Joint WMT 2012 Submission of the QUAERO Project

Proceedings of the 7th Workshop on Statistical Machine Translation, pages 322–329, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics Joint WMT 2012 Submission of the QUAERO Project ∗Markus Freitag, ∗Stephan Peitz, ∗Matthias Huck, ∗Hermann Ney, †Jan Niehues, †Teresa Herrmann, †Alex Waibel, ‡Le Hai-son, ‡Thomas Lavergne, ‡Alexandre Allauzen, §Bianka Buschbeck, §Josep Maria Crego, §Jean Senellart ∗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 QUAERO sub- mission to the WMT 2012 machine transla- tion evaluation. Four groups (RWTH Aachen University, Karlsruhe Institute of Technol- ogy, LIMSI-CNRS, and SYSTRAN) of the QUAERO project submitted a joint translation for the WMT German→English task. Each group translated the data sets with their own systems and finally the RWTH system combi- nation combined these translations in our final submission. Experimental results show im- provements of up to 1.7 points in BLEU and 3.4 points in TER compared to the best single system. 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 machine translation is mainly assigned to the four groups participating in this joint submission. The aim of this WMT submission was to show the quality of a joint translation by combining the knowledge of the four project partners. Each group develop and main- tain their own different machine translation system. These single systems differ not only in their general approach, but also in the preprocessing of training and test data. To take the advantage of these dif- ferences of each translation system, we combined all hypotheses of the different systems, using the RWTH system combination approach. This paper is structured as follows. In Section 2, the different engines of all four groups are in- troduced. In Section 3, the RWTH Aachen system combination approach is presented. Experiments with different system selections for system combi- nation are described in Section 4. Finally in Section 5, we discuss the results. 2 Translation Systems For WMT 2012 each QUAERO partner trained their systems on the parallel Europarl and News Com- mentary corpora. All single systems were tuned on the newstest2009 or newstest2010 development set. The newstest2011 dev set was used to train the system combination parameters. Finally, the newstest2008-newstest2010 dev sets were used to compare the results of the different system combina- tion settings. In this Section all four different system engines are presented. 2.1 RWTH Aachen Single Systems For the WMT 2012 evaluation the RWTH utilized RWTH’s state-of-the-art phrase-based and hierar- chical translation systems. GIZA++ (Och and Ney, 2003) was employed to train word alignments, lan- guage models have been created with the SRILM toolkit (Stolcke, 2002). 2.1.1 Phrase-Based System The phrase-based translation (PBT) system is similar to the one described in Zens and Ney (2008). After phrase pair extraction from the word-aligned parallel corpus, the translation probabilities are esti- mated …

Content truncated to 500 words