The Study On Text-based Emotion Cause Detection

Emotion, as human innate characteristics, plays a more and more important role in natural language processing,so emotion computation has positive significance in the study of cognitive science. And, with the development of the Internet, the huge number of information with people’s emotion and opinion is integrated into the environment of the Internet, so the computation and the research of emotion computation have urgent practical significance.Currently, the research of emotions focuses mainly on emotion analysis and emotion prediction, the number of deep-level emotion research like emotion cause detection is so few. Emotion cause detection is automatically detecting causes of emotion production and emotion transfer in text. In the process of constructing corpus, a lot of manual work is needed in corpus construction, so the lack of standard and public corpus leads to the underdevelopment of emotion cause detection’s rules and models; moreover, the emotion cause detection study is at the beginning stage, so models and features of the research is still in its infancy. The main parts of the paper are list below, and the first part is constructing corpus of emotion cause detection which is the foundation of the research, and mining statistical laws through analysis of the corpus; the second is that analyzing the corpus with statistical theory and designing 8 rules which improve the performance of the system with the rule-priority algorithm by 26.73%; at the same time, for handling rule conflicts, the research uses transformation-based error-driven algorithm to improve precision by 3.16%; the last is using statistical model combined with features, such as, linguistic features, distance feature, grammatical feature and so on, to improve the performance of the system by 7.92%, compared to rule-based system. The performance of the system with emotion cognitive knowledge and emotion semantic knowledge, mined from topic model is improved by 3.05%, compared to the system based on liguisitic features, distance and grammatical features.The main contributions of the study are list as follows. The first is the construction of emotion cause detection corpus which is the largest emotion cause detection corpus in weibo’s field and provides data foundation and statistical basis for the late model building; the second are designing rules for weibo text which can effectively improve the accuracy of system, and using the rule-priority algorithm and transformation-based error-driven algorithm to extract emotion cause; the last are taking emotion cause as classification and sequence tagging study, and combining emotion cognitive and semantic features, this part provides a reference for the future intensive emotion study based on cognitive knowledge.