Extracting Hierarchical Rules from a Weighted Alignment Matrix

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 1294–1303, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Extracting Hierarchical Rules from a Weighted Alignment Matrix Zhaopeng Tu, Yang Liu, Qun Liu and Shouxun Lin Key Lab. of Intelligent Info. Processing Institute of Computing Technology Chinese Academy of Sciences P.O. Box 2704, Beijing 100190, China {tuzhaopeng,yliu,liuqun,sxlin}@ict.ac.cn Abstract Word alignment is a fundamental step in machine translation. Current statistical machine translation systems suffer from a major drawback: they only extract rules from 1-best alignments, which adversely affects the rule sets quality due to align- ment mistakes. To alleviate this prob- lem, we extract hierarchical rules from weighted alignment matrix (Liu et al., 2009). Since the sub-phrase pairs would change the inside and outside areas in the weighted alignment matrix of the hi- erarchical rules, we propose a new algo- rithm to calculate the relative frequencies and lexical weights of hierarchical rules. To achieve a balance between rule table size and performance, we construct a scor- ing measure that incorporates both fre- quency and lexical weight to select the best target phrase for each source phrase. Experiments show that our approach im- proves BLEU score by ranging from 1.4 to 1.9 points over baseline for hierarchi- cal phrase-based, and 1.4 to 1.5 points for tree-to-string model. 1 Introduction Word alignment plays an important role in statis- tical machine translation (SMT). Most SMT sys- tems, not only phrase-based models (Och and Ney, 2004; Koehn et al., 2003; Xiong et al., 2006), but also syntax-based models (Chiang, 2005; Liu et al., 2006; Galley et al., 2006; Huang et al., 2006; Shen et al., 2008), usually extract rules from word aligned corpora. However, these systems suffer from a major drawback: they only extract rules from 1-best alignments, which adversely affects the rule sets quality due to alignment mistakes. Typically, syntax-based models are more sensi- tive to word alignments because they care about the development of China ’s economy zhongguo de jingji fazhan the development of China ’s economy zhongguo de jingji fazhan Figure 1: (a) One alignment of a sentence pair; (b) another alignment of the same sentence pair. Here coreless dots denote wrong links. inside (i.e., subtracted phrases). Figure 1(a) shows an alignment of a sentence pair. Since there is a wrong link (de, of), we could not extract many useful hierarchical rules such as (zhongguo X1 jingji, China X1 economy).To alleviate this prob- lem, a natural solution is to extract rules from n- best alignments (Venugopal et al., 2008). However, using n-best alignments still face two major challenges. First, n-best alignments have to be processed individually although they share many links, see (zhongguo, China) and (jingji, economy) in Figure 1. Second, regardless of prob- abilities of links in each alignment, numerous wrong rule would be extracted from n-best align- ments. For example, a wrong rule (X1 de jingji, of X1 ’s economy) would be extracted from the alignment in Figure 1(a). Since Liu et al. (2009) show that …

Content truncated to 500 words