Extracting Synonyms from Dictionary Definitions

International Conference RANLP 2009 - Borovets, Bulgaria, pages 471–477 Extracting Synonyms from Dictionary Definitions Tong Wang and Graeme Hirst Department of Computer Science University of Toronto Toronto, ON, M5S 3G4, Canada tong,gh@cs.toronto.edu Abstract We investigate the problem of extracting synonyms from dictionary definitions. Our premise for using def- inition texts in dictionaries is that, in contrast to free- texts, their composition usually exhibits more regular- ities in terms of syntax and style and thus, will pro- vide a better controlled environment for synonym ex- traction. We propose three extraction methods: two rule-based ones and one using the maximum entropy model; each method is evaluated on three experiments — by solving TOEFL synonym questions, by compar- ing extraction results with existing thesauri, and by la- beling synonyms in definition texts. Results show that simple rule-based extraction methods perform surpris- ingly well on solving TOEFL synonym questions; they actually out-perform the best reported lexicon-based method by a large margin, although they do not corre- late as well with existing thesauri. Keywords Lexical semantics, synonym extraction, dictionary definition mining, maximum entropy classification 1 Introduction 1.1 Motivation Synonymy is one of the classic lexical semantic relations based on which lexical semantic taxonomies such as Word- Net (Fellbaum et al., 1998) are constructed. Despite their usefulness in various NLP studies, such taxonomies are usually considered expensive resources since they are of- ten manually constructed. Consequently, much research effort has been devoted to automatically extracting words of certain semantic relations from various free-text corpora (e.g., Hearst 1992; Lin 1998, etc) or even building an entire taxonomy (e.g., Chodorow, Byrd, and Heidorn 1985). Meanwhile, identifying characteristic features for syn- onymy is non-trivial. Many studies in this direction have started out with intuitively appealing ideas, but in practice, there are always surprises for intuition. Dependency rela- tions used by Lin (1998) relate two nouns if, for example, they often serve as the object of the same verb. When it comes to adjectives, however, the relation established as such is no longer guaranteed to be strict synonymy, since two adjective antonyms may modify the same noun as well. As another example, when two words in one language are often translated into the same word in another language, it seems very natural to regard the two words as synonyms in the source language (Wu and Zhou, 2003), but the mapping on the lexical level between languages is far from bijective, which in turn leads to many exceptions to the hypothesis. The difficulties in finding features for strict synonymy partly come from the syntactic and stylistic diversity in free-texts — this is the motivation behind using dictionary definitions as a resource for synonym identification. Un- like the extraction strategy used by Hearst (1992), where hyponyms would necessarily follow the phrase such as: NP0 such as {NP1,NP2,...,(and|or)} NPn ∀i = 1,2,...,n,hyponym(NPi,NP0) there seems to be much fewer patterns, if any, to follow for using synonyms in free-text writing. In contrast, in the composition of dictionary definitions, there is usually much more regularity …

Content truncated to 500 words