Research On Emotion Recognition Of College Students Based On Forum Posts

The mental health of college students is a key concern in today’s society. Since the particularity of the age, living space and cultural level, the college students are more sensitive, prone to loneliness, depression or low self-esteem and other negative emotions. It will not only influence the individual’s learning and life, but also easily transmit to people around. The long-term negative emotional state can cause many serious consequences, for example, depression and anxiety are two of the most common negative emotions in college students, which will make people feel depressed, frustrated and helpless, or even lead to unpredictable results. Therefore, negative emotion individual identification and diagnosis is particularly important for physical and mental health of college students. For a long time, psychological scales are mainly used due to the complexity of emotions. This approach needs to disturb the individuals which are tested and the results will be affected by sensitivity factors of the experimental group. Sentiment analysis provides a new way to measure the emotion, which can be done without intervention to the people. It also provides a dynamic view to analyze and track emotional changes, therefore, sentiment analysis has better prospects for the research.This paper mainly focuses on the key techniques of negative emotion recognition of college forum texts. The purpose is to apply these techniques for identification of negative students, which will provide a reference for psychological intervention and counseling. As results of language diversity and briefness of emotional posts, the related research has to confront two difficulties:On the one hand, it is very hard to select the emotional features, because of the short length, wide topics and many nonstandard expressions of forum posts. On the other hand, due to the large number of emotional text and the number of different category of text is not balanced, seriously affect the recognition performance of sentiment classification. So it is urgent to develop an effective method to solve the problems above to improve recognition performance. Therefore, research works and innovations of this dissertation mainly include the following three aspects:Firstly, to deal with the problems of high-dimensionality and feature selection difficulties, a method which compounds the multi-type features to classify the positive and negative emotions of forum posts is presented. This method takes into account the complementarities of different type of features, and then combines four types of features, including words, multi-grams, co-occurrence words and word clusters to achieve better results. In this method, words clusters which were built in an innovative way by word vector clustering and were proved to be effective.Secondly, since the small number of certain emotional posts, to deal with the minority sentiment classification problems, we propose an Strategic Dynamic Subspace and Distance Based Under Sampling Method (SDS&DBUSM). In this method, random subspace based ensemble learning algorithm is using to establish the classification model and two improvements are proposed:the first one is that subspaces are generated by strategy selection instead of randomly sampling. In this strategy, the discriminative capabilities of features are calculated and effective features are more probable picked, which can guarantee the accuracy of the base classifier. The other improvement is that a distance based under sampling method is proposed to remove noise and redundancy samples of each subspaces, effectively improve the recognition performance. By contrast the single classifier method and traditional RSM method, SDS&DBUSM has proved to be more effective.Thirdly, a method for negative students diagnosis based on forum posts emotion analysis is presented. According to the behaviors of students in the forum, emotional related factors are mined firstly. Then they are weighted and used to calculate the students’ emotional value. At last, students are divided into different categories according to emotional values. Compared with the methods based on psychological scale, this method can improve the efficiency of individual emotional measurement, track the emotional changes of the subjects, and provide reference and support for the follow-up psychological counseling.