Combining Multiple Alignments to Improve Machine Translation

Word alignment is a critical component of machine translation systems. Various methods for word alignment have been proposed, and different models can produce significantly different outputs. To exploit the advantages of different models, we propose three ways to combine multiple alignments for machine translation: (1) alignment selection, a novel method to select an alignment with the least expected loss from multiple alignments within the minimum Bayes risk framework; (2) alignment refinement, an improved algorithm to refine multiple alignments into a new alignment that favors the consensus of various models; (3) alignment compaction, a compact representation that encodes all alignments generated by different methods (including (1) and (2) above) using a novel calculation of link probabilities. Experiments show that our approach not only improves the alignment quality, but also significantly improves translation performance by up to 1.96 BLEU points over single best alignments, and 1.28 points over merging rules extracted from multiple alignments individually.