Prototype equilibrium network with group emotional contagion for few-shot emotion recognition in conversation

Min Jiang, Mengdi Wang, Jun Kong

Abstract: Few-shot emotion recognition in conversation (FSERC) aims to classify the emotion of utterances in conversations with only a few labeled conversations. However, there is limited research on FSERC at present, and the few existing approaches for this task suffer from an imbalance of the utterance number of various classes. In addition, they ignore the impact of group emotional contagion and the representation of intra- and inter-conversation class prototypes. In this paper, we address these issues by proposing the prototype equilibrium network with group emotional contagion (ProEq-GEC). Firstly, we construct adaptive cross-sampling to form support set and query set samples in the way of adaptive classification and grouping sampling. Secondly, we introduce group emotional contagion and build a weighted directed acyclic graph neural network to capture the conversation context information more comprehensively by distinguishing the emotional impact of others and the speaker himself on the current utterance. Thirdly, we propose the prototype equilibrium network to enhance the representation of class prototypes by calculating intra- and inter-conversation class prototypes. Finally, experimental results show that our model is highly effective for the FSERC and significantly outperforms the existing models.