Breadth and Depth of Semantic Lexicons

Breadth and Depth of Semantic Lexicons is based on a 1996 workshop. It represents several fundamental lines of thinking that were, and remain, a useful organization of the issues in lexical semantics, and specifically in the organization and development of new generations of lexical corpora. The volume contains very useful background discussions of classic issues in lexical semantics (e.g., English count/mass nouns), and some rather striking claims about what should and should not be included in a lexicon. Several of the papers show progress along contemporary models of lexical semantics, in particular the various current theories of lexical rules. The book represents a period, ongoing today, that welcomes the convergence of linguistic semantic thought and its computational implications. The conclusion seems to be, as asserted both by Nirenburg and Raskin and by Helmreich and Farwell, that certain of the theoretical or heuristic linguistic models of lexical semantics are useful for formulating computational implementations, but remain profoundly problematic as linguistic theories. The seminal positions from which the papers generally proceed are not directly represented in the volume, primarily Levin's work on English verb classes (Levin 1993), and Pustejovsky's generative lexicon (Pustejovsky 1995). Nevertheless, the currency of these constructs forms the context for the organization of the workshop and the volume. These positions differ from each other, naturally, but each has the effect of claiming that there are generalizations that may be captured between lexical items and either forms or functions, and thus that the lexical entry itself may be simpler, underspecified for elements of reference, or resident in a smaller, nonpolysemous lexicon. Such theoretical elegance appears at first examination to benefit efficiency at implementation time, which is what attracts computational linguistics to these approaches. However, the difficulties encountered in trying to extend their coverage at the same time delimit the level of their usefulness for system development, while helping the theorists further the instantiation of the paradigms implied by their theories. By the juxtaposition of theory and implementation we become aware of the problems at the poles of the current models: the claims they make are so general that numerous exceptions arise in implementation (like the Levin classes); or they require such specificity that their limited coverage (e.g., the conversion of a count noun to a mass noun in the case that the referent is an animal fur) makes implementation unproductive.