LIMSI @ WMT11

Proceedings of the 6th Workshop on Statistical Machine Translation, pages 309–315, Edinburgh, Scotland, UK, July 30–31, 2011. c⃝2011 Association for Computational Linguistics LIMSI @ WMT11 Alexandre Allauzen H´el`ene Bonneau-Maynard Hai-Son Le Aur´elien Max Guillaume Wisniewski Franc¸ois Yvon Univ. Paris-Sud and LIMSI-CNRS B.P. 133, 91403 Orsay cedex, France Gilles Adda Josep M. Crego Adrien Lardilleux Thomas Lavergne Artem Sokolov LIMSI-CNRS B.P. 133, 91403 Orsay cedex, France Abstract This paper describes LIMSI’s submissions to the Sixth Workshop on Statistical Machine Translation. We report results for the French- English and German-English shared transla- tion tasks in both directions. Our systems use n-code, an open source Statistical Ma- chine Translation system based on bilingual n-grams. For the French-English task, we fo- cussed on finding efficient ways to take ad- vantage of the large and heterogeneous train- ing parallel data. In particular, using a sim- ple filtering strategy helped to improve both processing time and translation quality. To translate from English to French and Ger- man, we also investigated the use of the SOUL language model in Machine Trans- lation and showed significant improvements with a 10-gram SOUL model. We also briefly report experiments with several alternatives to the standard n-best MERT procedure, leading to a significant speed-up. 1 Introduction This paper describes LIMSI’s submissions to the Sixth Workshop on Statistical Machine Translation, where LIMSI participated in the French-English and German-English tasks in both directions. For this evaluation, we used n-code, our in-house Statistical Machine Translation (SMT) system which is open- source and based on bilingual n-grams. This paper is organized as follows. Section 2 pro- vides an overview of n-code, while the data pre- processing and filtering steps are described in Sec- tion 3. Given the large amount of parallel data avail- able, we proposed a method to filter the French- English GigaWord corpus (Section 3.2). As in our previous participations, data cleaning and filtering constitute a non-negligible part of our work. This includes detecting and discarding sentences in other languages; removing sentences which are also in- cluded in the provided development sets, as well as parts that are repeated (for the monolingual news data, this can reduce the amount of data by a fac- tor 3 or 4, depending on the language and the year); normalizing the character set (non-utf8 characters which are aberrant in context, or in the case of the GigaWord corpus, a lot of non-printable and thus in- visible control characters such as EOT (end of trans- mission)1). For target language modeling (Section 4), a stan- dard back-off n-gram model is estimated and tuned as described in Section 4.1. Moreover, we also in- troduce in Section 4.2 the use of the SOUL lan- guage model (LM) (Le et al., 2011) in SMT. Based on neural networks, the SOUL LM can handle an arbitrary large vocabulary and a high order marko- vian assumption (up to 10-gram in this work). Fi- nally, experimental results are reported in Section 5 both in terms of BLEU scores and translation edit rates (TER) measured on …

Content truncated to 500 words