| Bouillon, Pierrette and Federica Busa, ed. (2001) The Language of Word Meaning, Cambridge University Press, hardback, 387 pp., ISBN 0 521 78048 9, Studies in Natural Language Processing
Stella Markantonatou, Institute for Language and Speech Processing, Athens, Greece
This interesting book is an edited collection of papers having the Generative Lexicon Theory [JP1995] as a common thread. The book is divided into four Sections each one brilliantly introduced by the editors. The first Section, which is titled ''Linguistic Creativity and the Lexicon'', consists of four papers. The central question is: How does a theory of language explain the fact that words and utterances are used by humans in various contexts in a variety of senses? James Pustejovsky [JP1995] has introduced the concept of Polysemy and proposed an artificial language for representing it. The contributors to the first Section raise various important issues stemming from the study of this phenomenon and criticize Pustejovsky's approach.
The papers are:
(1) James McGilvray, Chomsky on the Creative Aspect of Language Use and Its Implications for Lexical Semantic Studies. Which is the object of study of the science of linguistics and what are the implications for lexical semantics? The author argues that a Science of Meaning can only be an Internalistic Syntactic Science and that referential approaches will never manage to make sense of the fact that humans produce unlimited linguistic output which is (almost always) appropriate to circumstances.
Furthermore, he argues that such a science must adopt an anthropocentric point of view in order to account for the fact that, after all, natural language expressions serve human needs. McGilvray works in the framework of minimalism; Underspecification, which is one of the cornerstones of the Generative Lexicon Theory, has no place in a framework where load on syntax is kept as minimal as possible.
(2) Jerry A. Fodor and Ernie Lepore, The Emptiness of the Lexicon: Critical Reflections on J. Pustejovsky's ''The Genarative Lexicon''. The authors favor a denotational approach to the Lexicon. They argue that entries in the Lexicon should not contain inferences. They strongly criticize Pustejovsky's use of the notion of Polysemy. They do not believe that words are semantically underspecified nor that their semantics is fixed in context. They argue that what eventually varies with, say, verbs, is not the semantics of the verb but the semantics of the VP headed by the particular verb. Therefore, they favor the listing of words/senses and the use of rules, which they conceive as a word's dowry, to determine the logical form of the phrases which are headed by the particular word.
(3) James Pustejovsky, Generativity and Explanation in Semantics: A Reply to Fodor and Lepor. The author defends Polysemy as (one of) the linguistic manifestation(s) of the faculty for generative categorization and compositional thought. This faculty, he says, is unique to humans. Then, he argues that Polysemy may be understood as the result of generative mechanisms, very much in the way syntax has been understood within the generative tradition. Such an approach will allow for (a) an explanation of the logical relations that hold among the various senses of polysemous words (2) the use of generative devices in order to express generalizations about the creative use of the words in terms of important semantic distinctions. In the more technical parts of the paper, the author also explains how the machinery he proposes adequately represents Polysemy phenomena and related generalizations.
(4) Yorick Wilks, The ''Fodor''-FODOR Fallacy Bites Back. The author argues that the Fodor & Lepore denotational semantics is not adequate both from the point of view of Artificial Intelligence (which is very much concerned with the manipulation of representations) and of a theory of the Language. The reason, he explains, is that the Fodor & Lepore approach does not actually contribute any information about the meaning of words. That is, in the Fodor & Lepor approach, according to the author, the lexicon is a simple listing of arbitrary strings with no information about their semantics. Instead, the meaning of the words may be understood as relationships among representational structures and the world as a whole.
The second Section, which is titled ''The Syntax of Word Meaning'', is mainly a technical one. The general principles of the upgraded Generative Lexicon Theory are introduced and then, the machinery is used to model a large range of frequent and challenging phenomena. An interesting discussion is carried out concerning the central claim of the Generative Lexicon Theory, namely, that there are general mechanisms which account for the multiple semantic interactions between words and their contexts. More particularly, the Section consists of seven papers as follows:
(5) James Pustejovsky, Type Construction and the Logic of Concepts. Here, certain innovations to the representational language of the Generative Lexicon Theory are described. More particularly, the concepts and mechanisms to construct an ontology for the representation of word meanings are explained. Type-coercion operations, which form a central mechanism for accounting for creativity within this theory, are formalised into a handful of well-defined types of operation.
(6) Jaques Jayez, Underspecification, Context Selection, and Generativity. This paper questions one of the central claims of Generative Lexicon Theory, namely that the variety of semantic relations between words and their contexts can be dealt with a number of generalised operations. These operations rely on the assumption that words are underspecified and the missing information is provided by the context. Evidence is brought from French (concerning the behavior of predicates such as ''attendre'' and ''suggester'') which supports the claim that words select for the context they appear in while a considerable amount of vagueness does not allow for an exact interpretation of the semantics of the resulting structures. There are at least two consequences of this fact: (a) the semantics of the resulting structure cannot be fully predicted (b) words tend to appear in restricted contexts. Such a simultaneously idiomatic and vague behavior can not be captured by representational languages which are based on generative mechanisms.
(7) Pierrette Bouillon and Federica Busa, Qualia and the Structuring of Verb Meaning. This paper takes over the discussion introduced in the previous paper and offers a detailed analysis of the behavior of the verb ''attendre''. It defends the view that it is possible to encode the challenging phenomena under question (and some more phenomena) by exploiting underspecification and the generative mechanisms made available by the Generative Lexicon Theory.
(8) Patrick Saint-Dizier, Sense Variation and Lexical Semantics Generative Operations. This paper suggests possible extensions to the machinery available to QUALIA for the case of adjectives and verbs. The importance of the TELIC role is stressed while it is suggested that a number of generalisations concerning sense variation can better be encoded with rules for type shifting rather than with general operations (as the Generative Lexicon Theory suggests).
(9) Salvador Climent, Individuation by Partitive Constructions in Spanish. Partitive constructions offer a way to individuate referents in a number of languages, Spanish included. Because it is the case that some kind of mutual selection between the partitive expression and the complement noun exists, the semantics of partitive constructions are better obtained with the co-composition mechanism of the Generative Lexicon Theory. The quale CONSTITUTIVE is shown to play an important role here.
(10) Laurence Danlos, Event Coreference in Causal Discourses. The extended event structure for causative verbs proposed in the Generative Lexicon Theory is exploited to study the semantics of discourses expressing direct causation. An event coreference relation is shown to exist in a discourse where the result is expressed by a causative verb in its transitive use. Two types of event coreference are claimed to exist: generalisation and particularisation. These notions are further used to formulate two demonstrably valid hypotheses about the structure and the interpretation of causal discourses.
The third Section, titled ''Interfacing the Lexicon'', discusses the possibility of treating ''deviant'' word uses, such as metaphors, with a generative mechanism, e.g. the one proposed by the Generative Lexicon Theory. The first three papers seem to agree that, to some extend, phenomena such as metaphor and metonymy can be captured by some combination of generative mechanisms. The fourth paper strongly questions this approach. A short description of the papers follows:
(1) Julius M. Moravcsik, Metaphor: Creative Understanding and the Generative Lexicon. The paper stresses that the metaphorical use of words preserves some important features of their literal use but not all of them. Words, when used metaphorically, can be thought to have underspecified semantics as compared to the semantics of their literal uses. The semantics of metaphorical structures, although it is about the world, is not compositionally derived. Instead, an element of subjectivity plays an important role and a holistic view is adopted. The author proposes a theory for representing lexical semantics. In this theory, lexical meanings are conceived of as functions of four specified factors (an idea which is close enough to the QUALIA role of the
Generative Lexicon Theory).
(2) Nicholas Asher and Alex Lascarides, Metaphor in Discourse. The paper offers a robust formalisation of metaphor phenomena, selecting the French and English Change of Location verbs as a case study. It divides the load of the interpretation of such structures between a lexical semantics component and a formal discourse semantics one. Lexical semantics is formalised with typed feature structures and inheritance and the formalisation is compatible with the HPSG framework. The QUALIA role is exploited. Formal discourse semantics is used to illustrate the interaction between lexical semantics and certain rhetorical relations such as Narration and Contrast. The main idea here is that words, in their metaphoric use, preserve those features that distinguish them in the word class they belong to.
(3) Jerry Hobbs, Syntax and Metonymy. The author relies on the Interpretation as Abduction framework to show that several phenomena, which are often attributed to the Syntax, are, in fact, instances of Metonymy: Metonymy is the phenomenon whereby an entity is referred to with the description of a functionally related entity. He demonstrates the capabilities of the method by offering a treatment of a wide range of phenomena such as: extraposed modifiers, ataxis, container nouns, distributive and collective readings, small clauses, assertion of grammatically subordinated information. At the bottomline, the paper demonstrates the utility of a framework where syntax, semantics and pragmatics are modeled in a uniform fashion.
(4) Adam Kilgarrif, Generative Lexicon Meets Corpus Data: The Case of Nonstandard Word Uses. The author questions the claim that the Generative Lexicon Theory, and any generative theory, can adequately deal with non-standard uses of words. He claims that lexical semantics often is not a matter of individual words but of the collocations they participate to. Furthermore, he claims that a large part of meanings is not computed each time they are used. Instead, these meanings are stored and retrieved when necessary - a claim which challenges the central role of generative mechanisms in the mental lexicon.
The fourth Section, which has the title ''Building Resources'', discusses aspects of the problem of constructing large scale lexica. It draws a lot on the experience of the project ''SIMPLE'' which largely adopted the Generative Lexicon Theory. The main issue discussed in this section is the construction of ontologies. It consists of three papers:
(1) Federica Busa, Nicoletta Calzolari and, Alessandro Lenci, Generative Lexicon and the SIMPLE Model: Developing Semantic Resources for NLP. The authors present the ontology developed in the SIMPLE project. This ontology further elaborates on the one proposed by the Generative Lexicon Theory. The aim is to provide such mechanisms for accommodating words in the ontology that are general enough to allow for creating large lexical resources for natural language processing.
(2) Nilda Ruimy, Elisabetta Gola and, Monica Monachini, Lexicography Informs Lexical Semantics: the SIMPLE Experience. The SIMPLE project developed a set of templates which helped lexicographers to accommodate senses without violating the general design of this multi-lingual lexical database. The authors explain how templates were constructed and select non-concrete nouns as a case study to show that theory and lexicographic practice provide feedback to each other.
(3) Piek Vossen, Condensed Meaning in EuroWordNet. The paper reports on the experience of constructing a multilingual lexical database, which was gained with the EuroWordNet project. Polysemy is accounted for in one language in a generative way by exploiting the hyponymy/hyperonymy relation. A ''global'' ontology is also defined to encode cross-lingual correspondences.
This is an interesting book well-addressed to a large audience in linguistics, computational linguistics and, to my view, philosophy of the language. It is actually a book about the Generative Lexicon Theory [JP1995] although its title (''The language of word meaning'') could be interpreted to allude to a more general discussion about existing artificial languages which have been proposed for the representation of lexical meaning (I will name only [Jackendoff 1990] but there are several more, of course).
However, in this volume, an interesting debate unfolds which eventually concerns serious issues of lexicography and linguistics in general. The debate is between scholars who more or less adopt and defend some central ideas of the Generative Lexicon Theory and scholars who argue against them. Among the hot topics discussed in the book are:
(1) The fact is that humans can successfully use the same words/structures with different contexts and in different situations. How does a theory of the language better capture this fact? In particular, is it at all possible to construct a representational theory for this purpose? Can this theory be similar to syntactic theories, that is, exploit generalisations and general purpose operations to account for this phenomenon? To this question, the Generative Lexicon Theory gives a positive answer. There are several researchers who claim that generalisation-based theories eventually miss out something because words tend to collocate and/or strongly select their contexts while the overall semantics of such structures does not seem to be compositionally derived. It is one of the merits of the book that this debate is well-presented and taken up from various points of view.
(2) Perhaps on the previous line of thought, is it possible to construct an adequate theory of metaphor and if yes, what would the burden be on lexical semantics and what on discourse analysis? The book includes very interesting contributions in this area. Still, before accounting for this issue, how do we know that something is a metaphor and not a literal use of a word? There may not exist an answer to the second question but, certainly there is no convincing answer in the book. Often contributors seem to know a priori which senses are literal and which are metaphorical. Certainly, this is not a book about metaphor but it is a book about Underspecification in the Lexicon which eventually addresses the issue of literal and non-literal uses of words. It would be an additional advantage of the book if a discussion existed about how the line is drawn between these two areas.
(3) Underspecification of lexical entries is a cornerstone of the Generative Lexicon Theory. But is it actually a panacea for representing the multi-faced semantic (and syntactic) behavior of words? The discussion is detailed and very illuminating. At the end of the book, I was convinced that underspecification can be exploited to constrain unwanted interpretations while allowing for a considerable degree of semantic flexibility but, it is not possible to use it in order to account for all meanings. The book is mainly of the theoretical persuasion.
However,it also touches upon the problems that have been faced by enterprises which aim to construct large computational dictionaries. The problem of constructing the ontologies which form the backbone of large scale lexica is discussed. This discussion is mainly located in Section Four but crucial information can also be found elsewhere in the book and especially in the papers by James Pustejovsky. As regards one or two papers in Section Four, I would observe that they look like papers which introduce a theory to an audience poorly informed about the adopted theoretical backbone, for instance, the audience in a conference of computer applications. However, this is a book about the Generative Lexicon Theory after all. Perhaps, those papers could have been further edited so as to keep in the introduction the absolutely relevant theoretical information only.
Last, but not least, my opinion is that the editors must be credited with an overall success (although, unfortunately, several typos have managed to escape the editorial eye) because both the selection of the topics and the overall organisation of the collection have resulted into an informative and stimulating text.
[JP1995] James Pustejovsky. 1995. The Generative Lexicon. MIT Press, Cambridge
[Jackendoff 1990]. Ray Jackendoff. 1990. Semantic Structures. MIT Press, Cambridge
Stella Markantonatou (1958) completed her PhD at the University of Essex in 1992. She has joined the Essex
Computational Linguistics Group for four years and has been in charge of the MT Department of ILSP since June 1998. She has taught in various Universities and has published in international journals and conferences. Her interests are in the areas of grammatical formalisms, lexical semantics and statistical approaches.
Dr. Stella Markantonatou
Head of the Machine Translation Department
Institute for Language and Speech Processing
Artemidos 6 & Epidavrou