Using N-gram based Features for Machine Translation System Combination

Conventional confusion network based system combination for machine translation (MT) heavily relies on features that are based on the measure of agreement of words in different translation hypotheses. This paper presents two new features that consider agreement of n-grams in different hypotheses to improve the performance of system combination. The first one is based on a sentence specific online n-gram language model, and the second one is based on n-gram voting. Experiments on a large scale Chinese-to-English MT task show that both features yield significant improvements on the translation performance, and a combination of them produces even better translation results.