Hybrid decoding: Decoding with partial hypotheses combination over multiple SMT systems

In this paper, we present hybrid decoding - a novel statistical machine translation (SMT) decoding paradigm using multiple SMT systems. In our work, in addition to component SMT systems, system combination method is also employed in generating partial translation hypotheses throughout the decoding process, in which smaller hypotheses generated by each component decoder and hypotheses combination are used in the following decoding steps to generate larger hypotheses. Experimental results on NIST evaluation data sets for Chinese-to-English machine translation (MT) task show that our method can not only achieve significant improvements over individual decoders, but also bring substantial gains compared with a state-of-the-art word-level system combination method.