The Universität Karlsruhe Translation System for the EACL-WMT 2009

Proceedings of the Fourth Workshop on Statistical Machine Translation , pages 80–84, Athens, Greece, 30 March – 31 March 2009. c⃝2009 Association for Computational Linguistics The Universit¨at Karlsruhe Translation System for the EACL-WMT 2009 Jan Niehues, Teresa Herrmann, Muntsin Kolss and Alex Waibel Universit¨at Karlsruhe (TH) Karlsruhe, Germany {jniehues,therrman,kolss,waibel}@ira.uka.de Abstract In this paper we describe the statistical machine translation system of the Univer- sit¨at Karlsruhe developed for the transla- tion task of the Fourth Workshop on Sta- tistical Machine Translation. The state-of- the-art phrase-based SMT system is aug- mented with alternative word reordering and alignment mechanisms as well as op- tional phrase table modifications. We par- ticipate in the constrained condition of German-English and English-German as well as in the constrained condition of French-English and English-French. 1 Introduction This paper describes the statistical MT system used for our participation in the WMT’09 Shared Translation Task and the particular language-pair- dependent variations of the system. We use stan- dard alignment and training tools and a phrase- based SMT decoder for creating state-of-the-art MT systems for our contribution in the transla- tion directions English-German, German-English, English-French and French-English. Depending on the language pair, the baseline system is augmented with part-of-speech (POS)- based short-range and long-range word reordering models, discriminative word alignment (DWA) and several modifications of the phrase table. Ex- periments with different system variants were con- ducted including some of those additional system components. Significantly better translation re- sults could be achieved compared to the baseline results. An overview of the system will follow in Sec- tion 2, which describes the baseline architecture, followed by descriptions of the additional system components. Translation results for the different languages and system variants are presented in Section 5. 2 Baseline System The core of our system is the STTK decoder (Vo- gel, 2003), a phrase-based SMT decoder with a local reordering window of 2 words. The de- coder generates a translation for the input text or word lattice by searching translation model and language model for the hypothesis that max- imizes phrase translation probabilities and target language probabilities. The translation model, i.e. the SMT phrase table is created during the training phase by a modified version of the Moses Toolkit (Koehn et al., 2007) applying GIZA++ for word alignment. Language models are built using the SRILM Toolkit. The POS-tags for the reorder- ing models were generated with the TreeTagger (Schmid, 1994) for all languages. 2.1 Training, Development and Test Data We submitted translations for the English- German, German-English, English-French and French-English tasks. All systems were trained on the Europarl and News Commentary corpora using the Moses Toolkit and apply 4-gram lan- guage models created from the respective mono- lingual News corpora. All feature weights are au- tomatically determined and optimized with respect to BLEU via MERT (Venugopal et al., 2005). For development and testing we used data pro- vided by the WMT’09, news-dev2009a and news- dev2009b, consisting of 1026 sentences each. 3 Word Reordering Model One part of our system that differs from the …

Content truncated to 500 words