Discriminative Training for Near-Synonym Substitution

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 1254–1262, Beijing, August 2010 Discriminative Training for Near-Synonym Substitution Liang-Chih Yu1, Hsiu-Min Shih2, Yu-Ling Lai2, Jui-Feng Yeh3 and Chung-Hsien Wu4 1Department of Information Management, Yuan Ze University 2Department of Mathematics, National Chung Cheng University 3Department of CSIE, National Chia-Yi University 4Department of CSIE, National Cheng Kung University Contact: lcyu@saturn.yzu.edu.tw Abstract Near-synonyms are useful knowledge re- sources for many natural language applica- tions such as query expansion for information retrieval (IR) and paraphrasing for text gen- eration. However, near-synonyms are not nec- essarily interchangeable in contexts due to their specific usage and syntactic constraints. Accordingly, it is worth to develop algorithms to verify whether near-synonyms do match the given contexts. In this paper, we consider the near-synonym substitution task as a classifica- tion task, where a classifier is trained for each near-synonym set to classify test examples into one of the near-synonyms in the set. We also propose the use of discriminative training to improve classifiers by distinguishing posi- tive and negative features for each near- synonym. Experimental results show that the proposed method achieves higher accuracy than both pointwise mutual information (PMI) and n-gram-based methods that have been used in previous studies. 1 Introduction Near-synonym sets represent groups of words with similar meaning, which are useful knowl- edge resources for many natural language appli- cations. For instance, they can be used for query expansion in information retrieval (IR) (Moldo- van and Mihalcea, 2000; Bhogal et al., 2007), where a query term can be expanded by its near- synonyms to improve the recall rate. They can also be used in an intelligent thesaurus that can automatically suggest alternative words to avoid repeating the same word in the composing of text when there are suitable alternatives in its synonym set (Inkpen and Hirst, 2006; Inkpen, 2007). These near-synonym sets can be derived from manually constructed dictionaries such as WordNet (called synsets) (Fellbaum, 1998), Eu- roWordNet (Rodríguez et al., 1998), or clusters derived using statistical approaches (Lin, 1998). Although the words in a near-synonym set have similar meaning, they are not necessarily interchangeable in practical use due to their spe- cific usage and collocational constraints. Pearce (2001) presented an example of collocational constraints for the context “ coffee”. In the given near-synonym set {strong, powerful}, the word “strong” is more suitable than “powerful” to fill the gap, since “powerful coffee” is an anti- collocation. Inkpen (2007) also presented several examples of collocations (e.g. ghastly mistake) and anti-collocations (e.g. ghastly error). Yu et al. (2007) described an example of the context mismatch problem for the context “ under the bay” and the near-synonym set {bridge, overpass, viaduct, tunnel} that represents the meaning of a physical structure that connects separate places by traversing an obstacle. The original word (target word) in the given context is “tunnel”, and cannot be substituted by the other words in the same set since all the substitu- tions are semantically implausible. Accordingly, it is worth to develop algorithms …

Content truncated to 500 words