BTG-based Machine Translation with Simple Reordering Model using Structured Perceptron

31st Pacific Asia Conference on Language, Information and Computation (PACLIC 31), pages 114–123 Cebu City, Philippines, November 16-18, 2017 Copyright c⃝2017 Hao Wang and Yves Lepage BTG-based Machine Translation with Simple Reordering Model using Structured Perceptron Hao Wang Graduate School of Information, Production and Systems, Waseda University oko ips@ruri.waseda.jp Yves Lepage Graduate School of Information, Production and Systems, Waseda University yves.lepage@waseda.jp Abstract In this paper, we present a novel statistical machine translation method which employs a BTG-based reordering model during decod- ing. BTG-based reordering models for pre- ordering have been widely explored, aiming to improve the standard phrase-based statisti- cal machine translation system. Less attention has been paid to incorporating such a reorder- ing model into decoding directly. Our reorder- ing model differs from previous models built using a syntactic parser or directly from anno- tated treebanks. Here, we train without using any syntactic information. The experiment re- sults on an English–Japanese translation task show that our BTG-based decoder achieves comparable or better performance than the more complex state-of-the-art SMT decoders. 1 Introduction The phrase-based method (Koehn et al., 2003) and the syntax-based method (Yamada and Knight, 2001) are two of the representative methods in sta- tistical machine translation (SMT). On the one hand, in the phrase-based model, the lexical reordering model is a crucial component, but it is often be crit- icized, especially when translating a language pair with widely divergent syntax like English-Japanese, as the na¨ıve distance-based lexical reordering model does not work so well when applied to longer re- orderings. On the other hand, in syntax-based SMT method, word reordering is implicitly addressed by translation rules. The performance is thus directly subject to the parsing errors of the syntactic parser. Syntax-based translation models are usually built from annotated treebanks to extract grammar rules for reordering (Genzel, 2010). Such reordering models are thus more difficult to train. Between these two models, some loose hierarchical structure models have been proposed: the hierarchical phrase- based model (Chiang, 2007) and or the Bracketing Transduction Grammar (BTG) based model (Wu, 1997). Compared with the hierarchical phrase-based model, the BTG model has many advantages like its simplicity. Also, its well-formed rules avoid ex- tracting a large number of rare or useless translation rules, as is the case of the hierarchical phrase-based model. In recent proposals, phrase-based statistical ma- chine translation has been shown to improve when BTG-based preordering is applied as a preprocess- ing (DeNero and Uszkoreit, 2011; Neubig et al., 2012; Nakagawa, 2015). The idea behind preorder- ing is to reduce the structural complexity. It is preferable to apply the reordering operations in ad- vance rather than during decoding as this benefits the word alignment step. In this paper, following (Xiong et al., 2008), we propose to incorporate the BTG-based reorder- ing model directly into the decoding step of a BTG-based SMT system using a simple Structured Perceptron (Rosenblatt, 1958; Collins and Roark, 2004). The rest of the paper is organized as follows. Section 2 briefly introduces previous BTG-based re- ordering methods …

Content truncated to 500 words