Soft Dependency Constraints for Reordering in Hierarchical Phrase-Based Translation

Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 857–868, Edinburgh, Scotland, UK, July 27–31, 2011. c⃝2011 Association for Computational Linguistics Soft Dependency Constraints for Reordering in Hierarchical Phrase-Based Translation Yang Gao, Philipp Koehn and Alexandra Birch School of Informatics University of Edinburgh Edinburgh, UK, EH8 9AB yanggao1119@gmail.com, pkoehn@inf.ed.ac.uk, a.birch@ed.ac.uk Abstract Long-distance reordering remains one of the biggest challenges facing machine translation. We derive soft constraints from the source de- pendency parsing to directly address the re- ordering problem for the hierarchical phrase- based model. Our approach significantly im- proves Chinese–English machine translation on a large-scale task by 0.84 BLEU points on average. Moreover, when we switch the tuning function from BLEU to the LRscore which promotes reordering, we observe total improvements of 1.21 BLEU, 1.30 LRscore and 3.36 TER over the baseline. On aver- age our approach improves reordering preci- sion and recall by 6.9 and 0.3 absolute points, respectively, and is found to be especially ef- fective for long-distance reodering. 1 Introduction Reordering, especially movement over longer dis- tances, continues to be a hard problem in statistical machine translation. It motivates much of the re- cent work on tree-based translation models, such as the hierarchical phrase-based model (Chiang, 2007) which extends the phrase-based model (Koehn et al., 2003) by allowing the so-called hierarchical phrases containing subphrases. The hierarchical phrase-based model captures the recursiveness of language without relying on syntac- tic annotation, and promises better reordering than the phrase-based model. However, Birch et al. (2009) find that although the hierarchical phrase- based model outperforms the phrase-based model in terms of medium-range reordering, it does equally poorly in long-distance reordering due to constraints to guarantee efficiency. Syntax-based models that use phrase structure constituent labels as non-terminals in their transfer rules, exemplified by that of Galley et al. (2004), produce smarter and syntactically motivated re- ordering. However, when working with off-the-shelf tools for parsing and alignment, this approach may impose harsh limits on rule extraction and requires serious efforts of optimization (Wang et al., 2010). An alternative approach is to augment the general hierarchical phrase-based model with soft syntactic constraints. Here, we derive three word-based, com- plementary constraints from the source dependency parsing, including: • A dependency orientation feature, trained with maximum entropy on the word-aligned par- allel data, which directly models the head- dependent orientation for source words; • An integer-valued cohesion penalty that com- plements the dependency orientation feature, and fires when a word is not translated with its head. It measures derivation well-formedness and is used to indirectly help reordering; • An auxiliary unaligned penalty feature that mit- igates search error given the other two features. We achieve significant improvements in terms of the overall translation quality and reordering behav- ior. To our knowledge we are the first to use the source dependency parsing to target the reordering problem for hierarchical phrase-based MT. 857 澳洲 是 与 北韩 有 邦交 的 少数 国家 之一 . pobj dobj prep Australia is with North Korea have …

Content truncated to 500 words