Methods Of Feature Extraction And Polarity Analysis For Chinese Opinion Mining

With the rapid development of Internet, web applications such as shopping online, micro blogging, blogging and BBS are becoming hot spots of interesting. Opinions and comments on online shopping nets, micro blogging, blogging and BBS are increasing with a surprising speed and are full of abundance information. For example, businesses can improve the production quality based on these feedback comments for a higher efficiency, while customers could select a desired product by the reviews, and managers could enhance the management efficiency. However, facing the huge numbers of opinions and reviews, how to how to handle of these unstructured data efficiently and effectively and how to obtain the useful information is a challenging and significant issue in the opinion mining.In this dissertation, researches are carried out on topic selection and polarity analysis of opinion mining as below:(1) We first introduce the concepts of opinion mining and main tasks of opinion mining, further summarize the related works and address the relationships of opining mining are pointed out as followed.(2) For existing methods of feature selection are mostly based on statistical, syntactical or template methods, which ignores the review characteristics and sentence structure information. Thus, we propose a new method of extracting opinion features in sentiment patterns called OFESP. Our method could present the review characteristic in sentiment patterns and satisfy the diversity of part of speech in Chinese word segmentation. Experiments show that OFESP is effective compared to traditional methods..(3) Further, existing methods of polarity analysis present inferiority in the positive opinion discrimination from negative opinion, and are unobvious in polarity hierarchy. To handle this issue, we propose a new method of word polarity analysis with polarity similarity measuring called POS. Our algorithm adds polarity primitives that are not considered moderately in the existing methods. Extensive results show that POS performs better in the accuracy of classifying different polarity and in the hierarchy of polarity value.(4) Last, the aforementioned work is integrated to an online opinion mining system called OURS, which provider the visual interface and the interaction interface.