Soft Syntactic Constraints for Hierarchical Phrase-Based Translation Using Latent Syntactic Distributions

Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 138–147, MIT, Massachusetts, USA, 9-11 October 2010. c⃝2010 Association for Computational Linguistics Soft Syntactic Constraints for Hierarchical Phrase-based Translation Using Latent Syntactic Distributions Zhongqiang Huang Institute for Advanced Computer Studies University of Maryland College Park, MD 20742 zqhuang@umiacs.umd.edu Martin ˇCmejrek and Bowen Zhou IBM T. J. Watson Research Center Yorktown Heights, NY 10598 {martin.cmejrek,zhou}@us.ibm.com Abstract In this paper, we present a novel approach to enhance hierarchical phrase-based machine translation systems with linguistically moti- vated syntactic features. Rather than directly using treebank categories as in previous stud- ies, we learn a set of linguistically-guided la- tent syntactic categories automatically from a source-side parsed, word-aligned parallel cor- pus, based on the hierarchical structure among phrase pairs as well as the syntactic structure of the source side. In our model, each X non- terminal in a SCFG rule is decorated with a real-valued feature vector computed based on its distribution of latent syntactic categories. These feature vectors are utilized at decod- ing time to measure the similarity between the syntactic analysis of the source side and the syntax of the SCFG rules that are applied to derive translations. Our approach maintains the advantages of hierarchical phrase-based translation systems while at the same time nat- urally incorporates soft syntactic constraints. 1 Introduction In recent years, syntax-based translation mod- els (Chiang, 2007; Galley et al., 2004; Liu et al., 2006) have shown promising progress in im- proving translation quality, thanks to the incorpora- tion of phrasal translation adopted from the widely used phrase-based models (Och and Ney, 2004) to handle local fluency and the engagement of syn- chronous context-free grammars (SCFG) to handle non-local phrase reordering. Approaches to syntax- based translation models can be largely categorized into two classes based on their dependency on anno- tated corpus (Chiang, 2007). Linguistically syntax- based models (e.g., (Yamada and Knight, 2001; Gal- ley et al., 2004; Liu et al., 2006)) utilize structures defined over linguistic theory and annotations (e.g., Penn Treebank) and guide the derivation of SCFG rules with explicit parsing on at least one side of the parallel corpus. Formally syntax-based mod- els (e.g., (Wu, 1997; Chiang, 2007)) extract syn- chronous grammars from parallel corpora based on the hierarchical structure of natural language pairs without any explicit linguistic knowledge or anno- tations. In this work, we focus on the hierarchi- cal phrase-based models of Chiang (2007), which is formally syntax-based, and always refer the term SCFG, from now on, to the grammars of this model class. On the one hand, hierarchical phrase-based mod- els do not suffer from errors in syntactic constraints that are unavoidable in linguistically syntax-based models. Despite the complete lack of linguistic guidance, the performance of hierarchical phrase- based models is competitive when compared to lin- guistically syntax-based models. As shown in (Mi and Huang, 2008), hierarchical phrase-based models significantly outperform tree-to-string models (Liu et al., 2006; Huang et al., 2006), even when at- tempts are made to alleviate parsing errors …

Content truncated to 500 words