The Sentiment Analysis About User Reviews Based LDA And Speech-Grammar Rules

The rapid development of e-commerce makes people’s lives more convenient. It also generates a large number of useful data like user comments. It is of great value doing sentiment analysis on this user comments. Firstly, these comments information are complex and are related to technology in many fields. Few mature models can directly process this data exists. Secondly, sentiment mining on user comments can achieve a win-win for service providers and users. Service providers find the product feedback through user comments, thereby improving the quality of the product. Users can make purchasing decisions by referencing the specific comments.A novel sentiment analysis method about product comments are proposed.The main contents include:(1) A coarse-grained sentiment analysis method based on sentiment words extraction and LDA feature representation was proposed. Firstly, sentiment words were extracted by sentiment lexicon. Then, sentiment words were as features based on LDA comments for modeling and returning the topics distribution. Finally, SVM classifier was used for conducted appraise binary classification. To deal with the problem of scarce comments’features, multiple classifications are achieved by repeatedly adjusting the size of the subjects, resulting in finding the best classification results. Experiments showed that this method was effective for comment appraise classification.(2) A fine-grained sentiment analysis method based on speech-grammar rules and LDA feature clustering was proposed. The evaluation objects are discovered by independently speech rules and grammar rules. Then combining these two rules the final evaluation objects are generated. Evaluation objects are clustered based on LDA. To achieve topics’sorting, small topic amount was set for computation of subject score and large topic amount was set for finding words’clusters. In order to verify the performance of our method, two types of new document were used for testing.Results showed that this method was effective for the aspects mining and topics cluster.