holmes (1993a, 1993b, 1993c, 1997) has published widely on sexism in English, e.g.
the epicene terms such as -man and he, gender-neutral terms like chairperson, and
sexist suffixes like -ess and -ette. Holmes and Sigley (2002), for example, used
Brown/LOB and Frown/FLOB/WWC to track social change in patterns of gender
marking between 1961 and 1991. They found that
 while women continue to be the linguistically marked gender, there is some
evidence to support a positive interpretation of many of the patterns identified in
the most recent corpora, since the relevant marked contexts reflect inroads made
by women into occupational domains previously considered as exclusively male.
(Holmes and Sigley 2002: 261)
While Holmes and Sigley (2002) approached gender marking from a diachronic
perspective, Baranowski (2002) approached the issue in a contrastive context.
Baranowski investigated the epicene pronominal usage of he, he or she and singular
they in two corpora of written English (one for British English and the other for
American English), and found that the traditional form he is no longer predominant
while singular they is most likely to be used. The form he or she is shown to be used
rather rarely. The study also reveals that American writers are more conservative in
their choice of a singular epicene pronoun. In gender studies like these, however, it is
important to evaluate and classify usages in context (cf. Holmes 1994), which can be
time-consuming and hard to decide sometimes.
In addition to sexism, femininity and sexual identity are two other important
areas of gender studies which have started to use corpus data. For example, Coates
(1999) used a corpus of women’s (and girls’) ‘backstage talk’ to explore their self-
presentation in contexts where they seem most relaxed and most off-record, focusing
on ‘those aspects of women’s backstage performance of self which do not fit
prevailing norms of femininity’ (Coates 1999: 65). Coates argued that the backstage
talk ‘provides women with an arena where norms can be subverted and challenged
and alternative selves explored’ while acknowledging ‘such talk helps to maintain the
heteropatriarchal order, by providing an outlet for the frustrations of frontstage
performance.’ Thorne and Coupland (1998: 234) studied, on the basis of a corpus of
200 lesbian and gay male dating advertisement texts, a range of discursive devices
and conventions used in formulating sexual/self-gendered identities. They also
discussed these discourse practices in relation to a social critique of contemporary gay
attitudes, belief and lifestyles in the UK. Baker (2004) undertook a corpus-based
keyword analysis of the debates over a Bill to equalize the age of sexual consent for
gay men with the age of consent for heterosexual sex at sixteen years in the House of
Lords in the UK between 1998 and 2000 (see unit 24 for further discussion of
keywords). Baker’s analysis uncovered the main lexical differences between
oppositional stances and helped to shed new light on the ways in which discourses of
homosexuality were constructed by the Lords. For example, it was found that
homosexual was associated with acts whereas gay was associated with identities.
While those who argued for the reform focused on equality and tolerance, those who
argued against it linked homosexuality to danger, ill health, crime and unnatural
behaviour.
While corpus-based sociolinguistic research has focused on language and gender,
corpora have also started to play a role in a wide range of more general issues in
sociolinguistics. For example, Banjo (1996) discussed the role that ICE-Nigeria is
expected to play in language planning in Nigeria; Florey (1998: 207) drew upon a
corpus of incantations ‘in order to address the issue of the extent to which specialised
sociocultural and associated linguistic knowledge persists in a context of language
shift’; Puchta and Potter (1999) analyzed question formats in a corpus of German
market research focus groups (i.e. ‘a carefully planned discussion designed to obtain
perceptions on a defined area of interest in a permissive, non-threatening
environment’, see Krueger 1994: 6) in an attempt ‘to show how elaborate questions
[i.e. ‘questions which include a range of reformulations and rewordings’ (Puchta and
Potter 1999: 314)] in focus groups are organized in ways which provide the kinds of
answers that the focus group moderators require’ (Puchta and Potter 1999: 332); de
Beaugrande (1998: 134) drew data from the Bank of English to show that terms like
stability and instability are not ‘self-consciously neutral’, but rather they are socially
charged to serve social interests. Dailey-O’Cain (2000) explored the sociolinguistic
distribution (sex and age) of focuser like (as in And there were like people blocking,
you know?) and quotative like (as in Maya’s like, ‘Kim come over here and be with
me and Brett’) as well as attitudes towards these markers.
In a more general context of addressing the debate over ideal vs. real language,
de Beaugrande (1998: 131) argues that sociolinguistics may have been affected,
during its formative stages, as a result of the long term tradition of idealizing language
and disconnecting it from speech and society. Unsurprisingly, sociolinguistics has
traditionally focused on phonological and grammatical variations in terms of ‘features
and rules’ (de Beaugrande 1998: 133). He observes that the use of corpus data can
bring sociolinguistics ‘some interesting prospects’ (de Beaugrande 1998: 137) in that
‘[r]eal data also indicate that much of the socially relevant variation within a language
does not concern the phonological and syntactic variations’ (de Beaugrande 1998:
133). In this sense, the
corpus can help sociolinguistics engage with issues and variations in usage that are
less tidy and abstract than phonetics, phonology, and grammar, and more
proximate to the socially vital issues of the day […] Corpus data can help us
monitor the ongoing collocational approximation and contestation of terms that
refer to the social conditions themselves and discursively position these in respect
to the interests of various social groups. (de Beaugrande 1998: 135)
With the increasing availability of corpora which encode rich sociolinguistic
metadata (e.g. the BNC), the corpus-based approach is expected to play a more
important role in sociolinguistics. To give an example of this new role, readers will
have an opportunity to explore, in case study 4 in Section C, the patterns of swearing
in modern British English along such dimensions as sex, age, social class of speakers
and writers encoded in the BNC.

10.12 Discourse analysis

Closely allied with sociolinguistics is discourse analysis (DA), especially critical

discourse analysis (CDA), which is mainly concerned with the studies of ideology,
power and culture (cf. Fairclough 1995). While both corpus linguistics and DA rely
heavily on real language, Leech (2000: 678-680) observes that there is ‘a cultural
divide’ between the two: while DA emphasizes the integrity of text, corpus linguistics
tends to use representative samples; while DA is primarily qualitative, corpus
linguistics is essentially quantitative; while DA focuses on the contents expressed by
language, corpus linguistics is interested in language per se; while the collector,
transcriber and analyst are often the same person in DA, this is rarely the case in
corpus linguistics; while the data used in DA is rarely widely available, corpora are
typically made widely available. It is also important to note that some terms used in
DA are defined differently from corpus linguistics. Apart from genre as noted
previously, for example, keywords in DA refers to words that have a particular
significance in a given discourse. The cultural divide, however, is now diminishing.
McEnery and Wilson (2001: 114) note that there are some important ‘points of
contact’ between DA and corpus linguistics: the common computer-aided analytic
techniques, and the great potential of standard corpora in DA as control data. Because
the corpus-based approach tends to obscure ‘the character of each text as a text’ and
‘the role of the text producer and the society of which they are a part’ (Hunston 2002:
110), some DA authors have avoided using corpus data. For example, Martin (1999:
52) argues that analyzing a lot of text from a corpus simultaneously would force the
analyst to lose ‘contact with text.’ Yet Stubbs (1997) and de Beaugrande (1999, 2001),
among many others, have insisted that corpora are indeed useful for studies of this
kind.
Specialized corpora are particularly useful in discourse analysis and most of the
recently published studies of ideology and culture are based on specialized corpora.
Political discourse is perhaps the most important and most widely used data in
discourse analysis. This is perhaps because politics is ‘[o]ne area of social life in
which the increasing salience of discourse has been especially apparent’ (Johnson,
Culpeper and Suhr 2003: 41). For example, Sotillo and Starace-Nastasi (1999)
undertook a critical discourse analysis on the basis of a corpus of 123 Letters to the
Editors (LEs) of two weekly newspapers written by candidates for political office,
their supporters, and opponents in an American working class town. They found that
gender and class markers were salient in the discourse of LEs. With regard to class,
there is an antagonistic dialogue between residents of the third ward (working class)
and those of the second and first wards (middle class): middle-class residents of the
first and second wards remain unsympathetic to the concerns of third-ward residents,
especially to their claims of a deteriorating quality of life. With respect to the saliency
of gender in LEs, qualitative differences were found between males and females in
writing style, lexical and syntactic choices, and tone of communication. For example,
men used more qualifiers and intensifiers than women, and women writers of LEs
were often less confrontational and more conciliatory than their male counterparts in
their criticism of those in power.
Teubert (2000) studied the language of Euroscepticism in Britain on the basis of
a corpus of texts downloaded from websites which take an antagonistic attitude
towards the European Union. Corpus analysis techniques like collocation and
phraseology enabled Teubert to make explicit what was implied but not stated by
Eurosceptics: only Britain in the whole of Europe is a true democracy with a truly
accountable government (Teubert 2000: 76-77). Similarly, Fairclough’s (2000)
comparative analysis of keywords (in the sense as used in corpus linguistics) in a
corpus of the British Prime Minister Blair’s speeches and other documents from New
Labour and a corpus of documents from Old Labour Party showed that the party has
changed its ideology, as reflected by its language.
Johnson, Culpeper and Suhr (2003) explored discourses of political correctness
(‘PC’) in a corpus of articles gathered from three broadsheet newspapers in the UK
between 1994 and 1999. Their frequency and (statistically defined) keyword analyses
showed that while the overall frequency of so-called ‘PC’-related terms (‘political
correctness’, ‘politically correct’, etc.) generally declined in the five-year period, there
was an interesting link between the frequency of such terms and the ways in which
they have been drawn upon as a means of framing debates over Blair and the Labour
Party throughout the period in question.
Saraceni (2003) analyzed two corpora of interviews and speeches related to the
war in Iraq in an attempt ‘to understand the extent to which, at least in linguistic terms,
the ideas of Blair and Bush may not be as alike as one might be tempted to believe.’
His analysis revealed some important differences in the ways in which Blair and Bush
use language and in what they actually say. While Bush’s rhetoric is typically right
wing, Blair’s discourse is more enigmatic, lacking many of the characteristics of
right-wing rhetoric but not typical of left-wing rhetoric either (Saraceni 2003: 12).
The marked contrast between words and actions in this case is a good example of a
complex issue which the corpus-based approach alone cannot resolve.
Partington (2003) provides a full, corpus-based account of the discourse of
White House press briefing, in an attempt ‘to show how it is possible to use
concordance technology and the detailed linguistic evidence available in corpora to
enhance the study of the discourse features of a particular genre of the language’
(Partington 2003: 3). The major corpus resource used by him is a corpus consisting of
48 briefings, amounting to approximately 250,000 words. The work presented in
Partington (ibid) represents an unusual contribution to corpus-based discourse
analysis because a large part of the book is devoted to devising ‘a suitable
methodology to study features of interaction in large bodies of texts, in corpora’
(Partington 2003: 5). Such methodologies are particularly important in the context of
most studies in discourse analysis undertaken so far having been based on corpora of
a number of single texts (e.g. Pardo 2001).
In addition to political discourse, corpora have been used in analyzing a number
of other types of discourse, for example, academic discourse (e.g. Piper 2000),
business discourse (e.g. Koller 2004), everyday demotic discourse (Carter and
McCarthy 2004), legal discourse (e.g. Graham 2001), media discourse (e.g. Downs
2002; Moore 2002; Pan 2002; Page 2003), medical discourse (e.g. Salager-Meyer,
Ariza and Zambrano 2003), and workshop discourse (e.g. Holmes and Marra 2002).
The works reviewed so far are all based mainly on specialized corpora, though
some of them (e.g. Piper 2000; Johnson, Culpeper and Suhr 2003; Partington 2003)
have used general corpora such as the BNC for comparative purposes. In contrast,
there has been far less work in discourse analysis that is based directly on general
corpora. There are a number of reasons for this. First, most discourse analysts prefer
to study whole texts – general corpora are typically composed of samples. Second,
with a few exceptions (e.g. the BNC), most general corpora have not encoded
variables required for discourse analysis (e.g. metadata relating to the language
producer). Third, most general corpora have not included spoken data for spoken
discourse analysis yet, as Partington (2003: 262) observes, some linguists only
consider spoken language as discourse. Finally, the field of discourse analysis has
historically been accustomed to analyzing a small number of single texts whereas
general corpora provide a much larger number of texts. There are, however, a number
of studies which are based on general corpora. For example, Stubbs (1996) gives
numerous examples of what he calls ‘cultural keywords’ in the Bank of English; de
Beaugrande (1999) compared the ideologies as reflected by ‘liberal’ and its
derivatives (e.g. ‘liberalism’, ‘liberalization’) in the UK and the US-based corpus
resources as well as in the Corpus of South African English (i.e. CSAE, which was
originally developed as part of the ICE corpus).
In conclusion, while the corpus-based approach to discourse analysis is still in its
infancy, corpora (either specialized or general) do present a real opportunity to
discourse analysis, because the automatic analysis of a large number of texts at one
time ‘can throw into relief the non-obvious in a single text’ (Partington 2003: 7). As
de Beaugrande (1999) comments:
Obviously, the methods for doing a ‘critical discourse analysis’ of corpus data are
far from established yet. Even when we have examined a fairly large set of
attestations, we cannot be certain whether our own interpretations of key items and
collocations are genuinely representative of the large populations who produced
the data. But we can be fairly confident of accessing a range of interpretative
 issues that is both wider and more precise than we could access by relying on our
own personal usages and intuitions. Moreover, when we observe our own
ideological position in contest with others, we are less likely to overlook it or take
it for granted. (de Beaugrande 1999: 287)

10.13 Stylistics and literary studies

Style is closely allied to registers/genres and dialects/language varieties (see unit 14),
because stylistic shifts in usage may be observed with reference to features associated
with either particular situations of use or particular groups of speakers (cf. Schilling-
Estes 2002: 375). In this section, we will consider only what Carter (1999: 195) calls
‘literary language’. Literariness is typically present in, but not restricted to literary
texts. However, given that most work in stylistics focuses upon literary texts, the
accent of this section will fall upon literary studies.
Stylisticians are typically interested in individual works by individual authors
rather than language or language variety as such. Hence while they may be interested
in computer-aided text analysis, the use of corpora in stylistics and literary studies
appears to be limited (cf. McEnery and Wilson 2001: 117). Nevertheless, as we will
see shortly, corpora and corpus analysis techniques are useful in a number of ways:
the study of prose style, the study of individual authorial styles and authorship
attribution, literary appreciation and criticism, teaching stylistics, and the study of
literariness in discourses other than literary texts have all been the focus of corpus-
based study.
As noted in unit 4.4.7, one of the focuses in the study of prose stylistics is the
representation of people’s speech and thoughts. Leech and Short (1981) developed an
influential model of speech and thought presentation, which has been used by many
scholars for literary and non-literary analysis (e.g. McKenzie 1987; Roeh and Nir
1990; Simpson 1993). The model was tested and further refined in Short, Semino and
Culpeper (1996), and Semino, Short and Culpeper (1997). Readers can refer to unit
4.4.7 for a description of the speech and thought categories in the model. Using this
model, Semino, Short and Wynne (1999) studied hypothetical words and thoughts in
contemporary British narratives; Short, Semino and Wynne (2002) explored the
notion of faithfulness in discourse presentation; Semino and Short (2004) provide a
comprehensive account of speech, thought and writing presentation in fictional and
non-fictional narratives.
The corpus-based approach has also been used to study the authorial styles of
individual authors. Corpora used in such studies are basically specialized. For
example, if the focus is on the stylistic shift of a single author, the corpus consists of
their early and later works, or works of theirs that belong to different genres (e.g.
plays and essays); if the focus is on the comparison of different authorial styles, the
corpus then consists of works by the authors under consideration. However, as
Hunston (2002: 128) argues, using a large, general corpus can provide ‘a means of
establishing a norm for comparison when discussing features of literary style.’ The
methodology used in studying authorial styles often goes beyond simple counting;
rather it typically relies heavily upon sophisticated statistical approaches such as
MF/MD (see unit 10.4; e.g. Watson 1994), Principal Component Analysis (e.g.
Binongo and Smith 1999a), and multivariate analysis (or more specifically, cluster
analysis, e.g. Watson 1999; Hoover 2003b). The combination of stylistics and
computation and statistics has given birth to a new interdisciplinary area referred to as
‘stylometry’ (Holmes 1998; Binongo and Smith 1999b), ‘stylometrics’ (Hunston 2002:
128), ‘computational stylistics’ (Merriam 2003), or ‘statistical stylistics’ (Hoover
2001, 2002).
Watson (1994) applied Biber’s (1988) MF/MD stylistic model in his critical
analysis of the complete prose works of the Australian Aboriginal author Mudrooroo
Nyoongah to explore a perceived diachronic stylistic shift. He found that Nyoongah
has shifted in style towards a more oral and abstract form of expression throughout his
career and suggested that this shift ‘may be indicative of Nyoongah’s steadily
progressive identification with his Aboriginality’ (Watson 1994: 280). In another
study of Nyoongah’s early prose fiction (five novels), Watson (1999) used cluster
analysis to explore the notion of involvement, more specifically eventuality (certainty
vs. doubt) and affect (positive vs. negative). The analysis grouped Wildcat, Sand and
Doin into one cluster and grouped Doctor and Ghost into another cluster, which
represent two very distinct styles. The first cluster is more affective and representative
of informal, unplanned language, using more certainty adverbs, certainty verbs and
expression of affect; in contrast, the second cluster is more typical of more structured,
integrated discourse, highlighted by a greater use of adjectives, in particular doubt
adjectives and negative affect adjectives, and a very low expression of affect.
Binongo and Smith (1999a, 1999b) applied Principal Component Analysis in
their studies of authorial styles. In Binongo and Smith (1999b), for example, the
authors studied the distribution of 25 prepositions in Oscar Wilde’s plays and essays.
They found that when the plays and essays are brought into a single analysis, the
difference in genre predominates over other factors, though the distinction is not
clear-cut, with a gradual change from plays to essays (Binongo and Smith 1999b:
785-786).
In addition to stylistic variation, authorship attribution is another focus of literary
stylistics. In a series of papers published in Literary and Linguistic Computing,
Hoover (2001, 2002, 2003a, 2003b) tested and modified cluster analysis techniques
which have traditionally been used in studies of stylistic variation and authorship
attribution. Hoover (2001) noted that cluster analyses of frequent words typically
achieved an accuracy rate of less than 90% for contemporary novels. Hoover (2002)
found that when frequent word sequences were used instead of frequent words, or in
addition to them, in cluster analyses, the accuracy often improved, sometimes
drastically. In Hoover (2003a), the author compared the accuracies when using
frequent words, frequent sequences and frequent collocations, and found that cluster
analysis based on frequent collocations provided a more accurate and robust method
for authorship attribution. Hoover (2003b) proposed yet another modification to
traditional authorship attribution techniques to measure stylistic variation. The new
approach takes into consideration locally frequent words, a modification which is
justified when one considers that authorship attribution focuses on similarities
persisting across differences whereas the study of style variation focuses on variations
of authorial style. Lexical choice is certainly part of authorial style. The modified
approach has achieved improved results on some frequently studied texts, including
Orwell’s 1984 and Golding’s The Inheritors. Readers can refer to Haenlein (1999) for
a full account of the corpus-based approach to authorship attribution.
Authorship is only one of the factors which affect stylistic variation. Merriam
(2003) demonstrated, on the basis of 14 texts by three authors, that three other factors,
proposed originally by Labbé and Labbé (2003), namely the vocabulary of the period,
treatment of theme and genre, also contributed to intertextual stylistic variation.
Louw (1997: 240) observed that ‘[t]he opportunity for corpora to play a role in
literary criticism has increased greatly over the last decade.’ He reported on a number
of examples from his students’ projects which showed that ‘corpus data can provide
powerful support for a reader’s intuition’ on the one hand while at the same time
providing ‘insights into aspects of “literariness”, in this case the importance of
collocational meaning, which has hitherto not been thought of by critics’ (Louw 1997:
247). Likewise, Jackson (1997) provided a detailed account of how corpora and
corpus analysis techniques can be used in teaching students about style.
While we have so far been concerned with literary texts, literariness is not
restricted to literature, as noted at the beginning of this section. Carter (1999)
explored, using the CANCODE corpus, the extent to which typically non-literary
discourses like everyday conversation can display literary properties. He concluded
that:
The opposition of literary to non-literary language is an unhelpful one and the
notion of literary language as a yes/no category should be replaced by one which
sees literary language as a continuum, a cline of literariness in language use with
some uses of language being marked as more literary than others. (Carter 1999:
207)

10.14 Forensic linguistics

The final example of the use of corpora and corpus analysis techniques which we will
consider in this section is forensic linguistics, the study of language related to court
trials and linguistic evidence. This is perhaps the most applied and exciting area
where corpus linguistics has started to play a role because court verdicts can very
clearly affect people’s lives. Corpora have been used in forensic linguistics in a
number of ways, e.g. in general studies of legal language (e.g. Langford 1999; Philip
1999) and courtroom discourses (e.g. Stubbs 1996; Heffer 1999; Szakos and Wang
1999; Cotterill 2001), and in the attribution of authorship of linguistic evidence. For
example, such texts as confession/witness statements (e.g. Coulthard 1993) and
blackmail/ransom/suicide notes related to specific cases (e.g. Baldauf 1999) have
been studied. Corpora have also been used in detecting plagiarism (e.g. Johnson 1997;
Woolls and Coulthard 1998).
Legal language has a number of words which either say things about doing and
happening (e.g. intention and negligence) or refer to doing things with words (e.g.
AGREE and PROMISE). Such key words are central to an understanding of the law but
are often defined obscurely in statutes and judgments. Langford (1999) used corpus
evidence to demonstrate how the meanings of words such as intention, recklessness
and negligence can be stated simply and clearly in words that anyone can understand.
When L2 data is involved, defining legal terms becomes a more challenging task.
Dictionaries are sometimes unhelpful in this regard. Philip (1999) showed how
parallel corpora, in this case, a corpus of European Community directives and
judgements, could be used to identify actual translation equivalents in Italian and
English.
Courtroom discourses are connected to the ‘fact-finding’ procedure, which
attempts to reconstruct reality through language, e.g. prosecutor’s presentation, the
eyewitness’s narratives, the defendant’s defence, and the judge’s summing up. As
people may choose to interpret language in different ways according to their own
conventions, experiences or purposes, the same word may not mean the same thing to
different people. Unsurprisingly, the prosecutor and the defendant produce conflicting
accounts of the same event. While the judge’s summing up and the eyewitness’s are
supposed to be impartial, studies show that they can also be evaluative.
 Stubbs (1996) gave an example based on his own experience in analyzing a
judge’s summing up in a real court case, which involved a man being accused of
hitting another man. The judge’s summing up used a number of words that had a
semantic preference for anger, e.g. aggravated, annoyed, irritation, mad and temper.
The judge also quoted the witness who claimed to have been hit, using the word
reeled four times. The word reeled was used with reference to the person being hit
falling backwards after he had allegedly been assaulted. If we look at how the word
REEL is used in the BNC, we can see that it is often used to connote violence or
confusion due to some sort of outside force. The word carries an implication that the
man was struck or pushed quite violently and is therefore likely to be remembered by
the jury because of the number of times it was repeated by the judge (who, being the
most important person in the court room, holds a lot of power and may be assumed to
be able to influence people), and because it paints quite a dramatic picture. Another
unusual aspect of the judge’s speech was his use of modal verbs, which are used
typically to indicate possibility or give permission. The judge used two modal verbs in
particular, may and might, a total of 31 times in his speech and the majority of these
occurred in phrases such as you may think that, you may feel, you may find and you
may say to yourselves… Stubbs found that only three of these could truly be said to
indicate possibility. In the other cases it was used to signal what the judge actually
thought about something. Given the importance of the judge in the courtroom, the
implication of phrases such as you may think can become ‘it would be reasonable or
natural for you to think that…’ or even ‘I am instructing you to think that…’.
Supported by corpus evidence, Stubbs claimed that in a number of ways, the judge
was using linguistic strategies in order to influence the jury.
While the court imposes severe constraints on the witness’s right to evaluate in
their narratives, the overall evaluative point of the narration is perhaps most clear in
this context. Heffer (1999) explored, on the basis of a small corpus of eyewitness
accounts in the trial of Timothy McVeigh, the ‘Oklahoma Bomber’, some of the
linguistic means by which lawyer and witness cooperate in direct examination to
circumvent the law of evidence and convey evaluation. He found that while witnesses
seldom evaluate explicitly, a combination of a careful examination strategy and
emotional involvement can result in highly effective narratives replete with evaluative
elements.
Cotterill (2001) explored the semantic prosodies in the prosecution and the
defence presented by both parties in the O. J. Simpson criminal trial, drawing upon
data from the Bank of English. The prosecution repeatedly exploited the negative
semantic prosodies of such terms as ENCOUNTER, CONTROL and cycle of in order to
deconstruct the professional image of Simpson as a football icon and movie star
wishing to ‘expose’ the other side of Simpson. Cotterill found that in the Bank of
English, ENCOUNTER typically refers to an inanimate entity and collocates with such
words as prejudice, obstacles, problems, a glass ceiling (used metaphorically to refer
to a barrier in one’s career), hazards, resistance, opposition, risks, all of which are
negative. The modifiers of resistance (stiff) and opposition (fierce) also indicate
violence. An analysis of the agents and objects of CONTROL in the corpus was also
revealing. Corpus evidence shows that the typical agents of CONTROL are authority
figures or representatives from government or official bodies (e.g. police), while the
objects of CONTROL often refer to something bad or dangerous (e.g. chemical weapons,
terrorist activities). It appears then, in this context, that CONTROL is legitimate only
when the controller has some degree of authority and when what is controlled is bad
or dangerous. Cotterill (2001: 299) suggested that the prosecutor was constructing
Simpson as a man who was entirely unjustified and unreasonable, and excessively
obsessed with discipline and authority. Another group of collocates of CONTROL in the
corpus refers to various emotional states or conditions. But in this context, it appears
that women tend to control their emotions while men tend to control their temper. In
this way, Simpson was portrayed as a violent and abusive husband who finally lost his
temper and murdered his emotionally vulnerable wife. The corpus shows that cycle of
collocates strongly with negative events and situations (e.g. violence and revenge
killings), and cycles tend to increase in severity over a long period of time. These two
characteristics were just what the prosecutor believed the Simpson case displayed
(Cotterill 2001: 301). The defence attorney, on the other hand, attempted to minimize
and neutralize the negative prosodies evoked by the prosecution through a series of
carefully selected lexical choices and the manipulation of semantic prosodies in his
response. For example, he repeatedly conceptualized Simpson’s assaults as ‘incidents’
(a relatively more neutral term), and used a series of verbal process nominalizations
(i.e. dispute, discussion and conversation) in his defence statement. Incidents only
occur at random rather than systemically. The Bank of English shows that at the top
of the collocate list of incident (MI, 4:4 window) is unrelated. The defence attorney
used the term incident to de-emphasize the systematical nature of Simpson’s attacks
and imply that Simpson only lost control and beat his wife occasionally and that these
events were unrelated. Nominalization not only de-emphasized Simpson’s role by
removing agency from a number of references to the attacks, it also turned a violent
actional event into a non-violent verbal event.
In the fact-finding procedure of court trials, the coherence of the defendant’s
account is an important criterion which may be used to measure its reliability. Szakos
and Wang (1999) presented a corpus-based study of coherence phenomena in the
investigative dialogues between judges and criminals. Their study was based on the
Taiwanese Courtroom Spoken Corpus, which includes 30 criminal cases with 17
different types of crimes. The authors demonstrated that word frequency patterns and
concordancing of corpus data could assist judges in finding out the truth and arriving
at fair judgments.
Another important issue in legal cases is to establish the authorship of a
particular text, e.g. a confession statement, a blackmail note, a ransom note, or a
suicide note. We have already discussed authorship attribution of literary texts (see
unit 10.13). The techniques used in those contexts, such as Principal Component
Analysis and cluster analysis, however, are rarely useful in forensic linguistics,
because the texts in legal cases are typically very short, sometimes only a few
hundred words. The techniques used in forensic linguistics are quite different from
those for authorship attribution of literary texts. Forensic linguists often rely on
comparing an anonymous incriminated text with a suspect’s writings and/or data from
general corpora.
Baldauf (1999), for example, reported on the work undertaken at the ‘linguistic
text analysis’ section of the Bundeskriminalamt (BKA) in Wiesbaden, Germany,
which has been dealing with the linguistic analysis of written texts, mainly in serious
cases of blackmail, for more than ten years. During this time a method has been
established that consists partly of computer-assisted research on a steadily growing
corpus of more than 1,500 authentic incriminated texts and partly of ad-hoc, case-
specific linguistic analysis.
Perhaps the most famous example of authorship attribution in forensic linguistics
is the case of Derek Bentley, who was hanged in the UK in 1953 for allegedly
encouraging his young companion Chris Craig to shoot a policeman. The evidence
that weighed against him was a confession statement which he signed in police
custody but later claimed at the trial that the police had ‘helped’ him produce.
Coulthard (see 1993, 1994) found that in Bentley’s confession, the word then was
unusually frequent: it occurred 10 times in his 582-word confession statement,
ranking as the 8th most frequent word in the statement. In contrast, the word ranked
58th in a corpus of spoken English, and 83rd in the Bank of English (on average once
every 500 words). Coulthard also examined six other statements, three made by other
witnesses and three by police officers, including two involved in the Bentley case.
The word then occurred just once in the witnesses’ 930-word statements whereas it
occurred 29 times – once in every 78 words in the police statements. Another
anomaly Coulthard noticed was the position of then. The sequence subject + then (e.g.
I then, Chris then) was unusually frequent in Bentley’s confession. For example, I
then occurred three times (once every 190 words) in his statement. In contrast, in a
1.5-million-word corpus of spoken English, the sequence occurred just nine times
(once every 165,000 words). No instance of I then was found in ordinary witness
statements, but nine occurrences were found in the police statement. The spoken data
in the Bank of English showed then I was ten times as frequent as I then. It appeared
that the sequence subject + then was characteristic of the police statement. Although
the police denied Bentley’s claim and said that the statement was a verbatim record of
what Bentley had actually said, the unusual frequency of then and its abnormal
position could be taken to be indicative of some intrusion of the policemen’s register
in the statement. The case was re-opened in 1993, 40 years after Derek was hanged.
Malcolm Coulthard, a forensic linguist, was commissioned to examine the confession
as part of an appeal to get a posthumous pardon for Derek Bentley by his family. The
appeal was initially rejected by the Home Secretary; but in 1998, another court of
appeal overthrew the original conviction and found Derek Bentley innocent. In 1999
the Home Secretary awarded compensation to the Bentley family.
An issue related to authorship distribution in forensic linguistics is plagiarism,
which is sometimes subject to civil or criminal legal action, and in the context of
education, subject to disciplinary action. Corpus analysis techniques have also been
used in detecting plagiarism. For example, Johnson (1997) carried out a corpus-based
study in which she compared lexical vocabulary and hapaxes (i.e. words that occur
only once) in student essays suspected of plagiarism in order to determine whether
those essays had been copied. Woolls and Coulthard (1998) demonstrated how a
series of corpus-based computer programs could be used to analyze texts of doubtful
or disputed authorship.
Readers can refer to Coulthard (1994), Heffer (1999) and Kredens (2000) for
further discussion of the use of corpora in forensic linguistics. While forensic
linguistics is a potentially promising area in which corpora can play a role, it may take
some time to persuade members of the legal profession to accept forensic linguistic
evidence. Yet in real life cases, Coulthard’s testimony helped to bring a happy ending
to the Bentley case. Other cases have been less successful, however. Stubbs’s
evidence against the judge’s biased summing up was not accepted by the Lord Chief
Justice who looked at the appeal. But whatever initial outcomes, forensic linguistics
needs to demonstrate that it can indeed arrive at correct answers so that the discipline
can gain more credibility. For this, more experimental tests need to be carried out
where linguists are given problems to solve where the answer is already known by an
independent judge.

10.15 What corpora cannot tell us

We have so far reviewed the use of corpora and corpus analysis techniques in a wide
range of areas of language studies. This review might give the misleading impression
that corpora are all-powerful and capable of solving all sorts of language problems.
But in fact, they are not. This section will briefly discuss a number of limitations of
the corpus-based approach to language studies. We will return to discuss the pros and
cons of using corpora in unit 12. For the moment, let us review the problems with
using corpora that we have noted so far.
First, corpora do not provide negative evidence. This means that they cannot tell
us what is possible or not possible. Everything included in a corpus is what language
users have actually produced. A corpus, however large or balanced, cannot be
exhaustive except in a very limited range of cases. Nevertheless, a representative
corpus can show what is central and typical in language.
Second, corpora can yield findings but rarely provide explanations for what is
observed. These explanations must be developed using other methodologies,
including intuition.
Third, the use of corpora as a methodology also defines the boundaries of any
given study. As we have emphasized throughout the book, the usefulness of corpora
in language studies depends upon the research question being investigated. As
Hunston (2002: 20) argues, ‘They are invaluable for doing what they do, and what
they do not do must be done in another way.’ It is also important, as will be seen in
units 13-16 of Section B as well as in Section C of this book, that readers learn how to
formulate research questions amenable to corpus-based investigation.
Finally, it is important to keep in mind that the findings based on a particular
corpus only tell us what is true in that corpus, though a representative corpus allows
us to make reasonable generalizations about the population from which the corpus
was sampled. Nevertheless, unwarranted generalizations can be misleading.
The development of the corpus-based approach as a tool in language studies has
been compared to the invention of telescopes in astronomy (Stubbs 1996: 231). If it is
ridiculous to criticize a telescope for not being a microscope, it is equally pointless to
criticize the corpus-based approach for not doing what it is not intended to do (Stubbs
1999).