An Effective Compositional Model for Lexical Alignment

An Effective Compositional Model for Lexical Alignment B´eatrice Daille Emmanuel Morin Universit´e de Nantes, LINA - FRE CNRS 2729 2, rue de la Houssini`ere, BP 92208 F-44322 Nantes cedex 03 � beatrice.daille,emmanuel.morin � @univ-nantes.fr Abstract The automatic compilation of bilingual dic- tionaries from comparable corpora has been successful for single-word terms (SWTs), but remains disappointing for multi-word terms (MWTs). One of the main problems is the insufficient coverage of the bilingual dic- tionary. Using the compositional translation method improved the results, but still shows some limits for MWTs of different syntac- tic structures. In this paper, we propose to bridge the gap between syntactic structures through morphological links. The results show a significant improvement in the com- positional translation of MWTs that demon- strate the efficiency of the morphologically based-method for lexical alignment. 1 Introduction Current research in the automatic compilation of bilingual dictionaries from corpora uses of compara- ble corpora. Comparable corpora gather texts shar- ing common features (domain, topic, genre, dis- course) without having a source text-target text re- lationship. They are considered by human transla- tors more trustworthy than parallel corpora (Bowker and Pearson, 2002). Moreover, they are available for any written languages and not only for pairs of lan- guages involving English. The compilation of spe- cialized dictionaries should take into account multi- word terms (MWTs) that are more precise and spe- cific to a particular scientific domain than single- word terms (SWTs). The standard approach is based on lexical context analysis and relies on the simple observation that a SWT or a MWT and its trans- lation tend to appear in the same lexical contexts. Correct results are obtained for SWTs with an ac- curacy of about 80% for the top 10-20 proposed candidates using large comparable corpora (Fung, 1998; Rapp, 1999; Chiao and Zweigenbaum, 2002) or 60% using small comparable corpora (D´ejean and Gaussier, 2002). In comparison, the results ob- tained for MWTs are disappointing. For instance, (Morin et al., 2007) have achieved 30% and 42% precision for the top 10 and top 20 candidates in a 0.84 million-word French-Japanese corpus. These results could be explained by the low frequency of MWTs compared to SWTs, by the lack of paral- lelism between the source and the target MWT ex- traction systems, and by the low performance of the alignment program. For SWTs, the process is in two steps: looking in a dictionary, and if no direct translation is available, starting the contextual anal- ysis. Looking in the dictionary gives low results for MWTs: 1% compared to 30% for French and 20% for Japanese SWTs (Morin and Daille, 2006). To ex- tend the coverage of the bilingual dictionary, an in- termediate step is added between looking in the dic- tionary and the contextual analysis that will propose several translation candidates to compare with the target MWTs. These candidate translations are ob- tained thanks to a compositional translation method (Melamed, 1997; Grefenstette, 1999). This method reveals some limits when MWTs in the source and the target …

Content truncated to 500 words