SweetieChat: A Strategy-Enhanced Role-playing Framework for Diverse Scenarios Handling Emotional Support Agent

TL;DR

This work introduces a strategy-enhanced role-playing framework to improve emotional support conversations by simulating diverse real-world interactions with three roles: Seeker, Strategy Counselor, and Supporter. The framework yields ServeForEmo, a large-scale dataset (3.7K+ dialogues, 62.8K+ utterances) used to fine-tune SweetieChat on open-domain scenarios, achieving more nuanced and tailored emotional support. Across automatic and human evaluations, SweetieChat demonstrates superior empathy, coherence, and strategic guidance compared to prior ESC models, and robustness improves when trained on combined data sources. The approach underscores the importance of incorporating explicit support strategies and role-based data generation to narrow the empathy gap between humans and AI in emotional support contexts, with practical impact for empathetic agents in real-world settings.

Abstract

Large Language Models (LLMs) have demonstrated promising potential in providing empathetic support during interactions. However, their responses often become verbose or overly formulaic, failing to adequately address the diverse emotional support needs of real-world scenarios. To tackle this challenge, we propose an innovative strategy-enhanced role-playing framework, designed to simulate authentic emotional support conversations. Specifically, our approach unfolds in two steps: (1) Strategy-Enhanced Role-Playing Interactions, which involve three pivotal roles -- Seeker, Strategy Counselor, and Supporter -- engaging in diverse scenarios to emulate real-world interactions and promote a broader range of dialogues; and (2) Emotional Support Agent Training, achieved through fine-tuning LLMs using our specially constructed dataset. Within this framework, we develop the \textbf{ServeForEmo} dataset, comprising an extensive collection of 3.7K+ multi-turn dialogues and 62.8K+ utterances. We further present \textbf{SweetieChat}, an emotional support agent capable of handling diverse open-domain scenarios. Extensive experiments and human evaluations confirm the framework's effectiveness in enhancing emotional support, highlighting its unique ability to provide more nuanced and tailored assistance.

Figures (16)

• Figure 1: Example responses from the Meta-LLaMA3.1-70B-Instruct and our SweetieChat. LLMs often give verbose and formulaic responses characterized by empathy + suggestions, resulting in a distinct 'AI flavor'. Conversely, SweetieChat excels in empathy and supportiveness, skillfully addressing and responding to user emotion needs.
• Figure 2: Overview of the proposed strategy-enhanced role-playing framework. Our framework incorporates 3 roles: Seeker, Strategy Counselor, and Supporter. We employ LLMs to simulate these roles and interact with each other like real-world emotional support conversations. Following the dialogue generation process, we develop the ServeForEmo dataset. Building on these foundations, we present SweetieChat, an emotional support agent capable of handling diverse open-domain scenarios.
• Figure 3: (a) Global inter-dialogue similarity statistics computed using TF-IDF vectors. (b) Inter-dialogue similarity statistics within the academic field. Lower similarity values indicate higher diversity.
• Figure 4: (a) Global similarity statistics of responses guided by the same support strategy, computed using TF-IDF features. (b) Similarity statistics of responses within the academic field under the same support strategy.
• Figure 5: (a) Distribution of strategies at different conversation progress on ServeForEmo dataset. (b) Distribution of unique strategies across different dialogues on ServeForEmo.