Learning chain for clause awareness: Triplex-contrastive learning for emotion recognition in conversations

Abstract

Emotion recognition in conversation (ERC) is essential for enhancing human-computer interaction and improving intelligent systems' emotional intelligence. Despite advancements, ERC still struggles to capture the complexity and nuances of emotions in dialogues. Traditional approaches rely on context modeling to improve text representations but often fail to fully capture emotions' continuous and intricate nature. While contrastive learning has emerged as a promising technique to enhance the discriminative power of text features, many methods still rely on discrete emotion labels, limiting their ability to model emotions' inherent continuity and relational structure. To address these challenges, we introduce the Chain of Triplex Contrastive Learning (CoTCL) framework, which progressively refines utterance representations, strengthens emotional distinctions, and incorporates contextual information from dialogues. CoTCL employs instance contrastive learning using dropout-based positive samples to improve the richness and separation of utterance features. Additionally, it introduces the Pleasure-Arousal-Dominance (PAD) space as a continuous representation framework, embedding utterances in a way that reduces confusion between similar emotions. This allows for more nuanced, relation-aware emotion modeling. Furthermore, CoTCL enhances contextual understanding by constructing an utterance graph, where nodes represent utterances and edges denote relationships. By integrating external knowledge from the COMET model and applying graph-based contrastive learning with edge perturbation and node masking, CoTCL improves contextual transmission, making the model more robust and adaptable to real-world dialogues. Experimental results demonstrate that CoTCL achieves state-of-the-art performance across multiple ERC benchmarks, highlighting the importance of continuous emotion modeling and context integration. By refining text representation, introducing a continuous emotion space, and leveraging external knowledge, CoTCL provides a strong foundation for advancing ERC research and improving real-world dialogue systems.