A decoding method of system combination based on hypergraph in SMT

Abstract

The word level system combination, which is better than phrase level and sentence level, has emerged as a powerful post-processing method for statistical machine translation (SMT). This paper first give the definition of HyperGraph(HG) as a kind of compact data structure in SMT, and then introduce simple bracket transduction grammar(SBTG) for hypergraph decoding. To optimize the more feature weights, we introduce minimum risk (MR) with deterministic annealing (DA) into the training criterion, and compare two classic training procedures in experiment. The deoding approaches of n-gram model based on hypergraph are shown to be superior to conventional cube pruning in the setting of the Chinese-to-English track of the 2008 NIST Open MT evaluation. © 2011 Springer-Verlag.