A novel dependency-to-string model for statistical machine translation

Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 216–226, Edinburgh, Scotland, UK, July 27–31, 2011. c⃝2011 Association for Computational Linguistics A Novel Dependency-to-String Model for Statistical Machine Translation Jun Xie, Haitao Mi and Qun Liu Key Laboratory of Intelligent Information Processiong Institute of Computing Technology Chinese Academy of Sciences P.O. Box 2704, Beijing 100190, China {junxie,htmi,liuqun}@ict.ac.cn Abstract Dependency structure, as a first step towards semantics, is believed to be helpful to improve translation quality. However, previous works on dependency structure based models typi- cally resort to insertion operations to complete translations, which make it difficult to spec- ify ordering information in translation rules. In our model of this paper, we handle this problem by directly specifying the ordering information in head-dependents rules which represent the source side as head-dependents relations and the target side as strings. The head-dependents rules require only substitu- tion operation, thus our model requires no heuristics or separate ordering models of the previous works to control the word order of translations. Large-scale experiments show that our model performs well on long dis- tance reordering, and outperforms the state- of-the-art constituency-to-string model (+1.47 BLEU on average) and hierarchical phrase- based model (+0.46 BLEU on average) on two Chinese-English NIST test sets without resort to phrases or parse forest. For the first time, a source dependency structure based model catches up with and surpasses the state-of-the- art translation models. 1 Introduction Dependency structure represents the grammatical relations that hold between the words in a sentence. It encodes semantic relations directly, and has the best inter-lingual phrasal cohesion properties (Fox, 2002). Those attractive characteristics make it pos- sible to improve translation quality by using depen- dency structures. Some researchers pay more attention to use de- pendency structure on the target side. (Shen et al., 2008) presents a string-to-dependency model, which restricts the target side of each hierarchical rule to be a well-formed dependency tree fragment, and em- ploys a dependency language model to make the out- put more grammatically. This model significantly outperforms the state-of-the-art hierarchical phrase- based model (Chiang, 2005). However, those string- to-tree systems run slowly in cubic time (Huang et al., 2006). Using dependency structure on the source side is also a promising way, as tree-based systems run much faster (linear time vs. cubic time, see (Huang et al., 2006)). Conventional dependency structure based models (Lin, 2004; Quirk et al., 2005; Ding and Palmer, 2005; Xiong et al., 2007) typically employ both substitution and insertion operation to complete translations, which make it difficult to specify ordering information directly in the transla- tion rules. As a result, they have to resort to either heuristics (Lin, 2004; Xiong et al., 2007) or sepa- rate ordering models (Quirk et al., 2005; Ding and Palmer, 2005) to control the word order of transla- tions. In this paper, we handle this problem by di- rectly specifying the ordering information in head- dependents rules that represent the source side as head-dependents relations and the target side as …

Content truncated to 500 words