Machine Translation by Modeling Predicate-Argument Structure Transformation

Proceedings of COLING 2012: Technical Papers, pages 3019–3036, COLING 2012, Mumbai, December 2012. Machine Translation by Modeling Predicate­Argument Structure Transformation Feifei ZHAI, Jiajun ZHANG, Yu ZHOU and Chengqing ZONG National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences, Beijing, China {ffzhai, jjzhang, yzhou, cqzong}@nlpr.ia.ac.cn ABSTRACT Machine translation aims to generate a target sentence that is semantically equivalent to the source sentence. However, most of current statistical machine translation models do not model the semantics of sentences. In this paper, we propose a novel translation framework based on predicate-argument structure (PAS) for its capacity on grasping the semantics and skeleton structure of sentences. By using PAS, the framework effectively models both semantics of languages and global reordering for translation. In the framework, we divide the translation process into 3 steps: (1) PAS acquisition: perform semantic role labeling (SRL) on the input sentences to acquire source-side PASs; (2) Transformation: convert source-side PASs to their target counterparts by predicate-aware PAS transformation rules; (3) Translation: first translate the predicate and arguments of PAS and then adopt a CKY-style decoding algorithm to translate the entire PAS. Experimental results show that our PAS-based translation framework significantly improves the translation performance. KEYWORDS: Predicate-argument structure; Semantic role labeling; PAS transformation; PAS- based translation 3019 1 Introduction Statistical machine translation (SMT) has made significant progress from word-based models (Brown et al., 1993) to phrase-based models (Koehn et al., 2003; Och and Ney, 2004) and syntax-based models (Galley et al., 2006; Liu et al., 2006; Marcu et al., 2006) over the past decades. However, the existing SMT models are always criticized for not modeling the semantics of languages. Furthermore, reordering is always one of the most difficult and important research problems in SMT. However, although current translation models are much good at local reordering1, most of them are weak to cope with global reordering2. The two weaknesses restrict current translation models a lot, which urges us to seek a new translation framework to model both the semantics of languages and global reordering. Formally, predicate-argument structure (PAS) is a structure that depicts the relationship between a predicate and its associated arguments, and it always indicates the semantic frame and skeleton structure of a sentence. From the characteristics of PAS, we can see that it provides not only a good semantic representation for modeling semantics, but also a skeleton structure for global reordering. Moreover, Fung et al. (2006) and Wu and Fung (2009b) have shown that PASs of the both sides are more consistent with each other than syntax structures. Considering current syntax- based translation models are always impaired by cross-lingual structure divergence (Eisner, 2003; Zhang et al., 2010), PAS will be a better alternative for building translation models. Therefore, in this paper, aiming at building a PAS-based translation framework, we propose a novel translation method based on PAS transformation. FIGURE 1 is an overview of our method. Specifically, we divide the entire translation process into 3 steps: (1) PAS acquisition: perform semantic role labeling (SRL) on the input …

Content truncated to 500 words