SMT and SPE Machine Translation Systems for WMT‘09

Proceedings of the Fourth Workshop on Statistical Machine Translation , pages 130–134, Athens, Greece, 30 March – 31 March 2009. c⃝2009 Association for Computational Linguistics SMT and SPE Machine Translation Systems for WMT’09 Holger Schwenk and Sadaf Abdul-Rauf and Lo¨ıc Barrault LIUM, University of Le Mans 72085 Le Mans cedex 9, FRANCE schwenk,abdul,barrault@lium.univ-lemans.fr Jean Senellart SYSTRAN SA 92044 Paris La D´efense cedex, FRANCE senellart@systran.fr Abstract This paper describes the development of several machine translation systems for the 2009 WMT shared task evaluation. We only consider the translation between French and English. We describe a sta- tistical system based on the Moses de- coder and a statistical post-editing sys- tem using SYSTRAN’s rule-based system. We also investigated techniques to auto- matically extract additional bilingual texts from comparable corpora. 1 Introduction This paper describes the machine translation sys- tems developed by the Computer Science labo- ratory at the University of Le Mans (LIUM) for the 2009 WMT shared task evaluation. This work was performed in cooperation with the company SYSTRAN. We only consider the translation be- tween French and English (in both directions). The main differences to the previous year’s system (Schwenk et al., 2008) are as follows: better us- age of SYSTRAN’s bilingual dictionary in the sta- tistical system, less bilingual training data, addi- tional language model training data (news-train08 as distributed by the organizers), usage of com- parable corpora to improve the translation model, and development of a statistical post-editing sys- tem (SPE). These different components are de- scribed in the following. 2 Used Resources In the frame work of the 2009 WMT shared trans- lation task many resources were made available. The following sections describe how they were used to train the translation and language models of the systems. 2.1 Bilingual data The latest version of the French/English Europarl and news-commentary corpus were used. We re- alized that the first corpus contains parts with for- eign languages. About 1200 such lines were ex- cluded.1 Additional bilingual corpora were avail- able, namely the Canadian Hansard corpus (about 68M English words) and an UN corpus (about 198M English words). In several initial exper- iments, we found no evidence that adding this data improves the overall system and they were not used in the final system, in order to keep the phrase-table small. We also performed ex- periments with the provided so-called bilingual French/English Gigaword corpus (575M English words in release 3). Again, we were not able to achieve any improvement by adding this data to the training material of the translation model. These findings are somehow surprising since it was eventually believed by the community that adding large amounts of bitexts should improve the translation model, as it is usually observed for the language model (Brants et al., 2007). In addition to these human generated bitexts, we also integrated a high quality bilingual dictio- nary from SYSTRAN. The entries of the dictio- nary were directly added to the bitexts. This tech- nique has the potential advantage that the dictio- nary words …

Content truncated to 500 words