Multi-agent and knowledge graph-based social and emotional learning question-answering

Abstract Social and emotional learning (SEL) education profoundly impacts children’s academic performance and behavioral development. However, current SEL practices encounter several obstacles, including inadequate professional training, limited access to information, and high implementation costs. Meanwhile, large language models show great potential in supporting emotional and mental health, but they still exhibit limitations such as insufficient accuracy and limited adaptability. To mitigate these issues in the SEL domain, we propose a multi-agent collaborative SEL question-answering framework driven by a complex knowledge graph to effectively facilitate SEL education in family settings. This framework achieves end-to-end SEL knowledge extraction and high-quality question-answering content generation through multi-agent division of labor and collaboration, significantly enhancing controllability, accuracy, and adaptability in semantic understanding and response generation. Based on this framework, we constructed a high-quality SEL question-answering dataset. Experimental results demonstrate the system’s superior performance in knowledge processing and generation. The constructed knowledge base and dataset have proven effective in supporting SEL research. To foster further research, both the code and dataset will be publicly available.