A Probability Model to Refine Word Alignments

Parallel texts have recently been recognized as a valuable resource for machine translation researchers. These texts, which present the same information in two or more languages, are a rich source of translation knowledge. However, the texts themselves are of limited use without a guide to indicate what portions of the two languages correspond to one another. The goal of word alignment algorithms is to link words in bilingual sentence pairs, in order to indicate which words are translations of one another. We present a statistical model for computing the probability of an alignment given a sentence pair. This model allows easy integration of context-specific features. A system called ProAlign has been built around this model to refine an existing word alignment. We show that ProAlign is competitive with the current state of the art in word alignment, and that its central probability model is better suited to alignment refinement than alternative models.