This paper presents novel approaches toreordering in phrase-based statistical machinetranslation. We perform consistentreordering of source sentences in trainingand estimate a statistical translationmodel. Using this model, we follow aphrase-based monotonic machine translationapproach, for which we develop an efficientand flexible reordering frameworkthat allows to easily introduce different reorderingconstraints. In translation, weapply source sentence reordering on wordlevel and use a reordering automaton as input.We show how to compute reorderingautomata on-demand using IBM or ITGconstraints, and also introduce two newtypes of reordering constraints. We furtheradd weights to the reordering automata.We present detailed experimental resultsand show that reordering significantly improvestranslation quality.
展开▼