Synchronous Constituent Context Model for Inducing Bilingual Synchronous Structures

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1865–1874, Dublin, Ireland, August 23-29 2014. Synchronous Constituent Context Model for Inducing Bilingual Synchronous Structures Xiangyu Duan Min Zhang∗ Qiaoming Zhu School of Computer Science & Technology, Soochow University {xiangyuduan;minzhang;qmzhu}@suda.edu.cn Abstract Traditional Statistical Machine Translation (SMT) systems heuristically extract synchronous structures from word alignments, while synchronous grammar induction provides better so- lutions that can discard heuristic method and directly obtain statistically sound bilingual syn- chronous structures. This paper proposes Synchronous Constituent Context Model (SCCM) for synchronous grammar induction. The SCCM is different to all previous synchronous grammar induction systems in that the SCCM does not use the Context Free Grammars to model the bilin- gual parallel corpus, but models bilingual constituents and contexts directly. The experiments show that valuable synchronous structures can be found by the SCCM, and the end-to-end ma- chine translation experiment shows that the SCCM improves the quality of SMT results. 1 Introduction Traditional Statistical Machine Translation (SMT) learns translation model from bilingual corpus that is sentence aligned. No large-scale hand aligned structures inside the parallel sentences are usually available to the SMT community, while the aligned structures are essential for training the translation model. Thus, various unsupervised methods had been explored to automatically obtain aligned structures inside the parallel sentences. Currently, the dominant method is a two step pipeline that obtains word alignments by unsupervised learning (Brown et al., 1993) at the first step, then obtains aligned structures at the second step by heuristically extracting all bilingual structures that are consistent with the word alignments. The second step in this two step pipeline is problematic due to its obtained aligned structures, whose counts are heuristically collected and violate valid translation derivations, while most SMT decoders perform translation via valid translation derivations. This problem leads to researches on synchronous grammar induction that discards the heuristic method and the two separate steps pipeline. Synchronous grammar induction aims to directly obtain aligned structures by using one statistically sound model. The aligned structures in synchronous grammar induction are hierarchical/syntax level (Cohn and Blunsom, 2009) synchronous structures, which can be modeled by Synchronous Context Free Grammars (SCFGs) (Cohn and Blunsom, 2009; Levenberg et al., 2012; Xiao et al., 2012; Xiao and Xiong, 2013) or a kind of SCFGs variant - Inversion Transduction Grammars (ITGs) (Neubig et al., 2011; Cohn and Haffari, 2013). Both SCFGs and ITGs are studied in recent years by using generative or discriminative modeling. This paper departs from using the above two traditional CFGs-based grammars, and proposes Syn- chronous Constituent Context Model (SCCM) which models synchronous constituents and contexts directly so that bilingual translational equivalences can be directly modeled. The proposed SCCM is inspired by researches on monolingual grammar induction, whose experience is valuable to the syn- chronous grammar induction community due to its standard evaluation on released monolingual tree- banks, while no hand annotated bilingual synchronous treebank is available for evaluating synchronous ∗Corresponding Author This work is licenced under a Creative Commons Attribution …

Content truncated to 500 words