Research On Emotion-Cause Pair Extraction Algorithm Based On Multi-Level Fusion Attention Neural Network

With the rapid development of artificial intelligence technology,the field of text emotion analysis has developed rapidly.A large amount of text information hides various emotions and attitudes that users want to express.Obtaining the potential information in the text plays a crucial role in business applications and government opinion management.However,most of the current research is based on coarse granularity,and the research on fine granularity is relatively immature.In response to this problem,some scholars have proposed a fine-grained emotion analysis task,namely,the emotion reason pair extraction task.This task aims to extract all emotions in the document and their corresponding reasons.There are two main challenges to improve its prediction accuracy.First,most existing methods ignore the relationship between emotion clause and cause clause when learning the expression of emotion and cause clause.How to accurately capture the relationship between them is a difficult point of research;Secondly,recent research has not been able to deal well with the potential relationship between emotion,reason extraction task and emotion-cause pair extraction task.At the same time,the existing algorithms ignore the syntax information contained in the document itself.How to accurately obtain the text grammatical information and the potential links between tasks is also another difficulty to be solved in this paper.In view of these two difficulties,this paper proposes two emotion-cause pair extraction models based on multi-level fusion attention.The first model is the emotional cause pair extraction model based on interactive attention(IA-ECPE).IA-ECPE is a stacking framework based on interactive attention.The lower network can provide information for the upper network to optimize the results.In the embedded part of the model,because the object is a document,hierarchical coding is used,from words to sentences,and then from sentences to documents.This model uses the Bi-LSTM and Bi-GRU two-level combination model to encode the sentences in the document to obtain the expression of emotion clauses and reason clauses.Then,input the obtained clause expression into the interactive attention module to learn the relationship between the emotion clause and the reason clause,and use the fusion mechanism to enhance the internal information of the emotion clause and the reason clause,so as to capture the complex associations at different levels.Finally,the obtained vector and position information are fused and predicted to obtain the final prediction result.The second model is based on SAP-ECPE to identify different types of emotional cause pairs,and reduce prediction errors while improving the efficiency of the model.This method mainly consists of three parts.The first is the span representation.Bi-LSTM and Bi-GRU are used to encode the sentences in the document at two levels,from words to sentences,and then from sentences to documents.The obtained sentence vectors are listed one by one,and each clause is used as an axis to focus on its various contextual information to strengthen the causal relationship between sentences.The second part is the span association matching part.Through the obtained sentence vector,take each clause as the fulcrum,pair the sentences in the document one by one in different spans,and integrate the span information into the vector representation.Finally,the prediction of emotional reasons.Integrate the predicted results of emotion and cause,relative position and other information,and then use multi-dimensional information interaction mechanism to select and predict emotion-cause pairs.