Event-driven Emotion Cause Detection From Text

With the rapid development of Internet technology,the mass online text data not only shows a wisdom crystallization,but also contains potential risks.Thus,the natural language processing based public sentiment monitoring,opinion extraction and emotion analys is are becoming more and more important.At present,the research in this area is shifting from the increasingly practical text emotion analysis to deeper analysis on emotion cause,that is from "to know what kind of emotion is expressed" to "what causes this emotion ",namely emotion cause extraction.The research on emotion cause extraction from text relies on both the algorithm imple mented and the available emotion cause annotated corpus.The lack of open emotion cause annotated corpus influences the research in this area.Thus,in this study,a medium-size emotion cause annotated corpus is firstly designed and constructed.O n this basis,the event-driven emotion cause extraction methods are investigated.The works in this study consist of the following three major components:Target to the lack of annotated corpus,in this study,an emotion cause annotated corpus on news text is designed and constructed.Based on the observation and analysis on the textual emotion cause expression,a set of complete and comprehensive annotation scheme is designed.Following this scheme,2,105 instances of emotional cause are manually selected from 15,687 news documents for constructing the emotion cause corpus through manually annotation.Using this corpus,the event-driven emotion cause detection methods are investigated.By analyzing the expression characteristics of e motion cause text,an abstraction method is investigated,which describes the external stimuli causing the production and change of emotion as event tuples.Next,the candidate emotion cause events identification algorithm which is based on dependency pars ing results,and the emotion cause extraction algorithm which is based on Support Vector Machine classifier with polynomial kernel,are investigated,respectively.The experimental results on our constructed emotion cause corpus show that this algorithm improves the F-measure of 3.34% compared to the baseline method.Considering the insufficient description capacity of event tuples,the event tuples are further transformed to event tree structures in order to improve the emotion cause projection from text to structures.By incorporating tree kernel and polynomial kernel,an improved emotion cause detection algorithm is designed and imple mented.The experimental results show that the F-measure achieved by this algorithm outperforms the baseline system for 10.61%.The proposed event-driven emotion cause detection method achieves the highest performance in the experiments,based on our knowledge,which benefits from the better abstraction and projection of emotion cause text.Meanwhile,the constructed C hinese emotion cause annotated corpus may be used as an open resource to promote the research in this area.