Alignment Link Projection Using Transformation-Based Learning

Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing (HLT/EMNLP), pages 185–192, Vancouver, October 2005. c⃝2005 Association for Computational Linguistics Alignment Link Projection Using Transformation-Based Learning Necip Fazil Ayan, Bonnie J. Dorr and Christof Monz Department of Computer Science University of Maryland College Park, MD 20742 {nfa,bonnie,christof}@umiacs.umd.edu Abstract We present a new word-alignment ap- proach that learns errors made by ex- isting word alignment systems and cor- rects them. By adapting transformation- based learning to the problem of word alignment, we project new alignment links from already existing links, using features such as POS tags. We show that our align- ment link projection approach yields a sig- nificantly lower alignment error rate than that of the best performing alignment sys- tem (22.6% relative reduction on English- Spanish data and 23.2% relative reduction on English-Chinese data). 1 Introduction Word-level alignment is a critical component of a wide range of NLP applications, such as construc- tion of bilingual lexicons (Melamed, 2000), word sense disambiguation (Diab and Resnik, 2002), pro- jection of language resources (Yarowsky et al., 2001), and statistical machine translation. Although word-level aligners tend to perform well when there is enough training data, the quality of word align- ment decreases as the size of training data de- creases. Moreover, word-alignment systems are of- ten tripped up by many-to-many correspondences, morphological language distinctions, paraphrased and free translations, and a high percentage of func- tion words (about 50% of the tokens in most texts). At the heart of the matter is a set of assumptions that word-alignment algorithms must make in order to reduce the hypothesis space, since word align- ment is an exponential problem. Because of these assumptions, learning algorithms tend to make sim- ilar errors throughout the entire data. This paper presents a new approach—Alignment Link Projection (ALP)—that learns common align- ment errors made by an alignment system and at- tempts to correct them. Our approach assumes the initial alignment system adequately captures certain kinds of word correspondences but fails to handle others. ALP starts with an initial alignment and then fills out (i.e., projects) new word-level alignment re- lations (i.e., links) from existing alignment relations. ALP then deletes certain alignment links associated with common errors, thus improving precision and recall. In our approach, we adapt transformation-based learning (TBL) (Brill, 1995; Brill, 1996) to the prob- lem of word alignment. ALP attempts to find an ordered list of transformation rules (within a pre- specified search space) to improve a baseline anno- tation. The rules decompose the search space into a set of consecutive words (windows) within which alignment links are added, to or deleted from, the initial alignment. This window-based approach ex- ploits the clustering tendency of alignment links, i.e., when there is a link between two words, there is frequently another link in close proximity. TBL is an appropriate choice for this problem for the following reasons: 1. It can be optimized directly with respect to an evaluation metric. 2. It learns rules that …

Content truncated to 500 words