A semantic structure-based emotion-guided model for emotion-cause pair extraction

Abstract

Emotion-cause pair extraction (ECPE) task aims to identify both emotion and cause clauses in a given document, which has received growing attention in recent years. Existing ECPE methods primarily focus on learning general clause representations by exploring interactions between words or clauses. Despite the remarkable achievements in these works, they neglect the distinct emotion types of clauses and semantic structures associated with each individual clause, which may affect the model's accuracy in grasping emotion and cause clues. In this paper, we propose a novel Semantic Structure-based Emotion-Guided (SEG) model, which integrates the information from both emotion types and semantic structures to capture relations between emotion and cause clauses. In specific, we design an emotion detection module to identify the emotion types of clauses and learn the emotion-enhanced clause representations. Furthermore, the module can help match clauses with similar emotion types that are more likely to form an emotion-cause pair, and promote the extraction performance of emotion-cause pairs. To model the complex semantic structures within the documents, we propose a semantic structure-based graph module to learn the relevant relationships of clauses and obtain semantic structure clause representations, which further facilitates the extraction of cause clauses. Experimental results on the public dataset demonstrate the effectiveness of our proposed SEG in the ECPE task when compared with state-of-the-art baselines.