An Incremental Three-pass System Combination Framework by Combining Multiple Hypothesis Alignment Methods

System combination has been applied successfully to various machine translation tasks in recent years. As is known, the hypothesis alignment method is a critical factor for the translation quality of system combination. To date, many effective hypothesis alignment metrics have been proposed and applied to the system combination, such as TER, HMM, ITER, IHMM, and SSCI. In addition, Minimum Bayes-risk (MBR) decoding and confusion networks (CN) have become state-of-the-art techniques in system combination. In this paper, we examine different hypothesis alignment approaches and investigate how much the hypothesis alignment results impact on system combination, and finally present a three-pass system combination strategy that can combine hypothesis alignment results derived from multiple alignment metrics to generate a better translation. Firstly, these different alignment metrics are carried out to align the backbone and hypotheses, and the individual CNs are built corresponding to each set of alignment results; then we construct a ‘super network’ by merging the multiple metric-based CNs to generate a consensus output. Finally a modified MBR network approach is employed to find the best overall translation. Our proposed strategy outperforms the best single confusion network as well as the best single system in our experiments on the NIST Chinese-to-English test set and the WMT2009 English-to-French system combination shared test set.