Tree Kernel-based SVM with Structured Syntactic Knowledge for BTG-based Phrase Reordering

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 698–707, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Tree Kernel-based SVM with Structured Syntactic Know- ledge for BTG-based Phrase Reordering Min Zhang Haizhou Li Institute for Infocomm Research 1 Fusionopolis Way,#21-01 Connexis (South Tower) Singapore 138632 {mzhang,hli}@i2r.a-star.edu.sg Abstract Structured syntactic knowledge is important for phrase reordering. This paper proposes us- ing convolution tree kernel over source parse tree to model structured syntactic knowledge for BTG-based phrase reordering in the con- text of statistical machine translation. Our study reveals that the structured syntactic fea- tures over the source phrases are very effective for BTG constraint-based phrase reordering and those features can be well captured by the tree kernel. We further combine the structured features and other commonly-used linear fea- tures into a composite kernel. Experimental re- sults on the NIST MT-2005 Chinese-English translation tasks show that our proposed phrase reordering model statistically signifi- cantly outperforms the baseline methods. 1 Introduction Phrase-based method (Koehn et al., 2003; Och and Ney, 2004; Koehn et al., 2007) and syntax- based method (Wu, 1997; Yamada and Knight, 2001; Eisner, 2003; Chiang, 2005; Cowan et al., 2006; Marcu et al., 2006; Liu et al., 2007; Zhang et al., 2007c, 2008a, 2008b; Shen et al., 2008; Mi and Huang, 2008) represent the state-of-the-art technologies in statistical machine translation (SMT). As the two technologies are complemen- tary in many ways, an interesting research topic is how to combine the strengths of the two me- thods. Many research efforts have been made to address this issue, which can be summarized into two ideas. One is to add syntax into phrase-based model while another one is to enhance syntax- based model to handle non-syntactic phrases. In this paper, we bring forward the first idea by studying the issue of how to utilize structured syntactic features for phrase reordering in a phrase-based SMT system with BTG (Bracketing Transduction Grammar) constraints (Wu, 1997). Word and phrase reordering is a crucial com- ponent in a SMT system. In syntax-based method, word reordering is implicitly addressed by trans- lation rules, thus the performance is subject to parsing errors to a large extent (zhang et al., 2007a) and the impact of syntax on reordering is difficult to single out (Li et al., 2007). In phrase- based method, local word reordering1 can be ef- fectively captured by phrase pairs directly while local phrase reordering is explicitly modeled by phrase reordering model and distortion model. Recently, many phrase reordering methods have been proposed, ranging from simple distance- based distortion model (Koehn et al., 2003; Och and Ney, 2004), flat reordering model (Wu, 1997; Zens et al., 2004), lexicalized reordering model (Tillmann, 2004; Kumar and Byrne, 2005), to hierarchical phrase-based model (Chiang, 2005; Setiawan et al., 2007) and classifier-based reor- dering model with linear features (Zens and Ney, 2006; Xiong et al., 2006; Zhang et al., 2007a; Xiong et al., 2008). However, one of the major limitations of these advances is the structured syntactic …

Content truncated to 500 words