Research On Text Emotion Classification By Incorporating Emotion Cause Detection

Social media is becoming the main platform for users to exchange views and express emotions.While users express their emotion on the platform,their interactions constantly affect the real world.Therefore,the research on emotion analysis and intention recognition based on natural language processing technology is becoming more and more important.Most existing text emotion computing researches focus on identifying and classifying the emotion by using the textual features related to emotion expression,but neglect to explore the causes of emotion and its use.So,it lacks the complete understanding and application of emotion related information for emotion classification.The existing emotion cause detection research may be camped into rule-based,machine learning based and deep learning based approaches.These three methods have some problems,such as low efficiency of rule manually construction,strong subjectivity of statistical features design and filtering,and unsatisfactory interpretability and controllability of the model,respectively.Therefore,this paper first studies the effective emotion cause detection method,and then studies the method of emotion cause detection and emotion classification method based on multi-task learning.The purpose is to make full use of the interaction information between emotional cause and emotional expression,and effectively improve the performance of emotional classification.This paper firstly studies the emotion cause detection method by incorporating knowledge regularization and hierarchical attention network.By using hierarchical structure and attention mechanism to model the relationship between discourse,clauses and emotion expression,the emotion cause can be better detected.At the same time,regularization constraints based on sentiment words and word position are designed to constrain the parameter of above model with prior knowledge.The experimental results on the EMNLP2016 Chinese emotion cause detection dataset show that the proposed method improves the F1 value by 2.08% compared with the existing state-of-the-art method.It achieves the highest known performance.Based on this,this study investigates a multi-task learning method for joint learning of emotion cause detection and emotion classification.By setting the task private layer and task shared layer,the specific features of the two tasks and the common features between them are simultaneously extracted.The same phrase training strategy is adopted to realize the information interaction between the two tasks.The experimental results on the NTCIR13-Chi sentiment classification dataset and the emotion cause detection dataset show that F1 values of emotion classification and emotion cause detection are increased by 1.05% and 0.27% respectively by using multi task learning method,which shows that the method incorporating emotion cause detection can effectively improve the performance of emotion classification.