LIUM’s SMT Machine Translation Systems for WMT 2012

Proceedings of the 7th Workshop on Statistical Machine Translation, pages 369–373, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics LIUM’s SMT Machine Translation Systems for WMT 2012 Christophe Servan, Patrik Lambert, Anthony Rousseau, Holger Schwenk and Lo¨ıc Barrault LIUM, University of Le Mans 72085 Le Mans cedex 9, FRANCE FirstName.LastName@lium.univ-lemans.fr Abstract This paper describes the development of French–English and English–French statisti- cal machine translation systems for the 2012 WMT shared task evaluation. We developed phrase-based systems based on the Moses de- coder, trained on the provided data only. Ad- ditionally, new features this year included im- proved language and translation model adap- tation using the cross-entropy score for the corpus selection. 1 Introduction This paper describes the statistical machine trans- lation systems developed by the Computer Science laboratory at the University of Le Mans (LIUM) for the 2012 WMT shared task evaluation. We only considered the translation between French and En- glish (in both directions). The main differences with respect to previous year’s system (Schwenk et al., 2011) are as follows: (i) use of more training data as provided by the organizers and (ii) better selection of the monolingual and parallel data according to the domain, using the cross-entropy difference with respect to in-domain and out-of-domain language models (Moore and Lewis, 2010). We kept some previous features: the improvement of the transla- tion model adaptation by unsupervised training, a parallel corpus retrieved by Information Retrieval (IR) techniques and finally, the rescoring with a con- tinuous space target language model for the trans- lation into French. These different points are de- scribed in the rest of the paper, together with a sum- mary of the experimental results showing the impact of each component. 2 Resources Used The following sections describe how the resources provided or allowed in the shared task were used to train the translation and language models of the sys- tem. 2.1 Bilingual data The latest version of the News-Commentary (NC) corpus and of the Europarl (Eparl) corpus (version 7) were used. We also took as training data a subset of the French–English Gigaword (109) corpus. This year we changed the filters applied to select this sub- set (see Sect. 2.4). We also included in the training data the test sets from previous shared tasks, that we called the ntsXX corpus and which was composed of newstest2008, newstest2009, newssyscomb2009. 2.2 Development data Development was initially done on newstest2010, and newstest2011 was used as internal test set (Sec- tion 3.1). The development and internal test sets were then (Section 4) switched (tuning was done on newstest2011 and internal evaluation on new- stest2010). The default Moses tokenization was used. However, we added abbreviations for the French tokenizer. All our models are case sensitive and include punctuation. The BLEU scores reported in this paper were calculated with the mteval-v13 tool and are case insensitive. 2.3 Use of Automatic Translations Available human translated bitexts such as the Eu- roparl or 109 corpus seem to be out-of domain for this task. We …

Content truncated to 500 words