Research On Emotional Analysis Of Topic - Oriented Microblogging

With the development of mobile terminal technology and the building of government micro-blog in 2014, the micro-blog’s development is faster and faster. Micro-blog has a large number of users group as a social media and a social networking platform. A hot topic always has tens of thousands of comments, and these comments always have the emotion of user. It’s helpful to public opinion analysis and marketing forecasting that analyzing the emotional tendency by research the short text. This paper does research on topic type micro-blog, and compares different data mining methods, and then we improve the algorithm of SVM.The research of the paper includes micro-blog objective and subjective judgment, the unsupervised classification that based on dictionary, the supervised classification that based on machine learning, the research of multi-class support vector machines algorithm, and the improvement of support vector machines. The emotion of micro-blog includes happiness, like, anger, sadness, fear, disgust, surprise and none.The first is the judgment of objective micro-blog or subjective micro-blog. We use part of speech feature and text feature vector to represent the micro-blog. Part of speech feature include emotional words, special symbols, degree adverbs, pronouns and 2-POS (two-Part-Of-Speech). We extract the text feature with the method of frequency, tf-idf and relative entropy. The analysis and comparison are made for the three methods.Then we identify the subjective sentences by Naive Bayes and Support Vector Machine. Better result has been achieved by SVM.Then we classify the Micro-blog texts that belong to the subjective cases. The first part is the unsupervised classification that based on dictionary. We use the emotion dictionary that arranged by Information Retrieval Laboratory of Dalian University of Technology. We expand the dictionary by using micro-blog expression, network words, and synonyms dictionary. The paper discusses the different methods in weighting emotion. The second part is the supervised classification that based on support vector machines, then we compare different method of the multi-class support vector machines algorithm, such as one-versus-one, one-versus-rest, binary tree method. We propose a new binary tree structure mode, and experiments show that it’s useful.In the end, we do research on the principle of support vector machines. We discuss the influence on multi-class support vector machines by different kernel function. We use a new grid-search method to find the optimum parameters of radial primary kernel function. The new method is improving efficiency and the optimum parameters that we searched are effective.