Investigating Complementary Methods for Verb Sense Pruning

m m Investigating Complementary Methods for Verb Sense Pruning Hongyan Jing and Vasileios Hatzivassiloglou and Rebecca Passonneau and Kathleen McKeown m u m Department of Computer Science 450 Computer Science Building Columbia University New York, N.Y. 10027 {hj ing, vh, becky, kathy}@¢s, columbia, edu Abstract We present an approach for tagging verb sense that combines a domain-independent method based on subcategorization and al- ternations with a domain-dependent meth- od utilizing statistically extracted verb clusters. Initial results indicate that verb senses can be pruned for highly polysemous verbs by up to 74% by the first method and by up to 85% by the second method. 1 Introduction Much work in natural language processing is predi- cated on the notion that linguistic usage varies suf- ficiently across different situations of language use that systems can be tailored to a particular sub- language variety (Kittredge and Lehrberger, 1982). Biber (1993) presents evidence that a corpus re- stricted to one or two language registers would ex- clude "much of the English language" by narrow- ing the lexicon, verb tense and aspect, and syntactic complexity. Such observations inform the increas- ing trend towards analysis of homogeneous corpora to identify linguistic constraints for use in systems intended to understand or generate coherent dis- course. Recent work in this vein includes identi- fication of lexical constraints from textual tutorial dialogue (Moser and Moore, 1995), constraints on illocutionary act type from spoken task-oriented di- alogue (Allen et al., 1995), prosodic constraints from spoken information-seeking monologues (Hirschberg and Nakatani, 1996), and constraints on referring ex- pressions from spoken narrative monologue (Passon- neau, 1996). Related work suggests that constraints of different types are interdependent (Biber, 1993; Passonneau and Litman, forthcoming), hence should be investigated together. Our ultimate goal is to de- velop methods to tag lexical semantic features in dis- course corpora in order to enhance extraction of con- straints of the sort just listed. Two types of inves- tigations that would undoubtedly be enhanced are explorations of the interrelation of lexical cohesion and global discourse structure (Morris and Hirst, 1991; Hearst, 1994), and identification of lexicaliza-: tion patterns for domain-specific concepts (Robin, 1994). In this paper, we propose a two-pronged approach to an initial step in lexical semantic tagging, prun- ing the search space for polysemous verbs. Rather than attempting to identify unique word senses, we aim for the more realistic goal of pruning sense in- formation. We will then incrementally evaluate the utility of tagging corpora with pruned sense sets for different types of discourse. We begin with verbs on the hypothesis that verb sense distinctions correlate with syntactic properties of verbs (Levin, 1993). Our initial results indicate that domain-independent syn- tactic information reduces potential verb senses for multiply polysemous verbs (five or more WordNet senses) by more than 50%. In Section 2, we outline our first method, based on domain-independent lex- ical knowledge, presenting results from an analysis of thousands of verbs. In the section following that, we present our complementary method, a technique utilizing verb clusters …

Content truncated to 500 words