Recurrent synchronization network for emotion-cause pair extraction

Abstract

Emotion-cause pair extraction is a fundamental task in emotion analysis, which aims to extract emotions and their corresponding causes from documents. Recent studies have employed two auxiliary tasks, namely emotion extraction and cause extraction, to facilitate the detection of emotion-cause pairs and their joint resolution in a multi-task learning framework. However, the implicit information-sharing mechanism in existing methods fails to fully utilize the rich interactive relations among these three tasks. In this study, we propose a recurrent synchronization network that explicitly models the interaction among different tasks. Our model performs multiple rounds of inference to detect emotions, causes, and emotion-cause pairs iteratively. Following each round of inference, the information from different modules is synchronized through explicit information transmission, allowing the three tasks to collaborate effectively. Extensive experiments demonstrate that our model can extract emotion-cause pairs more accurately, while significantly outperforming existing methods.