Research On Product Reviews Mining In Chinese Based On Reinforcement Of Features

With the rapid development of the electronic commerce business and the increase of online shopping customers, online product reviews as a way of reflect is dramatically increasing. Therefore, the automatic mining on these reviews is important for the potential customers, merchants and product manufacturers, and it has become the trend to use the machine learning methods to mining the reviews. However, the short length and weak signal to describe the characteristics of reviews makes application of machine learning methods difficult to apply in product review mining. Motivated by this, we focus on mining Chinese product reviews based on Reinforcement of Features in document-level sentiment classification and feature-based product opinion mining. The main contents are as follows:Based on the characteristics that feature information of review is weak, a method based on correlation features was introduced to analysis sentiment. In this method, to enhance the features of reviews, we used the correlation rules between feature items which were mined by FP-Growth algorithm. Finally, IG feature selection and SVM classify algorithm were used to classify sentiment of reviews. Experimental results show that the proposed method performs well, and its Micro-F1 and Macro-F1 are higher than those of traditional approach.Then, due to that there is not enough features information about those pairs, traditional single classifier is not satisfied. Therefore, we derived sentential information as feature of pairs from context besides traditional features such as lexical, part of speech, semantic and positional features. Besides of using these features of pairs, we built an ensemble classifier with weighted voting mechanism to identify the subjective relationship between feature-opinion pairs. Extensive studies on corpus of book and phone reviews in Chinese demonstrate that the introduction of sentential feature could improve the recall rate of classifier, and a weighted ensemble classifier also could achieve the higher value of F-measure with the trade-off between the accuracy and recall rate of sub-classifiers.Based on above methods, a product review mining system is constructed to analysis the review web page and review text.