Multi-task Emotion-cause Pair Extraction Based on KTSM

The purpose of sentiment-cause pair extraction is to extract sentiment-sentence and cause-sentence pairs. The existing two-stage model suffers from error propagation problems. In addition, the previous model did not well address the problem of unbalanced sample positions of emotion and reason. To solve the above problems, this paper proposes a new end-to-end multitask model of shared interaction based on Knowledge Graph and Transformer (KTSM). Firstly, it models the interactions between different tasks through multi-level shared modules to mine the shared information between the main task emotion-cause pair extraction and the two auxiliary tasks emotion extraction and cause extraction. Second, the appropriate labels are filtered and task-specific features are constructed based on the knowledge graph path length, enabling the model to focus on extracting pairs with corresponding emotion-cause relationships. On the ECPE benchmark dataset, experimental results show that the model in this paper achieves good performance, especially on imbalanced samples.