Joint decoding of multi-confusion-network in MT system combination

System combination has emerged as a powerful method for machine translation (MT). Inspired by the joint optimization, we re-designed the Incremental Indirect HMM (IHMM) alignment, which is one of the best hypothesis alignment methods for conventional MT system combination, in confusion network construction. This paper pursues a joint decoding strategy for combining outputs from multiple MT systems, where combine confusion network based feature including word alignment, word ordering, entropy and decoding based feature in a single log-linear model. The approaches of joint decoding based on multiple confusion networks are shown to be superior to incremental IHMM alignment in the setting of the Chinese-to-English track of the 2008 NIST Open MT evaluation.