Soft Dependency Matching for Hierarchical Phrase-based Machine Translation

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 2227–2236, Dublin, Ireland, August 23-29 2014. Soft Dependency Matching for Hierarchical Phrase-based Machine Translation1 Hailong Cao1, Dongdong Zhang2, Ming Zhou2 and Tiejun Zhao1 1Harbin Institute of Technology, Harbin, P.R. China 2Microsoft Research Asia, Beijing, P.R. China {hailong, tjzhao}@mtlab.hit.edu.cn {Dongdong.Zhang, mingzhou}@microsoft.com Abstract This paper proposes a soft dependency matching model for hierarchical phrase-based (HPB) machine translation. When a HPB rule is extracted, we enrich it with dependency knowledge automatically learnt from the training data. The dependency knowledge not only encodes the dependency relations between the components inside the rule, but also contains the dependency relations between the rule and its con- text. When a rule is applied to translate a sentence, the dependency knowledge is used to compute the syntactic structural consistency of the rule against the dependency tree of the sentence. We characterize the structure consistency by three features and integrate them into the standard SMT log-linear model to guide the translation process. Our method is evaluated on multiple Chinese-to-English machine transla- tion test sets. The experimental results show that our soft matching model achieves 0.7-1.4 BLEU points improvements over a strong baseline of an in-house implemented HPB translation system. 1 Introduction HPB model (Chiang, 2007) is widely used and has consistently delivered state-of-the-art performance. This model extends the phrase-based model (Koehn et al., 2003) by using the formal synchronous grammar to well capture the recursiveness of language during translation. In a formal synchronous grammar, the syntactic unit could be any sequence of contiguous terminals and non-terminals, which may not necessarily satisfy the linguistic constraints. HPB model is powerful to cover any format of translation pairs, but it might introduce ungrammatical rules and produce poor quality translations. To generate grammatical translations, lots of syntax-based models have been proposed by Galley et al. (2004), Liu et al. (2006), Huang et al. (2006), Mi et al. (2008), Shen et al. (2008), Xie et al. (2011), Zhang et al. (2008), etc. In these models, the syntactic units should be compatible with the syntactic structure of either the source sentence or the target sentence. These approaches can generate more grammatical translations by capturing the structural difference between language pairs. However, these models need special efforts to capture non-syntactic translation knowledge to improve the trans- lation performance. It is desired to combine the advantages of syntax-based models and the HPB model (Stein et al., 2010). There has been much work trying to improve HPB model by incorporating syntax information. Marton and Resnik (2008) leverage linguistic constituents to constrain the decoding softly. Some work go further to augment the non-terminals in HPB rules with syntactic tags which depend on the syntac- tic structure covered by the non-terminals (Zollmann and Venugopal, 2006; Chiang, 2010; Li et al., 2012; Huang et al., 2013). For example, given below HPB rules (1-4), the source non-terminal X could be refined into NP or PP as shown in rules (5-8) respectively. (1) <借 了 X, borrowed X> (2) …

Content truncated to 500 words