The LIG Machine Translation System for WMT 2010

Proceedings of the Joint 5th Workshop on Statistical Machine Translation and MetricsMATR, pages 161–166, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics The LIG machine translation system for WMT 2010 Marion Potet, Laurent Besacier and Herv´e Blanchon LIG Laboratory, GETALP Team University Joseph Fourier, Grenoble, France. Marion.Potet@imag.fr Laurent.Besacier@imag.fr Herve.Blanchon@imag.fr Abstract This paper describes the system submit- ted by the Laboratory of Informatics of Grenoble (LIG) for the fifth Workshop on Statistical Machine Translation. We participated to the news shared transla- tion task for the French-English language pair. We investigated differents techniques to simply deal with Out-Of-Vocabulary words in a statistical phrase-based ma- chine translation system and analyze their impact on translation quality. The final submission is a combination between a standard phrase-based system using the Moses decoder, with appropriate setups and pre-processing, and a lemmatized sys- tem to deal with Out-Of-Vocabulary con- jugated verbs. 1 Introduction We participated, for the first time, to the shared news translation task of the fifth Workshop on Ma- chine Translation (WMT 2010) for the French- English language pair. The submission was performed using a standard phrase-based trans- lation system with appropriate setups and pre- processings in order to deal with system’s un- known words. Indeed, as shown in (Carpuat, 2009), (Habash, 2008) and (Niessen, 2004), han- dling Ou-of-Vocabulary words with techniques like lemmatization, phrase table extension or mor- phological pre-processing is a way to improve translation quality. After a short presentation of our baseline system setups we discuss the effect of Out-Of-Vocabulary words in the system and in- troduce some ideas we chose to implement. In the last part, we evaluate their impact on translation quality using automatic and human evaluations. 2 Baseline System Setup 2.1 Used Resources We used the provided Europarl and News par- allel corpora (total 1,638,440 sentences) to train the translation model and the News monolin- gual corpora (48,653,884 sentences) to train the language model. The 2008 News test corpora (news-test2008; 2,028 sentences) was used to tune the produced system and last year’s test corpora (news-test2009; 3,027 sentences) was used for evaluation purposes. These corpora will be ref- ered to as Dev and Test later in the paper. As pre- processing steps, we applied the PERL scripts pro- vided with the corpora to lowercase and tokenise the data. 2.2 Language modeling The target language model is a standard n-gram language model trained using the SRI language modeling toolkit (Stocke, 2002) on the news monolingual corpus. The smoothing technique we applied is the modified Kneser-Ney discounting with interpolation. 2.3 Translation modeling The translation model was trained using the par- allel corpus described earlier (Europarl+News). First, the corpus was word aligned and then, the pairs of source and corresponding target phrases were extracted from the word-aligned bilingual training corpus using the scripts provided with the Moses decoder (Koehn et al., 2007). The re- sult is a phrase-table containing all the aligned phrases. This phrase-table, produced by the trans- lation modeling, is used to extract several transla- tions models. In our …

Content truncated to 500 words