Syntactic SMT Using a Discriminative Text Generation Model

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 177–182, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Syntactic SMT Using a Discriminative Text Generation Model Yue Zhang Kai Song∗ Linfeng Song∗ SUTD, Singapore NEU, China ICT/CAS, China yue zhang@sutd.edu.sg songkai.sk@alibaba-inc.com songlinfeng@ict.ac.cn Jingbo Zhu Qun Liu NEU, China CNGL, Ireland and ICT/CAS, China zhujingbo@mail.neu.edu.cn qliu@computing.dcu.ie Abstract We study a novel architecture for syntactic SMT. In contrast to the dominant approach in the literature, the system does not rely on translation rules, but treat translation as an unconstrained target sentence gen- eration task, using soft features to cap- ture lexical and syntactic correspondences between the source and target languages. Target syntax features and bilingual trans- lation features are trained consistently in a discriminative model. Experiments us- ing the IWSLT 2010 dataset show that the system achieves BLEU comparable to the state-of-the-art syntactic SMT systems. 1 Introduction Translation rules have been central to hierarchi- cal phrase-based and syntactic statistical machine translation (SMT) (Galley et al., 2004; Chiang, 2005; Liu et al., 2006; Quirk et al., 2005; Marcu et al., 2006; Shen and Joshi, 2008; Xie et al., 2011). They are attractive by capturing the recursiveness of languages and syntactic correspondences be- tween them. One important advantage of trans- lation rules is that they allow efficient decoding by treating MT as a statistical parsing task, trans- forming a source sentence to its translation via re- cursive rule application. The efficiency takes root in the fact that target word orders are encoded in translation rules. This fact, however, also leads to rule explosion, noise and coverage problems (Auli et al., 2009), which can hurt translation quality. Flexibility of function word usage, rich morphology and paraphrasing all add to the difficulty of rule extraction. In addition, restricting target word orders by hard translation rules can also hurt output fluency. ∗* Work done while visiting Singapore University of Technology and Design (SUTD) Figure 1: Overall system architecture. A potential solution to the problems above is to treat translation as a generation task, represent- ing syntactic correspondences using soft features. Both adequacy and fluency can potentially be im- proved by giving full flexibility to target synthe- sis, and leaving all options to the statistical model. The main challenge to this method is a signifi- cant increase in the search space (Knight, 1999). To this end, recent advances in tackling complex search tasks for text generation offer some so- lutions (White and Rajkumar, 2009; Zhang and Clark, 2011). In this short paper, we present a preliminary in- vestigation on the possibility of building a syn- tactic SMT system that does not use hard transla- tion rules, by utilizing recent advances in statisti- cal natural language generation (NLG). The over- all architecture is shown in Figure 1. Translation is performed by first parsing the source sentence, then transferring source words and phrases to their target equivalences, and finally synthesizing the target output. We choose dependency grammar for both the source and the …

Content truncated to 500 words