Ranking Translation Candidates Acquired from Comparable Corpora

International Joint Conference on Natural Language Processing, pages 401–409, Nagoya, Japan, 14-18 October 2013. Ranking Translation Candidates Acquired from Comparable Corpora Rima Harastani and B´eatrice Daille and Emmanuel Morin LINA UMR CNRS 6241 - University of Nantes 2 rue de la Houssini`ere, BP 92208 44322 Nantes, France {rima.harastani,beatrice.daille,emmanuel.morin}@univ-nantes.fr Abstract Domain-specific bilingual lexicons ex- tracted from domain-specific comparable corpora provide for one term a list of ranked translation candidates. This study proposes to re-rank these translation can- didates. We suggest that a term and its translation appear in comparable sen- tences that can be extracted from domain- specific comparable corpora. For a source term and a list of translation candidates, we propose a method to identify and align the best source and target sentences that contain the term and its translation can- didates. We report results with two lan- guage pairs (French-English and French- German) using domain-specific compara- ble corpora. Our method significantly im- proves the top 1, top 5 and top 10 preci- sions of a domain-specific bilingual lex- icon, and thus, provides a better user- oriented results. 1 Introduction Comparable corpora have been the subject of in- terest for extracting bilingual lexicons by sev- eral researchers (Rapp, 1995; Fung and Mcke- own, 1997; Rapp, 1999; Koehn and Knight, 2002; Morin et al., 2008; Bouamor et al., 2013, among others). Rapp (1995) was the first to suggest that if a word A co-occurs frequently with another word B in one language, then the translation of A and the translation of B should co-occur fre- quently in another language. Approaches emerg- ing from (Rapp, 1995) make different assump- tions to extract bilingual lexicon from comparable corpora. However, they are all based on the as- sumption that a translation pair shares some sim- ilar context in comparable corpora. We refer to such approaches that depend on co-occurrences of words to extract a bilingual lexicon by distribu- tional approaches. Results obtained from distribu- tional approaches vary according to many param- eters. For example, one of the parameters that im- pacts the performance of distributional approaches is the way the context of a word is defined. Vari- ous approaches defined contexts differently: win- dows (Rapp, 1999), sentences or paragraphs (Fung and Mckeown, 1997), or by taking into consid- eration syntax dependencies based on POS tags (Gamallo, 2007). However, the most common way the context of a word is defined is by choosing words within windows centered around the word (Laroche and Langlais, 2010), usually of small sizes (e.g. a window of size 3 is used by Rapp (1999)). Domain-specific comparable corpora have been used for bilingual terminology extraction. These corpora are of modest sizes since large domain- specific corpora are not available for many do- mains (Morin et al., 2008). As a matter of fact, distributional approaches perform best with large comparable corpora, and thus they often give lower precisions when applied to domain- specific comparable corpora (Chiao and Zweigen- baum, 2002). The goal of our work is to find translations of terms in …

Content truncated to 500 words