Improving Bilingual Terminology Extraction from Comparable Corpora via Multiple Word-Space Models

Improving Bilingual Terminology Extraction from Comparable Corpora via Multiple Word-Space Models Amir HAZEM, Emmanuel MORIN LINA, Universite de Nantes. 2 rue de la houssiniere 44000 Nantes, France. amir.hazem@univ-nantes.fr, emmanuel.morin@univ-nantes.fr Abstract There is a rich flora of word space models that have proven their efficiency in many different applications including information retrieval (Dumais et al., 1988), word sense disambiguation (Sch¨utze, 1993), various semantic knowledge tests (Lund et al., 1995; Karlgren and Sahlgren, 2001), and text categorization (Sahlgren and Karlgren, 2005). Based on the assumption that each model captures some aspects of word meanings and provides its own empirical evidence, we present in this paper a systematic exploration of the principal corpus-based word space models for bilingual terminology extraction from comparable corpora. We find that, once we have identified the best procedures, a very simple combination approach leads to significant improvements compared to individual models. Keywords: Comparable corpora, Bilingual lexicon extraction, word-space models 1. Introduction The distributional hypothesis which states that words with similar meanings tend to occur in similar contexts (Harris, 1954), has been extended to the bilingual scenario (Fung, 1998; Rapp, 1999). Hence, using comparable corpora, a translation of a source word can be found by identify- ing a target word with the most similar context. A pop- ular method often used as a baseline is the Standard Ap- proach (Fung, 1998). It consists of using the bag-of-words paradigm to represent words of source and target language by their context vector. After word contexts have been weighted using an association measure (the point-wise mu- tual information (Fano, 1961), the log-likelihood (Dunning, 1993), the discounted odds-ratio (Laroche and Langlais, 2010)), the similarity between a source word’s context vec- tor and all the context vectors in the target language is computed using a similarity measure (cosine (Salton and Lesk, 1968), Jaccard (Grefenstette, 1994)...). Finally, the translation candidates are ranked according to their sim- ilarity score. Many variants of the Standard Approach have been proposed. They can differ in context represen- tation (window-based, syntactic-based) (Gamallo, 2008), corpus characteristics (small, large, general or domain spe- cific...)(Chiao and Zweigenbaum, 2002; D´ejean and ´Eric Gaussier, 2002; Morin et al., 2007), type of words to translate (single word terms (SWTs) or multi-word terms (MWTs))(Rapp, 1999; Daille and Morin, 2005), words frequency (less frequent, rare...)(Pekar et al., 2006), etc. There exist other approaches for bilingual lexicon extrac- tion. D´ejean et al. (2002) introduce the extended approach to avoid the insufficient coverage of the bilingual dictionary required for the translation of source context vectors. A variation of the latter method based on centroid is proposed by (Daille and Morin, 2005). Haghighi et al. (2008) em- ploy dimension reduction using canonical component anal- ysis (CCA). The majority of the proposed approaches rely on context similarity. The starting point of context characterization is word co-occurrence statistics. It can provide a natural ba- sis for semantic representation. Corpus-based word space models allow to go from distributional statistics to a geo- metric representation that induce the semantic representa- tion of words from their …

Content truncated to 500 words