Lexical Structures For Linguistic Inference

In order to resolve metonymy and other violations of selectional restrictions between lexical items, a language understander must be able to infer relationships that do not have explicit lexical analogs in tile sentence. Although such inferencing has typically been relegated to the world knowledge portion of a natural language processing system, there is also evidence, from both theoretical analysis in compositional semantics and distributional analysis of corpus data, that some cases of metonymy may best be processed with respect to more specific lexical and syntactic constructions. In this paper, we argue how the richer vocabulary for lexical semantics proposed in Pustejovsky's "Generative Lexicon" theory allows one to explore the role of lexical information in such cases, and therefore sheds more light on the distinction between lexical inferences, which follow from defaults associated with lexical items and rules of composition, and pragmatic inferences, which depend on reasoning with respect to the context of the utterance. 1 I n t r o d u c t i o n Is a lexicon merely a set. of entry points into a conceptual, "real world" knowledge base, or does it have its own rich structure and rules of inference, related to, yet distinct from the body of world knowledge? Ill this paper, we explore this question from the point of view of the inferential machinery required to relate a verb and its arguments. We will argue that a rich lexical semantic structure makes it possible to express rules of composition with the degree of precision necessary to account for the many nuances of actual language use, capturing linguistic generalizations based on the semantic content of lexical items rather than relying solely on general purpose inferencing over an encyclopedic knowledge base. One of the fundamental properties of most computational lexicons is an accounting of the relationship between a verb and its arguments. Typically, each argument position of a verb is annotated with "selectional restrictions" which are intended to constrain the set of nouns which may legally fill the argument position 1. Such constraints are useful for word sense disambiguation, since semantically anomalous colnbinations of senses can be readily identified and discarded during parsing. A number of researchers have noted that such selectionai restrictions may legitimately be violated under certain circumstances, as ill the sentences below: (1) "The car drank gasoline." ([Wilks, 1978]) (2) "Ted played Bach." ([Fass, 1988]) 1 If nouns are organized into a conceptual "isa" hierarchy, then a selectlonal restriction may be no more than a reference to a position in the hierarchy.