Multiple confusion network application in MT system combination

Abstract

System combination has emerged as a powerful method for machine translation (MT). In the construction of word level confusion network (CN), the alignment and skeleton selection are two key issues of system combination. The paper introduces multi-CN for solving skeleton selection. We fail to yield the better performance through using simple prior score as CN-based feature, so we introduce more sophisticated CN-based and consensus-decoding-based features into combination framework to test multi-CN's validity. The approaches of multi-CN are shown to be superior to single-CN in the setting of the Chinese-to-English track of the 2008 NIST Open MT evaluation.