LIMSI @ WMT13

Proceedings of the Eighth Workshop on Statistical Machine Translation, pages 62–69, Sofia, Bulgaria, August 8-9, 2013 c⃝2013 Association for Computational Linguistics LIMSI @ WMT’13 Alexandre Allauzen1,2, Nicolas P´echeux1,2, Quoc Khanh Do1,2, Marco Dinarelli2, Thomas Lavergne1,2, Aur´elien Max1,2, Hai-Son Le3, Franc¸ois Yvon1,2 Univ. Paris-Sud1 and LIMSI-CNRS2 rue John von Neumann, 91403 Orsay cedex, France {firstname.lastname}@limsi.fr Vietnamese Academy of Science and Technology3, Hanoi, Vietnam lehaison@ioit.ac.vn Abstract This paper describes LIMSI’s submis- sions to the shared WMT’13 translation task. We report results for French-English, German-English and Spanish-English in both directions. Our submissions use n-code, an open source system based on bilingual n-grams, and continuous space models in a post-processing step. The main novelties of this year’s participation are the following: our first participation to the Spanish-English task; experiments with source pre-ordering; a tighter integra- tion of continuous space language mod- els using artificial text generation (for Ger- man); and the use of different tuning sets according to the original language of the text to be translated. 1 Introduction This paper describes LIMSI’s submissions to the shared translation task of the Eighth Workshop on Statistical Machine Translation. LIMSI partici- pated in the French-English, German-English and Spanish-English tasks in both directions. For this evaluation, we used n-code, an open source in- house Statistical Machine Translation (SMT) sys- tem based on bilingual n-grams1, and continuous space models in a post-processing step, both for translation and target language modeling. This paper is organized as follows. Section 2 contains an overview of the baseline systems built with n-code, including the continuous space mod- els. As in our previous participations, several steps of data pre-processing, cleaning and filter- ing are applied, and their improvement took a non- negligible part of our work. These steps are sum- marized in Section 3. The rest of the paper is de- voted to the novelties of the systems submitted this 1http://ncode.limsi.fr/ year. Section 4 describes the system developed for our first participation to the Spanish-English trans- lation task in both directions. To translate from German into English, the impact of source pre- ordering is investigated, and experimental results are reported in Section 5, while for the reverse di- rection, we explored a text sampling strategy us- ing a 10-gram SOUL model to allow a tighter in- tegration of continuous space models during the translation process (see Section 6). A final section discusses the main lessons of this study. 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 reordering step and a (monotonic) transla- tion step. Source reordering is based on a set of learned rewrite rules that non-deterministically re- order the input words. Applying these rules result in a finite-state graph of possible source reorder- ings, which is then searched for the best possible candidate translation. 2.1 Features Given a source sentence s of I words, the best translation hypothesis ˆt is defined as the sequence of J …

Content truncated to 500 words