The RALI Machine Translation System for WMT 2010

Proceedings of the Joint 5th Workshop on Statistical Machine Translation and MetricsMATR, pages 103–109, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics The RALI Machine Translation System for WMT 2010 St´ephane Huet, Julien Bourdaillet, Alexandre Patry and Philippe Langlais RALI - Universit´e de Montr´eal C.P. 6128, succursale Centre-ville H3C 3J7, Montr´eal, Qu´ebec, Canada {huetstep,bourdaij,patryale,felipe}@iro.umontreal.ca Abstract We describe our system for the translation task of WMT 2010. This system, devel- oped for the English-French and French- English directions, is based on Moses and was trained using only the resources sup- plied for the workshop. We report exper- iments to enhance it with out-of-domain parallel corpora sub-sampling, N-best list post-processing and a French grammatical checker. 1 Introduction This paper presents the phrase-based machine translation system developed at RALI in order to participate in both the French-English and English-French translation tasks. In these two tasks, we used all the corpora supplied for the con- straint data condition apart from the LDC Giga- word corpora. We describe its different components in Sec- tion 2. Section 3 reports our experiments to sub- sample the available out-of-domain corpora in or- der to adapt the translation models to the news domain. Section 4, dedicated to post-processing, presents how N-best lists are reranked and how the French 1-best output is corrected by a grammatical checker. Section 5 studies how the original source language of news acts upon translation quality. We conclude in Section 6. 2 System Architecture 2.1 Pre-processing The available corpora were pre-processed using an in-house script that normalizes quotes, dashes, spaces and ligatures. We also reaccentuated French words starting with a capital letter. We significantly cleaned up the parallel Giga word corpus (noted as gw hereafter), keeping 18.1 M of the original 22.5 M sentence pairs. For exam- ple, sentence pairs with numerous numbers, non- alphanumeric characters or words starting with capital letters were removed. Moreover, training material was tokenized with the tool provided for the workshop and truecased, meaning that the words occuring after a strong punctuation mark were lowercased when they be- longed to a dictionary of common all-lowercased forms; the others were left unchanged. In order to reduce the number of words unknown to the translation models, all numbers were serialized, i.e. mapped to a special unique token. The origi- nal numbers are then placed back in the translation in the same order as they appear in the source sen- tence. Since translations are mostly monotonic be- tween French and English, this simple algorithm works well most of the time. 2.2 Language Models We trained Kneser-Ney discounted 5-gram lan- guage models (LMs) on each available corpus us- ing the SRILM toolkit (Stolcke, 2002). These LMs were combined through linear interpola- tion: first, an out-of-domain LM was built from Europarl, UN and gw; then, this model was combined with the two in-domain LMs trained on news-commentary and news.shuffled, which will be referred to as nc and ns in the remainder of the article. Weights were fixed by optimizing the perplexity of a …

Content truncated to 500 words