Emotional Support with LLM-based Empathetic Dialogue Generation

Emotional Support Conversation (ESC) aims to provide empathetic and effective emotional assistance through dialogue, addressing the growing demand for mental health support. This paper presents our solution for the NLPCC 2025 Task 8 ESC evaluation, where we leverage large-scale language models enhanced by prompt engineering and finetuning techniques. We explore both parameter-efficient Low-Rank Adaptation and full-parameter fine-tuning strategies to improve the model’s ability to generate supportive and contextually appropriate responses. Our best model ranked second in the competition, highlighting the potential of combining LLMs with effective adaptation methods for ESC tasks. Future work will focus on further enhancing emotional understanding and response personalization to build more practical and reliable emotional support systems.

💡 Research Summary

The paper addresses the critical challenge of Emotional Support Conversation (ESC), a specialized domain of dialogue systems designed to provide empathetic and actionable emotional assistance to users facing mental health difficulties. As the demand for scalable mental health support grows, the authors present a highly effective approach for the NLPCC 2025 Task 8 ESC evaluation, where their proposed model achieved an impressive second-place ranking.

The core methodology revolves around the strategic adaptation of Large Language Models (LLMs) through a combination of advanced prompt engineering and sophisticated fine-tuning techniques. The researchers focused on optimizing the model’s ability to not only recognize emotional cues but also to generate “supportive” responses that facilitate emotional regulation. To achieve this, the study explores two distinct fine-tuning paradigms: Low-Rank Adaptation (LoRA) and Full-parameter Fine-tuning.

LoRA, a parameter-efficient fine-tuning (PEFT) method, was implemented to investigate how injecting trainable low-rank matrices into the transformer layers can enable domain-specific adaptation without the massive computational burden of updating all parameters. This approach preserves the foundational knowledge of the pre-trained LLM while specializing it for the nuances of emotional support. Conversely, the full-parameter fine-tuning strategy was employed to explore the potential of updating the entire weight matrix, aiming to capture the intricate linguistic patterns and deep contextual dependencies inherent in empathetic dialogue.

Furthermore, the integration of prompt engineering played a pivotal role in steering the model’s persona. By carefully crafting prompts, the researchers were able to instruct the LLM to adopt the role of a supportive counselor, ensuring that the generated text maintains an appropriate empathetic tone and follows the structural requirements of an effective emotional support intervention.

The success of this approach, evidenced by the second-place finish in a highly competitive global task, demonstrates that the synergy between efficient parameter adaptation and strategic prompting is a powerful tool for specialized NLP tasks. The paper concludes by outlining future research directions, emphasizing the need for enhanced emotional intelligence and response personalization. The ultimate goal is to develop highly reliable, personalized, and practical emotional support systems that can serve as a scalable supplement to human-led mental health services, bridging the gap between general-purpose AI and specialized psychological support agents.