LIMSI @ WMT12

Proceedings of the 7th Workshop on Statistical Machine Translation, pages 330–337, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics LIMSI @ WMT’12 Hai-Son Le1,2, Thomas Lavergne2, Alexandre Allauzen1,2, Marianna Apidianaki2, Li Gong1,2, Aur´elien Max1,2, Artem Sokolov2, Guillaume Wisniewski1,2, Franc¸ois Yvon1,2 Univ. Paris-Sud1 and LIMSI-CNRS2 rue John von Neumann, 91403 Orsay cedex, France {firstname.lastname}@limsi.fr Abstract This paper describes LIMSI’s submissions to the shared translation task. We report results for French-English and German-English in both directions. Our submissions use n-code, an open source system based on bilingual n-grams. In this approach, both the transla- tion and target language models are estimated as conventional smoothed n-gram models; an approach we extend here by estimating the translation probabilities in a continuous space using neural networks. Experimental results show a significant and consistent BLEU im- provement of approximately 1 point for all conditions. We also report preliminary experi- ments using an “on-the-fly” translation model. 1 Introduction This paper describes LIMSI’s submissions to the shared translation task of the Seventh Workshop on Statistical Machine Translation. LIMSI partic- ipated in the French-English and German-English tasks in both directions. For this evaluation, we used n-code, an open source in-house Statistical Machine Translation (SMT) system based on bilin- gual n-grams1. The main novelty of this year’s participation is the use, in a large scale system, of the continuous space translation models described in (Hai-Son et al., 2012). These models estimate the n-gram probabilities of bilingual translation units using neural networks. We also investigate an alter- native approach where the translation probabilities of a phrase based system are estimated “on-the-fly” 1http://ncode.limsi.fr/ by sampling relevant examples, instead of consider- ing the entire training set. Finally we also describe the use in a rescoring step of several additional fea- tures based on IBM1 models and word sense disam- biguation information. The rest of this paper is organized as follows. Sec- tion 2 provides an overview of the baseline systems built with n-code, including the standard transla- tion model (TM). The continuous space translation models are then described in Section 3. As in our previous participations, several steps of data pre- processing, cleaning and filtering are applied, and their improvement took a non-negligible part of our work. These steps are summarized in Section 5. The last two sections report experimental results ob- tained with the “on-the-fly” system in Section 6 and with n-code in Section 7. 2 System overview n-code implements the bilingual n-gram approach to SMT (Casacuberta and Vidal, 2004; Mari˜no et al., 2006; Crego and Mari˜no, 2006). In this framework, translation is divided in two steps: a source reorder- ing step and a (monotonic) translation step. Source reordering is based on a set of learned rewrite rules that non-deterministically reorder the input words. Applying these rules result in a finite-state graph of possible source reorderings, which is then searched for the best possible candidate translation. 2.1 Features Given a source sentence s of I words, the best trans- lation hypothesis ˆt is defined as the sequence of J words …

Content truncated to 500 words