A Study On Chinese Named Entity Recognition

Chinese Named Entity Recognition is a basic task of Natural Language Processing and also it is basis of some NLP tasks, such as machine translation, information retrieval, question answering and so on.Firstly, a model based on CRFs is built to do NER task. CRFs model, one of the best machine learning models, is widely used in NLP area and has a good performance. CRFs, an undirected graphical model, can avoid bias problems belong to direct graphical models, and at the same time, it can consider the information between correlative nodes. Analyzing the results obtained by sole CRFs, we find that the errors mainly happened in labels with low marginal probabilities. Two methods, statistical method and boundary template method, are introduced to correct the errors. If the marginal probability is greater than the given threshold, the test sample is recognized by CRFs; otherwise, one of these two methods is used. Experimental results show that the two hybrid methods have better performance than the CRFs method.Secondly, this paper introduces a new machine learning model, Max-Margin Markov Networks. It combines the advantages of both SVM and undirected graphical model. A novel method based on Max-Margin Markov Networks is presented in this paper to do Chinese location NER task and it obtains better results than CRFs and SVM.Lastly, a kind of CRFs model based on probability feature functions is presented. Probability feature functions are defined to replace binary functions, as to improve machine learning ability of system. Then NER tasks are employed to test machine learning ability of this improved CRFs. Experimental results show that methods based on improved CRFs are better than CRFs method.Our methods are expected to extend to other tasks of NLP area.