Limsi’s Statistical Translation Systems for WMT‘08

Proceedings of the Third Workshop on Statistical Machine Translation, pages 107–110, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics LIMSI’s statistical translation systems for WMT’08 Daniel Déchelotte, Gilles Adda, Alexandre Allauzen, Hélène Bonneau-Maynard, Olivier Galibert, Jean-Luc Gauvain, Philippe Langlais∗and François Yvon LIMSI/CNRS firstname.lastname@limsi.fr Abstract This paper describes our statistical machine translation systems based on the Moses toolkit for the WMT08 shared task. We address the Europarl and News conditions for the follow- ing language pairs: English with French, Ger- man and Spanish. For Europarl, n-best rescor- ing is performed using an enhanced n-gram or a neuronal language model; for the News condition, language models incorporate extra training data. We also report unconvincing re- sults of experiments with factored models. 1 Introduction This paper describes our statistical machine trans- lation systems based on the Moses toolkit for the WMT 08 shared task. We address the Europarl and News conditions for the following language pairs: English with French, German and Spanish. For Eu- roparl, n-best rescoring is performed using an en- hanced n-gram or a neuronal language model, and for the News condition, language models are trained with extra training data. We also report unconvinc- ing results of experiments with factored models. 2 Base System architecture LIMSI took part in the evaluations on Europarl data and on News data, translating French, German and Spanish from and to English, amounting a total of twelve evaluation conditions. Figure 1 presents the generic overall architecture of LIMSI’s transla- tion systems. They are fairly standard phrase-based ∗Univ. Montréal, felipe@iro.umontreal.ca Other Target text Target text Moses text Source or Translation model 4g language model 4g language model and extraction Rescoring $n$−best translations LM Interpolation Phrase pair extraction Neural network or or + News Co. Europarl Europarl Europarl News Co. sources Figure 1: Generic architecture of LIMSI’s SMT systems. Depending on the condition, the decoder generates ei- ther the final output or n-best lists. In the latter case, the rescoring incorporates the same translation features, except for a better target language model (see text). translation systems (Och and Ney, 2004; Koehn et al., 2003) and use Moses (Koehn et al., 2007) to search for the best target sentence. The search uses the following models: a phrase table, providing 4 scores and a phrase penalty, a lexicalized reordering model (7 scores), a language model score and a word penalty. These fourteen scores are weighted and lin- early combined (Och and Ney, 2002; Och, 2003); their respective weights are learned on development data so as to maximize the BLEU score. In the fol- lowing, we detail several aspects of our systems. 2.1 Translation models The translation models deployed in our systems for the europarl condition were trained on the provided Europarl parallel data only. For the news condition, they were trained on the Europarl data merged with 107 the news-commentary parallel data, as depicted on Figure 1. This setup was found to be more favor- able than training on Europarl data only (for obvious mismatching domain reasons) and …

Content truncated to 500 words