The Karlsruhe Institute of Technology Translation Systems for the WMT 2012

Proceedings of the 7th Workshop on Statistical Machine Translation, pages 349–355, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics The Karlsruhe Institute of Technology Translation Systems for the WMT 2012 Jan Niehues, Yuqi Zhang, Mohammed Mediani, Teresa Herrmann, Eunah Cho 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 WMT12 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, fine-grained part-of-speech (POS) and automatic cluster language models and discriminative word lexica. In addition, we explicitly handle out-of-vocabulary (OOV) words in German, if we have translations for other morphological forms of the same stem. Furthermore, we extended the POS-based reordering approach to also use information from syntactic trees. 1 Introduction In this paper, we describe our systems for the NAACL 2012 Seventh 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. In addition to the POS-based reordering model used in past years, for German-English we extended it to also use rules learned using syntax trees. The translation model was extended by the bilin- gual language model and a discriminative word lex- icon using a maximum entropy classifier. For the French-English and English-French translation sys- tems, we also used phrase table adaptation to avoid overestimation of the probabilities of the huge, but noisy Giga corpus. In the German-English system, we tried to learn translations for OOV words by ex- ploring different morphological forms of the OOVs with the same lemma. Furthermore, we combined different language models in the log-linear model. We used word- based language models trained on different parts of the training corpus as well as POS-based language models using fine-grained POS information and lan- guage models trained on automatic word clusters. The paper is organized as follows: The next sec- tion gives a detailed description of our systems in- cluding all the models. The translation results for all directions are presented afterwards and we close with a conclusion. 2 System Description For the French↔English systems the phrase table is based on a GIZA++ word alignment, while the systems for German↔English use a discriminative word alignment as described in Niehues and Vogel (2008). The language models are 4-gram SRI lan- guage models using Kneser-Ney smoothing trained by the SRILM Toolkit (Stolcke, 2002). The problem of word reordering is addressed with POS-based and tree-based reordering models as de- scribed in Section 2.3. The POS tags used in the reordering model are obtained using the TreeTagger (Schmid, 1994). The syntactic parse trees are gen- erated using the Stanford Parser (Rafferty and Man- ning, 2008). An in-house phrase-based decoder (Vogel, 2003) is used to perform translation. Optimization with 349 regard to the BLEU score is …

Content truncated to 500 words