A Maximum Entropy Approach to Combining Word Alignments

Proceedings of the Human Language Technology Conference of the North American Chapter of the ACL, pages 96–103, New York, June 2006. c⃝2006 Association for Computational Linguistics A Maximum Entropy Approach to Combining Word Alignments Necip Fazil Ayan and Bonnie J. Dorr Institute of Advanced Computer Studies (UMIACS) University of Maryland College Park, MD 20742 {nfa,bonnie}@umiacs.umd.edu Abstract This paper presents a new approach to combining outputs of existing word align- ment systems. Each alignment link is rep- resented with a set of feature functions extracted from linguistic features and in- put alignments. These features are used as the basis of alignment decisions made by a maximum entropy approach. The learning method has been evaluated on three language pairs, yielding significant improvements over input alignments and three heuristic combination methods. The impact of word alignment on MT quality is investigated, using a phrase-based MT system. 1 Introduction Word alignment—detection of corresponding words between two sentences that are translations of each other—is usually an intermediate step of statisti- cal machine translation (MT) (Brown et al., 1993; Och and Ney, 2003; Koehn et al., 2003), but also has been shown useful for other applications such as construction of bilingual lexicons, word-sense disambiguation, projection of resources, and cross- language information retrieval. Maximum entropy (ME) models have been used in bilingual sense disambiguation, word reordering, and sentence segmentation (Berger et al., 1996), parsing, POS tagging and PP attachment (Ratna- parkhi, 1998), machine translation (Och and Ney, 2002), and FrameNet classification (Fleischman et al., 2003). They have also been used to solve the word alignment problem (Garcia-Varea et al., 2002; Ittycheriah and Roukos, 2005; Liu et al., 2005), but a sentence-level approach to combining knowledge sources is used rather than a word-level approach. This paper describes an approach to combin- ing evidence from alignments generated by exist- ing systems to obtain an alignment that is closer to the true alignment than the individual align- ments. The alignment-combination approach (called ACME) operates at the level of alignment links, rather than at the sentence level (as in previous ME approaches). ACME uses ME to decide whether to include/exclude a particular alignment link based on feature functions that are extracted from the in- put alignments and linguistic features of the words. Since alignment combination relies on evidence from existing alignments, we focus on alignment links that exist in at least one input alignment. An important challenge in this approach is the selection of appropriate links when two aligners make differ- ent alignment choices. We show that ACME yields a significant relative error reduction over the input alignment systems and heuristic-based combinations on three different lan- guage pairs. Using a higher number of input align- ments and partitioning the training data into disjoint subsets yield further error-rate reductions. The next section briefly overviews ME models. Section 3 presents a new ME approach to combin- ing existing word alignment systems. Section 4 de- scribes the evaluation data, input alignments, and evaluation metrics. Section 5 presents experiments on three language pairs, upper bounds …

Content truncated to 500 words