A Correction Model for Word Alignments

Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 889–898, Edinburgh, Scotland, UK, July 27–31, 2011. c⃝2011 Association for Computational Linguistics A Correction Model for Word Alignments J. Scott McCarley, Abraham Ittycheriah, Salim Roukos, Bing Xiang, Jian-ming Xu IBM T.J. Watson Research Center 1101 Kitchawan Road, Rt. 134 Yorktown Heights, NY 10598 {jsmc,abei,roukos,bxiang,jianxu}@us.ibm.com Abstract Models of word alignment built as sequences of links have limited expressive power, but are easy to decode. Word aligners that model the alignment matrix can express arbitrary align- ments, but are difficult to decode. We pro- pose an alignment matrix model as a cor- rection algorithm to an underlying sequence- based aligner. Then a greedy decoding al- gorithm enables the full expressive power of the alignment matrix formulation. Improved alignment performance is shown for all nine language pairs tested. The improved align- ments also improved translation quality from Chinese to English and English to Italian. 1 Introduction Word-level alignments of parallel text are crucial for enabling machine learning algorithms to fully uti- lize parallel corpora as training data. Word align- ments appear as hidden variables in IBM Models 1- 5 (Brown et al., 1993) in order to bridge a gap be- tween the sentence-level granularity that is explicit in the training data, and the implicit word-level cor- respondence that is needed to statistically model lex- ical ambiguity and word order rearrangements that are inherent in the translation process. Other no- table applications of word alignments include cross- language projection of linguistic analyzers (such as POS taggers and named entity detectors,) a subject which continues to be of interest. (Yarowsky et al., 2001), (Benajiba and Zitouni, 2010) The structure of the alignment model is tightly linked to the task of finding the optimal alignment. Many alignment models are factorized in order to use dynamic programming and beam search for ef- ficient marginalization and search. Such a factoriza- tion encourages - but does not require - a sequential (often left-to-right) decoding order. If left-to-right decoding is adopted (and exact dynamic program- ming is intractable) important right context may ex- ist beyond the search window. For example, the link- age of an English determiner may be considered be- fore the linkage of a distant head noun. An alignment model that jointly models all of the links in the entire sentence does not motivate a par- ticular decoding order. It simply assigns comparable scores to the alignment of the entire sentence, and may be used to rescore the top-N hypotheses of an- other aligner, or to decide whether heuristic pertur- bations to the output of an existing aligner constitute an improvement. Both the training and decoding of full-sentence models have presented difficulties in the past, and approximations are necessary. In this paper, we will show that by using an ex- isting alignment as a starting point, we can make a significant improvement to the alignment by propos- ing a series of heuristic perturbations. In effect, we train a model to fix the errors of the …

Content truncated to 500 words