DFKI’s SMT System for WMT 2012

Proceedings of the 7th Workshop on Statistical Machine Translation, pages 382–387, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics DFKI’s SMT System for WMT 2012 David Vilar German Research Center for Artificial Intelligence (DFKI GmbH) Language Technology Lab Berlin, Germany david.vilar@dfki.de Abstract We describe DFKI’s statistical based submis- sion to the 2012 WMT evaluation. The sub- mission is based on the freely available ma- chine translation toolkit Jane, which supports phrase-based and hierarchical phrase-based translation models. Different setups have been tested and combined using a sentence selec- tion method. 1 Introduction In this paper we present DFKI’s submission for the 2012 MT shared task based on statistical ap- proaches. We use a variety of phrase-based and hi- erarchical phrase-based translation systems with dif- ferent configurations and enhancements and com- pare their performance. The output of the systems are later combined using a sentence selection mech- anism. Somewhat disappointingly the sentence se- lection hardly improves over the best single system. DFKI participated in the German to English and English to German translation tasks. Technical problems however hindered a more complete system for this last translation direction. This paper is organized as follows: Section 2 re- ports on the different single systems that we built for this shared task. Section 3 describes the sentence se- lection mechanism used for combining the output of the different systems. Section 4 concludes the paper. 2 Single Systems For all our setups we used the Jane toolkit (Vi- lar et al., 2010a), which in its current version sup- ports both phrase-based and hierarchical phrase- based translation models. In this Section we present the different settings that we used for the task. The bilingual training data used for training all systems was the combination of the provided Eu- roparl and News data. We also used two baseline 4- gram language models trained on the same Europarl training data and on the enhanced News Commen- tary monolingual training data. The newstest2010 dataset was used for optimization of the systems. 2.1 Phrase-based System The first system is a baseline phrase-based system trained on the available bilingual training data. Word alignments is trained using GIZA++ (Och and Ney, 2003), phrase extraction is performed with Jane us- ing standard settings, i.e. maximum source phrase length 6, maximum target phrase length 12, count features, etc. Consult the Jane documentation for more details. For reordering the standard distance- based reordering model is computed. Scaling factors are trained using MERT on n-best lists. 2.1.1 Verb reorderings Following (Popovi´c and Ney, 2006), for German to English translation, we perform verb reordering by first POS-tagging the source sentence and after- wards applying hand-defined rules. This includes rules for reordering verbs in subordinate clauses and participles. 2.1.2 Moore LM Moore and Lewis (2010) propose a method for filtering large quantities of out-of-domain language- model training data by comparing the cross-entropy 382 of an in-domain language model and an out-of- domain language model trained on a random sam- pling of the data. We followed this approach …

Content truncated to 500 words