Leveraging Two-Stream Cause-Effect Relation for Emotion-Cause Analysis

Abstract

Emotion Cause Analysis (ECA) is a task to analyze corresponding causes for emotions expressed in a text, utilizing explicit or implicit cause-effect relations between them. Previous research typically focused on sequence labeling methods to extract emotions and their corresponding causes, or clause pairing matrix methods to extract emotion-cause pairs directly. However, the sequence labeling methods fail to build mutual interactions between emotions and the causes, and the clause pairing matrix methods may confuse whether emotions are effects of the causes or reasons of other events. To address these limitations, we propose a method that enables the model to consider “how emotions are affected by the causes” and “How the causes lead to emotions”, namely the emotion-cause stream (ECS) and the cause-emotion stream (CES). We leverage discourse conjunctions as a bridge between these two streams, incorporating their discourse information to recognize two-way cause-effect relations, thereby enhancing the reasoning ability of our model. Further, we employ the conjunctions predicted by ChatGPT to assist our model in achieving better results, and our research demonstrates that our model achieves SOTA results in ECA and proves the superiority of our model.