基于反思型证据增强的知识图谱可解释问答框架

打开文本图片集

中图分类号：TP391.1 文献标志码：A 文章编号： 1000-5013（2026）02-0202-11

Abstract：To addressthe issue thatcurrent large language models exhibit strong multi-hop reasoning capabilities but lack of interpretability in knowledge graph question-answering tasks，a reflective evidence-enhanced explainable knowledge graph question answering （ReE-KGQA） framework is proposed. First，candidate semantic paths are generated using large language models，and a comprehensive scoring and verification strategy integrating immediate semantic relevance with graph structural connectivity is employed to select optimal reasoning paths as explainable evidence. Then，a joint optimization fine-tuning strategy for answer generation and path rationality is designed to simultaneously enhance question answering performance and reasoning interpretability. Finally，extensive evaluations are conducted on three commonly used benchmark datasets.Experimental results show that the ReE-KGQA framework outperforms existing mainstream methods across key metrics including Hits @1 ， F1 -score，and accuracy，achieving an average improvement of approximately 9% . Moreover，the generated reasoning paths exhibit favorable semantic readability. The proposed framework effectively improves both the accuracy and reliability of the answers while enhancing the interpretability of knowledge graph question answering.

Keywords：knowledge graph question answering； explainable reasoning；reflective evidence； large language model

知识图谱问答（knowledge graph question answering，KGQA）是自然语言处理与知识工程的重要交叉研究方向，其目标是将用户提出的自然语言问题映射为知识图谱中的实体与关系，并通过多步推理返回准确答案。（剩余15771字）