计算机科学

计算机科学 ›› 2026, Vol. 53 ›› Issue (10): 41-50.doi: 10.11896/jsjkx.260500139

• 大模型与多模态智能感知 • 上一篇    下一篇

SRG-RAG:融合图重排序与自反思机制的检索增强生成框架

孙天珍1, 张顺2, 李彤亮3, 李舟军1   

  1. 1 北京航空航天大学计算机学院 北京 100191
    2 中国人民公安大学信息网络安全学院 北京 100038
    3 北京信息科技大学计算机学院 北京 102206
  • 收稿日期:2026-05-27 修回日期:2026-07-04 发布日期:2026-10-13
  • 通讯作者: 李舟军(lizj@buaa.edu.cn)
  • 作者简介:(SY2306326@buaa.edu.cn)
  • 基金资助:
    国家自然科学基金(62276017,62406033,U1636211,61672081);复杂关键软件环境全国重点实验室(SKLCCSE-2024ZX-18)

SRG-RAG:Integrating Graph Re-ranking and Self-reflection into Retrieval-augmented Generation

SUN Tianzhen1, ZHANG Shun2, LI Tongliang3, LI Zhoujun1   

  1. 1 School of Computer Science and Engineering,Beihang University,Beijing 100191, China
    2 School of Information and Cyber Security,People’s Public Security University of China,Beijing 100038, China
    3 Computer School,Beijing Information Science and Technology University,Beijing 102206, China
  • Received:2026-05-27 Revised:2026-07-04 Online:2026-10-13
  • About author:SUN Tianzhen,born in 1991,master’s candidate.His main research interest is natural language processing based on large language models.
    LI Zhoujun,born in 1963,Ph.D,professor,is a senior member of CCF(No.06618S).His main research interests include artificial intelligence,natural language processing,and network and information security.
  • Supported by:
    National Natural Science Foundation of China(62276017,62406033,U1636211,61672081) and State Key Laboratory of Complex & Critical Software Environment(SKLCCSE-2024ZX-18).

PDF (PC)

94

摘要： 科技查新是评估科研项目新颖性与专利授权可行性的关键环节,整套比对流程对逻辑严谨性要求极高。传统检索增强生成范式在处理长文本技术文献时,常因“语义平滑”效应丢失微观工艺特征与拓扑关联;同时,在底层证据缺失的情境下,大模型易受参数化先验知识干扰而诱发级联幻觉,导致查新结论的事实性偏离。针对上述问题,提出一种融合图重排序与自反思机制的检索增强生成框架SRG-RAG。首先,该框架设计了多维意图解耦模型以提升文献初筛精度;其次,通过动态抽取技术特征子图并引入拓扑差分打分算法,对高混淆度查新方案完成精细化重排序;最后,构建基于伪相关反馈的自反思评估器,通过动态查询改写实现检索生成的闭环纠错。在自建的拓扑结构对齐评测基准TopoNovelty-Bench上的实验结果表明,SRG-RAG以Qwen2.5-72B-Instruct为底层基座,将事实幻觉率由基线的22.4%压降至6.2%,显著提升了查新报告的证据效力与判决准确性,为垂直领域的高信度生成任务提供了参考范式。

关键词: 检索增强生成, 科技查新, 图重排序, 自反思机制, 拓扑对齐

Abstract: Technology novelty search is critical for evaluating the innovation of research projects and patentability,where rigorous logical comparison poses significant challenges to generative artificial intelligence.Traditional retrieval-augmented generation pa-radigms frequently lose micro-level process features and topological correlations when processing long-form technical documents,as a result of the semantic smoothing effect.Meanwhile,in contexts where underlying evidence is insufficient,large language mo-dels are vulnerable to interference from parametric prior knowledge,which triggers cascading hallucinations and in turn causes factual deviations in novelty assessment conclusions.To address these issues,this paper proposes SRG-RAG,a retrieval-augmented generation framework integrating graph reranking and self-reflection mechanisms.Firstly,a multi-dimensional intent decoupling model is designed to improve initial screening precision.Secondly,by dynamically extracting technical feature subgraphs and introducing a topological differential scoring algorithm,fine-grained reranking of highly confusable technical schemes is achieved.Finally,a self-reflective evaluator based on pseudo-relevance feedback theory is constructed to implement closed-loop error correction via dynamic query rewriting.Experimental results on the self-built topological alignment benchmark TopoNovelty-Bench show that self-reflective retrieval-augmented generation,powered by Qwen2.5-72B-Instruct,reduces the factual hallucination rate from 22.4% to 6.2%,significantly enhancing the evidential validity and accuracy of novelty search reports.This work provides a reliable paradigm for high-fidelity generation tasks in vertical domains.

Key words: Retrieval-augmented generation, Technology novelty search, Graph reranking, Self-reflection mechanism, Topological alignment

中图分类号: 

  • TP391
[1] GAO Y,XIONG Y,GAO X,et al.Retrieval-augmented generation for large language models:A survey[J].arXiv:2312.10997,2023.
[2] HUANG L,YU W,MA W,et al.A survey on hallucination in large language models:Principles,taxonomy,challenges,and open questions[J].ACM Transactions on Information Systems,2025,43(2):1-55.
[3] TONMOY S,ZAMAN S,JAIN V,et al.A comprehensive survey of hallucination mitigation techniques in large language models[J].arXiv:2401.01313,2024.
[4] ASAI A,WU Z,WANG Y,et al.Self-rag:Learning to retrieve,generate,and critique through self-reflection[C]//International Conference on Learning Representations.2024:9112-9141.
[5] JIANG Z,XU F F,GAO L,et al.Active retrieval augmented generation[C]//Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing.2023:7969-7992.
[6] SHINN N,CASSANO F,GOPINATH A,et al.Reflexion:Language agents with verbal reinforcement learning[J].Advances in Neural Information Processing Systems,2023,36:8634-8652.
[7] EDGE D,TRINH H,CHENG N,et al.From local to global:Agraph rag approach to query-focused summarization[J].arXiv:2404.16130,2024.
[8] LIU N F,LIN K,HEWITT J,et al.Lost in the middle:Howlanguage models use long contexts[J].Transactions of the Association for Computational Linguistics,2024,12:157-173.
[9] DZIRI N,LU X,SCLAR M,et al.Faith and fate:Limits oftransformers on compositionality[J].Advances in Neural Information Processing Systems,2023,36:70293-70332.
[10] HUANG J,CHEN X,MISHRA S,et al.Large language models cannot self-correct reasoning yet[C]//International Conference on Learning Representations.2024:32808-32824.
[11] WANG X L,LI Y,MA C F,et al.A survey on retrieval-aug-mented generation(RAG):Methods and applications [J].Computer Science,2026,53(7):101-117.
[12] CHEN W L,ZHANG H B,JIN D Z.Design and implementation of knowledge base question answering system using retrieval-augmented generation based onLangChain [J].Computer Era,2026(2):51-56.
[13] WANG Y,HAN X X,ZHANG J X,et al.Multi-dimensional retrieval-augmented generation method for financial audit system [J/OL].Journal of Computer Applications.https://link.cnki.net/urlid/51.1307.tp.20260127.1141.002.
[14] KHATTAB O,ZAHARIA M.Colbert:Efficient and effectivepassage search via contextualized late interaction over bert[C]//Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval.2020:39-48.
[15] GAO T,YAO X,CHEN D.Simcse:Simple contrastive learning of sentence embeddings[C]//Proceedings of the 2021 Confe-rence on Empirical Methods in Natural Language Processing.2021:6894-6910.
[16] CHEN D,LIN Y,LI W,et al.Measuring and relieving the over-smoothing problem for graph neural networks from the topological view[C]//Proceedings of the AAAI Conference on Artificial Intelligence.2020:3438-3445.
[17] LEWIS P,PEREZ E,PIKTUS A,et al.Retrieval-augmentedgeneration for knowledge-intensive nlp tasks[J].Advances in Neural Information Processing Systems,2020,33:9459-9474.
[18] KARPUKHIN V,OGUZ B,MIN S,et al.Dense passage retrieval for open-domain question answering[C]//Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing(EMNLP).2020:6769-6781.
[19] CHEN J,XIAO S,ZHANG P,et al.Bge m3-embedding:Multi-lingual,multi-functionality,multi-granularity text embeddings through self-knowledge distillation[J].arXiv:2402.03216,2024.
[20] YANG A,LI A,YANG B,et al.Qwen3 technical report[J].arXiv:2505.09388,2025.
[21] GAO L,MA X,LIN J,et al.Precise zero-shot dense retrievalwithout relevance labels[C]//Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics.2023:1431-1443.
[22] WELLER O,LO K,WADDEN D,et al.When do generativequery and document expansions fail? A comprehensive study across methods,retrievers,and datasets[C]//Findings of the Association for Computational Linguistics:EACL 2024.2024:1827-1842.
[23] SONG W J,WANG L,XU C,et al.Historical materials know-ledge question answering system based on knowledge graph and retrieval-augmented generation [J].Data Analysis and Know-ledge Discovery,2026,10(3):100-112.
[24] LI J X,LI J T.Graph retrieval-augmented generation paradigm from the perspective of graph chain-of-thought [J].Computer Science,1-11.https://link.cnki.net/urlid/50.1075.tp.20260317.1922.025.
[25] JIN B,XIE C,ZHANG J,et al.Graph chain-of-thought:Augmenting large language models by reasoning on graphs[C]//Findings of the Association for Computational Linguistics:ACL 2024.2024:163-184.
[26] CHEN J,LIN H,HAN X,et al.Benchmarking large languagemodels in retrieval-augmented generation[C]//Proceedings of the AAAI Conference on Artificial Intelligence.2024:17754-17762.
[27] HUANG Y,BAI Y,ZHU Z,et al.C-eval:A multi-level multi-discipline chinese evaluation suite for foundation models[J].Advances in Neural Information Processing Systems,2023,36:62991-63010.
[28] ZHANG M,PRESS O,MERRILL W,et al.How language model hallucinations can snowball[J].arXiv:2305.13534,2023.
[29] XIE J,ZHANG K,CHEN J,et al.Adaptive chameleon or stubborn sloth:Revealing the behavior of large language models in knowledge conflicts[C]//International Conference on Learning Representations.2024:35623-35646.
[30] LI T L,LI Q F,HOU X,et al.Unsupervised dialogue topic segmentation algorithm based on utterance rewriting [J].Computer Science,2025,52(12):215-223.
[31] YAN S Q,GU J C,ZHU Y,et al.Corrective retrieval augmented generation[J].arXiv:2401.15884,2024.
[32] CUCONASU F,TRAPPOLINI G,SICILIANO F,et al.Thepower of noise:Redefining retrieval for rag systems[C]//Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval.2024:719-729.
[33] CRONEN-TOWNSEND S,ZHOU Y,CROFT W B.Predicting query performance[C]//Proceedings of the 25th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval.2002:299-306.
[34] ZHANG J Y,WANG T K,YAO C Y,et al.Construction and evaluation of intelligent question answering system for power knowledge base based on large language models [J].Computer Science,2024,51(12):286-292.
[35] VOORHEES E M.The TREC-8 question answering track re-port[R].Trec,1999:77-82.
[36] CLARK E,AUGUST T,SERRANO S,et al.All that’s ‘human’is not gold:Evaluating human evaluation of generated text[C]//Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing.2021:7282-7296.
[37] THAKUR N,REIMERS N,RÜCKLÉ A,et al.Beir:A heterogenous benchmark for zero-shot evaluation of information retrieval models[J].arXiv:2104.08663,2021.
[38] GAO L,CALLAN J.Condenser:a pre-training architecture for dense retrieval[C]//Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing.2021:981-993.
[39] SHI W,HAN X,LEWIS M,et al.Trusting your evidence:Hallucinate less with context-aware decoding[C]//Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics:Human Language Technologies.2024:783-791.
[40] MALLEN A,ASAI A,ZHONG V,et al.When not to trust lan-guage models:Investigating effectiveness of parametric and non-parametric memories[C]//Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics.2023:9802-9822.
[41] SHI F,CHEN X,MISRA K,et al.Large language models can be easily distracted by irrelevant context[C]//International Conference on Machine Learning.2023:31210-31227.
Viewed
Full text


Abstract

Cited
  Shared   
  Discussed   
No Suggested Reading articles found!
渝ICP备16013121号-4
地址：重庆市两江新区洪湖西路18号    邮编：401121  
电话：023-63500828（编辑部） 023-67039612（会议/专题） 023-67039623（财务/发票）
E-mail：jsjkx12@163.com
本系统由北京玛格泰克科技发展有限公司设计开发