Research On Sentiment Analysis For Web Short Texts Based On Topic Model

With the rapid development of Web 2.0, platform such as social network sites, microblog and BBS provide users with more convenient channels for publishing their own views and opinions. These social media data such as online reviews, microblog and so on usually contain a small number of text characters. Compared with traditional long texts, such as news, paper, etc., we call these social media data network short texts. These network short texts are usually full of users' sentiment implied with great value. Sentiment analysis of network short texts has attracted much attention. How to quickly and efficiently extract and analysis the sentiment polarity from network short texts has become a hot research topic. Topic sentiment models represented by Latent Dirichlet Allocation model have attracted much attention because of good performance and synchronization of analyzing topics and the sentiment in a document.Through investigation of literatures about LDA, we find that existing LDA topic sentiment models do not take user' comment behavior into consideration. It is impossible to analyze topic sentiment evolution of texts due to ignorance of time information hidden in social media. The fact that few models consider user relationship of social networking platform such as microblog leads to reducing the accuracy of text sentiment analysis. To address these issues above, three strategies are proposed and applied to the text sentiment analysis:1?For the insufficient of existing LDA topic sentiment models do not consider users' short comment behavior, we construct the new model TSCM, which is based on LDA topic sentiment models improved by adding sentence layer. Experimental results based on real corpus show that DTSCM topic sentiment model has higher text sentiment analysis accuracy.2?For the insufficient of existing LDA topic sentiment models can not analyze the topic sentiment evolution of texts and existing LDA topic evolutionary models can not analyze sentiment polarity of texts, we integrate time parameter into LDA topic sentiment model, and propose the model DTSCM. Experimental results based on real corpus show that the DTSCM can track the topic sentiment evolution of text.3?For the insufficient of existing LDA topic sentiment models do not consider the fact that the user relationship of microblog or other social networking platforms has effect on the accuracy of the text sentiment analysis well, the improved model SRTSM based on the user relationship of microblog is proposed. In SRTSM, user relationship distribution is integrated into LDA topic sentiment model. Experimental results based on real microblog corpus show that the new model can effectively improve the accuracy of microblog sentiment analysis.