Research On Emotion-cause Pair Extraction Based On Sequence Labeling And Transition Methods

In recent years,social media,such as Weibo,has made rapid development and is becoming an important part of social life.Social media texts often contain abundant emotional information.Automatic mining of emotion in text and discovering the causes of generation and transformation of emotion is helpful to deeply understand user emotion,which has important scientific research value and practical application prospect.Many existing researches on emotion cause detection are to automatically search and identify emotional causes after providing manually annotated emotional expressions.However,manually emotion annotation does not exist in real application scenarios.Therefore,the research on emotion-cause pair extraction,which identifies the emotion expression clause and the corresponding cause clause in the text,has important research and application values.The existing emotion-cause pair extraction methods are mainly divided into two-step pipeline structure method and one-step end-to-end structure method.The pipeline structure method has the problem of error propagation between each step,while the existing end-to-end methods mainly determine the relationship between clauses by constructing clause pair matrix,which has the disadvantages of high computational complexity and large search space.Therefore,based on end-to-end structure method,this thesis investigates the emotion-cause pair extraction methods based on sequence annotation and state transition,respectively,to achieve better extraction performance,and to improve the training and inference speed of the model.This thesis firstly studies the emotion-cause pair extraction method based on sequence labeling.By introducing a new label system,this method integrates the emotion extraction,emotion cause detection and emotion-cause pair reasoning process,so that the emotion-cause pair extraction problem is transformed into a binary sequence annotation problem at the clause level.Then,the model with context information is used to predict the label of each clause,which reduces the search space of the model while extracting the emotion-cause pairs.The experimental results on EMNLP2016 emotion cause detection dataset show that the proposed emotion-cause pair extraction method based on sequence labeling obtains significant performance improvement compared with the baselines.In the 1evaluation index,the emotion reason relationship improves the extraction performance by 2.26%.At the same time,the training and reasoning speed of this method are improved by 36%and 44%respectively.The F1 measure of emotion-cause pairs extraction is improved 2.26%.At the same time,the training and inference speed of this method are improved by 36% and 44%,respectively.Considering that the method based on sequence labeling has the defects of insufficient label coverage and excessive dependence on super parameters,this thesis proposes emotion-cause pair extraction method based on state transition.In this method,the relationship prediction of emotion clause and cause clause is integrated into the state transition action.The relationship between the emotion and the cause is determined in the process of prediction action,and the search space of the model is reduced by state transition,so as to improve the performance of this method.In addition,in order to effectively use the document structure information,the historical state of transition and succeeding clauses are also used to predict the state transition action.The experimental results show that the proposed method improves the 1 value of emotion-cause pair extraction for 2.49%,and achieves the known highest performance.