Study On Subjective Statements Screening Of Chinese Product Reviews Based On Features Combination

With the rapid development of network technology, the amount of Internet information growing fast, various business sites are emerging on the Internet. A large number of consumers will consult the evaluation published by other users, in order to guide their purchases. On the other hand, for manufacturers, analyzing product reviews is an effective way to obtain the feedback for the market research. Producers which can make great efforts to improve various aspects of product after analyzing product reviews. So the extraction of product reviews has becoming the hot issue in natural language processing, machine learning, information extraction and many other areas. While the product reviews are subjective statement to express the users’ views, opinions and feelings, so distinguishing the subjective and objective statement has become the primary subject of opinion mining.At present in the research field of subjective and objective statements for English classification has achieved some results and accumulated some method, subjective vocabularies and other features. However, due to the late start and the complexity of Chinese, this research is still in the developing stage. The thesis focuses on the following research.Firstly, this thesis presents the difference between Chinese product reviews and other texts and then some objective word features are selected for Chinese product reviews. We combine new subjective word features with existing features then construct subjective and objective word lists.Secondly, this thesis conducts the research to the N-pos model of the effect of different values, such as Bi-pos model, Tri-pos model, Tetra-pos model, Penta-pos model. The models are used as classification feature based on statistics.Thirdly, on the basis of syntactic analysis, the labels of syntactic structures are used as statistical features and we propose syntactic dependency features.Finally, this thesis combines subjective and objective word features with N-pos model features and syntactic dependency features as grammatical features, and then uses them to construct Naive Bayesian classifier.It is shown in the experiment results that using combined features achieves the best performance.