Shift-Reduce Word Reordering for Machine Translation

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1382–1386, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Shift-Reduce Word Reordering for Machine Translation Katsuhiko Hayashi†, Katsuhito Sudoh, Hajime Tsukada, Jun Suzuki, Masaaki Nagata NTT Communication Science Laboratories, NTT Corporation 2-4 Hikaridai, Seika-cho, Soraku-gun, Kyoto, 619-0237 Japan †hayashi.katsuhiko@lab.ntt.co.jp Abstract This paper presents a novel word reordering model that employs a shift-reduce parser for inversion transduction grammars. Our model uses rich syntax parsing features for word re- ordering and runs in linear time. We apply it to postordering of phrase-based machine trans- lation (PBMT) for Japanese-to-English patent tasks. Our experimental results show that our method achieves a significant improvement of +3.1 BLEU scores against 30.15 BLEU scores of the baseline PBMT system. 1 Introduction Even though phrase-based machine translation (PBMT) (Koehn et al., 2007) and tree-based MT (Graehl and Knight, 2004; Chiang, 2005; Galley et al., 2006) systems have achieved great success, many problems remain for distinct language pairs, including long-distant word reordering. To improve such word reordering, one promis- ing way is to separate it from the translation pro- cess as preordering (Collins et al., 2005; DeNero and Uszkoreit, 2011) or postordering (Sudoh et al., 2011; Goto et al., 2012). Many studies utilize a rule- based or a probabilistic model to perform a reorder- ing decision at each node of a syntactic parse tree. This paper presents a parser-based word reorder- ing model that employs a shift-reduce parser for in- version transduction grammars (ITG) (Wu, 1997). To the best of our knowledge, this is the first study on a shift-reduce parser for word reordering. The parser-based reordering approach uses rich syntax parsing features for reordering decisions. Our propoesd method can also easily define such . Source- ordered Target Sentence (HFE) . Source Sen- tence (J) . Target Sen- tence (E) . reordering Figure 1: A description of the postordering MT system. non-local features as the N-gram words of reordered strings. Even when using these non-local features, the complexity of the shift-reduce parser does not increase at all due to give up achieving an optimal solution. Therefore, it works much more efficient. In our experiments, we apply our proposed method to postordering for J-to-E patent tasks be- cause their training data for reordering have little noise and they are ideal for evaluating reordering methods. Although our used J-to-E setups need a language-dependent scheme and we describe our proposed method as a J-to-E postordering method, the key algorithm is language-independent and it can be applicable to preordering as well as postordering if the training data for reordering are available. 2 Postordering by Parsing As shown in Fig.1, postordering (Sudoh et al., 2011) has two steps; the first is a translation step that trans- lates an input sentence into source-ordered transla- tions. The second is a reordering step in which the translations are reordered in the target language or- der. The key to postordering is the second step. Goto et al. (2012) modeled the second step …

Content truncated to 500 words