Graph-Based Collective Lexical Selection for Statistical Machine Translation

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 1238–1247, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. Graph-Based Collective Lexical Selection for Statistical Machine Translation Jinsong Su1,2, Deyi Xiong3∗, Shujian Huang2, Xianpei Han4, Junfeng Yao1 Xiamen University, Xiamen, P.R. China1 State Key Laboratory for Novel Software Technology, Nanjing University, P.R. China2 Soochow University, Suzhou, P.R. China3 Institute of Software, Chinese Academy of Sciences, Beijing, P.R. China4 {jssu, yao0010}@xmu.edu.cn huangsj@nlp.nju.edu.cn dyxiong@suda.edu.cn xianpei@nfs.iscas.ac.cn Abstract Lexical selection is of great importance to statistical machine translation. In this paper, we propose a graph-based frame- work for collective lexical selection. The framework is established on a translation graph that captures not only local associ- ations between source-side content words and their target translations but also target- side global dependencies in terms of relat- edness among target items. We also in- troduce a random walk style algorithm to collectively identify translations of source- side content words that are strongly related in translation graph. We validate the ef- fectiveness of our lexical selection frame- work on Chinese-English translation. Ex- periment results with large-scale training data show that our approach significantly improves lexical selection. 1 Introduction Lexical selection, which selects appropriate trans- lations for lexical items on the source side, is a cru- cial task in statistical machine translation (SMT). The task is closely related to two factors: 1) asso- ciations of selected translations with lexical items on the source side, including corresponding source items and their neighboring words, and 2) depen- dencies1 between selected target translations and other items on the target side. As translation rules and widely-used n-gram language models can only capture local associ- ations and dependencies, we have witnessed in- ∗Corresponding author. 1Please note that dependencies in this paper are not nec- essarily syntactic dependencies. creasing efforts that attempt to incorporate non- local associations/dependencies into lexical selec- tion. Efforts using source-side associations mainly focus on the exploitation of either sentence-level context (Chan et al., 2007; Carpuat and Wu, 2007; Hasan et al., 2008; Mauser et al., 2009; He et al., 2008; Shen et al., 2009) or the utilization of document-level context (Xiao et al., 2011; Ture et al., 2012; Xiao et al., 2012; Xiong et al., 2013). In contrast, target-side dependencies attract little attention, although they have an important im- pact on the accuracy of lexical selection. The most common practice is to use language mod- els to estimate the strength of target-side depen- dencies (Koehn et al., 2003; Shen et al., 2008; Xiong et al., 2011). However, conventional n- gram language models are not good at capturing long-distance dependencies. Consider the exam- ple shown in Figure 1. As the translations of pol- ysemous words “w`ent´ı”, “ch´ıyˇou” and “l`ıchˇang” are far from each other, our baseline can only correctly translate “l`ıchˇang” as “stance”. It in- appropriately translates the other two words as “problem” and null, respectively, even with the support of an n-gram language model. If we could model long-distance dependencies among target translations of source words …

Content truncated to 500 words