A Tree-to-String Phrase-based Model for Statistical Machine Translation

CoNLL 2008: Proceedings of the 12th Conference on Computational Natural Language Learning, pages 143–150 Manchester, August 2008 A Tree-to-String Phrase-based Model for Statistical Machine Translation Thai Phuong Nguyen College of Technology Vietnam National University, Hanoi thainp@vnu.edu.vn Akira Shimazu1, Tu-Bao Ho2, Minh Le Nguyen1, and Vinh Van Nguyen1 1School of Information Science 2School of Knowledge Science Japan Advanced Institute of Science and Technology {shimazu,bao,nguyenml,vinhnv}@jaist.ac.jp Abstract Though phrase-based SMT has achieved high translation quality, it still lacks of generaliza- tion ability to capture word order differences between languages. In this paper we describe a general method for tree-to-string phrase- based SMT. We study how syntactic trans- formation is incorporated into phrase-based SMT and its effectiveness. We design syntac- tic transformation models using unlexicalized form of synchronous context-free grammars. These models can be learned from source- parsed bitext. Our system can naturally make use of both constituent and non-constituent phrasal translations in the decoding phase. We considered various levels of syntactic analy- sis ranging from chunking to full parsing. Our experimental results of English-Japanese and English-Vietnamese translation showed a significant improvement over two baseline phrase-based SMT systems. 1 Introduction Based on the kind of linguistic information which is made use of, syntactic SMT can be divided into four types: tree-to-string, string-to-tree, tree-to-tree, and hierarchical phrase-based. The tree-to-string ap- proach (Collins et al., 2005; Nguyen and Shimazu, 2006; Liu et al., 2006 and 2007) supposes that syn- tax of the source language is known. This approach can be applied when a source language parser is available. The string-to-tree approach (Yamada and Knight, 2001; Galley et al., 2006) focuses on syntactic modelling of the target language in cases it has syn- tactic resources such as treebanks and parsers. The tree-to-tree approach models the syntax of both lan- guages, therefore extra cost is required. The fourth approach (Chiang, 2005) constraints phrases under context-free grammar structure without any require- ment of linguistic annotation. In this paper, we present a tree-to-string phrase- based method which is based on synchronous CFGs. This method has two important properties: syntactic transformation is used in the decoding phase includ- ing a word-to-phrase tree transformation model and a phrase reordering model; phrases are the basic unit of translation. Since we design syntactic transforma- tion models using un-lexicalized synchronous CFGs, the number of rules is small1. Previous studies on tree-to-string SMT are different from ours. Collins et al. Collins et al. (2005) used hand crafted rules to carry out word reordering in the preprocessing phase but not decoding phase. Nguyen and Shimazu (2006) presented a more general method in which lexicalized syntactic reordering models based on PCFGs can be learned from source-parsed bitext and then applied in the preprocessing phase. Liu et al. (2006) changed the translation unit from phrases to tree-to-string align- ment templates (TATs) while we do not. TATs was represented as xRs rules while we use synchronous CFG rules. In order to overcome the limitation that TATs can not capture non-constituent phrasal transla- tions, Liu et al. (2007) proposed forest-to-string rules …

Content truncated to 500 words