Word Reordering Alignment for Combination of Statistical Machine Translation Systems

Word alignment is a basic and critical process in the Statistical Machine Translation (SMT). The previous work on word alignment mainly focuses on the training process to get the word mapping relation between the source sentences and target sentences. However, the word alignment for combination of SMT system outputs is also important, which aims to find the word correspondence between alternative translation hypotheses of a source language sentence. Unfortunately, it does not attract so much attention in SMT research. In this paper, we propose a novel word alignment approach to effectively address the word alignment between sentences with different valid word orders, which changes the order of the word sequences (called word reordering) of the output hypotheses to make the word order more exactly match the alignment reference. We present experimental results on the IWSLT'2008 challenge tasks with the combination of four state-of-the-art SMT systems outputs. The results show that our approach significantly improves the performance of the system combination.