Semantic Cohesion Model for Phrase-Based SMT

Proceedings of COLING 2012: Technical Papers, pages 867–878, COLING 2012, Mumbai, December 2012. Semantic Cohesion Model for Phrase-based SMT Minwei F ENG1 Weiwei SUN 2 Hermann N EY 1 (1) Human Language Technology and Pattern Recognition Group, Computer Science Department, RWTH Aachen University, Aachen, Germany (2) The MOE Key Laboratory of Computational Linguistics, Institute of Computer Science and Technology, Peking University, Beijing, China feng@cs.rwth-aachen.de, ws@pku.edu.cn, ney@cs.rwth-aachen.de ABSTRACT In this paper, we propose a novel semantic cohesion model. Our model utilizes the predicate- argument structures as soft constraints and plays the role as a reordering model in the phrase- based statistical machine translation system. We build a translation system with GALE data. Experimental results on the NIST02, NIST03, NIST04, NIST05 and NIST08 Chinese-English tasks show that our model improves the baseline system by 0.93 BLEU 0.98 TER on average. We also compare our method with a syntax-augmented model (Cherry, 2008), and demonstrate the importance of predicate-argument semantics in machine translation. KEYWORDS: statistical machine translation, semantic role labeling. 867 1 Introduction In recent years, there are growing interests in incorporating semantics into statistical machine translation (SMT) (Wu and Fung, 2009; Liu and Gildea, 2010; Gao and Vogel, 2011; Baker et al., 2012). Among all existing semantic representations, semantic role labeling (SRL) aims at automatically analyzing predicate-argument structures and can capture essential meaning of sentences. Since the seminal work of (Gildea and Jurafsky, 2002), quite a few researchers concentrate on resolving SRL with different machine learning methods. Nowadays, good SRL systems can be built based on accurate syntactic parsers for English, as well as many other languages. In this paper, we explore predicate-argument analysis of source sentences to improve phrase- based SMT systems. On one hand, the predicate-argument event layer of SRL captures global dependencies which is crucial for the MT output quality. On the other hand, the semantic role information contained in SRL also provide a good clue to the appropriateness of a phrase segment chosen by a translation system. Compared to the syntactic representation in both constituency and dependency formalisms, SRL focuses more on modeling the skeleton of a sentence. To exploit the two attributes of SRL, we propose two types of constraints which are implemented as two models in the decoder. The first model restrains the translation process so that it is consistent with the global dependency in the source sentence. The second model inspects the source sentence segmentation so that each source phrase is consistent with the semantic roles. We conduct experiments on the NIST02, NIST03, NIST04, NIST05 and NIST08 Chinese-English tasks. Experimental results show that our model improves the baseline system by 0.93 BLEU 0.98 TER on average. We also compare our method with a syntax-augmented model (Cherry, 2008), and demonstrate the importance of predicate-argument semantics in machine translation. The remainder of this paper is organized as follows: Section 2 describes the semantic soft constraints proposed for translation system. Section 3 introduces a related work (Cherry, 2008) as we compare our method with it. Section 4 provides the …

Content truncated to 500 words