Research On Product Reviews Analysis Based On Appraisal Expression

As the developing of Web 2.0 technogly, the conception of "user participated website content generation" have a great influnce in our daily lives. The Web 2.0 sites such as blog, SNS and microblog attract huge number of users. The user generated contents are growing as explosion. Among them, the product reviews which contain the opinionated data about the product have brought many users' attention. The customers are used to read the reviews before purchase a product and the sellers want to analyze the reviews to get useful feedbacks. Howerver, the amount of the reviews is huge, and the reviews are usually expressed informally. In order to analyze the reviews fast and efficiently, sentiment analysis or called opinion mining is essential.Target extraction and polarity analysis are two main tasks of sentiment analysis. We conducted researches on the two tasks respecitvely. We chose the product reviws as our corpus, and did two works sequentially: appraisal expression extaction and polarity analysis. We define the compositions or properties of the products as the features, and analyzed the review in the feature level such as wheather positive or negative about the features. Finally, we efficiently organize the analyzed results and showed them to users.Our works are mainly consisted of three parts:(1) We proposed a structure to present the appraisal expression patterns, and constructed the library automantically.The appraisal expression is the basic unit of product reviews. It is consisted of target and sentiment word. Currently the appraisal expressions are extracted by pattern recognization which uses flat syntax features to present the pattern and manually construct the pattern library. We proposed a new structure to present the appraisal expression patterns, which contains the structure information of parse tree, and is able to filter noise to achieve higher precision. We also used a method to construct pattern library automatically, so as to avoid the cost of manual construction.(2) We proposed an approximate pattern matching method based on convolution tree kernel to extract appraisal expressions.The commonly used exact matching method is hard to deal with the complicated patterns because of their diversity. So we propose an approximate matcing method, which uses convolution tree kernel to calculate the similarity between trees, to extract the appraisal expressions. The recall of the method is higher than former exact matching method. Further more, we proposed an approximate convolution tree kernel, improving the traditional convolution tree kernel, which increased the recall greatly with only slight decrease in precision.(3)We constructed a product feature ontology to support polarity analysis.Sentiment lexicon is an essential resource for polarity analysis. Though the exsiting sentiment lexicons contain the polarity of the sentiment words, but they didn't take into an important consideration that the polarity of the same word maybe depends on his afflicated targets and features, which means that a commonly used domain indepent lexicon is not well suited. On the other hand, the features extracted are not well organized to be shown to users.In order to solve the above two problems, we build a sentiment word taxonomy and a product feature ontology, and propose a method to combine these two sturcutres for the precise polarity analysis. The experiment result conducted on polarity classification showed the effectiveness of our method, besides, the product feature ontology is very useful to help organize the mining results.The conducted experiments have shown our methods in the paper are available and efficient.