The Distributional Inclusion Hypotheses and Lexical Entailment

Proceedings of the 43rd Annual Meeting of the ACL, pages 107–114, Ann Arbor, June 2005. c⃝2005 Association for Computational Linguistics The Distributional Inclusion Hypotheses and Lexical Entailment Maayan Geffet School of Computer Science and Engineering Hebrew University, Jerusalem, Israel, 91904 mary@cs.huji.ac.il Ido Dagan Department of Computer Science Bar-Ilan University, Ramat-Gan, Israel, 52900 dagan@cs.biu.ac.il Abstract This paper suggests refinements for the Distributional Similarity Hypothesis. Our proposed hypotheses relate the distribu- tional behavior of pairs of words to lexical entailment – a tighter notion of semantic similarity that is required by many NLP applications. To automatically explore the validity of the defined hypotheses we de- veloped an inclusion testing algorithm for characteristic features of two words, which incorporates corpus and web-based feature sampling to overcome data sparseness. The degree of hypotheses validity was then em- pirically tested and manually analyzed with respect to the word sense level. In addition, the above testing algorithm was exploited to improve lexical entailment acquisition. 1 Introduction Distributional Similarity between words has been an active research area for more than a decade. It is based on the general idea of Harris' Distributional Hypothesis, suggesting that words that occur within similar contexts are semantically similar (Harris, 1968). Concrete similarity measures com- pare a pair of weighted context feature vectors that characterize two words (Church and Hanks, 1990; Ruge, 1992; Pereira et al., 1993; Grefenstette, 1994; Lee, 1997; Lin, 1998; Pantel and Lin, 2002; Weeds and Weir, 2003). As it turns out, distributional similarity captures a somewhat loose notion of semantic similarity (see Table 1). It does not ensure that the meaning of one word is preserved when replacing it with the other one in some context. However, many semantic information-oriented applications like Question Answering, Information Extraction and Paraphrase Acquisition require a tighter similarity criterion, as was also demon- strated by papers at the recent PASCAL Challenge on Recognizing Textual Entailment (Dagan et al., 2005). In particular, all these applications need to know when the meaning of one word can be in- ferred (entailed) from another word, so that one word could substitute the other in some contexts. This relation corresponds to several lexical seman- tic relations, such as synonymy, hyponymy and some cases of meronymy. For example, in Ques- tion Answering, the word company in a question can be substituted in the text by firm (synonym), automaker (hyponym) or division (meronym). Un- fortunately, existing manually constructed re- sources of lexical semantic relations, such as WordNet, are not exhaustive and comprehensive enough for a variety of domains and thus are not sufficient as a sole resource for application needs1. Most works that attempt to learn such concrete lexical semantic relations employ a co-occurrence pattern-based approach (Hearst, 1992; Ravi- chandran and Hovy, 2002; Moldovan et al., 2004). Typically, they use a set of predefined lexico- syntactic patterns that characterize specific seman- tic relations. If a candidate word pair (like com- pany-automaker) co-occurs within the same sentence satisfying a concrete pattern (like " …companies, such as automakers"), then it is ex- …

Content truncated to 500 words