The LIGA (LIG/LIA) Machine Translation System for WMT 2011

Proceedings of the 6th Workshop on Statistical Machine Translation, pages 440–446, Edinburgh, Scotland, UK, July 30–31, 2011. c⃝2011 Association for Computational Linguistics The LIGA (LIG/LIA) Machine Translation System for WMT 2011 Marion Potet1, Rapha¨el Rubino2, Benjamin Lecouteux1, St´ephane Huet2, Herv´e Blanchon1, Laurent Besacier1 and Fabrice Lef`evre2 1UJF-Grenoble1, UPMF-Grenoble2 LIG UMR 5217 Grenoble, F-38041, France FirstName.LastName@imag.fr 2Universit´e d’Avignon LIA-CERI Avignon, F-84911, France FirstName.LastName@univ-avignon.fr Abstract We describe our system for the news com- mentary translation task of WMT 2011. The submitted run for the French-English direction is a combination of two MOSES-based sys- tems developed at LIG and LIA laboratories. We report experiments to improve over the standard phrase-based model using statistical post-edition, information retrieval methods to subsample out-of-domain parallel corpora and ROVER to combine n-best list of hypotheses output by different systems. 1 Introduction This year, LIG and LIA have combined their efforts to produce a joint submission to WMT 2011 for the French-English translation task. Each group started by developing its own solution whilst sharing re- sources (corpora as provided by the organizers but also aligned data etc) and acquired knowledge (cur- rent parameters, effect of the size of n-grams, etc.) with the other. Both LIG and LIA systems are stan- dard phrase-based translation systems based on the MOSES toolkit with appropriate carefully-tuned se- tups. The final LIGA submission is a combination of the two systems. We summarize in Section 2 the resources used and the main characteristics of the systems. Sec- tions 3 and 4 describe the specificities and report experiments of resp. the LIG and the LIA system. Section 5 presents the combination of n-best lists hypotheses generated by both systems. Finally, we conclude in Section 6. 2 System overview 2.1 Used data Globally, our system1 was built using all the French and English data supplied for the workshop’s shared translation task, apart from the Gigaword monolin- gual corpora released by the LDC. Table 1 sums up the used data and introduces designations that we follow in the remainder of this paper to refer to cor- pora. Four corpora were used to build translation models: news-c, euro, UN and giga, while three others are employed to train monolingual language models (LMs). Three bilingual corpora were de- voted to model tuning: test09 was used for the de- velopment of the two seed systems (LIG and LIA), whereas test08 and testcomb08 were used to tune the weights for system combination. test10 was finally put aside to compare internally our methods. 2.2 LIG and LIA system characteristics Both LIG and LIA systems are phrase-based trans- lation models. All the data were first tokenized with the tokenizer provided for the workshop. Kneser- Ney discounted LMs were built from monolingual corpora using the SRILM toolkit (Stolcke, 2002), while bilingual corpora were aligned at the word- level using GIZA++ (Och and Ney, 2003) or its multi-threaded version MGIZA++ (Gao and Vogel, 2008) for the large corpora UN and giga. Phrase table and lexicalized reordering models were built with MOSES (Koehn et al., 2007). Finally, …

Content truncated to 500 words