C-3MA: Tartu-Riga-Zurich Translation Systems for WMT17

Proceedings of the Conference on Machine Translation (WMT), Volume 2: Shared Task Papers, pages 382–388 Copenhagen, Denmark, September 711, 2017. c⃝2017 Association for Computational Linguistics C-3MA: Tartu-Riga-Zurich Translation Systems for WMT17 Mat¯ıss Rikters Faculty of Computing University of Latvia Riga, Latvia matiss@lielakeda.lv Chantal Amrhein University of Zurich Institute of Computational Linguistics Zurich, Switzerland chantal.amrhein@uzh.ch Maksym Del and Mark Fishel Institute of Computer Science University of Tartu Tartu, Estonia {maksym.del, fishel}@ut.ee Abstract This paper describes the neural machine translation systems of the University of Latvia, University of Zurich and Univer- sity of Tartu. We participated in the WMT 2017 shared task on news translation by building systems for two language pairs: English↔German and English↔Latvian. Our systems are based on an attentional encoder-decoder, using BPE subword seg- mentation. We experimented with back- translating the monolingual news corpora and filtering out the best translations as ad- ditional training data, enforcing named en- tity translation from a dictionary of par- allel named entities, penalizing over- and under-translated sentences, and combining output from multiple NMT systems with SMT. The described methods give 0.7 - 1.8 BLEU point improvements over our base- line systems. 1 Introduction We describe the neural machine translation (NMT) systems developed by the joint team of the Univer- sity of Latvia, University of Zurich and Univer- sity of Tartu (C-3MA). Our systems are based on an attentional encoder-decoder (Bahdanau et al., 2015), using BPE subword segmentation for open- vocabulary translation with a fixed vocabulary (Sennrich et al., 2016a). This paper is organized as follows: In Section 2 we describe our transla- tion software and baseline setups. Section 3 de- scribes our contributions for improving the base- line translations. Results of our experiments are summarized in Section 4. Finally, we conclude in Section 5. 2 Baseline Systems Our baseline systems were trained with two NMT and one statistical machine translation (SMT) framework. For English↔German we only trained NMT systems, for which we used Nematus (NT) (Sennrich et al., 2017). For English↔Latvian, apart from NT systems, we ad- ditionally trained NMT systems with Neural Mon- key (NM) (Helcl and Libovick`y, 2017) and SMT systems with LetsMT! (LMT) (Vasil¸jevs et al., 2012). In all of our NMT experiments we used a shared subword unit vocabulary (Sennrich et al., 2016b) of 35000 tokens. We clipped the gradient norm to 1.0 (Pascanu et al., 2013) and used a dropout of 0.2. Our models were trained with Adadelta (Zeiler, 2012) and after 7 days of training we per- formed early stopping. For training the NT models we used a maximum sentence length of 50, word embeddings of size 512, and hidden layers of size 1000. For decoding with NT we used beam search with a beam size of 12. For training the NM models we used a max- imum sentence length of 70, word embeddings and hidden layers of size 600. For decoding with NM a greedy decoder was used. Unfortunately, at the time when we performed our experiments the beam search decoder for NM was still under …

Content truncated to 500 words