CUNI submission in WMT17: Chimera goes neural

Proceedings of the Conference on Machine Translation (WMT), Volume 2: Shared Task Papers, pages 248–256 Copenhagen, Denmark, September 711, 2017. c⃝2017 Association for Computational Linguistics CUNI Submission in WMT17: Chimera Goes Neural Roman Sudarikov David Mareˇcek Tom Kocmi Duˇsan Variˇs Ondˇrej Bojar Charles University, Faculty of Mathematics and Physics Institute of Formal and Applied Linguistics Malostransk´e n´amˇest´ı 25, 118 00 Prague, Czech Republic surname@ufal.mff.cuni.cz Abstract This paper describes the neural and phrase-based machine translation systems submitted by CUNI to English-Czech News Translation Task of WMT17. We experiment with synthetic data for training and try several system combination tech- niques, both neural and phrase-based. Our primary submission CU-CHIMERA ends up being phrase-based backbone which in- corporates neural and deep-syntactic can- didate translations. 1 Introduction The paper describes CUNI submissions for English-to-Czech WMT 2017 News Translation Task. We experimented with several neural ma- chine translation (NMT) systems and we further developed our phrase-based statistical machine translation system Chimera, which was our pri- mary system last year (Tamchyna et al., 2016). This year, we planned our setup in a way that would allow us to experiment with neural system combination. To this end, we reserved the pro- vided English-Czech parallel data for the training of the system combination and trained our “indi- vidual forward systems” on almost only synthetic data. The structure of the paper is the following. In Section 2, we provide an overview of the relatively complex setup. Section 3 details how the train- ing data for all the systems were prepared, includ- ing the description of MT systems used for back- translation. Section 4 is devoted to our individual forward translation systems, each of which could actually serve as a submission to the translation task. We do not stop there and train system com- binations in Section 5. In Section 6, we present the systems we actually submitted to WMT17 and we conclude by Section 8. 2 Setup Overview Our setup this year is motivated by the ability to use all the parallel data for system combination training. The overall sequence of system training is the following: 1. Use available monolingual data and last year’s systems to prepare a synthetic parallel corpus using “back translation” (Section 3). 2. Train “individual forward systems” on this synthetic corpus (Section 4). 3. Apply individual forward systems to the source side of the genuine parallel data. 4. Train a (neural) system combination on this dataset (Section 5). 5. Apply individual forward systems to the test set and apply the trained combination system to their output (Section 5). Each of the steps is fully described in the respective section of this paper. By “back- translated” data we mean that for English- to-Czech translation task, we created a syn- thetic English-Czech parallel corpus by “back- translating” Czech monolingual data into English. To distinguish back-translation Czech-to-English systems and the English-to-Czech systems to be submitted, we will call Czech-to-English systems “back-translation systems” and English-to-Czech systems “forward(-translation) systems”. 3 Data Preparation The section describes the data used for training …

Content truncated to 500 words