Improving Tree-to-Tree Translation with Packed Forests

Proceedings of the 47th Annual Meeting of the ACL and the 4th IJCNLP of the AFNLP, pages 558–566, Suntec, Singapore, 2-7 August 2009. c⃝2009 ACL and AFNLP Improving Tree-to-Tree Translation with Packed Forests Yang Liu and Yajuan L¨u and Qun Liu Key Laboratory of Intelligent Information Processing Institute of Computing Technology Chinese Academy of Sciences P.O. Box 2704, Beijing 100190, China {yliu,lvyajuan,liuqun}@ict.ac.cn Abstract Current tree-to-tree models suffer from parsing errors as they usually use only 1- best parses for rule extraction and decod- ing. We instead propose a forest-based tree-to-tree model that uses packed forests. The model is based on a probabilis- tic synchronous tree substitution gram- mar (STSG), which can be learned from aligned forest pairs automatically. The de- coder finds ways of decomposing trees in the source forest into elementary trees us- ing the source projection of STSG while building target forest in parallel. Compa- rable to the state-of-the-art phrase-based system Moses, using packed forests in tree-to-tree translation results in a signif- icant absolute improvement of 3.6 BLEU points over using 1-best trees. 1 Introduction Approaches to syntax-based statistical machine translation make use of parallel data with syntactic annotations, either in the form of phrase structure trees or dependency trees. They can be roughly divided into three categories: string-to-tree mod- els (e.g., (Galley et al., 2006; Marcu et al., 2006; Shen et al., 2008)), tree-to-string models (e.g., (Liu et al., 2006; Huang et al., 2006)), and tree-to- tree models (e.g., (Eisner, 2003; Ding and Palmer, 2005; Cowan et al., 2006; Zhang et al., 2008)). By modeling the syntax of both source and tar- get languages, tree-to-tree approaches have the po- tential benefit of providing rules linguistically bet- ter motivated. However, while string-to-tree and tree-to-string models demonstrate promising re- sults in empirical evaluations, tree-to-tree models have still been underachieving. We believe that tree-to-tree models face two major challenges. First, tree-to-tree models are more vulnerable to parsing errors. Obtaining syntactic annotations in quantity usually entails running automatic parsers on a parallel corpus. As the amount and domain of the data used to train parsers are relatively limited, parsers will inevitably output ill-formed trees when handling real-world text. Guided by such noisy syntactic in- formation, syntax-based models that rely on 1-best parses are prone to learn noisy translation rules in training phase and produce degenerate trans- lations in decoding phase (Quirk and Corston- Oliver, 2006). This situation aggravates for tree- to-tree models that use syntax on both sides. Second, tree-to-tree rules provide poorer rule coverage. As a tree-to-tree rule requires that there must be trees on both sides, tree-to-tree mod- els lose a larger amount of linguistically unmoti- vated mappings. Studies reveal that the absence of such non-syntactic mappings will impair transla- tion quality dramatically (Marcu et al., 2006; Liu et al., 2007; DeNeefe et al., 2007; Zhang et al., 2008). Compactly encoding exponentially many parses, packed forests prove to be an excellent fit for alleviating the above two problems (Mi et al., 2008; Mi and Huang, 2008). In this paper, …

Content truncated to 500 words