Modeling Syntactic and Semantic Structures in Hierarchical Phrase-based Translation

Proceedings of NAACL-HLT 2013, pages 540–549, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Modeling Syntactic and Semantic Structures in Hierarchical Phrase-based Translation Junhui Li University of Maryland College Park, USA lijunhui@umiacs.umd.edu Philip Resnik University of Maryland College Park, USA resnik@umd.edu Hal Daum´e III University of Maryland College Park, USA hal@umiacs.umd.edu Abstract Incorporating semantic structure into linguistics-free translation model is chal-a lenging, since semantic structures are closely tied to syntax. In this paper, we propose a two-level approach to exploiting predicate-argument structure reordering in a hierarchical phrase-based translation model. First, we introduce linguistically motivated constraints into a hierarchical model, guiding translation phrase choices in favor of those that respect syntactic boundaries. Second, abased on such translation phrases, we propose predicate-argument structure reordering model that predicts reordering not only between an argument and its predicate, but also between two arguments. Experiments on Chinese-to-English translation demonstrate that both advances significantly improve translation accuracy. 1 Introduction Hierarchical phrase-based (HPB) translation mod- els (Chiang, 2005; Chiang, 2007) that utilize syn- chronous context free grammars (SCFG) have been widely adopted in statistical machine translation (SMT). Although formally syntactic, such models rarely respect linguistically-motivated syntax, and have no formal notion of semantics. As a re- sult, they tend to produce translations containing both grammatical errors and semantic role confu- sions. Our goal is to take advantage of syntactic and semantic parsing to improve translation qual- ity of HPB translation models. Rather than intro- ducing semantic structure into the HPB model di- rectly, we construct an improved translation model by incorporating linguistically motivated syntactic constraints into a standard HPB model. Once the translation phrases are linguistically constrained, we are able to propose a predicate-argument reorder- ing model. This reordering model aims to solve two problems: ensure that arguments are ordered properly after translation, and to ensure that the proper argument structures even exist, for instance in the case of PRO-drop languages. Experimental results on Chinese-to-English translation show that both the hard syntactic constraints and the predicate- argument reordering model obtain significant im- provements over the syntactically and semantically uninformed baseline. In principle, semantic frames (or, more specifi- cally, predicate-argument structures: PAS) seem to be a promising avenue for translational modeling. While languages might diverge syntactically, they are less likely to diverge semantically. This has previously been recognized by Fung et al. (2006), who report that approximately 84% of semantic role mappings remained consistent across transla- tions between English and Chinese. Subsequently, Zhuang and Zong (2010) took advantage of this consistency to jointly model semantic frames on Chinese/English bitexts, yielding improved frame recognition accuracy on both languages. While there has been some encouraging work on integrating syntactic knowledge into Chiang’s HPB model, modeling semantic structure in a linguisti- cally naive translation model is a challenge, because the semantic structures themselves are syntactically motivated. In previous work, Liu and Gildea (2010) model the reordering/deletion of source-side seman- tic roles in a tree-to-string translation model. While it is natural to include semantic structures in a tree- based translation …

Content truncated to 500 words