Semantic-Based Text Sentiment Analysis

With the arrival of web2.0and mobile Internet, the network is not only an important way to access to information, but also a platform to express users’ opinion. A large number of users express their ideas in weibo, BBS and e-commerce platform. There is a lot of sentiment information which can be found. Research on mining this information is a hot problem in natural language processing. And sentiment analysis is an important part. It is needed in the analysis of public opinion, product comments, intellective shopping guide and other fields.Two methods have been proposed by domestic and foreign scholars on sentiment analysis. One of these two methods is based on semantic methods, and another is based on machine learning. The method used in this paper is based on semantic analysis. After giving the general process of sentiment analysis based on semantic, the paper proposed some improvements in the aspect of appraisal expression and sentiment calculation.The main work in this paper are as follows:First, the algorithm to recognizing appraisal expression based on Tri-training. The appraisal expression consists of the modifier phrase and the phrase to be modified. Different phrase to be modified with different modifier phrase will express different sentiment tendencies. There are three classifications of different models, i.e. SVM, CRF and MaxEnt model classifier. Result show that the method is better than using single classifier.Second, different dictionaries are the basis of the method based on semantic. The paper used both privative dictionary, adverbs of degree dictionary, field evaluation and conjunctions dictionary to make the dictionaries be more comprehensive in sentiment analysis.Third, analyze the sentiment in the phrase level and sentence level, then calculate the sentiment of whole text. It includes strength calculation and sentiment analysis of phrases level. We put forward the computational formula of the sentiment type of the phrase level using the dictionaries building in this paper. Combining the evaluation phrase and evaluation objects and identifying the conjunction in the sentence to calculate the sentiment result of the sentence. Then gain the sentiment result of the text by the results of sentences and the paragraphs and their weight. The improved method achieves a good result in the experiment.