TLDR: This research paper details an innovative approach to Emotional Support Conversation (ESC) using large language models (LLMs). The authors, Shiquan Wang, Ruiyu Fang, Zhongjiang He, Shuangyong Song, and Yongxiang Li, from China Telecom Corp Ltd, developed a solution for the NLPCC 2025 Task 8 ESC evaluation. Their method combines structured prompt engineering with parameter-efficient Low-Rank Adaptation (LoRA) and full-parameter fine-tuning on Qwen2.5 models. This strategy significantly improved the models’ ability to generate empathetic and contextually appropriate responses, leading to a second-place ranking in the competition. The work highlights the strong potential of LLMs in mental health support and aims to further enhance emotional understanding and personalization in future systems.
In an era where mental well-being is increasingly prioritized, the demand for accessible emotional support is growing. Addressing this critical need, a recent research paper titled Emotional Support with LLM-based Empathetic Dialogue Generation introduces a significant advancement in Emotional Support Conversation (ESC) systems. This work, presented by Shiquan Wang, Ruiyu Fang, Zhongjiang He, Shuangyong Song, and Yongxiang Li from the Institute of Artificial Intelligence (TeleAI), China Telecom Corp Ltd, details their innovative approach to leveraging large language models (LLMs) for providing empathetic and effective emotional assistance through dialogue.
The core objective of ESC is to alleviate emotional distress and offer support through conversational interactions. While traditional methods have made strides, they often fall short in capturing the subtle nuances of human emotions, leading to responses that can feel rigid or lack genuine emotional resonance. This paper tackles these challenges by exploring the powerful capabilities of LLMs, enhanced through sophisticated prompt engineering and fine-tuning techniques.
The Methodology: Tailoring LLMs for Empathy
The researchers adopted a comprehensive methodology centered on LLMs, specifically the Qwen2.5-72B-Instruct and Qwen2.5-7B-Instruct models. Their approach involved two key strategies: prompt engineering and model fine-tuning.
Prompt engineering played a crucial role in guiding the LLM to understand user emotions and generate appropriate responses. A structured prompt template was designed, explicitly defining the model’s role as an empathetic assistant. This template included components such as Role Definition, Task Objective, User Profile (to enable personalized generation based on user background and emotional struggles), and Response Guidelines (specifying tone, empathy strategies, and risk avoidance). This multi-faceted design significantly improved the model’s ability to grasp emotional needs and maintain high-quality, context-aware responses.
For model fine-tuning, the team explored both parameter-efficient Low-Rank Adaptation (LoRA) and full-parameter fine-tuning. LoRA was applied to the larger Qwen2.5-72B-Instruct model, allowing for efficient adaptation by inserting trainable low-rank matrices while keeping original weights frozen. This method proved effective in enhancing emotional support capabilities with high efficiency. For the smaller Qwen2.5-7B-Instruct model, full-parameter fine-tuning was employed. While more resource-intensive, this approach offered greater expressive capacity, further improving the model’s ability to recognize and respond to nuanced emotional expressions, leading to more delicate and natural support responses.
Also Read:
- EmoSApp: Your Personal Offline AI for Mental Well-being
- Unpacking Emotion Hierarchies in Large Language Models
Performance and Future Directions
The effectiveness of their solution was rigorously tested in the NLPCC 2025 Task 8 ESC evaluation. The dataset used simulated real-world emotional support scenarios, with 1,500 dialogue instances split into training, validation, and test sets. Various automatic evaluation metrics were used, including BLEU-4, METEOR, ROUGE-L, Vector Extrema, and Distinct-2/3, alongside a human-centric G-Score evaluated by GPT-4.
The experimental results were highly promising. All fine-tuned models significantly outperformed the base model. Notably, the Qwen2.5-72B-Instruct model with LoRA (rank=32, alpha=64) achieved the highest total score on the validation set. Their best submission, which incorporated additional validation data during training, secured second place in the official competition. This outstanding performance underscores the potential of combining LLMs with effective adaptation methods for ESC tasks.
Looking ahead, the researchers plan to focus on further enhancing emotional understanding, personalization, and response safety. The goal is to build even more practical and reliable emotional support systems that can truly make a difference in real-world applications, offering a beacon of hope for individuals seeking conversational emotional assistance.
