The RWTH Aachen German-English Machine Translation System for WMT 2015

Proceedings of the Tenth Workshop on Statistical Machine Translation, pages 158–163, Lisboa, Portugal, 17-18 September 2015. c⃝2015 Association for Computational Linguistics. The RWTH Aachen German-English Machine Translation System for WMT 2015 Jan-Thorsten Peter, Farzad Toutounchi, Joern Wuebker and Hermann Ney Human Language Technology and Pattern Recognition Group Computer Science Department RWTH Aachen University D-52056 Aachen, Germany <surname>@cs.rwth-aachen.de Abstract This paper describes the statistical machine translation system developed at RWTH Aachen University for the German→English translation task of the EMNLP 2015 Tenth Workshop on Statistical Machine Translation (WMT 2015). A phrase-based machine transla- tion system was applied and augmented with hierarchical phrase reordering and word class language models. Further, we ran discriminative maximum expected BLEU training for our system. In addition, we utilized multiple feed-forward neural network language and translation models and a recurrent neural network language model for reranking. 1 Introduction For the WMT 2015 shared translation task1, RWTH utilized a state-of-the-art phrase-based translation system. We participated in the German→English translation task. The system included a hierarchical reordering model, a word class (cluster) language model, and discrimina- tive maximum expected BLEU training. Further, we reranked the nbest lists produced by our sys- tem with three feed-forward neural network mod- els and a recurrent neural language model. This paper is structured as follows: First, we briefly describe our preprocessing pipeline for the language pair German→English in Section 2, which is based on our 2014 pipeline. Next, morpho-syntactic analysis for preprocessing the data is described in Section 2.3. Different align- ment methods are discussed in Section 3. In Sec- tion 4, we present a summary of all methods used in our submission. More details are given about 1http://www.statmt.org/wmt15/ translation-task.html the language models (Section 4.2), maximum ex- pected BLEU training (Section 4.4), the hierarchi- cal reordering model (Section 4.5), feed-forward neural network training (Section 4.6), and recur- rent neural network language model (Section 4.7). Experimental results are discussed in Section 5. We conclude the paper in Section 6. 2 Preprocessing In this section we briefly describe our preprocess- ing pipeline, which is a modification of our WMT 2014 German→English preprocessing pipeline (Peitz et al., 2014). 2.1 Categorization We worked on the categorization of the digits and written numbers for the translation task. All writ- ten numbers were categorized. As the training data and also the test sets contain several errors for numbers in the source as well as in the target part, we put effort into producing correct English num- bers. 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 sepa- rator. 2.2 Remove Foreign Languages The WMT German→English Common Crawl cor- pora contains bilingual sentence pairs with non- German source or non-English target sentences. By using an ASCII filtering, we removed all sen- tences with more than 5% non-ASCII characters from the Common Crawl corpus. Chinese, Arabic and Russian are among the languages which can be easily filtered by deleting the sentences con- taining …

Content truncated to 500 words