Joint WMT 2013 Submission of the QUAERO Project

Proceedings of the Eighth Workshop on Statistical Machine Translation, pages 185–192, Sofia, Bulgaria, August 8-9, 2013 c⃝2013 Association for Computational Linguistics Joint WMT 2013 Submission of the QUAERO Project ∗Stephan Peitz, ∗Saab Mansour, ∗Matthias Huck, ∗Markus Freitag, ∗Hermann Ney, †Eunah Cho, †Teresa Herrmann, †Mohammed Mediani, †Jan Niehues, †Alex Waibel, ‡Alexandre Allauzen, ‡Quoc Khanh Do, §Bianka Buschbeck, §Tonio Wandmacher ∗RWTH Aachen University, Aachen, Germany †Karlsruhe Institute of Technology, Karlsruhe, Germany ‡LIMSI-CNRS, Orsay, France §SYSTRAN Software, Inc. ∗surname@cs.rwth-aachen.de †firstname.surname@kit.edu ‡firstname.lastname@limsi.fr §surname@systran.fr Abstract This paper describes the joint submis- sion of the QUAERO project for the German→English translation task of the ACL 2013 Eighth Workshop on Statisti- cal Machine Translation (WMT 2013). The submission was a system combina- tion of the output of four different transla- tion systems provided by RWTH Aachen University, Karlsruhe Institute of Technol- ogy (KIT), LIMSI-CNRS and SYSTRAN Software, Inc. The translations were joined using the RWTH’s system com- bination approach. Experimental results show improvements of up to 1.2 points in BLEU and 1.2 points in TER compared to the best single translation. 1 Introduction QUAERO is a European research and develop- ment program with the goal of developing multi- media and multilingual indexing and management tools for professional and general public applica- tions (http://www.quaero.org). Research in ma- chine translation is mainly assigned to the four groups participating in this joint submission. The aim of this submission was to show the quality of a joint translation by combining the knowledge of the four project partners. Each group develop and maintain their own different machine translation system. These single systems differ not only in their general approach, but also in the preprocess- ing of training and test data. To take advantage of these differences of each translation system, we combined all hypotheses of the different systems, using the RWTH system combination approach. This paper is structured as follows. First, the different engines of all four groups are introduced. In Section 3, the RWTH Aachen system combina- tion approach is presented. Experiments with dif- ferent system selections for system combination are described in Section 4. This paper is concluded in Section 5. 2 Translation Systems For WMT 2013, each QUAERO partner trained their systems on the parallel Europarl (EPPS), News Commentary (NC) corpora and the web- crawled corpus. All single systems were tuned on the newstest2009 and newstest2010 development set. The newstest2011 development set was used to tune the system combination parameters. Fi- nally, on newstest2012 the results of the different system combination settings are compared. In this Section, all four different translation engines are presented. 2.1 RWTH Aachen Single System For the WMT 2013 evaluation, RWTH utilized a phrase-based decoder based on (Wuebker et al., 2012) which is part of RWTH’s open-source SMT toolkit Jane 2.1 1. GIZA++ (Och and Ney, 2003) was employed to train a word alignment, language models have been created with the SRILM toolkit (Stolcke, 2002). After phrase pair extraction from the word- aligned parallel corpus, the translation probabil- ities are estimated by relative frequencies. The …

Content truncated to 500 words