School of Information and Control Engineering, Qingdao University of Technology, Qingdao 266520, China
Show less
文章历史+
Received
Published
Revised
2025-05-27
2026-05-08
2025-11-26
Online Date
2026-08-06
PDF (5864K)
摘要
传统检索增强生成(retrieval-augmented generation,RAG)方法在处理需要多跳推理和全局理解的复杂查询时存在明显不足,比如在教育领域的跨章节问答等场景中表现欠佳。针对以上问题,提出一种基于知识图谱的多跳检索增强生成机制(graph-enhanced retrieval-augmented generation for multi-hop QA, GRAG-MH)。该机制首先从原始文档中提取实体及关系构建基础知识图谱,其次为紧密关联的实体组生成摘要节点,形成层次化的摘要增强知识图谱。在处理查询时,通过动态问题分解将复杂问题拆解为子问题序列,利用摘要增强知识图谱实现多跳邻居扩展检索,最后通过自洽性投票方法整合多个推理路径的结果。实验结果表明,在HotpotQA和2WikiMultihopQA两个多跳问答基准数据集上,GRAG-MH方法的F1分别为53.7%和50.7%,较传统RAG方法提升超过50%。通过可解释的多路径推理过程与层次化知识表示相结合,该方法显著提升了问答系统的准确度。
Abstract
Traditional retrieval-augmented generation (RAG) methods exhibit significant limitations in handling complex queries requiring multi-hop reasoning and global understanding, such as cross-chapter question answering in educational contexts. To address these issues, a graph-enhanced retrieval-augmented generation mechanism for multi-hop QA (GRAG-MH) was proposed. Firstly, entities and relations were extracted from the original documents to construct a basic knowledge graph. Then, summary nodes were generated for closely related entity groups, forming a hierarchical summary-enhanced knowledge graph. During query processing, complex questions were decomposed into sub-question sequences through dynamic question decomposition. Multi-hop neighbor expansion retrieval was performed using the summary-enhanced knowledge graph. Finally, the results from multiple reasoning paths were integrated via a self-consistency voting method. Experimental results demonstrate that GRAG-MH achieves F1 scores of 53.7% and 50.7% on the multi-hop QA benchmark datasets HotpotQA and 2WikiMultihopQA, respectively, outperforming traditional RAG methods by over 50%. By combining interpretable multi-path reasoning with hierarchical knowledge representation, the proposed method significantly improves the accuracy of QA systems.
ZHOU Wei, LIN Hong-sheng, LI Hai-peng, LIU Lei, YANG Xin.
Multi-hop Retrieval-augmented Generation Method Fusing Hierarchical Knowledge Graphs and Dynamic Decomposition[J].
Science Technology and Engineering, 2026, 26(13): 5594-5603 DOI:10.12404/j.issn.1671-1815.2503930
近年来,大语言模型[1](large language models, LLMs)在智能教育领域[2-3]展现出变革性潜力。其通过海量数据训练获得的语言理解与生成能力,能够为用户提供高效的知识获取体验,比如个性化学习内容的动态生成。这种强大的语义处理能力使得LLMs成为智能时代重要的知识交互媒介,为教育辅助、专业咨询等场景提供了新的技术范式。
知识图谱增强的信息检索与内容生成技术已成为人工智能领域的重要研究方向,其核心在于利用图结构的语义关联特性突破传统检索增强生成的局限性。与传统RAG基于文本片段的检索机制不同,基于知识图谱的检索增强生成通过从结构化知识图谱中检索与查询语义相关的图元素(包括实体节点、关系边、推理路径或连通子图),实现了对复杂语义关系的精准捕捉。这种基于图结构的检索范式能够显示建模文本间的深层关联,显著提升了多跳推理任务中的信息召回率和准确率。典型的工作包括GraphRAG[32]通过从预构建的知识图谱中检索子图或路径来增强生成,但依赖静态摘要节点导致冗余信息泛滥,且缺乏动态多跳推理能力。G-Retriever[33]创新性地融合图神经网络、大语言模型和RAG框架,通过引入软提示技术增强图语义理解,并采用奖赏收集斯坦纳树(PCST)算法实现最优子图检索;KGP[34]构建了跨文档的知识图谱索引系统,其节点表征文档段落或结构化元素(如表格),边则编码段落间的语义关联和文档结构关系,有效支持了多文档场景下的知识推理。在知识图谱问答方面,CoK (chain of knowledge)[35]通过检索相关事实并将其作为提示集成到输入序列中,显著提升了问答性能。当前子图检索技术主要分为两类:基于规则的方法(如RA-Sim[36]构建专利-短语图进行特定领域检索,MHGRN[29]采用两跳邻居子图策略)和路径优先方法(如SR[37]利用RoBERTa模型识别多推理路径后合并实体)。
WuDi, LiHuan, ChenXu. Analysis of the impact of general artificial intelligence models on educational applications[J]. Open Education Research, 2023, 29(2): 19-25.
[4]
CarruanaM A, AlarioH C, DelgadoK C. Smart education: a review and future research directions[C]// The Proceedings of the 13th International Conference on Ubiquitous Computing and Ambient Intelligence. Toledo: MDPI, 2019: 57-67.
[5]
BanerjeeS, AgarwalA, SinglaS. LLMs will always hallucinate, and we need to live with this[J]. Intelligent Systems and Applications, 2025, 1554: 624-648.
[6]
MaX, GongY, HeP, et al. Query rewriting for retrieval-augmented large language models[C]// The Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. Singapore: ACL, 2023: 5303-5315.
LiShengfei, LuYujie, ChenXiaoying, et al. Implementation of a knowledge Q&A model for construction project management based on retrieval-augmented generation[J]. Science Technology and Engineering, 2025, 25(25): 10840-10849.
DongXingtong, ChenShihong, ChenShuxin. Design of an optimized algorithm for text duplicate detection in natural language processing[J]. Science Technology and Engineering, 2022, 22(3): 1091-1097.
WangShuai, HeWenchun, WangFudi, et al. Question-answering system of large language models integrating knowledge graphs and vector retrieval[J]. Science Technology and Engineering, 2024, 24(32): 13902-13910.
XuNa, ChenXi, YangJianping, et al. Research status, challenges, and trends of large language models in the field of civil engineering[J]. Science Technology and Engineering, 2025, 25(21): 8773-8783.
ZhangHeyi, WangXin, HanLifan, et al. Research on question-answering systems integrating large language models with knowledge graphs[J]. Computer Science and Exploration, 2023, 17(10): 2377-2388.
AchiamT, AdlerS, AgarwalS, et al. The LLaMA3 herd of models[J]. ArXiv: 2407.21783, 2024.
[20]
WeiJ, TayY, BommasaniR, et al. Emergent abilities of large language models under continued pre-training for language adaptation[C]// Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics. Vienna: ACL, 2025: 32174-32186.
[21]
SuD, XuY, WinataG I, et al. Generalizing question answering system with pre-trained language model fine-tuning[C]// Proceedings of the 2nd workshop on machine reading for question answering. Hong Kong: ACL, 2019: 203-211.
[22]
LewisM, LiuY, GoyalN, et al. Bart: denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension[C]// Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. Stroudsburg: ACL, 2020: 7871-7880.
[23]
LiJ, TangT, ZhaoW X, et al. Pre-trained language models for text generation: a survey[J]. ACM Computing Surveys, 2024, 56(9): 1-39.
[24]
ZhangH, SongH, LiS, et al. A survey of controllable text generation using transformer-based pre-trained language models[J]. ACM Computing Surveys, 2023, 56(3): 1-37.
[25]
VrandečičD, KrötzschM. Wikidata: a free collaborative knowledgebase[J]. Communications of the ACM, 2014, 57(10): 78-85.
[26]
SuchanekM, KasneciG, WeikumG. Yago: a core of semantic knowledge[C]// Proceedings of the 16th International Conference on World Wide Web. New York: ACM, 2007: 697-706.
[27]
CarlsonA, BetteridgeJ, KisielB, et al. Toward an architecture for never-ending language learning[C]// Proceedings of the AAAI Conference on Artificial Intelligence. Atlanta: AAAI, 2010: 1306-1313.
[28]
ZhangJ, ChenB, ZhangL, et al. Neural, symbolic and neural-symbolic reasoning on knowledge graphs[J]. AI Open, 2021, 214-235.
[29]
WangX, GaoT, ZhuZ, et al. KEPLER: a unified model for knowledge embedding and pre-trained language representation[J]. Transactions of the Association for Computational Linguistics, 2021, 9: 176-194.
[30]
MelnykI, DogninP, DasP. Grapher: multi-stage knowledge graph construction using pretrained language models[C]// NeurIPS 2021 Workshop DGMs and Applications. San Diego: OpenReview, 2021: 120-131.
[31]
JiangJ, ZhouK, ZhaoW, 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. Kigali: ICLR, 2023: 34362-34373.
[32]
HuY, LeiZ, ZhangZ, et al. Grag: Graph retrieval-augmented generation[C]// Findings of the Association for Computational Linguistics. Albuquerque: ACL, 2025: 4145-4157.
[33]
WangX, KapanipathiP, MusaR, et al. Improving natural language inference using external knowledge in the science questions domain[C]// Proceedings of the 33rd AAAI Conference on Artificial Intelligence. Honolulu: AAAI, 2019: 7208-7215.
[34]
LinY, ChenX, ChenJ, et al. Kagnet: knowledge-aware graph networks for commonsense reasoning[C]// Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing. Hong Kong: ACL, 2019: 2829-2839.
[35]
FengY, ChenX, LinB Y, et al. Scalable multi-hop relational reasoning for knowledge-aware question answering[C]// Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing. Stroudsburg: ACL, 2020: 1295-1309.
[36]
SunY, ShiQ, QiL, et al. JointLK: joint reasoning with language models and knowledge graphs for commonsense question answering[C]//Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. Seattle: ACL, 2022: 5049-5060.
[37]
ZhangX, BosselutA, YasunagaM, et al. Greaselm: graph reasoning enhanced language models for question answering[C]// Proceedings of the 10th International Conference on Learning Representations. Appleton: ICLR, 2022: 24765-24781.
[38]
EdgeD, TrinhH, ChengN, et al. From local to global: a graph rag approach to query-focused summarization[J]. ArXiv: 2404.16130, 2024.
[39]
HeX, TianY, SunY, et al. G-retriever: retrieval-augmented generation for textual graph understanding and question answering[C]// Proceedings of the 38th International Conference on Neural Information Processing Systems. Red Hook: Curran Associates Inc, 2024: 132876-132907.
[40]
WangY, LipkaN, RossiR, et al. Knowledge graph prompting for multi-document question answering[C]// Proceedings of the AAAI Conference on Artificial Intelligence. Vancouver: AAAI, 2024: 19206-19214.
[41]
LiX, ZhaoR, ChiaY K, et al. Chain of knowledge: a framework for grounding large language models with structured knowledge bases[C]// Proceedings of the 12th International Conference on Learning Representations. Vienna: ICLR, 2024: 55930-55953.
[42]
PengZ, YangY. Connecting the dots: inferring patent phrase similarity with retrieved phrase graphs[C]// Findings of the Association for Computational Linguistics. Mexico: ACL, 2024: 1877-1890.
[43]
ZhangJ, ZhangX, YuJ, 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. Dublin: ACL, 2022: 5773-5784.
[44]
KimD, KimB, HanD, et al. AutoRAG: automated framework for optimization of retrieval augmented generation pipeline[J]. ArXiv: 2410.20878, 2024.
[45]
LiuN, LinK, HewittJ, et al. Lost in the middle: How language models use long contexts[C]// Transactions of the Association for Computational Linguistics. Cambridge: MIT Press, 2024: 157-173.