Towards better Machine Translation Quality for the German-English Language Pairs

Proceedings of the Third Workshop on Statistical Machine Translation, pages 139–142, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics Towards better Machine Translation Quality for the German–English Language Pairs Philipp Koehn Abhishek Arun Hieu Hoang School of Informatics University of Edinburgh pkoehn@inf.ed.ac.uk a.arun@sms.ed.ac.uk h.hoang@sms.ed.ac.uk Abstract The Edinburgh submissions to the shared task of the Third Workshop on Statistical Machine Translation (WMT-2008) incorporate recent advances to the open source Moses system. We made a special effort on the German– English and English–German language pairs, leading to substantial improvements. 1 Introduction Edinburgh University participated in the shared task of the Third Workshop on Statistical Machine Trans- lation (WMT-2008), which is partly funded by the EUROMATRIX project, which also funds our work. In this project, we set out to build machine trans- lation systems for all language pairs of official EU languages. Hence, we also participated in the shared task in all language pairs. For all language pairs, we used the Moses decoder (Koehn et al., 2007), which follows the phrase-based statistical machine translation approach (Koehn et al., 2003), with default settings as a starting point. We recently added minimum Bayes risk de- coding and reordering constraints to the decoder. We achieved consistent increase in BLEU scores with these improvements, showing gains of up to 0.9% BLEU on the 2008 news test set. Most of our efforts were focused on the language pairs German–English and English–German. For both language pairs, we explored language-specific and more general improvements, resulting in gains of up to 1.5% BLEU for German–English and 1.4% BLEU for English–German. 2 Recent Improvements Over the last months, we added minimum Bayes risk decoding and additional reordering constraints to the Moses decoder. The WMT-2008 shared task offered the opportunity to assess these components over a large range of language pairs and tasks. For all our experiments, we trained solely on the Europarl corpus, which allowed us to treat the 2007 news commentary test set (nc-test2007) as a stand- in for the 2008 news test set (news-2008), for which we have no in-domain training data. This may have resulted in lower performance due to less (and very relevant) training data, but it also allowed us to opti- mize for a true out-of-domain test set. The baseline training uses Moses default param- eters. We use a maximum sentence length of 80, a phrase translation table with the five traditional fea- tures, lexicalized reordering, and lowercase training and test data. All reported BLEU scores are not case- sensitive, computed using the NIST tool. 2.1 Minimum Bayes Risk Decoding Minimum Bayes risk decoding was proposed by Ku- mar and Byrne (2004). Instead of selecting the trans- lation with the highest probability, minimum Bayes risk decoding selects the translation that is most sim- ilar to the highest scoring translations. Intuitively, this avoid the selection of an outlier as the best trans- lation, since the decision rule prefers translations that are similar to other high-scoring translations. Minimum Bayes risk decoding is defined as: eMBR = argmaxe …

Content truncated to 500 words