| 摘 要: 现有的人工智能情感支持对话研究多集中于单次会谈,无法满足实际应用中情感支持的持续性需求。针对这一研究空白,提出 MSESChat———一种支持多次会谈的情感支持对话合成数据框架。MSESChat设计了求助者、助人者和规划者三个智能体,以真实案例数据为基础,通过 AI智能体角色扮演生成长周期、多轮次的情感支持对话数据集。利用该数据集对开源的大语言模型进行监督微调。实验结果显示,微调后的模型在情感支持效果和内容连贯性方面较基线模型有显著提升。 |
| 关键词: 情感支持 大语言模型 合成数据 智能体 |
|
中图分类号: TP18;TP391
文献标识码: A
|
| 基金项目: 教育部人文社会科学研究一般项目(23YJCZH28123YJCZH281);上海市哲学社会科学规划课题(2022ZGL010) |
|
| Artificial Intelligence Emotional Support Dialogue for Long-Term Multiple Sessions |
|
SHI Jiajie, YIN Pei, HE Zihao
|
(Business School, University of Shanghai for Science & Technology, Shanghai 200093, China)
gejsshi@163.com; pyin@usst.edu.cn; 2556689265@qq.com
|
| Abstract: Current research on AI emotional support dialogues predominantly focuses on single-session interactions,failing to address the persistent demand for sustained emotional support in practical applications. To bridge this
research gap, this paper proposes MSESChat—a synthetic dialogue framework designed for multi-session emotional support. MSESChat incorporates three AI agents: the help-seeker, the helper, and the planner. Leveraging real-world case data, these agents engage in role-playing to generate a long-term, mult-i session emotional support dialogue dataset. The paper further fine-tunes open-source large language models using this dataset. Experimental results demonstrate that the fine-tuned models significantly outperform baseline models in both emotional support effectiveness and content coherence. |
| Keywords: emotional support large language models synthetic data agents |
