Constituent Reordering and Syntax Models for English-to-Japanese Statistical Machine Translation

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 626–634, Beijing, August 2010 Constituent Reordering and Syntax Models for English-to- Japanese Statistical Machine Translation Young-Suk Lee IBM Research ysuklee@us.ibm.com Bing Zhao IBM Research zhaob@us.ibm.com Xiaoqiang Luo IBM Research xiaoluo@us.ibm.com Abstract We present a constituent parsing-based reordering technique that improves the performance of the state-of-the-art Eng- lish-to-Japanese phrase translation sys- tem that includes distortion models by 4.76 BLEU points. The phrase transla- tion model with reordering applied at the pre-processing stage outperforms a syn- tax-based translation system that incor- porates a phrase translation model, a hi- erarchical phrase-based translation model and a tree-to-string grammar. We also show that combining constituent re- ordering and the syntax model improves the translation quality by additional 0.84 BLEU points. 1 Introduction Since the seminal work by (Wu, 1997) and (Ya- mada and Knight, 2001), there have been great advances in syntax-based statistical machine translation to accurately model the word order distortion between the source and the target lan- guages. Compared with the IBM source-channel mod- els (Brown et al., 1994) and the phrase transla- tion models (Koehn et al., 2003), (Och and Ney, 2004) which are good at capturing local reorder- ing within empirical phrases, syntax-based mod- els have been effective in capturing the long- range reordering between language pairs with very different word orders like Japanese-English (Yamada and Knight, 2001), Chinese-English (Chiang, 2005) and Urdu-English (Zollmann et al. 2008), (Callison-Burch et al. 2010). However, (Xu et al., 2009) show that apply- ing dependency parsing-based reordering as pre- processing (pre-ordering hereafter) to phrase translation models produces translation qualities significantly better than a hierarchical phrase- based translation model (Hiero hereafter) im- plemented in (Zollman and Venugopal, 2006) for English-to-Japanese translation. They also report that the two models result in comparable translation qualities for English-to- Korean/Hindi/Turkish/Urdu, underpinning the limitations of syntax-based models for handling long-range reordering exhibited by the strictly head-final Subject-Object-Verb (SOV) order languages like Japanese and the largely head- initial Subject-Verb-Object (SVO) order lan- guages like English. In this paper, we present a novel constituent parsing-based reordering technique that uses manually written context free (CFG hereafter) and context sensitive grammar (CSG hereafter) rules. The technique improves the performance of the state-of-the-art English-to-Japanese phrase translation system that includes distortion models by 4.76 BLEU points. The phrase trans- lation model with constituent pre-ordering con- sistently outperforms a syntax-based translation system that integrates features from a phrase translation model, Hiero and a tree-to-string grammar. We also achieve an additional 0.84 BLEU point improvement by applying an ex- tended set of reordering rules that incorporate new rules learned from the syntax model for decoding. The rest of the paper is organized as follows. In Section 2, we summarize previous work re- lated to this paper. In Section 3, we give an overview of the syntax model with which we compare the performance of a phrase translation 626 model with pre-ordering. We also discuss a chart-based decoder used in all of our experi- ments. In Section 4, …

Content truncated to 500 words