Phrase-based Machine Translation using Multiple Preordering Candidates

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 1419–1428, Osaka, Japan, December 11-17 2016. Phrase-based Machine Translation using Multiple Preordering Candidates Yusuke Oda† Taku Kudo‡ Tetsuji Nakagawa‡ Taro Watanabe‡ †Nara Institute of Science and Technology 8916-5 Takayama-cho, Ikoma-shi, Nara 630-0192, Japan ‡Google Japan Inc. 6-11-1 Roppongi, Minato-ku, Tokyo 106-6108, Japan oda.yusuke.on9@is.naist.jp {taku, tnaka, tarow}@google.com Abstract In this paper, we propose a new decoding method for phrase-based statistical machine transla- tion which directly uses multiple preordering candidates as a graph structure. Compared with previous phrase-based decoding methods, our method is based on a simple left-to-right dynamic programming in which no decoding-time reordering is performed. As a result, its runtime is very fast and implementing the algorithm becomes easy. Our system does not depend on specific preordering methods as long as they output multiple preordering candidates, and it is trivial to employ existing preordering methods into our system. In our experiments for translating diverse 11 languages into English, the proposed method outperforms conventional phrase-based decoder in terms of translation qualities under comparable or faster decoding time. 1 Introduction One of the main problem of phrase-based statistical machine translation (PBMT) (Koehn et al., 2003; Och and Ney, 2004) is handling the difference of word orders between source and target languages. Decoding-time reordering models (Koehn et al., 2005; Zens and Ney, 2006; Galley and Manning, 2008) measure positional relationship between each phrase at the decoding time. However, reordering models have a common problem in that it is difficult to take global information in the source sentence into account, and as a result the decoder may generate grammatically incorrect word orders. In addition, using reordering models demands a complicated decoding algorithm, in which the decoder has to consider concatenations of source phrases with arbitrary orders. On the other hand, preordering methods (Xia and McCord, 2004; Isozaki et al., 2010; Neubig et al., 2012; Nakagawa, 2015) change word orders of source sentence to be close to the target sentence before starting the decoding process. These methods can use global information in the source sentence and may generate grammatically correct reordering results. However, previous PBMT methods with preordering usually take only one-best preordered sentence and it is difficult to avoid the noise of the input caused by the errors from preordering methods. One of the trivial way to avoid preordering errors is to obtain N-best preordering candidates, translate each candidate one-by-one and select the most probable result (Li et al., 2007; Zhu, 2014). This method has an obvious problem on computation time because the decoding process is executed N times. An- other way to resolve preordering errors is combining a preordering method and decoding-time reordering models. However, it is not trivial to integrating these methods in a single system while comprehending their interactions. In this study, we propose a new phrase-based decoding method which employs multiple preordering candidates for a source sentence. Our method first encodes multiple preordering candidates as a compact graph structure (we call it preordering lattice), and generates …

Content truncated to 500 words