Handling Ambiguities of Bilingual Predicate-Argument Structures for Statistical Machine Translation

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 1127–1136, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics Handling Ambiguities of Bilingual Predicate-Argument Structures for Statistical Machine Translation 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 Predicate-argument structure (PAS) has been demonstrated to be very effective in improving SMT performance. However, since a source- side PAS might correspond to multiple differ- ent target-side PASs, there usually exist many PAS ambiguities during translation. In this pa- per, we group PAS ambiguities into two types: role ambiguity and gap ambiguity. Then we propose two novel methods to handle the two PAS ambiguities for SMT accordingly: 1) in- side context integration; 2) a novel maximum entropy PAS disambiguation (MEPD) model. In this way, we incorporate rich context in- formation of PAS for disambiguation. Then we integrate the two methods into a PAS- based translation framework. Experiments show that our approach helps to achieve sig- nificant improvements on translation quality. 1 Introduction Predicate-argument structure (PAS) depicts the relationship between a predicate and its associat- ed arguments, which indicates the skeleton struc- ture of a sentence on semantic level. Basically, PAS agrees much better between two languages than syntax structure (Fung et al., 2006; Wu and Fung, 2009b). Considering that current syntax- based translation models are always impaired by cross-lingual structure divergence (Eisner, 2003; Zhang et al., 2010), PAS is really a better repre- sentation of a sentence pair to model the bilin- gual structure mapping. However, since a source-side PAS might correspond to multiple different target-side PASs, there usually exist many PAS ambiguities during translation. For example, in Figure 1, (a) and (b) carry the same source-side PAS <[A0]1 [Pred(是)]2 [A1]3> for Chinese predicate “是”. However, in Figure 1(a), the corresponding target-side-like PAS is <[X1] [X2] [X3]>, while in Figure 1(b), the counterpart target-side-like PAS1 is <[X2] [X3] [X1]>. This is because the two PASs play different roles in their corresponding sentences. Actually, Figure 1(a) is an independ- ent PAS, while Figure 1(b) is a modifier of the noun phrase “中国 和 俄罗斯”. We call this kind of PAS ambiguity role ambiguity. 中国 和 俄罗斯 两个 大国 是 [ A0 ]1 [Pred]2[ A1 ]3 ， being , should … two major countries [ X3 ] [X2] China and Russia [ X1 ] 应 … 防洪 首要 的 任务 是 [ A0 ]1 [ A1 ]3 [Pred]2 flood prevention is the primary mission [ X1 ] [ X2 ] [ X3 ] 奥运村 的 位置 对 运动员 是 最 好 的 [ A0 ]1 [ A1 ]3 [Pred]2 the location of the olympic village for athletes is the best [X2][ X3 ] [ X1 ] (a) (c) (b) Figure 1. An example of ambiguous PASs. Meanwhile, Figure 1 also depicts another kind of PAS ambiguity. From Figure 1, we can see that (a) and (c) get the same source-side PAS and target-side-like …

Content truncated to 500 words