Sentiment Classification For Online Reviews Based On Complex Network

With WEB technology and E-Business developing, online shopping is becoming trend for consumers. Now, Consumers share their shopping experience by writing online reviews, and they also express their preference and experience to product or serve. There is some sentiment message for product, seller, manufacturers and logistics, which is important link in online shopping. All these is important and significant for merchant to improving products and service quality improvement, so that business can improve satisfaction of consumers, at the same time, these reviews also contribute consumers to make decision. So that research for online reviews increasing quickly. In these research, sentiment analysis for online reviews is becoming hot issue, and sentiment classification is the main task in sentiment analysis so that it is attracted.The research of online reviews'sentiment classification is principal line in this paper, and a new algorithm is proposed using complex network. There are three parts we study as follows:Firstly, because of the public domain dictionary emotion do not contains all words that in professional field, cyberspeak, new words. So we propose that using the public domain dictionary emotion as seed emotion dictionary and using adjective words in reviews corpus as candidate sentiment words, and calculate the relevancy of candidate sentiment words and seed emotional words based on point mutual information theory. So that we can extend public sentiment word dictionary as professional semantic lexicon.Secondly, because of traditional sentiment classification algorithm ignore word order and lack of syntactic analysis for online reviews. So we proposed a new sentiment classification algorithm based on directed network. The directed network is created for online reviews, and the model of sentiment classification is proposed by network topological properties. First, the directed network is built for online review based on co-occurrence theory, and then mining sub-network which has sentiment information, the Weber-Fechner Law is introduced for sub-network because of excursion caused by adverb and negative. A new approach called DNSA(Directed network and syntactic analysis) is created for calculating similarity of reviews based on above all. The similarity of review in test set and train set is computed, and the most of class as the class in the top K.Thirdly, because of traditional sentiment classification algorithm based on vector space model ignore features of semantic information and text structural information, result in comment on emotional resources missing problem. we propose a new algorithm for feature selection of sentiment classification for online review based on complex network. It makes better for semantic relativity between feature words, so that more emotive information is got. The relational complex network is created for candidate feature by complex network theory. Then the part and overall important of nodes is considered. We use degree centrality, closeness centrality and betweeness centrality to measure important of nodes for select sentiment classification feature, the algorithm is named NTFS(Complex network feature selection). At last, SVM, NNET, NB, which are classical classifier, are used for online reviews sentiment classification.At last, hotel reviews and phone reviews as the experimental data for confirming the effectiveness of the two algorithm. The experiment results show that, these two kinds of sentiment classification algorithm can effectively realize online reviews of emotion classification, and these two sentiment classification algorithm can improve the performance of sentiment classification significantly.