Paraphrase Alignment for Synonym Evidence Discovery

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 403–411, Beijing, August 2010 Paraphrase Alignment for Synonym Evidence Discovery Gintar˙e Grigonyt˙e Saarland University g.grigonyte@hmf.vdu.lt Jo˜ao Cordeiro, Ga¨el Dias, Rumen Moraliyski HULTIG University of Beira Interior {jpaulo,ddg,rumen}@di.ubi.pt Pavel Brazdil LIAAD/FEP University of Porto pbrazdil@liaad.up.pt Abstract We describe a new unsupervised approach for synonymy discovery by aligning para- phrases in monolingual domain corpora. For that purpose, we identify phrasal terms that convey most of the concepts within domains and adapt a methodol- ogy for the automatic extraction and align- ment of paraphrases to identify para- phrase casts from which valid synonyms are discovered. Results performed on two different domain corpora show that gen- eral synonyms as well as synonymic ex- pressions can be identified with a 67.27% precision. 1 Introduction Synonymy is a specific type of a semantic re- lationship. According to (Sowa and Siekmann, 1994), a synonym is a word (or concept) that means the same or nearly the same as another word (or concept). It has been observed that words are similar if their contexts are similar (Fre- itag et al., 2005) and so synonymy detection has received a lot of attention during the last decades. However, words used in the same context are not necessarily synonyms and can embody dif- ferent semantic relationships such as hyponyms, meronyms or co-hyponyms (Heylen et al., 2008). In this paper, we introduce a new unsupervised methodology for synonym detection by extract- ing and aligning paraphrases on normalized do- main corpora1. In particular, we study a specific structure within aligned paraphrases, paraphrase 1By normalized, we intend that phrasal terms have been previously identified. casts, from which valid synonyms are discovered. In fact, we propose a new approach based on the idea that synonyms are substitutable words within a given domain corpus. Results performed on two different domain corpora, the Corpus of Computer Security (COCS) and the Corpus of Cancer Re- search (COCR), show that general synonyms as well as synonymic expressions can be identified with a 67.27% precision performance. 2 Related Work Automatic synonymy detection has been tackled in a variety of ways which we explain as follows. 2.1 Pattern-based Approaches This approach to information extraction is based on a technique called selective concept extraction as defined by (Riloff, 1993). Selective concept extraction is a form of text skimming that selec- tively processes relevant text while effectively ig- noring surrounding text that is thought to be ir- relevant to the domain. The pioneer of pattern- based approaches (Hearst, 1992) has introduced lexico-syntactic patterns to automatically acquire given word semantic relationships. Specific pat- terns like ”X and other Y” or ”X such as Y” were used for hypernym-hyponym detection. Later, the idea was extended and adapted for synonymy by other researchers such as (Roark and Charniak, 1998), (Caraballo, 1999) and (Maynard and Pe- ters, 2009). In general, manual pattern definition is time consuming and requires linguistic skills. Usually, systems based on lexico-syntactic pat- terns perform with very high precision, but low recall …

Content truncated to 500 words