Using Syntax to Improve Word Alignment Precision for Syntax-Based Machine Translation

Proceedings of the Third Workshop on Statistical Machine Translation, pages 44–52, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics Using Syntax to Improve Word Alignment Precision for Syntax-Based Machine Translation Victoria Fossum Dept. of Computer Science University of Michigan Ann Arbor, MI 48104 vfossum@umich.edu Kevin Knight Information Sciences Institute University of Southern California Marina del Rey, CA 90292 knight@isi.edu Steven Abney Dept. of Linguistics University of Michigan Ann Arbor, MI 48104 abney@umich.edu Abstract Word alignments that violate syntactic cor- respondences interfere with the extraction of string-to-tree transducer rules for syntax- based machine translation. We present an algorithm for identifying and deleting incor- rect word alignment links, using features of the extracted rules. We obtain gains in both alignment quality and translation quality in Chinese-English and Arabic-English transla- tion experiments relative to a GIZA++ union baseline. 1 Introduction 1.1 Motivation Word alignment typically constitutes the first stage of the statistical machine translation pipeline. GIZA++ (Och and Ney, 2003), an implementation of the IBM (Brown et al., 1993) and HMM (?) alignment models, is the most widely-used align- ment system. GIZA++ union alignments have been used in the state-of-the-art syntax-based statistical MT system described in (Galley et al., 2006) and in the hierarchical phrase-based system Hiero (Chiang, 2007). GIZA++ refined alignments have been used in state-of-the-art phrase-based statistical MT sys- tems such as (Och, 2004); variations on the refined heuristic have been used by (Koehn et al., 2003) (diag and diag-and) and by the phrase-based system Moses (grow-diag-final) (Koehn et al., 2007). GIZA++ union alignments have high recall but low precision, while intersection or refined align- ments have high precision but low recall.1 There are two natural approaches to improving upon GIZA++ alignments, then: deleting links from union align- ments, or adding links to intersection or refined alignments. In this work, we delete links from GIZA++ union alignments to improve precision. The low precision of GIZA++ union alignments poses a particular problem for syntax-based rule ex- traction algorithms such as (Quirk et al., 2005; Gal- ley et al., 2006; Huang et al., 2006; Liu et al., 2006): if the incorrect links violate syntactic corre- spondences, they force the rule extraction algorithm to extract rules that are large in size, few in number, and poor in generalization ability. Figure 1 illustrates this problem: the dotted line represents an incorrect link in the GIZA++ union alignment. Using the rule extraction algorithm de- scribed in (Galley et al., 2004), we extract the rules shown in the leftmost column (R1–R4). Rule R1 is large and unlikely to generalize well. If we delete the incorrect link in Figure 1, we can extract the rules shown in the rightmost column (R2–R9): Rule R1, the largest rule from the initial set, disappears, and several smaller, more modular rules (R5–R9) re- place it. In this work, we present a supervised algorithm that uses these two features of the extracted rules (size of largest rule and total number of rules), as well as a handful of structural and lexical …

Content truncated to 500 words