Research Of The Chinese Review Sentiment Classification By Means Of Transfer Learning Strategy

With the development of the Internet, the network has changed the way people to express their opinions. Especially with the e-commerce spring up, more and more reviews of products are posted at merchant sites and many other user-generated content sites. This online word-of-mouth behavior represents new and measurable sources of information for the other customers and manufactures. The business value of these reviews is more and more highlighted. While, the information sources on the Web are diverse, and the opinions may hide in the huge volume of blogs, and finding opinion sources and extracting the positive and negative information on the Web is still a formidable task. Sentiment classification grows out of this need. As an important branch of the text mining, the sentiment classification has got an explosive growth of both researches in academia and applications in the industry for its tremendous value for practical applications.Since the products change rapidly in the age of technology, we may need mining the opinions for the new product frequently. However, the new product reviews usually may lack of labels. Since the machine learning methods can’t get the satisfactory results and labeling by human is time consuming, how to get the customers’ opinions with the no-labeled reviews is still an unresolved problem.The transfer learning is the technique that uses the knowledge learned from other tasks to assist with the target task. So we combine the machine learning with the transfer learning for the sentiment classification with few labeled reviews or even no labeled reviews.In this dissertation, firstly, inspired by the way that human can make use of the sentiment information of a word already known and the semantic information to judge the emotion for a new word; we propose the feature-presentation transfer learning method based on the semantic information to solve the sentiment classification task without any labeled reviews in the target domain. The method transfers the word’s indication to a class in the source domain to the target domain by the semantic similarity and semantic relevancy. And then classify the unlabeled reviews in the target domain with the class space model.For the second, we improve the instance transfer learning method-TraAdaBoost algorithm for the few labeled target reviews sentiment classification task. We build a two-phase selection approach for the selection of related reviews in the source domain to assist training the classifier. Our method balances the accuracies of the positive class and the negative class which is more practical for the users.Finally, we conclude our work and put forward some directions for the further research.