Emotion-cause pair extraction via knowledge-driven multi-classification and graph-based position embedding

Emotion-Cause Pair Extraction (ECPE) is an important yet challenging task, focused on concurrently extracting emotional clauses and their corresponding causal clauses. Despite notable improvements in extraction performance over the past few years, three critical issues have been overlooked: (1) The binary classification of emotional and causal clauses may neglect the emotional semantics. In reality, there exist four types of clauses: emotion, cause, emotion-cause, and none-emotion-cause. (2) The integration of prior knowledge concerning sentiment information and causal information has been lacking. (3) The positional information of emotion-cause pairs may be adversely affected by imprecise emotional hypotheses and unbalanced document lengths. To tackle these challenges, we propose a new knowledge-driven multi-classification sub-task aimed at classifying clauses into the four mentioned types. Additionally, we introduce graph-based position embedding to capture relevant positional information. Experimental results underscore the effectiveness of our approach in addressing these issues.