| With the popularization and development of text mining techniques,simple text sentiment analysis is no longer sufficient to meet the growing demands of various complex life and social scenarios.Managers and decision-makers are more concerned about the deeper reasons behind the emotions expressed in texts.Emotion-Cause Pair Extraction(ECPE)has attracted extensive attention as it identifies emotional clauses and corresponding cause clauses in texts to form emotion-cause pairs without relying on emotion annotations.Compared to emotion cause identification,ECPE addresses the issue of insufficient emotion labeling information and is better suited for different application scenarios,making it a research hotspot in the field of sentiment analysis.Introduced in 2019,the current research on ECPE mainly focuses on exploring the relationship between emotional clauses and cause clauses,overlooking the fine-grained interactions between emotion-cause pairs and the causal relationships within the context of documents.Additionally,the relative positional distribution of emotion-cause pairs in existing datasets is imbalanced,and most methods overlook the recognition of samples with distant and position-insensitive relative distances,resulting in poor generalization ability for positioninsensitive data.This study explores and investigates the task of emotion-cause pair identification from the perspective of the characteristics of interactive representations of emotion-cause pairs and the influence of document context on different emotion causes,combining it with the probability distribution of relative positions.(1)If multiple emotion-cause pairs appear in a document,the interactions between different pairs can improve the effectiveness of emotion-cause pair identification.This study proposes the ECPE-HGA model(Model of Emotion-Cause Pair Extraction based on Heterogeneous Graph Attention Network).It models the shared characteristics of different emotion-cause pairs that share the same clauses using a heterogeneous graph to aggregate information between nodes and achieve interaction between emotion causes.It also uses a conditional normalization network to model the directional causal relationship between emotional clauses and cause clauses,obtaining emotion-cause pairs oriented towards emotion and cause,respectively.Finally,the probabilities of both types are integrated,partially alleviating the issue of data imbalance.Experimental results show that the proposed ECPEHGA outperforms existing methods by at least 1.53% in terms of F1 score.(2)Recognizing causal relationships requires considering the specific context of the current document,as emotion-cause pairs composed of different clauses correspond to different contextual backgrounds.This study proposes the ECPE-MCA model(Model of Emotion-Cause Pair Extraction based on Multiple Context Aware).It uses segmented convolutional neural networks to extract different contextual background information for different emotion-cause pairs.By combining gate units,it dynamically aggregates semantic information from different document segments,incorporating the overall context of the document into each emotion-cause pair to learn richer event features and enhance the model’s understanding of causal relationships between clauses.Furthermore,addressing the bias in the positional relationships within the current dataset,this study introduces a prior probability distribution calculated from relative positional relationships.It serves as a regulatory factor to modify the loss function of the ECPE-MCA model,guiding the model to learn the causal relationships of position-sensitive samples and reducing the model’s reliance on relative positional information.Experimental results show that without using relative positional information,the proposed ECPE-MCA model improves F1 score by 0.33% compared to the state-of-the-art model.On biased and unbiased datasets,it achieves F1 score improvements of1.68% and 6.28%,respectively,compared to the MGSAG model,which mitigates position bias in emotion-cause pair recognition tasks,confirming the robustness and generalization ability of the ECPE-MCA model in the task of emotion-cause pair identification. |
