Improved Tree-to-String Transducer for Machine Translation

Proceedings of the Third Workshop on Statistical Machine Translation, pages 62–69, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics Improved Tree-to-string Transducer for Machine Translation Ding Liu and Daniel Gildea Department of Computer Science University of Rochester Rochester, NY 14627 Abstract We propose three enhancements to the tree- to-string (TTS) transducer for machine trans- lation: first-level expansion-based normaliza- tion for TTS templates, a syntactic align- ment framework integrating the insertion of unaligned target words, and subtree-based n- gram model addressing the tree decomposi- tion probability. Empirical results show that these methods improve the performance of a TTS transducer based on the standard BLEU- 4 metric. We also experiment with semantic labels in a TTS transducer, and achieve im- provement over our baseline system. 1 Introduction Syntax-based statistical machine translation (SSMT) has achieved significant progress during recent years, with two threads developing simul- taneously: the synchronous parsing-based SSMT (Galley et al., 2006; May and Knight, 2007) and the tree-to-string (TTS) transducer (Liu et al., 2006; Huang et al., 2006). Synchronous SSMT here denotes the systems which accept a source sentence as the input and generate the translation and the syntactic structure for both the source and the translation simultaneously. Such systems are sometimes also called TTS transducers, but in this paper, TTS transducer refers to the system which starts with the syntax tree of a source sentence and recursively transforms the tree to the target language based on TTS templates. In synchronous SSMT, TTS templates are used similar to the context free grammar used in the stan- dard CYK parser, thus the syntax is part of the output and can be thought of as a constraint on the transla- tion process. In the TTS transducer, since the parse tree is given, syntax can be thought of as an addi- tional feature of the input to be used in the transla- tion. The idea of synchronous SSMT can be traced back to Wu (1997)’s Stochastic Inversion Transduc- tion Grammars. A systematic method for extract- ing TTS templates from parallel corpora was pro- posed by Galley et al. (2004), and later binarized by Zhang et al. (2006) for high efficiency and ac- curacy. In the other track, the TTS transducer orig- inated from the tree transducer proposed by Rounds (1970) and Thatcher (1970) independently. Graehl and Knight (2004) generalized the tree transducer to the TTS transducer and introduced an EM al- gorithm to estimate the probability of TTS tem- plates based on a bilingual corpus with one side parsed. Liu et al. (2006) and Huang et al. (2006) then used the TTS transducer on the task of Chinese- to-English and English-to-Chinese translation, re- spectively, and achieved decent performance. Despite the progress SSMT has achieved, it is still a developing field with many problems un- solved. For example, the word alignment com- puted by GIZA++ and used as a basis to extract the TTS templates in most SSMT systems has been observed to be a problem for SSMT (DeNero and Klein, 2007; …

Content truncated to 500 words