Semantic Role Features for Machine Translation

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 716–724, Beijing, August 2010 Semantic Role Features for Machine Translation Ding Liu Department of Computer Science University of Rochester Daniel Gildea Department of Computer Science University of Rochester Abstract We propose semantic role features for a Tree-to-String transducer to model the re- ordering/deletion of source-side semantic roles. These semantic features, as well as the Tree-to-String templates, are trained based on a conditional log-linear model and are shown to significantly outperform systems trained based on Max-Likelihood and EM. We also show significant im- provement in sentence fluency by using the semantic role features in the log-linear model, based on manual evaluation. 1 Introduction Syntax-based statistical machine translation (SSMT) has achieved significant progress during recent years (Galley et al., 2006; May and Knight, 2007; Liu et al., 2006; Huang et al., 2006), showing that deep linguistic knowledge, if used properly, can improve MT performance. Semantics-based SMT, as a natural extension to SSMT, has begun to receive more attention from researchers (Liu and Gildea, 2008; Wu and Fung, 2009). Semantic structures have two major advantages over syntactic structures in terms of helping machine translation. First of all, semantic roles tend to agree better between two languages than syntactic constituents (Fung et al., 2006). This property motivates the approach of using the consistency of semantic roles to select MT outputs (Wu and Fung, 2009). Secondly, the set of semantic roles of a predicate models the skeleton of a sentence, which is crucial to the readability of MT output. By skeleton, we mean the main structure of a sentence including the verbs and their arguments. In spite of the theoretical potential of the semantic roles, there has not been much success in using them to improve SMT systems. Liu and Gildea (2008) proposed a semantic role based Tree-to-String (TTS) transducer by adding semantic roles to the TTS templates. Their ap- proach did not differentiate the semantic roles of different predicates, and did not always improve the TTS transducer’s performance. Wu and Fung (2009) took the output of a phrase-based SMT sys- tem Moses (Koehn et al., 2007), and kept permut- ing the semantic roles of the MT output until they best matched the semantic roles in the source sen- tence. This approach shows the positive effect of applying semantic role constraints, but it requires re-tagging semantic roles for every permuted MT output and does not scale well to longer sentences. This paper explores ways of tightly integrating semantic role features (SRFs) into an MT system, rather than using them in post-processing or n- best re-ranking. Semantic role labeling (SRL) sys- tems usually use sentence-wide features (Xue and Palmer, 2004; Pradhan et al., 2004; Toutanova et al., 2005); thus it is difficult to compute target- side semantic roles incrementally during decoding. Noticing that the source side semantic roles are easy to compute, we apply a compromise approach, where the target side semantic roles are generated by projecting the source side semantic roles us- ing …

Content truncated to 500 words