Be Affective, Not Just Cognitive - Towards Imparting Pertinent Empathy in Dialogue Agents

ICLR 2026 Conference Submission22435 Authors

Keywords: Dialogue and Interactive Systems, Empathetic Response Generation, Affective Empathy, Pre-Training, Synthetic Data

Abstract: Empathetic Response Generation (ERG) has gained significant attention in diverse areas but still faces challenges that hinder its effectiveness. These challenges include $1$) the lack of affective empathy in existing works, where they exhibit cognitive empathy (feel $\textit{for}$ user); $2$) generate generic responses, where agents address an emotion with monotonous replies; $3$) have limited user relatability. To tackle these issues, we propose incorporating affective empathy in models through additional pre-training. We introduce a benchmark dataset and its collection mechanism, that helps curate an $8.5$GB dataset, enabling the agent to truly feel $\textit{with}$ user. Using this pre-trained model, our framework EMPATH enhances ERG by reducing generic responses. This is achieved by a novel loss function that involves both conversation history and golden response. EMPATH also enhances user relatability by accounting for multiple emotions and their underlying causes via explainability. Through extensive experimentation, we demonstrate the effectiveness of our dataset on our proposed framework and other existing approaches. Additionally, we depict EMPATH's superior performance in ERG on benchmark datasets across various metrics.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 22435