Transformation from Discontinuous to Continuous Word Alignment Improves Translation Quality

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 147–152, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Transformation from Discontinuous to Continuous Word Alignment Improves Translation Quality Zhongjun He1 Hua Wu1 Haifeng Wang1 Ting Liu2 1 Baidu Inc., No. 10, Shangdi 10th Street, Beijing, 100085, China 2 Harbin Institute of Technology, Harbin, China {hezhongjun,wu hua,wanghaifeng}@baidu.com tliu@ir.hit.edu.cn Abstract We present a novel approach to im- prove word alignment for statistical ma- chine translation (SMT). Conventional word alignment methods allow discontin- uous alignment, meaning that a source (or target) word links to several target (or source) words whose positions are dis- continuous. However, we cannot extrac- t phrase pairs from this kind of align- ments as they break the alignment con- sistency constraint. In this paper, we use a weighted vote method to transform dis- continuous word alignment to continuous alignment, which enables SMT system- s extract more phrase pairs. We carry out experiments on large scale Chinese- to-English and German-to-English trans- lation tasks. Experimental results show statistically significant improvements of BLEU score in both cases over the base- line systems. Our method produces a gain of +1.68 BLEU on NIST OpenMT04 for the phrase-based system, and a gain of +1.28 BLEU on NIST OpenMT06 for the hierarchical phrase-based system. 1 Introduction Word alignment, indicating the correspondence between the source and target words in bilingual sentences, plays an important role in statistical machine translation (SMT). Almost all of the SMT models, not only phrase-based (Koehn et al., 2003), but also syntax-based (Chiang, 2005; Liu et al., 2006; Huang et al., 2006), derive translation knowledge from large amount bilingual text anno- tated with word alignment. Therefore, the quality of the word alignment has big impact on the qual- ity of translation output. Word alignments are usually automatically ob- tained from a large amount of bilingual training corpus. The most widely used toolkit for word alignment in SMT community is GIZA++ (Och and Ney, 2004), which implements the well known IBM models (Brown et al., 1993) and the HM- M model (Vogel and Ney, 1996). Koehn et al. (2003) proposed some heuristic methods (e.g. the “grow-diag-final” method) to refine word align- ments trained by GIZA++. Another group of word alignment methods (Liu et al., 2005; Moore et al., 2006; Riesa and Marcu, 2010) define feature functions to describe word alignment. They need manually aligned bilingual texts to train the mod- el. However, the manually annotated data is too expensive to be available for all languages. Al- though these models reported high accuracy, the GIZA++ and “grow-diag-final” method are domi- nant in practice. However, automatic word alignments are usu- ally very noisy. The example in Figure 1 shows a Chinese and English sentence pair, with word alignment automatically trained by GIZA++ and the “grow-diag-final” method. We find many er- rors (dashed links) are caused by discontinuous alignment (formal definition is described in Sec- tion 2), a source (or target) word linking to sev- eral discontinuous target (or …

Content truncated to 500 words