Abstract: Statistical machine translation(SMT)system is trained with large-scale and domain-mixed parallel corpus, when the data for training and testing are not belonged to the same domain, the translation quality usually drops dramatically. To solve this problem, we proposed a novel approach to adapt the translation model based on semantic distribution similarity of translation pair. The approach firstly obtained word representations both in source and target language, and then built mapping to link the different vector space. With the mapping function the semantic k-nearest neighbors of source language in the target vector space can be easily obtained. Based on the semantic distribution of k neighbors in the general domain space, we computed phrases translation similarity in the domain of interest. The similarities are then integrated into the decoder engine, in order to enhance the adaption ability of common translation model. Experiments on English to Chinese translation tasks show that the optimized translation systems build on our method outperform the baseline system by 0.67 and 0.56 BLUE points on news and science-technology test sets respectively.

Key words: translation model, word representation, semantic distribution, domain adaptation