AC-EIC: addressee-centered emotion inference in conversations

Xingle Xu, Shi Feng, Yuan Cui, Yifei Zhang, Daling Wang

Abstract: The emotional reactions of users to the dialogue context can guide the dialogue system to generate more satisfactory responses. Compared to the traditional task of Emotion Recognition in Conversation (ERC), the task of Emotion Inference in Conversations (EIC) is more challenging as it aims to infer the addressee’s emotional reactions to the context when the addressee’s utterances are unknown. Previous studies on EIC mainly focus on dialogue history information, neglecting the crucial role of the addressee as the subject of in emotion inference. In this paper, we propose an Addressee-Centered Emotion Inference in Conversations (AC-EIC) method, which can understand the dialogue history supplemented by commonsense knowledge and emotional knowledge based on the addressee’s personality. Additionally, due to the scarcity of character personality data, we manually collect the personality information of characters from three commonly used EIC datasets, expanding the original dialogue dataset. The experimental results show that AC-EIC achieves the new state-of-the-art performance on multiple datasets, demonstrating that our method can make more accurate inferences by focusing more on the addressee. Additionally, we also found that the mixed use of different types of knowledge has a positive impact on EIC tasks.

External IDs:dblp:journals/mlc/XuFCZW25