Synonym Acquisition Using Bilingual Comparable Corpora

International Joint Conference on Natural Language Processing, pages 1077–1081, Nagoya, Japan, 14-18 October 2013. Synonym Acquisition Using Bilingual Comparable Corpora Daniel Andrade Masaaki Tsuchida Takashi Onishi Kai Ishikawa Knowledge Discovery Research Laboratories, NEC Corporation, Nara, Japan {s-andrade@cj, m-tsuchida@cq, t-onishi@bq, k-ishikawa@dq}.jp.nec.com Abstract Various successful methods for synonym acquisition are based on comparing con- text vectors acquired from a monolingual corpus. However, a domain-specific cor- pus might be limited in size and, as a con- sequence, a query term’s context vector can be sparse. Furthermore, even terms in a domain-specific corpus are sometimes ambiguous, which makes it desirable to be able to find the synonyms related to only one word sense. We introduce a new method for enriching a query term’s con- text vector by using the context vectors of a query term’s translations which are ex- tracted from a comparable corpus. Our ex- perimental evaluation shows, that the pro- posed method can improve synonym ac- quisition. Furthermore, by selecting ap- propriate translations, the user is able to prime the query term to one sense. 1 Introduction Acquiring synonyms or near synonyms is impor- tant for various applications in NLP, like, for ex- ample, paraphrasing and recognizing textual en- tailment (Bentivogli et al., 2009). For these tasks, lexical resources like WordNet are often used to improve performance. Although these resources provide good coverage in general domains, they lack vocabulary specific to certain domains. Other problems are the limited avail- ability and size of lexical resources for languages other than English. 1 As a consequence, various previous works (Grefenstette, 1994) among others, suggest to acquire synonyms and other semantically 1For example, the coverage of words for the English WordNet is 147,278 words, whereas for Japanese WordNet’s coverage is only 93,834 words. related words automatically from a monolingual corpus. The key assumption is that semantically similar words occur in similar context. In general the larger the size of the monolin- gual corpus, the better and more detailed, we can extract the context, or context vectors for each rel- evant word (Curran and Moens, 2002). However, in a specific domain, the given monolingual cor- pus might be limited to a small size which leads to sparse context vectors. Another problem is that even for a specific domain, words can be ambigu- ous, which makes it unclear for which sense we are searching a synonym. For example, in the au- tomobile domain, the ambiguous Japanese word バルブ (bulb, valve) has the synonyms 電球 (bulb) or 弁 (valve), depending on the meaning intended by the user.2 Our work tries to overcome both of these prob- lems by enriching a context vector of a query word using the context vectors of its translations ob- tained from a comparable corpus in a different lan- guage. This way, some of the zero entries of a sparse context vector can be filled, and also dis- ambiguation of the query word is possible. For ex- ample, if the desired query word is バルブ (bulb, valve), the user can select the auxiliary translation …

Content truncated to 500 words