Emotion-Cause Pair Extraction Based On Attention Interpretation

A large amount of user sentiment information exists in social media.Enterprise decision makers not only want to understand user emotion,but also want to understand the corresponding reasons,which can help them think and optimize products from the user level.In this demand,emotion cause detection has become one of the tasks with important value and application prospects in natural language processing,and gradually developed into emotion-cause pair extraction task(ECPE).The traditional two-stage ECPE model has the problem that the model error transmission cannot be corrected.Although the later end-to-end model solves this problem,there are still some problems such as insufficient utilization of Document event features and mixed factors.Therefore,this thesis studies the emotion-cause pair from the perspective of event feature mining and confounder eliminating.The main research contents are as follows:(1)A classification feature encoding method of emotion clause is proposed.Due to the lack of effective sentiment label information in the current emotion clause extraction process,the model fails to make full use of emotion features to extract cause.Therefore,this thesis adds an emotion analysis task to ECPE,combines the characteristics of emotion clause with the prediction matrix,and uses the attention mechanism to obtain the emotion category characteristics containing document event information.(2)A causal event context extractor based on the convolution neural network(CNN)is proposed.The current ECPE models do not make full use of the event information of background clause in pairing,just as reading comprehension does not provide text paragraphs,the features obtained by the model are more one-sided.Therefore,based on CNN,this thesis integrates the event features of background clause into the pairing process.Firstly,the emotion category features are taken as the convolution kernel of CNN,which is called emotion convolution kernel(EK).Then,all clauses of the document are input into CNN.In this way,EK will scan all clauses of a document in turn during convolution operation,so that the model can learn richer event characteristics and causality.(3)The mapping of long-distance emotional causality is studied by temporal convolution network(TCN).In order to improve the pairing quality of long-distance emotion-cause pairs,based on point(2),this thesis uses the characteristics of hole convolution in TCN to expand the text receptive field.It helps to connect and mine the characteristics of cross clause events,so as to construct long-distance causal mapping.(4)Based on the front door criterion theory,an improved attention framework is proposed to eliminate causal confounder.The confounder in the emotion cause document makes a false causal relationship between the emotion cause and the background clause.Therefore,based on the characteristics that the front door criterion does not need to obtain confusion factors,this thesis improves the attention mechanism,and proposes self document attention and cross document attention,so that the original unsupervised attention distribution process can be effectively guided by causality,so as removing the influence of confusion factors.The results show that the front door criterion can effectively remove the cause of confusion in the document,and therefore improves the F1 score of ECPE.