Research On Chinese Short-text Sentiment Analysis

With the rapid development of Internet, online forums and micro blog providepeople with a broader information exchange platform. Most of the information is reviewsof online commodity, film or television works, public figures, and hot events, whichcontain huge commercial values and social values. To analyze this information has agreat practical significance. Sentiment analysis or opinion mining refers to theapplication of natural language processing, computational linguistics, and text analyticsto identify and extract subjective information in source materials. Generally speaking,sentiment analysis aims to determine the attitude of a speaker or a writer with respect tosome topic or the overall contextual polarity of a document. Now Chinese sentimentanalysis is still in its infancy, and the technology is not mature, so there is still a lot ofwork to do.In this paper, we take MicroBlog messages for study and analysis the emotionaltendencies of sentence-level Chinese short text, the main study involves the followingthree aspects. First, we build a multi-class sentiment dictionary. By analysis the polarityof sentiment words, we propose that sentiment words should be divided into the staticsentiment words, dynamic sentiment words and sentiment phrases, we aslo study thepolarity discrimination of dynamic sentiment word. Secondly, we propose the rulefiltering and machine learning subjective sentence recognition method. we use contextsliding window algorithm to recognition some of the subjective sentences, thenextraction the feature of subjective sentence and use machine learning classification torecognition the other subjective sentences. Finally, we propose multi-strategy subjectivesentence polarity classification method. According to the type of subjective sentence, weuse different methods to recognition its sentiment polarity. These methods includingdictionary-based method, machine learning classification method and dependency-basedmethod, and we use them in collaborative.Experiment is based on the proposed sentiment analysis method. with ICTCLAS,HIT dependency parsing tools and LIBSVM Kit, we build a text sentiment analysissystem. we use NLP&CC2012evaluation corpus as the test data for short textsentence-level sentiment analysis. The experimental results show that the proposedmethod has greatly improved on the precision and recall rate of short text sentimentanalysis than before.