Research On Opinion Extraction And Classification Technologies For Product Review Mining

With the Internet and E-commerce development and popularity, Web has largely changed the way that customers feedback opinions. Customers give their reviews online after purchasing and using these products. The reviews show their positive or negative attitude toward product's performance and function. Manufacturers and customers will get a great deal of useful information by analyzing products reviews. However, products reviews on the web are mass and structure less data. Manufacturers and customers get information from product review only through manual means to read, and it is a time-consuming and error-prone process. Therefore, product reviews mining was emerged. Customer opinion extraction and polarity classification is one of key technologies for product review mining,and it have great research value.This paper intensive studies products feature extraction technology and customer opinion extraction and polarity classification technologies. The main works of this paper include the following:Firstly, this paper introduced the research background, significance and the situation of study at home and abroad of product review mining,then explains the key techniques of the product reviews mining. At the same time, review data preprocessing solution is proposed.Secondly, we study and analyze two methods for feature word extraction ,one is based on statistical way,the other is based on pattern extraction and match。Then,we combine two methods for high-frequency and low-frequency feature words extraction. First of all,Word sequences that contain noun will be extracted from reviews as candidate features, and then three pruning rules are defined used to obtain high-frequency feature word. After that, we use pattern extraction and matching approach to identify low-frequency feature words. Experimental results show the new method for high frequency and low frequency words is more effective than the feature word extraction method that use association rule to mining frequency item sets and prune by support. Finally, we divided the feature words extracted into common feature words and specific feature words. A product features database also been built.Thirdly,a method for feature-opinion pairs extraction based on dependency relations is proposed. We extracted part of speeches and dependency relations of words from review to generate dependency relation and part of speech pairs at first. Then, we use specific dependency relation and part of speech pairs to generate feature-opinion pairs. By this way, we can find the relationships between customer opinions and product features.In addition, the paper give a way to compute polarity strength of feature-opinion pairs,then proposes a customer opinion polarity classification method according to polarity strength. First, we build a Chinese polarity dictionary by using HowNet and other resources. Second, we apply this dictionary to calculate the polarity strength of feature-opinion pairs. Moreover, polarity classification method and polarity strength calculate method of products, features and reviews based on the polarity of feature-opinion pairs are proposed. Experiments show that these polarity classification methods based on polarity dictionary and feature-opinion pairs have achieved good results for Chinese product review mining.In the end, this paper concludes by summarizing the research and indicating its future work.