Dependency Forest for Statistical Machine Translation

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 1092–1100, Beijing, August 2010 Dependency Forest for Statistical Machine Translation Zhaopeng Tu † Yang Liu † Young-Sook Hwang ‡ Qun Liu † Shouxun Lin † †Key Lab. of Intelligent Info. Processing ‡HILab Convergence Technology Center Institute of Computing Technology C&I Business Chinese Academy of Sciences SKTelecom {tuzhaopeng,yliu,liuqun,sxlin}@ict.ac.cn yshwang@sktelecom.com Abstract We propose a structure called dependency forest for statistical machine translation. A dependency forest compactly represents multiple dependency trees. We develop new algorithms for extracting string-to- dependency rules and training depen- dency language models. Our forest-based string-to-dependency system obtains sig- nificant improvements ranging from 1.36 to 1.46 BLEU points over the tree-based baseline on the NIST 2004/2005/2006 Chinese-English test sets. 1 Introduction Dependency grammars have become increasingly popular in syntax-based statistical machine trans- lation (SMT). One important advantage of depen- dency grammars is that they directly capture the dependencies between words, which are key to re- solving most parsing ambiguities. As a result, in- corporating dependency trees proves to be effec- tive in improving statistical machine translation (Quirk et al., 2005; Ding and Palmer, 2005; Shen et al., 2008). However, most dependency-based translation systems suffer from a major drawback: they only use 1-best dependency trees for rule extraction, dependency language model training, and decod- ing, which potentially introduces translation mis- takes due to the propagation of parsing errors (Quirk and Corston-Oliver, 2006). While the treelet system (Quirk et al., 2005) takes a de- pendency tree as input, the string-to-dependency system (Shen et al., 2008) decodes on a source- language string. However, as we will show, the string-to-dependency system still commits to us- ing degenerate rules and dependency language models learned from noisy 1-best trees. To alleviate this problem, an obvious solu- tion is to offer more alternatives. Recent studies have shown that SMT systems can benefit from widening the annotation pipeline: using packed forests instead of 1-best trees (Mi and Huang, 2008), word lattices instead of 1-best segmenta- tions (Dyer et al., 2008), and weighted alignment matrices instead of 1-best alignments (Liu et al., 2009). Along the same direction, we propose a struc- ture called dependency forest, which encodes ex- ponentially many dependency trees compactly, for dependency-based translation systems. In this pa- per, we develop two new algorithms for extracting string-to-dependency rules and for training depen- dency language models, respectively. We show that using the rules and dependency language models learned from dependency forests leads to consistent and significant improvements over that of using 1-best trees on the NIST 2004/2005/2006 Chinese-English test sets. 2 Background Figure 1 shows a dependency tree of an English sentence he saw a boy with a telescope. Arrows point from the child to the parent, which is often referred to as the head of the child. For example, in Figure 1, saw is the head of he. A dependency tree is more compact than its constituent counter- part because there is no need to build a large su- perstructure over a sentence. Shen et …

Content truncated to 500 words