Research On Emotion Cause Extraction Based On Deep Learning

Emotion-cause extraction(ECE)is an important subtask in the field of sentiment analysis,and it has received attention from academia and industry because of its wide application prospects.The research goal of emotion-based cause extraction is to find the cause for the emotion expressed in a given document.Because of the complex structure of the text,making the task of emotion cause extraction very challenging.At present,most of the existing methods only focus on modeling the relationship between emotional clauses and cause clauses,while ignoring some other important text features.On the other hand,some studies have shown that emotion-based cause extraction needs to manually annotate the emotion expressed in the document,which greatly limits the application scenarios.Therefore,after completing the research on emotion-based cause extraction,we further try to conduct joint extraction of emotion and cause,that is,to extract the emotion and cause from the end-to-end without emotion labeling in document,which is also called the emotion-cause pair extraction(ECPE).In response to the above problems,relevant research has been carried out.This paper proposes a new method of emotional cause extraction based on deep learning technology,combining attention mechanism and pre-training language model.The main research contents are as follows:(1)A network model based on multi-layer attention mechanism is proposed for emotion cause extraction.Firstly,the pre-training language model is used to encode the text to obtain the word level features.Then,combined with the attention mechanism,the semantic features of clauses in the document are extracted by using the bidirectional long short-term memory network.After analysis,there is not only semantic relationship between emotion clause and reason clause,but also obvious relative position relationship in the document.Therefore,this paper combines the characteristics of relative position relationship,and uses self-attention mechanism to model the semantic relationship between emotion clause and cause clause.Finally,in order to alleviate the imbalance between positive and negative samples in the dataset,the focus loss function is introduced.Experimental results show that the model can effectively improve the performance of emotion cause clause extraction.(2)This paper proposes a multi task learning framework based on transformer for joint extraction of emotion and cause,which can achieve joint extraction of emotion reason pairs without emotion annotation.At present,most of the researches on emotion cause extraction adopt the method of extracting emotion clause and cause clause respectively,and then associating them.In this method,the correctness of emotion cause pair extraction depends too much on the result of clause extraction,and the result of emotion cause pair extraction cannot be fed back to the task of clause extraction,resulting in task propagation error.Therefore,a neural network model based on transformer is proposed to achieve the end-to-end extraction of emotion and cause.On the other hand,in order to further explore the relationship between different subtasks,this paper constructs different multi task models.The experimental results show that there is a significant correlation between different subtasks,and the emotion classification task can significantly improve the performance of emotion cause pair extraction task.