Abstract: Improved hidden Markov model was used to align words and solve the inconsistency between word alignment and phrase structures. Translation rules were extracted based on aligned phrases and English phrase trees. An extended CYK一CYK algorithm was used as the decoder and a two-pass-decoding algorithm was proposed for intergrating the language model during decoding, which can decode non-Chomsky normal form. The experimental results show the 13I_EU score of improved HMM is higher than the score of HMM, and the translation quality of translation rules is bet- ter than phrase-based machine transtion. The BLEU score of two-pass-decoding algorithm is close to the score of cube prune algorithm and decoding time costs less.

Key words: Statistical machine translation,Improved hidden markov modcl(HMM), I}ranslation ru1c,CYK- algorithm, I3IFI1