Semantic Domains and Linguistic Theory

Alfio Gliozzo
University of Rome Tor Vergata
Department of Computer Science, System and Production
00133 Roma (Italy)
[email protected]

Abstract 1 Introduction

The predominant view in lexical semantic is the

This paper is about the relations between Saussure’s structural semantics (de Saussure, 1922),
the concept of Semantic Domain and the claiming that a word meaning is determined by the
“Theory of Semantic Fields”, a structural “horizontal” paradigmatic and the “vertical” syntag-
model for lexical semantics proposed by matic relations between that word and others in the
Jost Trier at the beginning of the last cen- whole language (Lyons, 1977). Structural assump-
tury. The main limitation of the Trier’s no- tions are also widely adopted in Computational Lin-
tion is that it does not provide an objective guistic. For example, many machine readable dic-
criterion to aggregate words around fields, tionaries describe the word senses by means of se-
making the overall model too vague, and mantic networks representing relations among terms
then unuseful for computational purposes. (e.g. W ORD N ET (Miller, 1990)). The main limita-
The notion of Semantic Domain improves tion of the “radical” structuralist view is that it is al-
that of Semantic Field by providing such most impossible to describe the associations among
a criterion. In particular, the structural- all the possible terms in a natural language, because
ist approach in semantics has been con- the huge number of concepts and semantic relations
nected to the Wittgenstein’s meaning-is- among them.
use assumption, providing an objective
criterion to infer Semantic Domains from The Semantic Fields Theory (Trier, 1931) goes
corpora relying on a lexical coherence as- a step further in the structural approach to lexical
sumption. The task based evaluation we semantics by introducing an additional aggregation
did for our claims shows that the notion level and by delimiting to which extend paradig-
of Semantic Domains is effective because matic relations hold. The basic assumption of this
it allows to define an uniform methodol- theory is that the lexicon is structured into Semantic
ogy to deal with many different Natural Fields: semantic relations among concepts belong-
Language Processing tasks. In the paper ing to the same field are very dense, while concepts
we discuss the epistemological issues con- belonging to different fields are typically unrelated.
cerning the possibility of adopting a task In fact, a word meaning is established only by the
based methodology to support linguistic network of relations among the terms of its field.
theory, showing the case study of Seman- Another property of great interest is that there exists
tic Domains in Computational Linguistics a strong correspondence among Semantic Fields of
as a paradigmatic example for our claims. different languages, while such a strong correspon-
dence cannot be established among the terms them-
selves.
 It has been observed that the main limitation of Wittgenstein’s meaning-is-use assumption. Section
the Trier’s notion is that it does not provide an objec- 4 describes the concept of Semantic domains as
tive criterion to aggregate words around fields, mak- the confluence of both perspectives, highlighting its
ing the overall model too vague, and then unuseful technological impact in developing state-of-the-art
for computational purposes. The notion of Seman- systems for NLP, while Section 5 conclude the pa-
tic Domain improves that of Semantic Field by pro- per discussing the possibility of adopting the indirect
viding such a criterion. In particular, the structural- task based evaluation to support linguistic theory.
ist approach in semantics has been connected to the
meaning-is-use assumption introduced by Ludwig 2 The Theory of Semantic Fields
Wittgenstein in his celebrated “Philosophical Inves- Semantic Domains are a matter of recent interest in
tigations” (Wittgenstein, 1965). A word meaning is Computational Linguistics (Magnini and Cavaglià,
its use into the concrete “form of life” where it is 2000; Magnini et al., 2002; Gliozzo et al., 2005a),
adopted, i.e. the linguistic game, in the Wittgen- even though their basic assumptions are inspired
stein’s terminology. Frequently co-occurring words from a long standing research direction in structural
in texts are then associated to the same linguistic linguistics started in the beginning of the last cen-
game. It follows that fields can be identified from a tury and widely known as “The Theory of Semantic
corpus based analysis of the lexicon, exploiting the Fields” (Lyons, 1977). The notion of Semantic Field
connections between linguistic games and Seman- has proved its worth in a great volume of studies,
tic Fields already depicted. The notion of Seman- and has been mainly put forward by Jost Trier (Trier,
tic Domain arises from this convergence, providing 1931), whose work is credited with having “opened
an objective criterion to identify semantically related a new phase in the history of semantics”(Ullmann,
words in texts, supported by a lexical coherence as- 1957).
sumption, that we empirically corroborated in text in In that work, it has been claimed that the lexicon
the earlier stages of our work. is structured in clusters of very closely related con-
The notion of Semantic Domain is intimately re- cepts, lexicalized by sets of words. Word senses are
lated to several phenomena in the language at both a determined and delimitated only by the meanings of
lexical and a textual level. At a lexical level Seman- other words in the same field. Such clusters of se-
tic Domains can be used as a (shallow) model for mantically related terms have been called Semantic
lexical ambiguity and variability, while at a textual Fields1 , and the theory explaining their properties is
level semantic domains provide meaningful topic known as “The theory of Semantic Fields” (Vassi-
taxonomies that can be used to group texts into se- lyev, 1974).
mantic clusters. In addition, the inherent multilin- Semantic Fields are conceptual regions shared out
gual nature of semantic domains allows an uniform amongst a number of words. Each field is viewed as
representation of both the lexicon and the texts in a partial region of the whole expanse of ideas that is
most of the natural languages. covered by the vocabulary of a language. Such ar-
Exploiting Semantic Domains for Natural Lan- eas are referred to by groups of semantically related
guage Processing (NLP) allowed us to improve sen- words, i.e. the Semantic Fields. Internally to each
sibly the state-of-the-art in all those tasks in which field, a word meaning is determined by the network
they have been applied, providing and indirect evi- of relations established with other words.
dence to support their linguistic properties. The ma- There exists a strong correspondence among Se-
jor goal of this paper is to discuss the possibility of mantic Fields of different languages, while such a
adopting such a task based methodology to support strong correspondence cannot be established among
linguistic theory, showing the case study of Seman- the terms themselves. For example, the field of
tic Domains in computational linguistics as a suc- C OLORS is structured differently in different lan-
cessfully paradigmatic example of our methodology.
1
The paper is structured as follows. Section 2 is There is no agreement on the terminology adopted by dif-
ferent authors. Trier uses the German term wortfeld (literally
about the Semantic Fields Theory while Section 3 “word field” or ”lexical field” in Lyons’ terminology) to denote
concerns the relations between this theory and the what we call here semantic field.
guages, and sometimes it is very difficult, if not im- objective criterion to characterize Semantic Fields.
possible, to translate name of colors, even whether The solutions we have found in the literature (Weis-
the chromatic spectrum perceived by people in dif- gerber, 1939; Porzig, 1934; Coseriu, 1964) rely on
ferent countries (i.e. the conceptual field) is the very obscure notions, of scarse interest from a com-
same. Some languages adopt many words to de- putational point of view. To overcome such a lim-
note the chromatic range to which the English term itation, in this section we introduce the concept of
white refers, distinguishing among different de- Semantic Domain (see Section 4).
grees of “whiteness” that have not a direct transla- The notion of Semantic Domain improves that of
tion in English. Anyway, the chromatic range cov- Semantic Fields by connecting the structuralist ap-
ered by the C OLORS fields of different languages is proach in semantics to the meaning-is-use assump-
evidently the same. The meaning of each term is tion introduced by Ludwig Wittgenstein in his cele-
defined in virtue of its oppositions with other terms brated “Philosophical Investigations” (Wittgenstein,
of the same field. Different languages have differ- 1965). A word meaning is its use into the concrete
ent distinctions, but the field of C OLORS itself is a “form of life” where it is adopted, i.e. the linguis-
constant among all the languages. tic game, in the Wittgenstein’s terminology. Words
Another implication of the Semantic Fields The- are then meaningful only if they are expressed into
ory is that words belonging to different fields are ba- concrete and situated linguistic games that provide
sically unrelated. In fact, a word meaning is estab- the conditions for determining the meaning of natu-
lished only by the network of relations among the ral language expressions. To illustrate this concept,
terms of its field. As far as paradigmatic relations are Wittgenstein provided a clarifying example describ-
concerned, two words belonging to different fields ing a very basic linguistic game: “. . . Let us imag-
are then un-related. This observation is crucial form ine a language . . . The language is meant to serve
a methodological point of view. The practical ad- for communication between a builder A and an as-
vantage of adopting the Semantic Field Theory in sistant B. A is building with building-stones; there
linguistics is that it allows a large scale structural are blocks, pillars, slabs and beams. B has to pass
analysis of the whole lexicon of a language, other- the stones, and that in the order in which A needs
wise infeasible. In fact, restricting the attention to a them. For this purpose they use a language consist-
particular field is a way to reduce the complexity of ing of the words block, pillar, slab, beam. A calls
the overall task of finding relations among words in them out; – B brings the stone which he has learnt to
the whole lexicon, that is evidently quadratic in the bring at such-and-such a call. – Conceive of this as a
number of words. complete primitive language.” (Wittgenstein, 1965)
The main limitation of the Trier’s theory is that it We observe that the notions of linguistic game and
does not provide any objective criterion to identify Semantic Field show many interesting connections.
and delimitate Semantic Fields in the language. The They approach the same problem from two different
author himself admits “what symptoms, what char- points of view, getting to a similar conclusion. Ac-
acteristic features entitle the linguist to assume that cording to Trier’s view, words are meaningful when
in some place or other of the whole vocabulary there they belong to a specific Semantic Field, and their
is a field? What are the linguistic considerations that meaning is determined by the structure of the lexi-
guide the grasp with which he selects certain ele- con in the field. According to Wittgenstein’s view,
ments as belonging to a field, in order then to exam- words are meaningful when there exists a linguis-
ine them as a field?” (Trier, 1934). The answer to tic game in which they can be formulated, and their
this question is an issue opened by the Trier’s work. meaning is exactly their use. In both cases, meaning
arises from the wider contexts in which words are
3 Semantic Fields and the meaning-is-use located.
view Words appearing frequently into the same linguis-
tic game are likely to be located into the same field.
In the previous section we have pointed out that the In the previous example the words block, pillar,
main limitation of the Trier’s theory is the gap of an slab and beam have been used in a common lin-
guistic game, while they clearly belong to the Se- provide an objective criterion to distinguish and de-
mantic Field of BUILDING INDUSTRY. This exam- limitate fields from a corpus based analysis of lexi-
ple suggests that the notion of linguistic game pro- cal co-occurences in texts. We refer to this particular
vides a criterion to identify and to delimitate Se- view on Semantic Fields by using the name Seman-
mantic Fields. In particular, the recognition of the tic Domains. The concept of Semantic Domain is
linguistic game in which words are typically formu- the main topic of this work, and it will be illustrated
lated can be used as a criterion to identify classes more formally in the following section.
of words composing lexical fields. The main prob-
lem of this assumption is that it is not clear how to 4 Semantic Domains
distinguish linguistic games between each other. In
fact, linguistic games are related by a complex net- In our usage, Semantic Domains are common areas
work of similarities, but it is not possible to iden- of human discussion, such as E CONOMICS, P OL -
tify a set of discriminating features that allows us ITICS , L AW , S CIENCE , which demonstrate lexical
to univocally recognize them. “I can think of no coherence. The Semantic Domain associated to a
better expression to characterize these similarities particular field is the set of domain specific terms
than ‘family resemblances’; for the various resem- belonging to it, and it is characterized by a set of do-
blances between members of a family: build, fea- main words whose main property is to co-occur in
tures, colour of eyes, gait, temperament, etc. etc. texts.
overlap and criss-cross in the same way. - And I An approximation to domains are Subject Field
shall say: ‘games’ form a family” ((Wittgenstein, Codes, used in Lexicography (e.g. in (Procter,
1965), par. 67). 1978)) to mark technical usages of words. Although
We observe that linguistic games are naturally re- this information is useful for sense discrimination, in
flected in texts, allowing us to detect them from a dictionaries it is typically used only for a small por-
word distribution analysis on a large scale corpus. tion of the lexicon. W ORD N ET D OMAINS (Magnini
In fact, according to Wittgenstein’s view, the content and Cavaglià, 2000) is an attempt to extend the cov-
of any text is located into a specific linguistic game, erage of domain labels within an already existing
otherwise the text itself would be meaningless. Texts lexical database, W ORD N ET (Fellbaum, 1998). As
can be perceived as open windows through which we a result W ORD N ET D OMAINS can be considered an
can observe the connections among concepts in the extension of W ORD N ET in which synsets have been
real world. Frequently co-occurring words in texts manually annotated with one or more domain labels,
are then associated to the same linguistic game. selected from a hierarchically organized set of about
It follows that the set of concepts belonging to a two hundred labels.
particular field can be identified from a corpus based W ORD N ET D OMAINS represents the first attempt
analysis of the lexicon, exploiting the connections to provide an exhaustive systematization of the con-
between linguistic games and Semantic Fields al- cept of Semantic Field and its connections to the
ready depicted. For example, the two words fork textual interpretation depicted in section 3. It al-
and glass are evidently in the same field. A corpus lowed us to start an empirical investigation about
based analysis shows that they frequently co-occur the connections between the textual and the lexical
in texts, then they are also related to the same lin- counterparts of Semantic Domains. First we concen-
guistic game. On the other hand, it is not clear what trated on corroborating a lexical-coherence assump-
would be the relation among water and algorithm, tion, claiming that a great percentage of the concepts
if any. They are totally unrelated simply because expressed in the same text belong to the same do-
the concrete situations (i.e. the linguistic games) in main. Lexical coherence is then a basic property of
which they occur are in general distinct. It reflects most of the texts expressed in any natural language
on the fact that they are often expressed in different and it allows us to disambiguate words in context
texts, then they belong to different fields. by associating domain specific senses to them. Oth-
Our proposal is then to merge the notion of lin- erwise stated, words taken out of context show do-
guistic game and that of Semantic Field, in order to main polysemy, but, when they occur into real texts,
their polysemy is solved by the relations among their the disambiguation process. DDD is performed by
senses and the domain specific concepts occurring in selecting the word sense whose Semantic Domain
their contests. maximize the similarity with the domain of the con-
Intuitively, texts may exhibit somewhat stronger text in which the word is located. For example, the
or weaker orientation towards specific domains, but word virus is ambiguous between its Biology and
it seems less sensible to have a text that is not related Computer Science senses, and can be disam-
to at least one domain. In other words, it is difficult biguated by assigning the correct domain to the con-
to find a “generic” text. This intuition is largely sup- texts where it actually occurs. Results clearly shows
ported by our data: all the texts in SemCor 2 (Landes that domain information is crucial for WSD, allow-
et al., 1998) exhibit concepts belonging to a small ing our system to improve the state-of-the-art for un-
number of relevant domains, demonstrating the do- supervised WSD.
main coherence of the lexical-concepts expressed in The main conclusion of that work was that Se-
the same text. In particular, 34.5 % of nouns in co- mantic Domains play a dual role in linguistic de-
occurring in the same texts in SemCor are annotated scription. One role is characterizing word senses
with the same domain label, while about 40% refer (i.e. lexical-concepts), typically by assigning do-
to generic concepts. The conclusion of this exper- main labels to word senses in a dictionary or lex-
iment is that there exists a strong tendency for the icon. On the other hand, at a text level, Semantic
lexicon in texts to be aggregate around a specific do- Domains are clusters of texts regarding similar top-
main. As we will see later in the paper, such a ten- ics/subjects. They can be perceived as collections
dency should be presupposed to allow lexical disam- of domain specific texts, in which a generic corpus
biguation. is organized. Examples of Semantic Domains at the
Then we investigated the relations between Se- text level are the subject taxonomies adopted to or-
mantic Domains and lexical ambiguity and vari- ganize books in libraries.
ability, the two most basic and pervasive phenom- The generality of these results encouraged us to
ena characterizing lexical semantics. The differ- extend the range of applicability of our assump-
ent senses of ambiguous words should be neces- tions, leading to the definition of a large number of
sarily located into different domains, because they NLP techniques relying on the common theoretical
are characterized by different relations with different framework provided by Semantic Domains in com-
words. On the other hand, variability can be mod- putational linguistics (Gliozzo, 2005). For brevity,
eled by observing that synonymous terms refer to we will not describe into details all these results,
the same concepts, then they will necessarily belong limiting ourselves to enumerate the range of appli-
to the same domain. Thus, the distribution of words cability of domain driven techniques in NLP: Word
among different domains is a relevant aspect to be Sense Disambiguation (Gliozzo et al., 2005b), Text
taken into account to identify word senses. Under- Categorization (Gliozzo and Strapparava, 2005b),
standing words in contexts is mainly the operation of Term Categorization (D’Avanzo et al., 2005), On-
locating them into the appropriate semantic fields. tology Learning (Gliozzo, 2006) and Multilinguality
To corroborate these assumptions we developed (Gliozzo and Strapparava, 2005a).
a Word Sense Disambiguation (WSD) procedure In all those tasks state-of-the-art results have been
relying on domain information only, named Do- achieved by following the common methodology
main Driven Disambiguation (DDD) (Magnini et al., of acquiring Domain Models from texts by means
2001; Gliozzo et al., 2004). The underlying hypoth- of a common corpus based technique, inspired and
esis of the DDD approach is that information pro- motivated by the Trier’s theory and by its connec-
vided by domain labels offers a natural way to es- tion to the meaning-is-use assumption. In particu-
tablish associations among word senses in a certain lar we adopted an approach based on Latent Seman-
text fragment, which can be profitably used during tic Analysis to acquire domain models from corpora
describing the application domain, and we assumed
2
Semcor is a subportion of the Brown corpus annotated by the principal components so acquired be mapped to
WordNet senses. a set of semantic domains. Latent Semantic Analysis
has been performed on a term-by-document matrix fits perfectly this view, preventing us from apply-
capturing only co-occurrency information among ing the traditional linguistic epistemology to com-
terms in texts, with the aim of demonstrating our putational linguistics. In fact, we are interested in
meaning-is-use assumptions. Then we exploited do- exploiting the language in concrete and situated lin-
main based representations to index both terms and guistic games rather than representing it in an inten-
texts, adopting a semi-supervised learning paradigm sional way. From this point of view, the task based
based on kernel methods. Empirical results showed support we have given to our claims is a strong evi-
that domain based representations performs better dence to conclude that Semantic Domains are com-
than standard bag-of-words commonly adopted for putational models for lexical semantics.
retrieval purposed, allowing a better generalization Anyhow, my opinion is just a minor contribu-
over the training data (i.e. improving the learning tion to stimulate a larger epistemological debate in-
curve in all the supervised tasks in which they have volving linguists, cognitive scientists, philosophers,
been applyied), and allowing the definition of hy- computer scientists, engineers, among the others.
brid similarity measures to compare terms and texts, I hope that my research will contribute to stimu-
as expected from the notion of Semantic Domain. late this debate and to find a way to escape from
the “empasse” caused by the vicious distinction be-
5 Conclusion tween empirical and theoretical methods character-
In this paper we explicitly depicted the connections izing the research in computational linguistics in the
between the use of Semantic Domains in NLP and last decade.
the linguistic theory motivating them. Understand-
ing these relations provided us an useful guideline to References
lead our research, leading to the definition of state-
E. Coseriu. 1964. Pour une sémantique diachronique
of-the-art techniques for a wide range of tasks. Hav-
structurale. Travaux de Linguistique et de Littérature,
ing in mind a clear picture of the semantic phenom- 2(1):139–186.
ena we where modeling allowed us to identify the
E. D’Avanzo, A. Gliozzo, and C. Strapparava. 2005. Au-
correct applications, to predict the results of the ex-
tomatic acquisition of domain information for lexical
periments and to motivate them. concepts. In Proceedings of the 2nd MEANING work-
Nonetheless, several questions arises when look- shop, Trento, Italy.
ing at semantic domains from an epistemological
F. de Saussure. 1922. Cours de linguistique générale.
point of view: Payot, Paris.
1. is the concept of semantic domain a computa- C. Fellbaum. 1998. WordNet. An Electronic Lexical
tional theory for lexical semantics? Database. MIT Press.
A. Gliozzo and C. Strapparava. 2005a. Cross language
2. do we have enough empirical evidence to sup- text categorization by acquiring multilingual domain
port our linguistic claims? models from comparable corpora. In n Proceedings
of the ACL Workshop on Building and Using Parallel
3. is the task based evaluation a valid epistemo- Texts.
logical framework to corroborate linguistic the-
A. Gliozzo and C. Strapparava. 2005b. Domain ker-
ory? nels for text categorization. In Proceedings of the
Ninth Conference on Computational Natural Lan-
My personal point of view is that the task based guage Learning (CoNLL-2005), pages 56–63.
evaluation is probably the only objective support we
A. Gliozzo, C. Strapparava, and I. Dagan. 2004. Unsu-
can provide to linguistic theory, and especially to all pervised and supervised exploitation of semantic do-
those issues that are more intimately related to lex- mains in lexical disambiguation. Computer Speech
ical semantics. The basic motivation is that com- and Language, 18(3):275–299.
putational linguistics is also a branch of Artificial A. Gliozzo, C. Giuliano, and C. Strapparava. 2005a. Do-
Intelligence, and then it is subjected to the behav- main kernels for word sense disambiguation. In Pro-
ioral Turing test. The meaning-is-use assumption ceedings of the 43rd annual meeting of the Association
 for Computational Linguistics (ACL-05), pages 403– L. Wittgenstein. 1965. Philosophical Investigations.
410, Ann Arbor, Michigan, June. The Macmillan Company, New York.
A. Gliozzo, C. Giuliano, and C. Strapparava. 2005b. Do-
main kernels for words sense disambiguation. In to
appear in proc. of ACL-2005.
A. Gliozzo. 2005. Semantic Domains in Computational
Linguistics. Ph.D. thesis, University of Trento.
A. Gliozzo. 2006. The god model. In proceedings of
EACL-06.
S. Landes, C. Leacock, and R. I. Tengi. 1998. Building
semantic concordances. In C. Fellbaum, editor, Word-
Net: An Electronic Lexical Database. The MIT Press.
J. Lyons. 1977. Semantics. Cambridge University Press.
B. Magnini and G. Cavaglià. 2000. Integrating subject
field codes into WordNet. In Proceedings of LREC-
2000, Second International Conference on Language
Resources and Evaluation, pages 1413–1418, Athens,
Greece, June.
B. Magnini, C. Strapparava, G. Pezzulo, and A. Gliozzo.
2001. Using domain information for word sense dis-
ambiguation. In Proceedings of SENSEVAL-2: Sec-
ond International Workshop on Evaluating Word Sense
Disambiguation System, pages 111–114, Toulose,
France, July.
B. Magnini, C. Strapparava, G. Pezzulo, and A. Gliozzo.
2002. The role of domain information in word
sense disambiguation. Natural Language Engineer-
ing, 8(4):359–373.
G. Miller. 1990. An on-line lexical database. Interna-
tional Journal of Lexicography, 13(4):235–312.
W. Porzig. 1934. Wesenhafte bedeutungsbeziehungen.
Beiträge zur Geschichte der deutschen Sprache und
Literatur, 58.
Procter. 1978. Longman Dictionary of Contemporary
English.
J. Trier. 1931. Der deutsche Wortschatz im Sinnbezirk
des Verstandes. Heidelberg.
J. Trier. 1934. Das sprachliche feld. eine auseinander-
setzung. Neue Fachrbücher für Wissenschaft und Ju-
gendbildung, 10:428–449.
S. Ullmann. 1957. The Principles of Semantics. Black-
well, Oxford.
L.M. Vassilyev. 1974. The theory of semantic fields: a
survey. Linguistics, 137:79–93.
L. Weisgerber. 1939. Vom inhaltlichen aufbau des
deutschen wortschatzes. Wortfeldforschung, pages
193–225.