Nowadays statistical machine translation shows its benefits and has received much attention. In this paper, phrase-based statistical machine translation was carefully studied. Improved Hidden Markov Model(HMM) was used to align words and solve the inconsistency between word alignment and phrase structures, and can serve word alignment better. Translation rules were extracted based on aligned phrases and English phrase trees. CYK+, an improved CYK algorithm, as adopted as the decoder to decode non-Chomsky translation rules; Two-round-decoding algorithm was proposed to integrate the language model during decoding. The experiment results showed the BLEU score of improved HMM was higher than the score of HMM, so it follows that the translation system based on translation rules has more stable translation effect on different data collection.