The Karlsruhe Institute of Technology Translation Systems for the WMT 2011

Proceedings of the 6th Workshop on Statistical Machine Translation, pages 379–385, Edinburgh, Scotland, UK, July 30–31, 2011. c⃝2011 Association for Computational Linguistics The Karlsruhe Institute of Technology Translation Systems for the WMT 2011 Teresa Herrmann, Mohammed Mediani, Jan Niehues and Alex Waibel Karlsruhe Institute of Technology Karlsruhe, Germany firstname.lastname@kit.edu Abstract This paper describes the phrase-based SMT systems developed for our participation in the WMT11 Shared Translation Task. Translations for English↔German and English↔French were generated using a phrase-based translation system which is extended by additional models such as bilingual and fine-grained POS language models, POS-based reordering, lattice phrase extraction and discriminative word alignment. Furthermore, we present a special filtering method for the English-French Giga corpus and the phrase scoring step in the training is parallelized. 1 Introduction In this paper we describe our systems for the EMNLP 2011 Sixth Workshop on Statistical Ma- chine Translation. We participated in the Shared Translation Task and submitted translations for English↔German and English↔French. We use a phrase-based decoder that can use lattices as input and developed several models that extend the stan- dard log-linear model combination of phrase-based MT. These include advanced reordering models and corresponding adaptations to the phrase extraction process as well as extension to the translation and language model in form of discriminative word alignment and a bilingual language model to ex- tend source word context. For English-German, lan- guage models based on fine-grained part-of-speech tags were used to address the difficult target lan- guage generation due to the rich morphology of Ger- man. We also present a filtering method directly ad- dressing the problems of web-crawled corpora, which enabled us to make use of the French-English Giga corpus. Another novelty in our systems this year is the parallel phrase scoring method that re- duces the time needed for training which is espe- cially convenient for such big corpora as the Giga corpus. 2 System Description The baseline systems for all languages use a trans- lation model that is trained on EPPS and the News Commentary corpus and the phrase table is based on a GIZA++ word alignment. The language model was trained on the monolingual parts of the same corpora by the SRILM Toolkit (Stolcke, 2002). It is a 4-gram SRI language model using Kneser-Ney smoothing. The problem of word reordering is addressed us- ing the POS-based reordering model as described in Section 2.4. The part-of-speech tags for the re- ordering model are obtained using the TreeTagger (Schmid, 1994). An in-house phrase-based decoder (Vogel, 2003) is used to perform translation and optimization with regard to the BLEU score is done using Minimum Error Rate Training as described in Venugopal et al. (2005). During decoding only the top 20 translation options for every source phrase were considered. 2.1 Data We trained all systems using the parallel EPPS and News Commentary corpora. In addition, the UN corpus and the Giga corpus were used for training 379 the French-English systems. Optimization was done for most languages using the news-test2008 data set and news-test2010 was …

Content truncated to 500 words