The RWTH Aachen Machine Translation System for WMT 2010

Proceedings of the Joint 5th Workshop on Statistical Machine Translation and MetricsMATR, pages 93–97, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics The RWTH Aachen Machine Translation System for WMT 2010 Carmen Heger, Joern Wuebker, Matthias Huck, Gregor Leusch, Saab Mansour, Daniel Stein and Hermann Ney RWTH Aachen University Aachen, Germany surname@cs.rwth-aachen.de Abstract In this paper we describe the statisti- cal machine translation system of the RWTH Aachen University developed for the translation task of the Fifth Workshop on Statistical Machine Translation. State- of-the-art phrase-based and hierarchical statistical MT systems are augmented with appropriate morpho-syntactic en- hancements, as well as alternative phrase training methods and extended lexicon models. For some tasks, a system combi- nation of the best systems was used to gen- erate a final hypothesis. We participated in the constrained condition of German- English and French-English in each trans- lation direction. 1 Introduction This paper describes the statistical MT system used for our participation in the WMT 2010 shared translation task. We used it as an opportunity to in- corporate novel methods which have been investi- gated at RWTH over the last year and which have proven to be successful in other evaluations. For all tasks we used standard alignment and training tools as well as our in-house phrase- based and hierarchical statistical MT decoders. When German was involved, morpho-syntactic preprocessing was applied. An alternative phrase- training method and additional models were tested and investigated with respect to their effect for the different language pairs. For two of the language pairs we could improve performance by system combination. An overview of the systems and models will fol- low in Section 2 and 3, which describe the base- line architecture, followed by descriptions of the additional system components. Morpho-syntactic analysis and other preprocessing issues are cov- ered by Section 4. Finally, translation results for the different languages and system variants are presented in Section 5. 2 Translation Systems For the WMT 2010 Evaluation we used stan- dard phrase-based and hierarchical translation sys- tems. Alignments were trained with a variant of GIZA++. Target language models are 4-gram lan- guage models trained with the SRI toolkit, using Kneser-Ney discounting with interpolation. 2.1 Phrase-Based System Our phrase-based translation system is similar to the one described in (Zens and Ney, 2008). Phrase pairs are extracted from a word-aligned bilingual corpus and their translation probability in both di- rections is estimated by relative frequencies. Ad- ditional models include a standard n-gram lan- guage model, phrase-level IBM1, word-, phrase- and distortion-penalties and a discriminative re- ordering model as described in (Zens and Ney, 2006). 2.2 Hierarchical System Our hierarchical phrase-based system is similar to the one described in (Chiang, 2007). It allows for gaps in the phrases by employing a context-free grammar and a CYK-like parsing during the de- coding step. It has similar features as the phrase- based system mentioned above. For some sys- tems, we only allowed the non-terminals in hierar- chical phrases to be substituted with initial phrases as in …

Content truncated to 500 words