The DCU-ICTCAS MT system at WMT 2014 on German-English Translation Task

Proceedings of the Ninth Workshop on Statistical Machine Translation, pages 136–141, Baltimore, Maryland USA, June 26–27, 2014. c⃝2014 Association for Computational Linguistics The DCU-ICTCAS MT system at WMT 2014 on German-English Translation Task Liangyou Li∗, Xiaofeng Wu∗, Santiago Cort´es Va´ıllo∗ Jun Xie†, Andy Way∗, Qun Liu∗† ∗ CNGL Centre for Global Intelligent Content, School of Computing Dublin City University, Dublin 9, Ireland † Key Laboratory of Intelligent Information Processing, Institute of Computing Technology Chinese Academy of Sciences, Beijing, China {liangyouli,xiaofengwu,scortes,away,qliu}@computing.dcu.ie xiejun@ict.ac.cn Abstract This paper describes the DCU submis- sion to WMT 2014 on German-English translation task. Our system uses phrase- based translation model with several pop- ular techniques, including Lexicalized Reordering Model, Operation Sequence Model and Language Model interpolation. Our final submission is the result of sys- tem combination on several systems which have different pre-processing and align- ments. 1 Introduction On the German-English translation task of WMT 2014, we submitted a system which is built with Moses phrase-based model (Koehn et al., 2007). For system training, we use all provided German-English parallel data, and conducted sev- eral pre-processing steps to clean the data. In ad- dition, in order to improve the translation quali- ty, we adopted some popular techniques, includ- ing three Lexicalized Reordering Models (Axel- rod et al., 2005; Galley and Manning, 2008), a 9- gram Operation Sequence Model (Durrani et al., 2011) and Language Model interpolation on sev- eral datasets. And then we use system combina- tion on several systems with different settings to produce the final outputs. Our phrase-based systems are tuned with k-best MIRA (Cherry and Foster, 2012) on development set. We set the maximum iteration to be 25. The Language Models in our systems are trained with SRILM (Stolcke, 2002). We trained Corpus Filtered Out (%) Bilingual 7.17 Monolingual (English) 1.05 Table 1: Results of language detection: percentage of filtered out sentences a 5-gram model with Kneser-Ney discounting (Chen and Goodman, 1996). In the next sections, we will describe our system in detail. In section 2, we will explain our pre- processing steps on corpus. Then in section 3, we will describe some techniques we have tried for this task and the experiment results. In section 4, our final configuration for submitted system will be presented. And we conclude in the last section. 2 Pre-processing We use all the training data for German-English translation, including Europarl, News Commen- tary and Common Crawl. The first thing we no- ticed is that some Non-German and Non-English sentences are included in our training data. So we apply Language Detection (Shuyo, 2010) for both monolingual and bilingual corpora. For mono- lingual data (only including English sentences in our task), we filter out sentences which are detect- ed as other language with probability more than 0.999995. And for bilingual data, A sentence pair is filtered out if the language detector detect- s a different language with probability more than 0.999995 on either the source or the target. The filtering results are given in Table 1. In our experiment, German …

Content truncated to 500 words