Lexical Knowledge Representation and Natural Language Processing

Pustejovsky, J. and B. Boguraev, Lexical knowledge representation and natural language processing, Artificial Intelligence 63 (1993) 193-223. Traditionally, semantic information in computational lexicons is limited to notions such as selectional restrictions or domain-specific constraints, encoded in a "static" representation. This information is typically used in natural language processing by a simple knowledge manipulation mechanism limited to the ability to match valences of structurally related words. The most advanced device for imposing structure on lexical information is that of inheritance, both at the object (lexical items) and meta (lexical concepts) levels of lexicon. In this paper we argue that this is an impoverished view of a computational lexicon and that, for all its advantages, simple inheritance lacks the descriptive power necessary for characterizing fine-grained distinctions in the lexical semantics of words. We describe a theory of lexical semantics making use of a knowledge representation framework that offers a richer, more expressive vocabulary for lexical information. In particular, by performing specialized inference over the ways in which aspects of knowledge structures of words in context can be composed, mutually compatible and contextually relevant lexical components of words and phrases are highlighted. We discuss the relevance of this view of the lexicon, as an explanatory device accounting for language creativity, as well as a mechanism underlying the implementation of open-ended natural language processing systems. In particular, we demonstrate how lexical ambiguity resolution--now an integral part of the same procedure that creates the semantic interpretation of a sentence itself--becomes a process not of selecting from a pre-determined set of senses, but of highlighting certain lexical properties brought forth by, and relevant to, the current context. Correspondence to: J. Pustejovsky, Computer Science Department, Brandeis University, Waltham, MA 02254, USA. E-mail: [email protected]. 0004-3702/93/$ 06.00 ~) 1993 Elsevier Science Publishers B.V. All rights reserved 194 J. Pustejovsky, B. Boguraev 1. Inadequacies of lexical representations In this paper we introduce a theory of computational lexical semantics making use of a knowledge representation framework which offers a rich and expressive vocabulary for lexical information. The motivations for this work are at least twofold. Overall, we are concerned with explaining the creative use of language; we consider the lexicon to be the key repository holding much of the information underlying this phenomenon. More specifically, however, it is the notion of a constantly evolving lexicon that we are trying to emulate; this is in contrast to currently prevalent views of static lexicon design, where the set of contexts licensing the use of words is determined in advance, and there is no formal mechanism offered for expanding this set. The traditional organization of lexicons in natural language processing (NLP) systems assumes that word meaning can be exhaustively defined by an enumerable set of senses per word. Computational lexicons, to date, generally tend to follow this organization. As a result, whenever natural language interpretation tasks face the problem of lexical ambiguity, a particular approach to disambiguation is warranted. The system attempts to select the most appropriate "definition" available under the lexical entry for any given word; the selection process is driven by matching sense characterizations against contextual factors. One disadvantage of such a design follows from the need to specify, ahead of time, the contexts in which a word might appear; failure to do so results in incomplete coverage. Furthermore, dictionaries and lexicons currently are of a distinctly static nature: the division into separate word senses not only precludes permeability; it also fails to account for the creative use of words in novel contexts. We argue below that the framework for representation of lexical knowledge developed here is superior to current lexical entry formats--both in terms of expressiveness of notation and the kinds of interpretive operations it is capable of supporting. Rather than taking a "snapshot" of language at any moment of time and freezing it into lists of word sense specifications, the model of the lexicon proposed here does not preclude extensibility: it is open-ended in nature and accounts for the novel, creative, uses of words in a variety of contexts by positing procedures for generating semantic expressions for words on the basis of particular contexts. Adopting such a model presents a number of benefits. From the point of view of a language user, a rich and expressive lexicon can explain aspects of learnability. From the point of view of a natural language processing system, it can offer improvements in robustness of coverage. Such benefits stem from the fact that the model offers a scheme for explicitly encoding lexical knowledge at several levels of generalization. As we discuss below, factoring these along different dimensions makes it possible for NLP systems to carry Lexical knowledge representation 195 out semantically intensive--and hitherto complex--tasks. For instance, a consequence of adopting the representation model presented below as the basis of semantic interpretation is that some classically difficult problems in lexical ambiguity are resolved by viewing them from a different perspective. In particular, we illustrate how, by making lexical ambiguity resolution an integral part of a uniform semantic analysis procedure, the problem is rephrased in terms of dynamic interpretation of a word in context; this is in contrast to current frameworks which select among a static, predetermined set of word senses, and do so separately from constructing semantic representations for larger text units. There are several methodological motivations for importing tools developed for the computational representation and manipulation of knowledge into the study of word meaning, or lexical semantics. Generic knowledge representation (KR) mechanisms, such as inheritance structures or rule bases, can--and have been--used for encoding of linguistic information. However, not much attention has been paid to the notion of what exactly constitutes such linguistic information; this has been especially true in the context of developing operational NLP systems. Traditionally, the application area of knowledge representation formalisms has been the domain of general world knowledge. By shifting the focus to a level below that of words (or lexical concepts) we are now able to abstract the notion of lexical meaning away from world knowledge, as well as from other semantic influences such as discourse and pragmatic factors. Such a process of abstraction is an essential prerequisite for the principled creation of lexical entries. Furthermore, we argue below that judicious use of KR tools enriches the semantics of lexical expressions, while preserving felicitous partitioning of the information space. Keeping lexical meaning separate from other linguistic factors, as well as from general world knowledge is a methodologically sound principle; nonetheless, we maintain that all of these should be referenced by a lexical entry. The mechanisms developed for multi-faceted representation of knowledge in information-rich artificial intelligence contexts facilitate systematic incorporation of world knowledge into the lexicon. On the other hand, these mechanisms also make it possible to maintain the boundary between lexical and common sense knowledge. Additionally, KR tools allow us to concisely state the kinds of generalizations about the systematic patterning of words, which the theory presented below is largely concerned with. In particular, on the basis of a formal language for stating such generalizations, we evolve and propose a set of guidelines for capturing and expressing these within the format of lexical entries. The interplay of these capabilities--multiple levels of representation for the different kinds of lexical information, systematic reduction of information in the entries, imposing structure both on the lexicon as a whole and on individual entries--offers a novel way of capturing multiple word senses 196 J. Pustejovsl~v, B. Boguraev through richer composition. In essence, such capabilities are the base components of a generative language whose domain is that of lexical knowledge. The interpretive aspect of this language embodies a set of principles for richer composition of components of word meaning. In a manner explained later in this paper, semantic expressions for word meaning in context are constructed by a fixed number of generative devices (cf. Pustejovsky [44] ). Such devices operate on a core set of senses (with greater internal structure than hitherto assumed); through composition, an extended set of word senses is obtained when individual lexical items are considered jointly with others in larger phrases. The language presented below thus becomes an expressive tool for capturing lexical knowledge, without presupposing finite sense enumeration. In the remainder of this paper we illustrate a particular theory of computational lexical semantics, which promotes the notion of a generative lexicon (Pustejovsky [44] ). By way of setting the scene, we discuss certain types of lexical ambiguity, and demonstrate how traditional methods of ambiguity resolution fail to scale up for these (and other) cases. We then develop a model of semantic interpretation embodying richer methods of compositionality, and outline a framework for adequately representing lexical semantics in a knowledge base. We also analyze the effect this has on the size of a lexical knowledge base as a whole, and on the size of individual entries. Issues of organization and content of computational lexicons, and in particular, modeling word meaning via lexical definitions, are directly relevant not only to NLP but also to language acquisition and large-scale lexicon population. The generative theory of lexical semantics imposes a strong focus on current efforts to derive lexical data from large on-line text resources (dictionaries and corpora): it not only offers a uniform representational framework for expressing the data extracted by the tools and methods of computational lexicography (cf. Boguraev and Briscoe, [10]) but also offers guidance on the kinds of lexical data--or distinctions in the lexical behavior of words-which should be sought in such resources (cf. Boguraev [9], Anick and Pustejovsky [3]). In the final section of this paper we sketch the use of a KR formalism to capture a particular regularity of word behavior--lexical transfer as observed in real data--and make it available to an NLP sys-