Reordering Phrase-based machine translation over chunks

The paper presents a new method for reordering in phrase based statistical machine translation (PBMT). Our method is based on previous chunk-level reordering methods for PBMT. First, we parse the source language sentence to a chunk tree, according to the method developed by [16]. Second, we apply a series of transformation rules which are learnt automatically from the parallel corpus to the chunk tree over chunk level. Finally, we integrate a global reordering model directly in a decoder as a graph of phrases, and solve the overlapping phrase and chunk problem. The experimental results with English-Vietnamese pairs show that our method outperforms the baseline PBMT in both accuracy and speed