To address the issues of static knowledge and factual hallucinations in large language models (LLMs), as well as the incomplete coverage and delayed updates in knowledge graphs, a collaborative reasoning framework integrating LLMs and knowledge graphs is proposed via dynamic knowledge retrieval. A closed-loop mechanism is designed to implement dynamic knowledge retrieval, collaborative reasoning, and incremental updating. Firstly, a subgraph-enhanced retrieval strategy is adopted, which extracts relevant subgraphs from the knowledge graph through question decomposition and multi-hop reasoning to improve the recall under complex queries. Secondly, semantically relevant passages are retrieved from unstructured texts through a multi-dimensional hybrid weighting method to supplement missing knowledge. Finally, LLMs are utilized to integrate multi-source knowledge for reasoning, generating answers while extracting new triples for dynamically updating the knowledge graph. Experiments demonstrate that the proposed method can effectively handle complex multi-hop questions, reduce factual hallucinations, and ensure timeliness and continuous evolution of the knowledge base through closed-loop updating.
近年来,以GPT(Generative Pre-Training)系列为代表的大语言模型(Large Language Models, LLMs)在自然语言理解、生成与推理任务中展现出卓越性能,推动了人工智能技术在各领域的广泛应用。然而,大语言模型依赖其预训练语料中的静态知识,难以捕捉现实世界的动态变化,且在生成过程中易出现“事实幻觉”问题,这严重限制了其在金融、医疗等高可靠性场景中的应用[1]。
通过验证的高置信度新知识将被结构化为标准三元组添加回知识图谱中,并支持附加时间属性。例如,确认“USS Carl Vinson的母港为圣迭戈海军基地”后,系统生成三元组(USS Carl Vinson, has_homeport, San Diego Naval Base),并可标注有效时间(如valid_from: 2025-Q1)。同时,更新记录(推理路径、证据来源、时间戳、置信度等)将作为元数据保存,形成可审计的知识演化轨迹。
本研究在WebQSP(WebQuestionsSP)和CWQ(Complex Web Questions)两个经典知识问答数据集上进行评估。两者均基于Freebase知识图谱构建,并提供自然语言问题及其对应的标准SPARQL查询。数据集统计信息如表1所示,外部知识源采用Wikidata作为补充文档数据库。
ZhangJ, ZhangX K, YuJ F, et al. Subgraph retrieval enhanced model for multi-hop knowledge base question answering[C]//Proceedings of the 60th annual meeting of the association for computational linguistics. Stroudsburg, USA: ACL, 2022: 5773-5784.
[2]
JiangJ H, ZhouK, DongZ C, et al. StructGPT: a general framework for large language model to reason over structured data[C]//Proceedings of the 2023 conference on empirical methods in natural language processing. Stroudsburg, USA: ACL, 2023:9237-9251.
[3]
SunJ, XuC, TangL, et al. Think-on-graph: deep and responsible reasoning of large language model on knowledge graph[C]//Proceedings of the 12th international conference on learning representations. Appleton, USA: ICLR, 2024: 3868-3898.
[4]
ShuY H, YuZ W, LiY H, et al. TIARA: multi-grained retrieval for robust question answering over large knowledge base[C]//Proceedings of the 2022 conference on empirical methods in natural language processing. Stroudsburg, USA: ACL, 2022: 8108-8121.
[5]
LuoH R, E H, TangZ C, et al. ChatKBQA: a generate-then-retrieve framework for knowledge base question answering with fine-tuned large language models[C]//Findings of the association for computational linguistics: ACL 2024. Stroudsburg, USA: ACL, 2024:2039-2056.
[6]
BhutaniN, ZhengX Y, JagadishH V. Learning to answer complex questions over knowledge bases with query composition[C]//Proceedings of the 28th ACM international conference on information and knowledge management. New York, USA: ACM, 2019: 739-748.
[7]
LanY S, JiangJ. Query graph generation for answering multi-hop complex questions from knowledge bases[C]//Proceedings of the 58th annual meeting of the association for computational linguistics. Stroudsburg, USA: ACL, 2020:969-974.
[8]
MaS J, XuC J, JiangX H, et al. Think-on-graph 2.0: deep and faithful large language model reasoning with knowledge-guided retrieval augmented generation[C]//Proceedings of the 13th international conference on learning representations. Appleton, USA: ICLR, 2025:52782-52806.
[9]
GuoZ R, XiaL H, YuY H, et al. LightRAG: simple and fast retrieval-augmented generation[C]//Findings of the association for computational linguistics: EMNLP 2025. Stroudsburg, USA: ACL, 2025: 10746-10761.
[10]
YuH Q, McQuadeF. RAG-KG-IL: a multi-agent hybrid framework for reducing hallucinations and enhancing LLM reasoning through RAG and incremental knowledge graph learning integration. [PP/OL] arXiv (2025-03-14)[2026-01-28].
[11]
ZhangY T, ChenL X, LiS H, et al. Way to specialist: closing loop between specialized LLM and evolving domain knowledge graph[C]//Proceedings of the 31st ACM SIGKDD conference on knowledge discovery and data mining V.1. New York, USA: ACM, 2025:1996-2007.
[12]
YuD H, ZhangS, NgP, et al. DecAF: joint decoding of answers and logical forms for question answering over knowledge bases[C]//Proceedings of the 11th international conference on learning representations. Appleton, USA: ICLR,2023. DOI: 10.48550/arXiv.2210.00063 .
[13]
JiangJ H, ZhouK, ZhaoW X, et al. UniKGQA: unified retrieval and reasoning for solving multi-hop question answering over knowledge graph[C]//Proceedings of the 11th international conference on learning representations. Appleton, USA: ICLR, 2023. DOI: 10.48550/arXiv.2212.00959 .
[14]
LiuY, YavuzS, MengR, et al. Uni-parser: unified semantic parser for question answering on knowledge base and database[C]//Proceedings of the 2022 conference on empirical methods in natural language processing. Stroudsburg, USA: ACL, 2022:8858-8869.
[15]
LanY S, WangS H, JiangJ. Knowledge base question answering with topic units[C]//Proceedings of the twenty-eighth international joint conference on artificial intelligence. California, USA: IJCAI, 2019: 5046-5052.
[16]
CaoS L, ShiJ X, YaoZ J, et al. Program transfer for answering complex questions over knowledge bases[C]//Proceedings of the 60th annual meeting of the association for computational linguistics. Stroudsburg, USA: ACL, 2022:8128-8140.
[17]
ChenY R, LiH Y, QiG L, et al. Outlining and filling: hierarchical query graph generation for answering complex questions over knowledge graphs[J]. IEEE Transactions on Knowledge and Data Engineering, 2023,35(8):8343-8357.