In this paper, the authors modified the classical HMM (hidden Markovmodel) method so as to train the model whenever the new corpus is introduced and no matter the corpus is tagged or not. Therefore, the information can be learned from the corpusto improve the accuracy of the model.