The RWTH Aachen Machine Translation System for IWSLT 2016

The RWTH Aachen Machine Translation System for IWSLT 2016 Jan-Thorsten Peter, Andreas Guta, Nick Rossenbach, Miguel Grac¸a, and Hermann Ney Human Language Technology and Pattern Recognition Group Computer Science Department RWTH Aachen University Aachen, Germany <surname>@cs.rwth-aachen.de Abstract This work describes the statistical machine translation (SMT) systems of RWTH Aachen University developed for the evaluation campaign of International Workshop on Spo- ken Language Translation (IWSLT) 2016. We have par- ticipated in the MT track for the German→English lan- guage pair employing our state-of-the-art phrase-based sys- tem, neural machine translation implementation and our joint translation and reordering decoder. Furthermore, we have ap- plied feed-forward and recurrent neural language and trans- lation models for reranking. The attention-based approach has been used for reranking the n-best lists for both phrase- based and hierarchical setups. On top of these systems, we make use of system combination to enhance the translation quality by combining individually trained systems. 1. Introduction We describe the statistical machine translation (SMT) sys- tems developed by RWTH Aachen University for the eval- uation campaign of IWSLT 2016. We participated in the machine translation (MT) track in the Talk and MSLT task for the German→English language pair. The combination of multiple machine translation systems has proven to be highly effective on this task. The individual engines include state-of-the-art neural machine translation (NMT), phrase- based translation (PBT) and the joint translation and reorder- ing (JTR) systems. Each of these is a sound system includ- ing preprocessing, translation and postprocessing. As a final step, we combine all of these systems using our system com- bination implementation. The preprocessing used in this work is described in Sec- tion 2. In Section 3 we describe the translation software and setups we have made use of. It is split up into the description of our NMT (Section 3.1), PBT (Section 3.3) and JTR sys- tem (Section 3.3). Section 3.4 describes the applied language models and Section 3.5 the long short term memory (LSTM) language and translation models used in rescoring. Section 3.6 explains the system combination pipeline applied on top of the individual systems. Our experiments for each track are summarized in Section 4 and we conclude with Section 5. 2. Preprocessing In this section we briefly describe our preprocessing pipeline, which is equal to our IWSLT 2015 German→English prepro- cessing pipeline [1]. 2.1. Categorization We worked on the categorization of the digits and written numbers for the translation task. All written numbers were categorized. As the training data and also the test sets con- tain several errors for numbers in the source as well as in the target part, we put effort into producing correct English numbers. In addition, ’,’ and ’.’ marks were inverted in most cases, as in German the former mark is the decimal mark and the latter is the thousand separator. 2.2. Compound Splitting and POS-based Word Reorder- ing We reduced the source vocabulary size for the German→English translation and split the German com- pound words with the frequency-based method described in [2]. …

Content truncated to 500 words