Using Hypothesis Selection Based Features for Confusion Network MT System Combination

摘要

This paper describes the development operated into MANY, an open source system combination software based on confusion networks developed at LIUM. The hypotheses from Chinese-English MT systems were combined with a new version of the software. MANY has been updated in order to use word confidence score and to boost n-grams occurring in input hypotheses. In this paper we propose either to use an adapted language model or adding some additional features in the decoder to boost certain n-grams probabilities. Experimental results show that the updates yielded significant improvements in terms of BLEU score.