Applying class triggers in Chinese pos tagging based on maximum entropy model

A method of applying class triggers in Chinese POS tagging based on Maximum Entropy model is proposed in this paper. First of all, Feature template of "word->word/tat" is used to extract the triggers from corpus and the triggers that we extracted are added into the Maximum Entropy model as a new kind of feature. Then, the average mutual information is applied to make feature selection and the semantic lexicon is used to build class triggers to overcome sparseness problem. Meanwhile, A solution based on experience to deal with over-fitting problem in model training is presented. Finally, the performance of the system is evaluated on a manually annotated POS tag corpus. The experiment demonstrates that the method can provide increase of accuracy of POS tagging from 94% to 96%, compared our new model with HMM model that is smoothed by absolute smoothing.