基于自适应引用验证与双阶段融合重排的自动化文献综述生成方法

doi:10.11871/jfdc.issn.2096-742X.2026.05.014

数据与计算发展前沿 ›› 2026, Vol. 8 ›› Issue (5): 177-189.

CSTR: 32002.14.jfdc.CN10-1649/TP.2026.05.014

doi: 10.11871/jfdc.issn.2096-742X.2026.05.014

基于自适应引用验证与双阶段融合重排的自动化文献综述生成方法

龚厚骅^1,²(),刘子涵^1,²,张云飞^1,²,黎建辉^1,^3,^*(),朴英超¹,裴昶华¹

¹ 中国科学院计算机网络信息中心，北京 100083
² 中国科学院大学，北京 100049
³ 南京大学，前沿科学学院，江苏苏州 215163

收稿日期:2025-10-29 出版日期:2026-10-20 发布日期:2026-10-20
通讯作者: ^*黎建辉（E-mail: lijh@nju.edu.cn）
作者简介:龚厚骅，中国科学院计算机网络信息中心，硕士研究生，主要研究方向为大语言模型应用。
本文承担工作为：论文撰写与系统研发
GONG Houhua is a master’s student at the Computer Network Information Center, Chinese Academy of Sciences. His main research interest include the application of large language models.
In this paper, he is responsible for paper composition and system development. E-mail: hhgong@cnic.cn|黎建辉，南京大学教授兼任可持续发展大数据国际研究中心副主任委员，以及国际科学理事会数据委员会（CODATA）副主席（2018-2023）。研究方向为地球科学和可持续发展目标领域的大数据、云计算和人工智能研究。
本文承担工作为：论文方向指导与审阅、项目管理
LI Jianhui is a professor at the Nanjing University and Deputy Director General of the International Research Center of Big Data for Sustainable Development Goals (CBAS), and served as Vice President of the CODATA from 2018 to 2023. His research interests include big data analytics, cloud computing, and artificial intelligence with applications to earth science and sustainable development goals.
In this paper, he is responsible for guidance and review of the paper, and project management. E-mail: lijh@nju.edu.cn
基金资助:
可持续发展大数据国际研究中心运行项目（可持续发展目标大模型开发与应用）(E53Z0505T)

An Automated Literature Review Generation Method Based on Adaptive Citation Verification and Two-Stage Fusion Rerank

GONG Houhua^1,²(),LIU Zihan^1,²,ZHANG Yunfei^1,²,LI Jianhui^1,^3,^*(),PIAO Yingchao¹,PEI Changhua¹

¹ Computer Network Information Center, Chinese Academey of Sciences, Beijing 100083, China
² University of Chinese Academy of Sciences, Beijing 100049, China
³ School of Frontier Sciences, Nanjing University, Suzhou, Jiangsu 215163, China

Received:2025-10-29 Online:2026-10-20 Published:2026-10-20

摘要/Abstract

摘要：

【目的】随着学术文献数量的指数级增长，人工撰写高质量综述不仅耗时费力，而且容易出现遗漏与偏差。近年来，虽有研究尝试利用大语言模型实现综述的自动化生成以提升效率，但在高质量参考文献覆盖度与引用准确性方面，仍与人工撰写存在明显差距。【方法】 针对上述问题，本文提出了一种基于自适应引用验证与双阶段融合重排的自动化文献综述生成方法。首先，通过查询扩展获取与输入主题高度相关的主题描述，并基于该描述检索相关文献；其次，设计双阶段重排机制，以筛选出与研究主题关联度更高的学术文献；最后，在完成大纲与正文生成后，引入自适应引用验证机制，对引用内容进行动态核验与调整，从而显著提升引用的准确性。【结果】为验证方法的有效性，本文构建了覆盖可持续发展、计算机科学和生物医药三个领域、共计104,321篇论文的领域知识库，并从引用质量、内容质量和参考文献质量三个维度开展评估。实验结果表明，本方法在文献覆盖度与引用准确性方面均优于现有方法（如AutoSurvey），显著提升了自动化综述的学术可靠性。

关键词: 大语言模型, 自动化文献综述生成, RAG（检索增强生成）, 长文本生成

Abstract:

[Objective] With the exponential growth of academic publications, manual writing of high-quality literature reviews has become increasingly time-consuming and prone to omissions and biases. In recent years, some studies have attempted to leverage large language models (LLMs) to automate the review-writing process and improve efficiency. However, challenges remain in achieving comprehensive coverage of high-quality references and ensuring citation accuracy compared to human-authored reviews. [Methods] To address these limitations, this paper proposes an automated literature review generation method based on adaptive citation verification and two-stage fusion reranking. First, query expansion is used to obtain topic descriptions that are highly relevant to the input theme. The topic descriptions are then employed to retrieve related documents. Second, a two-stage fusion reranking mechanism is designed to identify academic papers with stronger topic relevance. Finally, after generating the outline and main text, an adaptive citation verification mechanism dynamically validates and refines citations to significantly enhance citation accuracy. [Results] To evaluate the effectiveness of the proposed method, we constructed a domain knowledge base comprising 104,321 papers across three domains: sustainable development, computer science, and biomedicine, and conducted evaluations from three dimensions: citation quality, content quality, and reference quality. Experimental results demonstrate that our method outperforms existing approaches (e.g., AutoSurvey) in both literature coverage and citation accuracy, thereby substantially improving the academic reliability of automated literature reviews.

Key words: large language model (LLM), automated literature review generation, retrieval-augmented generation (RAG), long text generation

龚厚骅, 刘子涵, 张云飞, 黎建辉, 朴英超, 裴昶华. 基于自适应引用验证与双阶段融合重排的自动化文献综述生成方法[J]. 数据与计算发展前沿, 2026, 8(5): 177-189.

GONG Houhua, LIU Zihan, ZHANG Yunfei, LI Jianhui, PIAO Yingchao, PEI Changhua. An Automated Literature Review Generation Method Based on Adaptive Citation Verification and Two-Stage Fusion Rerank[J]. Frontiers of Data and Computing, 2026, 8(5): 177-189, https://cstr.cn/32002.14.jfdc.CN10-1649/TP.2026.05.014.

图/表 8

图1

算法1

自动化文献综述生成方法"

Algorithm1: The pipline of our method

Input: 用户输入主题T;论文数据库D;
OutPut: 文献综述生成结果F
// 双阶段融合重排
Function TwoStageFusionRerank(P)：
${\mathit{r}}_{\mathit{t}\mathit{i}\mathit{m}\mathit{e}}\leftarrow $RankByTime(P);
${\mathit{r}}_{\mathit{c}\mathit{i}\mathit{t}\mathit{e}}\leftarrow $RankByCitations(P);
${\mathit{r}}_{\mathit{V}\mathit{e}\mathit{c}}\leftarrow $RankByVector(P);
${\mathit{r}}_{\mathit{f}\mathit{u}\mathit{s}\mathit{e}\mathit{d}}\leftarrow $FuseByAvgInverseRank(r_time,r_cite,r_Vec);
Return LLMRerank(r_fused);
Begin
// 阶段一：查询扩展与初始检索
$\mathit{q}\mathit{u}\mathit{e}\mathit{r}\mathit{y}\leftarrow $QueryExpand(T);
${\mathit{P}}_{\mathit{i}\mathit{n}\mathit{i}\mathit{t}}\leftarrow $Retrieve(query,D);
// 阶段二：一级与二级大纲生成
基于P_init生成一级大纲集合$\left\{{\mathit{O}}_{\mathit{i}}\right\}$;
for每个一级大纲O_ido
${\mathit{P}}_{\mathit{i}}\leftarrow $Retrieve(d_i,D);
${\mathit{P}}_{\mathit{i}}^{\mathit{r}\mathit{e}\mathit{r}\mathit{a}\mathit{n}\mathit{k}\mathit{e}\mathit{d}}\leftarrow \mathit{T}\mathit{w}\mathit{o}\mathit{S}\mathit{t}\mathit{a}\mathit{g}\mathit{e}\mathit{F}\mathit{u}\mathit{s}\mathit{i}\mathit{o}\mathit{n}\mathit{R}\mathit{e}\mathit{r}\mathit{a}\mathit{n}\mathit{k}\left({\mathit{P}}_{\mathit{i}}\right)$;
基于P_i^reranked生成对应二级大纲集合$\left\{{\mathit{O}}_{\mathit{i}\mathit{j}}\right\}$;
end for
// 阶段三：章节内容并行生成
for 每个二级大纲O_ij（并行）do
for子章节描述d_ijdo
${\mathit{P}}_{\mathit{i}\mathit{j}}\leftarrow $Retrieve(d_ij,D);
${\mathit{P}}_{\mathit{i}\mathit{j}}^{\mathit{r}\mathit{e}\mathit{r}\mathit{a}\mathit{n}\mathit{k}\mathit{e}\mathit{d}}\leftarrow $TwoStageFusionRerank(P_ij);
基于P_ij^reranked生成子章节内容C_ij;
end for
end for
${\mathit{F}}_{\mathit{d}\mathit{r}\mathit{a}\mathit{f}\mathit{t}}\leftarrow $MergeSections(C_ij);
// 阶段四：自适应引用验证
$\mathit{S},{\mathit{P}}_{\mathit{r}\mathit{e}\mathit{l}\mathit{a}\mathit{t}\mathit{e}\mathit{d}}\leftarrow \mathit{E}\mathit{x}\mathit{t}\mathit{r}\mathit{a}\mathit{c}\mathit{t}\mathit{C}\mathit{i}\mathit{t}\mathit{e}\mathit{d}\mathit{S}\mathit{e}\mathit{n}\mathit{t}\mathit{e}\mathit{n}\mathit{c}\mathit{e}\mathit{s}\left({\mathit{F}}_{\mathit{d}\mathit{r}\mathit{a}\mathit{f}\mathit{t}}\right)$;
for 每个句子$\mathit{s}\in \mathit{S}$及其对应文献 $\mathit{p}\in {\mathit{P}}_{\mathit{r}\mathit{e}\mathit{l}\mathit{a}\mathit{t}\mathit{e}\mathit{d}}$do
for $\mathit{t}\leftarrow 1$ to K do
check$\leftarrow $CheckCitationConsistency(s,p);
if check==true then
break;
${\mathit{P}}_{\mathit{r}\mathit{e}\mathit{f}}\leftarrow $Retrieve(s,D);
$\mathit{s}\leftarrow $RefineSentenceWithEvidence(s,P_ref,LLM);
end for
end for
$\mathit{F}\leftarrow $MergeSentences(F_draft,S);
Return F;
End

算法1

表1

表2

图2

表3

图3

图4

参考文献 19

[1]	ACHIAM J, ADLER S, AGARWAL S, et al. Gpt-4 technical report[J/OL]. 2023. https://doi.org/10.48550/arXiv.2303.08774.
[2]	GAO Y, XIONG Y, GAO X, et al. Retrieval-augmented generation for large language models: A survey[J/OL]. 2023. https://doi.org/10.48550/arXiv.2312.10997.
[3]	FAN W, DING Y, NING L, et al. A survey on rag meeting llms: Towards retrieval-augmented large language models[C]. Proceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining. 2024: 6491-6501.
[4]	OpenAI. Introducing deep research[J/OL]. 2024.https://openai.com/index/introducing-deep-research.
[5]	WANG Y, GUO Q, YAO W, et al. Autosurvey: Large language models can automatically write surveys[J]. Advances in Neural Information Processing Systems, 2024, 37: 115119-115145.
[6]	ASAI A, HE J, SHAO R, et al. Openscholar: Synthesizing scientific literature with retrieval-augmented lms[J/OL]. 2024. https://doi.org/10.48550/arXiv.2411.14199.
[7]	JI Z, LEE N, FRIESKE R, et al. Survey of hallucination in natural language generation[J]. ACM Computing Surveys, 2023, 55(12): 1-38.
[8]	LEE S, HUH S Y, MCNIEL R D. Automatic generation of concept hierarchies using WordNet[J]. Expert Systems with Applications, 2008, 35(3): 1132-1144.
[9]	HU Y, WAN X. Automatic generation of related work sections in scientific papers: an optimization approach[C]. Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP). 2014: 1624-1633.
[10]	JHA R, JBARA A A, QAZVINIAN V, et al. NLP-driven citation analysis for scientometrics[J]. Natural Language Engineering, 2017, 23(1): 93-130.
[11]	NASAR Z, JAFFRY S W, MALIK M K. Information extraction from scientific articles: a survey[J]. Scientometrics, 2018, 117(3): 1931-1990.
[12]	SUN X, ZHUGE H. Automatic generation of survey paper based on template tree[C]. 2019 15th International Conference on Semantics, Knowledge and Grids (SKG). IEEE, 2019: 89-96.
[13]	TANG X, DUAN X, CAI Z G. Large Language Models for Automated Literature Review: An Evaluation of Reference Generation, Abstract Writing, and Review Composition[J/OL]. 2024. https://doi.org/10.48550/arXiv.2412.13612.
[14]	DENNSTäDT F, ZINK J, PUTORA P M, et al. Title and abstract screening for literature reviews using large language models: an exploratory study in the biomedical domain[J]. Systematic Reviews, 2024, 13(1): 158.
[15]	HU Y, LI Z, ZHANG Z, et al. Hireview:Hierarchical taxonomy-driven automatic literature review generation[J/OL]. 2024. https://doi.org/10.48550/arXiv.2410.03761.
[16]	YAN X, FENG S, YUAN J, et al. Surveyforge: On the outline heuristics, memory-driven generation, and multi-dimensional evaluation for automated survey writing[J/OL]. 2025. https://doi.org/10.48550/arXiv.2503.04629.
[17]	WANG B, XU C, ZHAO X, et al. Mineru: An open-source solution for precise document content extraction[J]. 2024. https://doi.org/10.48550/arXiv.2409.18839.
[18]	LIU A, FENG B, XUE B, et al. Deepseek-v3 technical report[J/OL]. 2024. https://doi.org/10.48550/arXiv.2412.19437
[19]	TEAM Q. Qwen2 technical report[J/OL]. 2024. https://doi.org/10.48550/arXiv.2412.15115.

Topic	文章类型	Methods	GenerationRecall	Info. density	Structure	Relevance	Coverage	Claim Recall	Claim Precision
SDG	研究型	本方法	0.22	5.70	95.17	92.29	88.84	0.86	0.84
	研究型	AutoSurvey	0.13	5.56	93.33	92.00	87.00	0.44	0.43
	综述型	本方法	0.14	5.64	95.33	92.08	87.67	0.92	0.88
	综述型	AutoSurvey	0.05	5.34	95.00	92.00	87.50	0.53	0.52
CS	研究型	本方法	0.22	6.04	94.17	92.00	88.67	0.95	0.89
	研究型	AutoSurvey	0.15	6.12	95.00	92.08	88.17	0.59	0.58
	综述型	本方法	0.12	5.98	94.83	92.08	89.50	0.94	0.90
	综述型	AutoSurvey	0.09	5.87	94.17	92.00	89.50	0.60	0.59
BioMedicine	研究型	本方法	0.41	6.27	95.20	92.00	88.40	0.92	0.91
	研究型	AutoSurvey	0.47	6.24	95.80	92.70	90.00	0.46	0.45
	综述型	本方法	0.28	6.00	95.33	87.00	85.67	0.82	0.80
	综述型	AutoSurvey	0.29	5.89	94.33	92.00	85.17	0.48	0.48

领域	文章类型	Methods	Retrieve Precision	Retrieve Recall	Generation Precision	Generation Recall
SDG	研究型	本方法	0.07	0.26	0.08	0.22
	研究型	AutoSurvey	0.05	0.16	0.07	0.13
	综述型	本方法	0.18	0.15	0.15	0.14
	综述型	AutoSurvey	0.13	0.06	0.14	0.05
	研究型	本方法	0.06	0.24	0.08	0.22
CS	研究型	AutoSurvey	0.04	0.21	0.03	0.15
	综述型	本方法	0.34	0.18	0.33	0.12
	综述型	AutoSurvey	0.28	0.09	0.30	0.09
BioMedicine	研究型	本方法	0.06	0.43	0.12	0.41
	研究型	AutoSurvey	0.10	0.51	0.14	0.47
	综述型	本方法	0.33	0.42	0.47	0.28
	综述型	AutoSurvey	0.29	0.3	0.43	0.29

Methods	Retrieve Recall	Generation Recall	Claim Recall	Claim Precision
本方法	0.28	0.23	0.88	0.87
本方法w/o QE	0.23	0.19	0.87	0.85
本方法w/o TR	0.26	0.22	0.9	0.89
本方法w/o ACV	0.28	0.23	0.56	0.54

基于自适应引用验证与双阶段融合重排的自动化文献综述生成方法

An Automated Literature Review Generation Method Based on Adaptive Citation Verification and Two-Stage Fusion Rerank

RichHTML

PDF

可视化

摘要/Abstract

引用本文

使用本文

图/表 8

参考文献 19

相关文章 15

编辑推荐

Metrics

本文评价

[1]	郝文玉, 刘伟, 高思远, 梁凌, 游博. 基于大模型和RAG技术的能源产业链知识问答系统设计与实现[J]. 数据与计算发展前沿, 2026, 8(4): 214-221.
[2]	周晓芳, 鲍彦荣, 顾荣杰, 王昕. 基于多智能体协作和复杂网络理论的贸易事件知识图谱构建[J]. 数据与计算发展前沿, 2026, 8(4): 234-245.
[3]	马乐, 张然, 韩颐堃, 于诗睿, 王在田, 宁致远, 张静涵, 许萍, 李鹏江, 乔子越, 琚玮, 陈冲, 王东杰, 刘鲲鹏, 汪澎洋, 王鹏飞, 傅衍杰, 刘春江, 吕长天. 向量数据库：存储与检索技术综述[J]. 数据与计算发展前沿, 2026, 8(4): 42-71.
[4]	张梓涵,杨婉霞,赵翔,周蓓蓓,王培龙. 基于大语言模型的饮食评论知识抽取与推荐生成方法[J]. 数据与计算发展前沿, 2026, 8(3): 217-232.
[5]	王丹琳, 唐云祁. 零样本语音合成技术研究进展[J]. 数据与计算发展前沿, 2026, 8(2): 204-214.
[6]	王成,曾诗容,王楚文,蒋昌俊. 大语言模型驱动的市场交易行为模拟[J]. 数据与计算发展前沿, 2026, 8(1): 2-13.
[7]	武智晖,黄绍晗,张逸飞,齐家兴,肖智文,曾畅,栾钟治. 基于检索增强的日志问答系统[J]. 数据与计算发展前沿, 2026, 8(1): 64-76.
[8]	吴建华, 刘振宇, 曾睿, 王文瑄, 易勇, 王世轶. 基于LoRA微调大模型在网络安全等级保护测评报告质量检测中的应用研究[J]. 数据与计算发展前沿, 2025, 7(6): 111-123.
[9]	董伟, 廖佳纯, 姚思诚, 陈海粟, 阚苏南. 大语言模型赋能的面向微数据脱敏的敏感信息识别分析方法[J]. 数据与计算发展前沿, 2025, 7(6): 35-43.
[10]	刘典玉,刘青凯,肖雨阳,王杰. CAE-Bench: 面向结构力学仿真的大语言模型评估基准[J]. 数据与计算发展前沿, 2025, 7(4): 155-168.
[11]	屈志勇,王晓光,周纯葆,史源香,乔嘉伟. 面向国产超算系统的大模型训练优化方法[J]. 数据与计算发展前沿, 2025, 7(2): 120-129.
[12]	王子健, 李凯, 曹荣强, 周纯葆. 基于检查点的大模型弹性训练方法研究[J]. 数据与计算发展前沿, 2025, 7(1): 135-151.
[13]	韦一金, 陈彦清, 王秀东, 樊景超. 基于大语言模型的《中国小麦品种志》信息提取[J]. 数据与计算发展前沿, 2025, 7(1): 175-185.
[14]	马秋平, 张琪, 赵晓凡. 图表问答研究综述[J]. 数据与计算发展前沿, 2025, 7(1): 19-37.
[15]	裴炳森,李欣,蒋章涛,刘明帅. 基于大语言模型的司法文本摘要生成与评价技术研究[J]. 数据与计算发展前沿, 2024, 6(6): 62-73.