Building Specialized Bilingual Lexicons Using Word Sense Disambiguation

International Joint Conference on Natural Language Processing, pages 952–956, Nagoya, Japan, 14-18 October 2013. Building Specialized Bilingual Lexicons Using Word Sense Disambiguation Dhouha Bouamor CEA, LIST, Vision and Content Engineering Laboratory, 91191 Gif-sur-Yvette CEDEX France dhouha.bouamor@cea.fr Nasredine Semmar CEA, LIST, Vision and Content Engineering Laboratory, 91191 Gif-sur-Yvette CEDEX France nasredine.semmar@cea.fr Pierre Zweigenbaum LIMSI-CNRS, F-91403 Orsay CEDEX France pz@limsi.fr Abstract This paper presents an extension of the standard approach used for bilingual lex- icon extraction from comparable corpora. We study the ambiguity problem revealed by the seed bilingual dictionary used to translate context vectors and augment the standard approach by a Word Sense Dis- ambiguation process. Our aim is to iden- tify the translations of words that are more likely to give the best representation of words in the target language. On two spe- cialized French-English and Romanian- English comparable corpora, empirical ex- perimental results show that the proposed method consistently outperforms the stan- dard approach. 1 Introduction Over the years, bilingual lexicon extraction from comparable corpora has attracted a wealth of re- search works (Fung, 1998; Rapp, 1995; Chiao and Zweigenbaum, 2003). The main work in this re- search area could be seen as an extension of Har- ris’s distributional hypothesis (Harris, 1954). It is based on the simple observation that a word and its translation are likely to appear in similar con- texts across languages (Rapp, 1995). Based on this assumption, the alignment method, known as the standard approach builds and compares con- text vectors for each word of the source and target languages. A particularity of this approach is that, to enable the comparison of context vectors, it requires the existence of a seed bilingual dictionary to translate source context vectors. The use of the bilingual dictionary is problematic when a word has sev- eral translations, whether they are synonymous or polysemous. For instance, the French word action can be translated into English as share, stock, law- suit or deed. In such cases, it is difficult to iden- tify in flat resources like bilingual dictionaries, wherein entries are usually unweighted and un- ordered, which translations are most relevant. The standard approach considers all available trans- lations and gives them the same importance in the resulting translated context vectors indepen- dently of the domain of interest and word ambigu- ity. Thus, in the financial domain, translating ac- tion into deed or lawsuit would probably introduce noise in context vectors. In this paper, we present a novel approach which addresses the word ambiguity problem ne- glected in the standard approach. We introduce a use of a WordNet-based semantic similarity mea- sure permitting the disambiguation of translated context vectors. The basic intuition behind this method is that instead of taking all translations of each seed word to translate a context vector, we only use the translations that are more likely to give the best representation of the context vec- tor in the target language. We test the method on two comparable corpora specialized on the Breast Cancer domain, for the French-English …

Content truncated to 500 words