Weighted Alignment Matrices for Statistical Machine Translation

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 1017–1026, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Weighted Alignment Matrices for Statistical Machine Translation Yang Liu , Tian Xia , Xinyan Xiao and Qun Liu Key Laboratory of Intelligent Information Processing Institute of Computing Technology Chinese Academy of Sciences P.O. Box 2704, Beijing 100190, China {yliu,xiatian,xiaoxinyan,liuqun}@ict.ac.cn Abstract Current statistical machine translation sys- tems usually extract rules from bilingual corpora annotated with 1-best alignments. They are prone to learn noisy rules due to alignment mistakes. We propose a new structure called weighted alignment matrix to encode all possible alignments for a par- allel text compactly. The key idea is to as- sign a probability to each word pair to in- dicate how well they are aligned. We de- sign new algorithms for extracting phrase pairs from weighted alignment matrices and estimating their probabilities. Our ex- periments on multiple language pairs show that using weighted matrices achieves con- sistent improvements over using n-best lists in significant less extraction time. 1 Introduction Statistical machine translation (SMT) relies heav- ily on annotated bilingual corpora. Word align- ment, which indicates the correspondence be- tween the words in a parallel text, is one of the most important annotations in SMT. Word-aligned corpora have been found to be an excellent source for translation-related knowledge, not only for phrase-based models (Och and Ney, 2004; Koehn et al., 2003), but also for syntax-based models (e.g., (Chiang, 2007; Galley et al., 2006; Shen et al., 2008; Liu et al., 2006)). Och and Ney (2003) indicate that the quality of machine transla- tion output depends directly on the quality of ini- tial word alignment. Modern alignment methods can be divided into two major categories: generative methods and dis- criminative methods. Generative methods (Brown et al., 1993; Vogel and Ney, 1996) treat word alignment as a hidden process and maximize the likelihood of bilingual training corpus using the expectation maximization (EM) algorithm. In contrast, discriminative methods (e.g., (Moore et al., 2006; Taskar et al., 2005; Liu et al., 2005; Blunsom and Cohn, 2006)) have the freedom to define arbitrary feature functions that describe var- ious characteristics of an alignment. They usu- ally optimize feature weights on manually-aligned data. While discriminative methods show supe- rior alignment accuracy in benchmarks, genera- tive methods are still widely used to produce word alignments for large sentence-aligned corpora. However, neither generative nor discriminative alignment methods are reliable enough to yield high quality alignments for SMT, especially for distantly-related language pairs such as Chinese- English and Arabic-English. The F-measures for Chinese-English and Arabic-English are usually around 80% (Liu et al., 2005) and 70% (Fraser and Marcu, 2007), respectively. As most current SMT systems only use 1-best alignments for ex- tracting rules, alignment errors might impair trans- lation quality. Recently, several studies have shown that offer- ing more alternatives of annotations to SMT sys- tems will result in significant improvements, such as replacing 1-best trees with packed forests (Mi et al., 2008) and replacing …

Content truncated to 500 words