The RWTH Machine Translation System for WMT 2009

Proceedings of the Fourth Workshop on Statistical Machine Translation , pages 66–69, Athens, Greece, 30 March – 31 March 2009. c⃝2009 Association for Computational Linguistics The RWTH Machine Translation System for WMT 2009 Maja Popovi´c, David Vilar, Daniel Stein, Evgeny Matusov and Hermann Ney RWTH Aachen University Aachen, Germany Abstract RWTH participated in the shared transla- tion task of the Fourth Workshop of Sta- tistical Machine Translation (WMT 2009) with the German-English, French-English and Spanish-English pair in each transla- tion direction. The submissions were gen- erated using a phrase-based and a hierar- chical statistical machine translation sys- tems with appropriate morpho-syntactic enhancements. POS-based reorderings of the source language for the phrase-based systems and splitting of German com- pounds for both systems were applied. For some tasks, a system combination was used to generate a final hypothesis. An ad- ditional English hypothesis was produced by combining all three final systems for translation into English. 1 Introduction For the WMT 2009 shared task, RWTH submit- ted translations for the German-English, French- English and Spanish-English language pair in both directions. A phrase-based translation system en- hanced with appropriate morpho-syntactic trans- formations was used for all translation direc- tions. Local POS-based word reorderings were ap- plied for the Spanish-English and French-English pair, and long range reorderings for the German- English pair. For this language pair splitting of German compounds was also applied. Spe- cial efforts were made for the French-English and German-English translation, where a hierarchi- cal system was also used and the final submis- sions are the result of a system combination. For translation into English, an additional hypothesis was produced as a result of combination of the final German-to-English, French-to-English and Spanish-to-English systems. 2 Translation models 2.1 Phrase-based model We used a standard phrase-based system similar to the one described in (Zens et al., 2002). The pairs of source and corresponding target phrases are ex- tracted from the word-aligned bilingual training corpus. Phrases are defined as non-empty contigu- ous sequences of words. The phrase translation probabilities are estimated using relative frequen- cies. In order to obtain a more symmetric model, the phrase-based model is used in both directions. 2.2 Hierarchical model The hierarchical phrase-based approach can be considered as an extension of the standard phrase- based model. In this model we allow the phrases to have “gaps”, i.e. we allow non-contiguous parts of the source sentence to be translated into pos- sibly non-contiguous parts of the target sentence. The model can be formalized as a synchronous context-free grammar (Chiang, 2007). The model also included some additional heuristics which have shown to be helpful for improving translation quality, as proposed in (Vilar et al., 2008). The first step in the hierarchical phrase extrac- tion is the same as for the phrased-based model. Having a set of initial phrases, we search for phrases which contain other smaller sub-phrases and produce a new phrase with gaps. In our sys- tem, we restricted the number of non-terminals for each hierarchical phrase to a maximum of two, …

Content truncated to 500 words