Context Dynamics Tracking and Curriculum Learning for Emotion Recognition in Conversation

Emotion Recognition in Conversation (ERC) has emerged as a research hotspot in fields such as chatbots and question-answering systems. In the domain of ERC, the complex dynamics of conversations necessitate sophisticated models to address the subtleties of human communication. Traditional ERC models often struggle to capture the evolution of emotional states and intentions within dialogues, especially in complex and extended conversations. To tackle these challenges, this study introduces a comprehensive framework based on curriculum learning, incorporating a Context Dynamics Tracker (CDT), a Speaker Information Module, and a curriculum learning strategy. The CDT module is designed to exploit the capabilities of multi-head attention mechanisms, allowing the model to grasp both spatial and temporal contexts by simultaneously focusing on various aspects of the dialogue history and the current discourse, which effectively enhances the model's ability to perceive long-range dependencies and subtle cues within dialogues. Our Speaker Information Module dynamically simulates the influence, reactions, and intentions of speakers using a bidirectional GRU network. By integrating external knowledge, enabling the model to capture the dynamic features of speakers and their interactions, thus deepening the understanding of how speakers influence each other and respond to conversational cues. Lastly, the curriculum learning strategy addresses the challenges of class imbalance and complexity in ERC tasks. This adaptive learning approach ensures effective training of the model across various emotional scenarios, enhancing its overall performance and reliability. Experimental results demonstrate that our model achieves competitive outcomes in ERC tasks.