NRC Machine Translation System for WMT 2017

Proceedings of the Conference on Machine Translation (WMT), Volume 2: Shared Task Papers, pages 330–337 Copenhagen, Denmark, September 711, 2017. c⃝2017 Association for Computational Linguistics NRC Machine Translation System for WMT 2017 Chi-kiu Lo Samuel Larkin Boxing Chen Darlene Stewart National Research Council Canada 1200 Montreal Road, Ottawa, ON K1A 0R6, Canada FirstName.LastName@nrc-cnrc.gc.ca Colin Cherry Roland Kuhn George Foster∗ Google Research Montreal, Canada fosterg@google.com Abstract We describe the machine translation sys- tems developed at the National Research Council of Canada (NRC) for the Russian- English and Chinese-English news trans- lation tasks of the Second Conference on Machine Translation (WMT 2017). We conducted several experiments to explore the best baseline settings for neural ma- chine translation (NMT). In the Russian- English task, to our surprise, our best- performing system is one that rescores phrase-based statistical machine transla- tion outputs using NMT rescoring fea- tures. On the other hand, in the Chinese- English task, which has far more paral- lel training data, NMT is able to outper- form SMT significantly. The NRC MT systems is the best constrained system in Russian-English (out of nine participants) and the fourth best constrained system in Chinese-English (out of twenty partici- pants) in WMT 2017 human evaluation. 1 Introduction We present NRC’s submission to the Russian- English and Chinese-English news translation task of WMT 2017. In contrast to last year, when we participated in the Russian-English task only, with our well-developed phrase-based statistical ma- chine translation system (Lo et al., 2016; Larkin et al., 2010; Foster et al., 2009), this year we built large-scale state-of-the-art neural machine transla- tion (NMT) systems for these two language pairs to facilitate further understanding and discussion of NMT. Russian-English and Chinese-English are both challenging language pairs for machine transla- ∗Work performed while at NRC. tion. Russian is a highly inflectional and free word order language. The skewed Russian to English word type ratio introduces a data sparsity prob- lem that cannot be solved by discarding word in- flections, since they play an important role in dis- ambiguating the meaning of sentences. Chinese does not have clear word boundaries. The number of Chinese word types created by automatic word segmentation software is high, while naive char- acter segmentation would result in a skewed Chi- nese to English sentence length ratio. These char- acteristics make it difficult for machine translation systems to learn the correct association between words in Chinese and English. Since this was the first time we deployed NMT models in an evaluation, we first tried to repli- cate the results of previous work (Sennrich et al., 2016a). Our NMT systems are based on Nema- tus (Sennrich et al., 2017). We used automatic back-translation (Sennrich et al., 2016b) of a sub- selected monolingual News corpus as additional training data, and all the training data is seg- mented into subword units using BPE (Sennrich et al., 2016c). We also experimented with perva- sive dropout as implemented in Nematus. For Russian-English, our WMT16 PBMT sys- tem scored higher than all the NMT …

Content truncated to 500 words