The Research Of Online Review Emotional Tendency Identification Based On Convolutional Neural Networks

As the Internet continues to grow in both size and importance, the quantity and impact of online reviews continually increases. Online reviews are often the primary factor in a customer's decision to purchase a product or service, and are valuable source of information that can be used to determine public opinion on these products or services. Mining emotional information carried by online reviews, analyzing emotional tendencies of customers, is very significant to online retailers and service providers.So far, researchers have made many researches in the field of sentiment orientation analysis, and promoted the progress of sentiment analysis. Because the online reviews contain customer's emotional information, and there is no doubt that natural language processing research itself is a of great challenge task; in addition, it is inevitable that review spam exists in online reviews. As a result, several difficult problems should be solved urgently in the following:(1) Review spam can mislead the opinion mining system, and reduce the system accuracy. How to solve the problem of review spam is one of the problems confronting orientation analysis.(2) Because online reviews contain many complex language phenomenon, such as new words, negative words, fixed collocations, abbreviations. Whether we can build a reasonable online review sentiment analysis model is the key to improve orientation analysis.In order to solve the problems mentioned above, this paper established the following two research content, and finally made progress. The major works of this paper are listed as follows:(1) Aiming at the problem of review spam, this paper proposed a method of review spam detection based on ensemble. In this method, we construct feature vector based on online review text and statistical information. We regarded logistic regression as secondary learner, and combined logistic regression, support vector machine, random forests and neural networks with stacking. This method focused on the seventh Chinese Opinion Analysis Evaluation (COAE 2015) in online review spam detection task, and was first among all teams attending this task.(2) In view of the characteristics of online review, this paper proposed a method of online review orientation analysis based on convolutional neural networks. In this method, we tried to avoid task-specific features, took character level embedding as input, and used convolutional neural networks to discover relevant features to the tasks. The experiment results indicated that the convolutional neural networks we built could improve the online review orientation analysis.