Abstract
We introduce a method for learning to reorder source sentences. In our approach, sentences are transformed into new sequences of words aimed at reducing non-local reorderings in phrase translation. The method involves automatically extracting instances of structural divergences from sentence pairs, and automatically learning lexicalized grammatical rules probabilistically encoded with bilingual word-order relations. At run-time, source sentences are reordered by applying the rules prior to phrase-based machine translation systems. Experiments show that our method cleanly captures systematic similarities and differences in languages' grammars, resulting in substantial improvement over state-of-the-art phrase-based translation systems. © 2009 IEEE.