The JHU Machine Translation Systems for WMT 2017

Proceedings of the Conference on Machine Translation (WMT), Volume 2: Shared Task Papers, pages 276–282 Copenhagen, Denmark, September 711, 2017. c⃝2017 Association for Computational Linguistics The JHU Machine Translation Systems for WMT 2017 Shuoyang Ding† Huda Khayrallah† Philipp Koehn† Matt Post‡ Gaurav Kumar† and Kevin Duh‡ †Center for Language and Speech Processing ‡Human Language Technology Center of Excellence Johns Hopkins University {dings, huda, phi}@jhu.edu, {post, gkumar, kevinduh}@cs.jhu.edu Abstract This paper describes the Johns Hop- kins University submissions to the shared translation task of EMNLP 2017 Sec- ond Conference on Machine Translation (WMT 2017). We set up phrase-based, syntax-based and/or neural machine trans- lation systems for all 14 language pairs of this year’s evaluation campaign. We also performed neural rescoring of phrase- based systems for English-Turkish and English-Finnish. 1 Introduction The JHU 2017 WMT submission consists of phrase-based systems, syntax-based systems and neural machine translation systems. In this paper we discuss features that we integrated into our sys- tem submissions. We also discuss lattice rescoring as a form of system combination of phrase-based and neural machine translation systems. The JHU phrase-based translation systems for our participation in the WMT 2017 shared trans- lation task are based on the open source Moses toolkit (Koehn et al., 2007) and strong baselines of our submission last year (Ding et al., 2016). The JHU neural machine translation systems were built with the Nematus (Sennrich et al., 2016c) and Marian (Junczys-Dowmunt et al., 2016) toolkits. Our lattice rescoring experiments are also based on a combination of these three toolkits. 2 Phrase-Based Model Baselines Although the focus of research in machine transla- tion has firmly moved onto neural machine trans- lation, we still built traditional phrase-based statis- tical machine translation systems for all language pairs. These submissions also serve as a baseline of where neural machine translation systems stand with respect to the prior state of the art. Our systems are very simmilar to the JHU sys- tems from last year (Ding et al., 2016). 2.1 Configuration We trained our systems with the following set- tings: a maximum sentence length of 80, grow- diag-final-and symmetrization of GIZA++ align- ments, an interpolated Kneser-Ney smoothed 5- gram language model with KenLM (Heafield, 2011) used at runtime, hierarchical lexicalized re- ordering (Galley and Manning, 2008), a lexically- driven 5-gram operation sequence model (OSM) (Durrani et al., 2013) with 4 count-based sup- portive features, sparse domain indicator, phrase length, and count bin features (Blunsom and Os- borne, 2008; Chiang et al., 2009), a distortion limit of 6, maximum phrase-length of 5, 100-best trans- lation options, compact phrase table (Junczys- Dowmunt, 2012) minimum Bayes risk decoding (Kumar and Byrne, 2004), cube pruning (Huang and Chiang, 2007), with a stack-size of 1000 during tuning and 5000 during test and the no- reordering-over-punctuation heuristic (Koehn and Haddow, 2009). We optimize feature function weights with k-best MIRA (Cherry and Foster, 2012). We used POS and morphological tags as addi- tional factors in phrase translation models (Koehn and Hoang, 2007) for the German-English lan- guage pairs. …

Content truncated to 500 words