A Novel Method for the N-best Generation of Phrased-Based SMT

Phrase based statistic MT (PBSMT) is an important milestone in SMT. In the PBSMT, finding proper weights is an important issue. Currently, the MERT is widely adapted to optimize the weights of the models for various SMT approaches. For PBSMT, the N-best generation is an important factor affecting the performance of MERT according to our experimental results. In previous method on N-best generation of PBSMT, the word lattice is employed which can reserve exponential in polynomial time. However, in this method, the hypotheses which are similar to the best hypotheses probably obtain high score as well, which may further encourages the similarity of the N-best. In this paper, a novel N-best generation method is proposed to retain considerable divergence for PBSMT. According the experimental results, our method is significant better than previous widely adopted method in Moses.