An academic literature review automatic generation method based on literature knowledge graph and multi-agent cooperation

Abstract

The present disclosure is a kind of academic literature review automatic generation method based on literature knowledge graph and multi-agent cooperation, comprising: literature knowledge graph construction; multi-role agent collaborative review content generation; content generation optimization based on citation inference chain.The present application solves the cross-literature theme abstraction and hierarchical problem by constructing the literature knowledge graph that fuses citation network and semantic entity, and explicitly modeling research context and theme evolution;Adopting multi-role agent division of labor and cooperation mechanism, simulating the "search-writing-review" process of human scientific research team, significantly improving the comprehensiveness, criticality and structural consistency of the review;Introducing citation inference chain and self-feedback reinforcement learning, with continuous optimization capability, effectively inhibiting knowledge illusion, and improving academic standard alignment degree.The present application does not need to rely on a large amount of artificial marking review data, and can end-to-end generate high-quality, traceable, standard cited academic review, significantly reducing the cognitive load of researchers in the literature sorting stage.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of artificial intelligence and scientific research automation, and in particular to a method for automatically generating academic literature reviews based on literature knowledge graphs and multi-agent collaboration. Background Technology

[0002] In the research process, literature review writing is a crucial step in constructing knowledge background, identifying research gaps, and formulating research hypotheses. However, with the explosive growth in the number of academic papers worldwide, researchers face enormous information overload and cognitive burden when writing reviews. A large number of documents are scattered across different databases, disciplines, and research topics, making manual reading, screening, and integration extremely time-consuming. The rise of interdisciplinary research further exacerbates this challenge. Against this backdrop, automated review generation technology has been proposed, aiming to assist researchers in efficiently completing reviews through semantic understanding and text generation capabilities.

[0003] While large language models have demonstrated powerful text generation capabilities, they still suffer from fundamental limitations in academic review tasks. First, they are insufficient in handling thematic abstraction and hierarchical modeling across multiple documents, often only able to provide partial summaries of a single or small number of documents, resulting in generated content with disjointed themes and loose logic. Second, due to the highly specialized nature of academic literature, models are prone to domain illusions in the absence of explicit knowledge alignment mechanisms, such as incorrect citations, fabricated conclusions, or methodological confusion, severely impacting the factual reliability and traceability of the review. Third, existing methods often employ single-agent or single-round generation models, lacking explicit control over the global structure of long texts. Finally, current systems generally lack a self-feedback optimization mechanism similar to the human "evaluation-modification-rewriting" process, failing to perform multi-dimensional evaluation and iterative correction of the generated content's quality, leading to unstable output quality that fails to meet the rigor and depth requirements of real-world research scenarios.

[0004] Therefore, it is necessary to improve one or more of the problems existing in the above-mentioned related technical solutions.

[0005] It should be noted that this section is intended to provide background or context for the technical solutions of this disclosure as set forth in the claims. The description herein does not constitute an admission that it is prior art simply because it is included in this section. Summary of the Invention

[0006] The purpose of this invention is to provide an automatic generation method for academic literature reviews based on literature knowledge graphs and multi-agent collaboration, thereby overcoming, to at least to some extent, one or more problems caused by the limitations and defects of related technologies.

[0007] This invention first provides a method for automatically generating academic literature reviews based on literature knowledge graphs and multi-agent collaboration, including: S1, Literature Knowledge Graph Construction: From the collection of academic papers related to the target research topic, a small sample semantic entity extraction is performed using a large language model to extract the structured elements of each paper; through semantic embedding and DBSCAN clustering, cross-paper entities are disambiguated and fused to construct a literature knowledge graph with the paper title as the central node and multiple types of academic elements as attribute nodes. S2, Multi-role intelligent agent collaborative review content generation: Set up and utilize multi-role intelligent agents to perform topic import, outline generation, parallel writing and global integration based on literature knowledge graph, and generate a reference list. S3, Content Generation Optimization Based on Citation Inference Chain: Define the citation inference chain as a structured inference paradigm for review generation, construct preference pair data based on the citation inference chain, construct a CRC-annotated dataset based on high-quality review samples, and use the preference alignment method to perform reinforcement learning on the agent to generate high-quality reviews.

[0008] In this invention, S1 includes the following specific steps: S1-1, Structured parsing and semantic entity extraction of academic papers: The original academic papers are structured and parsed. The text of each chapter is input into the large language model. The model is guided by a few sample prompt words to output semantic entities in a predefined JSON format. After extracting semantic entity samples, they are manually labeled. S1-2, Entity Cleaning, Disambiguation and Fusion: Performing semantic-level clustering and merging of candidate semantic entities; S1-3, Knowledge Graph Construction and Database Import: The cleaned semantic entities and the original paper titles are used as nodes to construct a directed graph, which is then imported into the database to form a literature knowledge graph.

[0009] In this invention, in S1-1, semantic entities include research fields, research questions, research methods, models, and tasks.

[0010] In this invention, S1-2 includes the following specific steps: The Qwen3-embedding-0.6B model is used to generate a 1024-dimensional semantic vector for each semantic entity; The DBSCAN clustering algorithm is used to perform density clustering on semantic vectors, grouping semantically similar entities into a unified node and filtering out invalid or low-frequency entities.

[0011] In this invention, S2 includes the following specific steps: S2-1, Topic Import and Initial Search: The searcher agent generates heuristic keywords based on the topic, retrieves multiple relevant documents from the paper database, and constructs an initial paper pool; S2-2, Outline Drafting and Optimization: The editor agent generates a first-level outline based on the document knowledge graph, each writer agent refines the sub-outlines, the editor agent evaluates the logic and coverage of the sub-outlines, and provides feedback on revisions, iteratively optimizing the sub-outlines; S2-3, Parallel writing of sub-chapter: Each writer agent dynamically retrieves literature from the paper pool according to the assigned sub-chapter title, generates a first draft with citation marks, the editor agent scores the first draft from multiple dimensions, and each writer agent revises the first draft according to the score; S2-4, Global Integration and Reference Generation: The editor agent standardizes the terminology and style of the revised draft, and the searcher agent generates a reference list based on the citation tags.

[0012] In this invention, S3 includes the following specific steps: S3-1: Extract cited paragraphs from expert reviews and annotate them using GPT-4o to transform the analytical thinking of expert reviews into a structured, uniformly formatted citation reasoning chain. S3-2, generate citation reasoning chains for the same topic using both ordinary and labeled prompts, forming preference pairs (rejection samples, acceptance samples); S3-3 uses the SimPO preference alignment algorithm to fine-tune the writer agent based on preference pairs. By minimizing the SimPO loss function, the probability of generating high-quality citation inference chains is increased and the probability of generating low-quality inference chains is reduced. Finally, a high-quality review is generated through the writer agent.

[0013] In this invention, the method further includes: The generated results are evaluated using precision, recall, and F1 score.

[0014] The technical solution provided by this invention may include the following beneficial effects: This invention presents an automatic academic literature review generation method based on a literature knowledge graph and multi-agent collaboration. Firstly, it constructs a literature knowledge graph integrating citation networks and semantic entities to explicitly model research context and topic evolution, solving the challenges of cross-document topic abstraction and hierarchy. Secondly, it employs a multi-role agent collaborative mechanism, simulating the "retrieval-writing-review" process of a human research team, significantly improving the comprehensiveness, criticality, and structural consistency of the review. Thirdly, it introduces citation reasoning chains and self-feedback reinforcement learning, enabling the system to continuously optimize, effectively suppressing knowledge illusions and improving alignment with academic norms. This invention does not rely on a large amount of manually annotated review data; it only requires the original paper PDF to generate high-quality, traceable academic reviews with standardized citations end-to-end, significantly reducing the cognitive load on researchers during the literature review stage and providing core technical support for intelligent research under the AI ​​for Science paradigm. Attached Figure Description

[0015] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure. It is obvious that the drawings described below are merely some embodiments of this disclosure, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0016] Figure 1 The flowchart illustrates the automatic generation method for academic literature reviews based on literature knowledge graphs and multi-agent collaboration in this invention. Figure 2 This diagram illustrates the framework of the automatic generation method for academic literature reviews based on literature knowledge graphs and multi-agent collaboration in this invention. Figure 3 This diagram illustrates the document knowledge graph construction process in this invention. Figure 4 This diagram illustrates the multi-agent collaboration process in this invention. Figure 5 This diagram illustrates the process of constructing and training citation inference chain preference data in this invention. Detailed Implementation

[0017] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that this disclosure will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0018] Furthermore, the accompanying drawings are merely illustrative diagrams of embodiments of this disclosure and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities.

[0019] This example implementation provides a method for automatically generating academic literature reviews based on literature knowledge graphs and multi-agent collaboration. Please refer to [link / reference]. Figure 1 and Figure 2 This method may include: S1-S3, as follows: S1, Literature Knowledge Graph Construction: From the collection of academic papers related to the target research topic, a small sample semantic entity extraction is performed using a large language model to extract the structured elements of each paper; through semantic embedding and DBSCAN clustering, cross-paper entities are disambiguated and fused to construct a literature knowledge graph with the paper title as the central node and multiple types of academic elements as attribute nodes. S2, Multi-role intelligent agent collaborative review content generation: Set up and utilize multi-role intelligent agents to perform topic import, outline generation, parallel writing and global integration based on literature knowledge graph, and generate a reference list. S3, Content Generation Optimization Based on Citation Inference Chain: Define the citation inference chain as a structured inference paradigm for review generation, construct preference pair data based on the citation inference chain, construct a CRC-annotated dataset based on high-quality review samples, and use the preference alignment method to perform reinforcement learning on the agent to generate high-quality reviews.

[0020] In this embodiment, a literature knowledge graph integrating citation networks and semantic entities is constructed to explicitly model research context and topic evolution, solving the challenges of cross-document topic abstraction and hierarchy. Secondly, a multi-role intelligent agent collaborative mechanism is adopted to simulate the "retrieval-writing-review" process of a human research team, significantly improving the comprehensiveness, criticality, and structural consistency of the review. Thirdly, citation reasoning chains and self-feedback reinforcement learning are introduced, enabling the system to continuously optimize, effectively suppressing knowledge illusions and improving alignment with academic norms. This invention does not rely on a large amount of manually annotated review data; it only requires the original paper PDF to generate high-quality, traceable academic reviews with standardized citations end-to-end, significantly reducing the cognitive load on researchers during the literature review stage and providing core technical support for intelligent research under the AI ​​for Science paradigm.

[0021] The following is for reference. Figures 1-5 The specific process of each step in the above embodiments will be explained.

[0022] Please refer to Figure 3 S1 includes the following specific steps: S1-1, Structured parsing and semantic entity extraction of academic papers: The original academic papers are structured and parsed. The text of each chapter is input into the large language model. The model is guided by a few sample prompt words to output semantic entities in a predefined JSON format. After extracting semantic entity samples, they are manually labeled. S1-2, Entity Cleaning, Disambiguation and Fusion: Semantic-level clustering and merging of candidate semantic entities; specifically, the DBSCAN clustering algorithm is used to perform density clustering on semantic vectors, merging semantically similar entities into unified nodes, and filtering out invalid or low-frequency entities; S1-3, Knowledge Graph Construction and Database Import: The cleaned semantic entities and the original paper titles are used as nodes to construct a directed graph, which is then imported into the database to form a literature knowledge graph.

[0023] The document knowledge graph construction process includes: designing structured prompts for the title, abstract, introduction, and experimental and conclusion sections of each paper to guide the large language model to output semantic entities in a predefined JSON format; using the Qwen3-embedding-0.6B model to perform high-dimensional vectorization of candidate entities, and implementing cross-paper semantic entity clustering based on cosine similarity and the DBSCAN algorithm; using paper titles, fields, keywords, questions, methods, models, tasks, datasets, and indicators as graph nodes, defining 15 types of directed relations such as "belongs to", "solves", "adopts", "proposes", and "experiments on", and storing all nodes and relations in the Neo4j graph database to form a queryable and reasonable academic knowledge graph.

[0024] In step S1-1, to support subsequent knowledge graph construction, the original PDF papers first need to be structured and parsed. This embodiment uses the GROBID tool to automatically convert each paper to TEI-XML format, extracting structured chapter content such as title, abstract, introduction, methods, experiments, and conclusions. Subsequently, the text of each chapter is input into a large language model (e.g., Qwen3-7B), and the model is guided by few-sample prompts to output semantic entities in a predefined JSON format. Specifically, for each paper: Using title + abstract + introduction + conclusion as input, extract research field, research problem, research method, model, and task. Using the title and experiment section as input, extract the dataset and evaluation metrics.

[0025] To control the quality of extracted semantic entities, after the initial entity extraction, over a hundred erroneous samples (including invalid data, format errors, vague or ambiguous descriptions, etc.) were manually labeled, and the XLNet-based classifier was fine-tuned as an automatic error detector. Semantic entities classified as invalid were directly removed, and entities with format or semantic errors were resubmitted to the large language model for secondary extraction and correction.

[0026] In S1-2, due to differences in how different papers describe the same concept (e.g., "reasoning compression" vs. "reasoning acceleration"), directly using the original strings would lead to redundancy in the knowledge graph. Therefore, this embodiment performs semantic-level clustering and merging of candidate semantic entities. First, the Qwen3-embedding-0.6B model is used to generate a 1024-dimensional semantic vector for each semantic entity; then, the DBSCAN clustering algorithm (parameters: ε=0.18, MinPts=2) is used to perform density clustering on the high-dimensional vectors, merging semantically similar entities into unified nodes. After clustering, invalid or low-frequency entities (such as "a certain method" or "a model") are filtered out. This process significantly improves the compactness of the knowledge graph and its consistency across papers.

[0027] In S1-3, the cleaned semantic entities and the original paper titles are used as nodes. A directed graph is constructed based on the relationship types shown in Table 1 (such as "Title —solves→ Problem" and "Method —works on→ Task") and imported into the Neo4j graph database to form a literature knowledge graph.

[0028] Table 1. Relationship Types of Semantic Entities

[0029] This step extracts structured elements such as research questions, methods, models, tasks, and datasets from a collection of academic papers related to the target research topic using a large language model. Then, it disambiguates and fuses cross-paper entities through semantic embedding and DBSCAN clustering, constructing a literature knowledge graph with the paper title as the central node and multiple academic elements as attribute nodes.

[0030] Please refer to Figure 4 S2 includes the following specific steps: S2-1, Topic Import and Initial Search: The searcher agent generates heuristic keywords based on the topic, retrieves multiple relevant documents from the paper database, and constructs an initial paper pool; S2-2, Outline Drafting and Optimization: The editor agent generates a first-level outline based on the document knowledge graph, each writer agent refines the sub-outlines, the editor agent evaluates the logic and coverage of the sub-outlines, and provides feedback on revisions, iteratively optimizing the sub-outlines; S2-3, Parallel writing of sub-chapter: Each writer agent dynamically retrieves literature from the paper pool according to the assigned sub-chapter title, generates a first draft with citation marks, the editor agent scores the first draft from multiple dimensions, and each writer agent revises the first draft according to the score; S2-4, Global Integration and Reference Generation: The editor agent standardizes the terminology and style of the revised draft, and the searcher agent generates a reference list based on the citation tags.

[0031] This embodiment sets up three core intelligent agents: Fetcher, responsible for multi-stage document retrieval and allocation; Writer, responsible for sub-chapter content generation and revision; and Editor, responsible for outline formulation, quality assessment, and overall integration.

[0032] Each agent is implemented based on the Qwen2.5-14B-Instruct large language model and is equipped with dedicated prompt word templates and tool interfaces (such as Faiss vector retrieval and citation verification modules). Then, the four-stage collaboration process of S2-1 to S2-4 above is executed.

[0033] Specifically, first, Fetcher generates heuristic keywords (such as "pruning" and "verification mechanism") based on the topic and performs multiple rounds of academic literature searches, such as retrieving hundreds of relevant articles from a paper database to build an initial paper pool. Second, Editor generates a first-level review outline based on a literature knowledge graph; each Writer refines the sub-outlines for each chapter; Editor evaluates the logic and coverage, provides feedback on revisions, iterates and optimizes the content, and forms the final structured outline. Then, each Writer dynamically retrieves literature from the paper pool based on the assigned sub-chapter titles and generates a first draft with standardized citation marks; Editor scores the draft based on dimensions such as logic, completeness, and citation accuracy, and Writers revise based on feedback. Finally, Editor reads through the entire text, unifies terminology and style, and ensures there are no logical conflicts; Fetcher generates a reference list that conforms to academic standards based on the citation marks.

[0034] In the multi-role intelligent agent collaborative process, the retrieval agent performs a secondary retrieval during the sub-chapter writing stage to ensure sufficient information in the local content; the editor adopts a multi-dimensional scoring mechanism to quantitatively evaluate the structural rationality, critical analysis, and citation norms; and the writer conducts multiple rounds of iterative revisions based on feedback.

[0035] Please refer to Figure 5 S3 includes the following specific steps: S3-1: Extract cited paragraphs from expert reviews and annotate them using GPT-4o to transform the analytical thinking of expert reviews into a structured, uniformly formatted citation reasoning chain. S3-2, generate citation reasoning chains for the same topic using both ordinary and labeled prompts, forming preference pairs (rejection samples, acceptance samples); S3-3 uses the SimPO preference alignment algorithm to fine-tune the writer agent based on preference pairs. By minimizing the SimPO loss function, the probability of generating high-quality citation inference chains is increased and the probability of generating low-quality inference chains is reduced. Finally, a high-quality review is generated through the writer agent.

[0036] Specifically, firstly, the data is constructed according to the citation inference chain paradigm: citation paragraphs are extracted from expert reviews and annotated using GPT-4o according to the four-stage template. Contribution: Analyzing the contribution of individual articles; Connect: Building cross-document associations; Rebuild: Reconstruct the expression; Integrate: to integrate into a coherent paragraph.

[0037] The annotation process transforms the analytical approach of expert reviews into a structured, standardized citation reasoning chain.

[0038] Secondly, citation reasoning chains are generated for the same topic using both ordinary and labeled prompts, forming preference pairs (rejection samples, acceptance samples).

[0039] Then, the SimPO algorithm was used to fine-tune the Writer model, with the optimization goal of maximizing the generation probability of accepted samples, thereby improving logical coherence and academic standardization.

[0040] The implicit reward function of SimPO can be expressed as:

[0041] in, Indicates the SimPO reward value; The scaling factor representing the reward difference is used to control the intensity of preference optimization. This represents the log probability that the model generates the complete response y given input x; This indicates the response length and is used to eliminate the length bias of long sequences in probability accumulation.

[0042] High-quality response to the same input x and low-quality response The preference probability can be expressed as:

[0043] in, This indicates the selected high-quality responses, also known as chosen samples. This indicates a low-quality response that was rejected, also known as a rejected sample. " indicates a preference relationship; Represents the Sigmoid function; r(x, ) and r(x, These represent the reward values ​​for high-quality and low-quality responses, respectively. This represents the target reward interval, used to enhance the model's ability to distinguish between good and bad responses.

[0044] The final loss function of SimPO can be expressed as:

[0045] in, The SimPO training loss is represented by E; E represents the expected value of the samples in the preference dataset. Represents a preference dataset; (x, , ) represents a training sample triple in the preference dataset; and These represent the sequence lengths of high-quality and low-quality responses, respectively.

[0046] By minimizing The Writer model can increase the probability of generating high-quality citation inference chains and reduce the probability of generating low-quality inference chains.

[0047] In the early generation paradigm based on citation inference chains, the model's preferences were trained by constructing preference pairs of "low-quality citation inference chains (rejected samples)" and "high-quality citation inference chains (accepted samples)". The writer agent samples the citation inference chains during the generation process and updates the strategy based on a large model preference alignment method (such as SimPO) to avoid generating content with factual errors and logical breaks.

[0048] Based on the above embodiments, the generation effect can also be verified through experiments.

[0049] To illustrate the practical effectiveness of the technical solution presented in this application, this embodiment verifies the three core technical aspects: literature knowledge graph retrieval and review outline planning, multi-role intelligent agent collaborative text generation, and citation reasoning chain-driven self-feedback optimization. Experimental data are derived from a collection of natural language processing and computer science papers in the applicant's research. Among the evaluation metrics, precision, recall, and F1 score are used to: measure the retrieval capability of the literature knowledge graph; logical progression, completeness, and redundancy are used to measure the review outline structure; and coverage, relevance, structure, relationality, and citation quality are used to measure the academic quality of the generated review. The following experiments are for illustrative purposes only and do not constitute a limitation on the scope of protection of this application.

[0050] First, we conducted an experiment on the effectiveness of literature knowledge graph retrieval.

[0051] This experiment was conducted to verify the effectiveness of knowledge graph construction and entity fusion in S1 in enhancing the ability to retrieve academic documents under multiple constraints. The experiment used the NLPPaper paper query dataset and a hybrid test set of QASPER and NLP-paper-to-QA-generation, and selected knowledge graph-enhanced retrieval methods as a comparison. The experimental results are shown in Table 2.

[0052] Table 2. Results of the map retrieval experiment

[0053] As shown in Table 2, on the NLPPaper dataset, the method described in this application improves the precision, recall, and F1 score by 2.6 percentage points, 6.7 percentage points, and 4.6 percentage points, respectively, compared to the knowledge graph augmentation retrieval method; and on the mixed test set, the improvements are 5.4 percentage points, 2.1 percentage points, and 2.5 percentage points, respectively. These results demonstrate that by sharing semantic entities and relational constraints, this application can more stably recall sets of papers that meet complex academic constraints.

[0054] Secondly, an experiment was conducted to assess the effectiveness of the framework generation.

[0055] This experiment was conducted to verify whether knowledge graph retrieval results could further support the overall structure planning of the review. The experiment covered 60 natural language processing topics, including 30 broad topics and 30 detailed topics. The system only input topic text, without additional manual outlines, and was compared with a single-prompt large-model approach that did not incorporate a literature knowledge graph. The experimental results are shown in Table 3.

[0056] Table 3 Results of the vascularization experiment

[0057] As shown in Table 3, the method presented in this application is significantly superior to the single-hint large model scheme in terms of logical progression and completeness, while maintaining a low redundancy rate of 0.58%. These results demonstrate that literature knowledge graphs can not only be used for retrieval but also provide stable support for the hierarchical organization, topic coverage, and structural planning of review sections.

[0058] Then, we conducted an experiment on collaborative text generation by multiple intelligent agents.

[0059] This experiment aims to verify whether the collaborative mechanism among the three agents (retrieval agent, writer, and editor) in S2 can improve the quality of review articles. The experiment covers 60 topics in computer science, including 30 macro-level topics and 30 fine-grained topics. The comparison is between pure LLM and RAG generation methods, i.e., schemes that complete review articles solely through retrieval enhancement and single-model generation. The results of the collaborative experiment are shown in Table 4.

[0060] Table 4 Results of the Collaborative Experiment

[0061] As shown in Table 4, the average content quality score of the method presented in this application reached 7.55, higher than that of the pure LLM and RAG schemes; the structural index reached 8.51, an improvement of 2.06 points compared to the pure LLM and RAG schemes. Meanwhile, the citation quality of the method presented in this application increased from 78.31% to 81.48%, and the generation speed reached 15.05 articles / hour. These results demonstrate that multi-role division of labor and parallel writing processes can improve generation efficiency while enhancing structural control and citation reliability.

[0062] Finally, a citation inference chain experiment was conducted.

[0063] This experiment was conducted to verify whether the citation inference chain and SimPO preference alignment training in S3 can further enhance the ability to integrate cross-document relationships and reduce the cost of multiple revisions. The experiment was carried out on another set of 60 topics to avoid overlap with the aforementioned experimental topics. The comparison objects were a multi-agent collaborative scheme without CRC training and the complete scheme of this application. The experimental results are shown in Table 5.

[0064] Table 5 Experimental Results

[0065] As shown in Table 5, while maintaining relatively stable coverage and structure, the complete solution of this application improved the relationality index from 7.02 to 7.44, the average content quality from 7.55 to 7.69, and the citation quality from 81.48% to 83.12%. The weighted average iteration count was used to characterize the overall revision rounds for different generation tasks. The calculation first divided the number of iterations into four intervals: 1-30, 31-70, 71-100, and over 100, with 15, 50, 85, and 115 as representative values ​​for each interval, respectively. Then, a weighted average was calculated based on the frequency of occurrence in each interval. Simultaneously, the weighted average iteration count decreased by 35.16%, and the average token consumption decreased by 51.19%. These results indicate that the citation reasoning chain can distill editorial feedback into trainable intermediate reasoning states, enabling the Writer model to better establish the connection between "citation—argument—overall context," thereby reducing the overhead of repeated revisions.

[0066] In summary, the experimental results show that the method proposed in this application has stable technical effects in terms of literature knowledge organization, review structure planning, collaborative generation of main text, citation quality control and optimization. It can provide effective improvements to existing automatic review generation methods, such as unstable structure, insufficient cross-document association, weak citation support and high cost of multiple revisions.

[0067] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

[0068] It should be noted that the installation of image acquisition and personal identification equipment in public places involved in this application is necessary for maintaining public safety, complies with relevant national regulations, and is accompanied by prominent warning signs. The collected personal images and identification information can only be used for the purpose of maintaining public safety and not for other purposes; or the images, personal identification data, etc. in this application are all legally and compliantly obtained or collected with the individual's separate consent.

Claims

1. A method for automatically generating academic literature reviews based on literature knowledge graphs and multi-agent collaboration, characterized in that, include: S1, Literature Knowledge Graph Construction: From the collection of academic papers related to the target research topic, a large language model is used to extract semantic entities from a small number of samples and extract the structured elements of each paper. By using semantic embedding and DBSCAN clustering to disambiguate and fuse cross-paper entities, a literature knowledge graph is constructed with the paper title as the central node and multiple academic elements as attribute nodes. S2, Multi-role intelligent agent collaborative review content generation: Set up and utilize multi-role intelligent agents to perform topic import, outline generation, parallel writing and global integration based on literature knowledge graph, and generate a reference list. S3, Content Generation Optimization Based on Citation Inference Chain: Define the citation inference chain as a structured inference paradigm for review generation, construct preference pair data based on the citation inference chain, construct a CRC-annotated dataset based on high-quality review samples, and use the preference alignment method to perform reinforcement learning on the agent to generate high-quality reviews.

2. The method for automatically generating academic literature reviews based on literature knowledge graphs and multi-agent collaboration as described in claim 1, characterized in that, S1 includes the following specific steps: S1-1, Structured parsing and semantic entity extraction of academic papers: The original academic papers are structured and parsed. The text of each chapter is input into the large language model. The model is guided by a few sample prompt words to output semantic entities in a predefined JSON format. After extracting semantic entity samples, they are manually labeled. S1-2, Entity Cleaning, Disambiguation and Fusion: Performing semantic-level clustering and merging of candidate semantic entities; S1-3, Knowledge Graph Construction and Database Import: The cleaned semantic entities and the original paper titles are used as nodes to construct a directed graph, which is then imported into the database to form a literature knowledge graph.

3. The method for automatically generating academic literature reviews based on literature knowledge graphs and multi-agent collaboration as described in claim 2, characterized in that, In S1-1, semantic entities include research fields, research questions, research methods, models, and tasks.

4. The method for automatically generating academic literature reviews based on literature knowledge graphs and multi-agent collaboration as described in claim 2, characterized in that, S1-2 includes the following specific steps: The Qwen3-embedding-0.6B model is used to generate a 1024-dimensional semantic vector for each semantic entity; The DBSCAN clustering algorithm is used to perform density clustering on semantic vectors, grouping semantically similar entities into a unified node and filtering out invalid or low-frequency entities.

5. The method for automatically generating academic literature reviews based on literature knowledge graphs and multi-agent collaboration according to claim 1, characterized in that, S2 includes the following specific steps: S2-1, Topic Import and Initial Search: The searcher agent generates heuristic keywords based on the topic, retrieves multiple relevant documents from the paper database, and constructs an initial paper pool; S2-2, Outline Drafting and Optimization: The editor agent generates a first-level outline based on the document knowledge graph, each writer agent refines the sub-outlines, the editor agent evaluates the logic and coverage of the sub-outlines, and provides feedback on revisions, iteratively optimizing the sub-outlines; S2-3, Parallel writing of sub-chapter: Each writer agent dynamically retrieves literature from the paper pool according to the assigned sub-chapter title, generates a first draft with citation marks, the editor agent scores the first draft from multiple dimensions, and each writer agent revises the first draft according to the score; S2-4, Global Integration and Reference Generation: The editor agent standardizes the terminology and style of the revised draft, and the searcher agent generates a reference list based on the citation tags.

6. The method for automatically generating academic literature reviews based on literature knowledge graphs and multi-agent collaboration according to claim 1, characterized in that, S3 includes the following specific steps: S3-1: Extract cited paragraphs from expert reviews and annotate them using GPT-4o to transform the analytical thinking of expert reviews into a structured, uniformly formatted citation reasoning chain. S3-2, generate citation reasoning chains for the same topic using both ordinary and labeled prompts, forming preference pairs (rejection samples, acceptance samples); S3-3 uses the SimPO preference alignment algorithm to fine-tune the writer agent based on preference pairs. By minimizing the SimPO loss function, the probability of generating high-quality citation inference chains is increased and the probability of generating low-quality inference chains is reduced. Finally, a high-quality review is generated through the writer agent.

7. The method for automatically generating academic literature reviews based on literature knowledge graphs and multi-agent collaboration according to any one of claims 1 to 6, characterized in that, The method further includes: The generated results are evaluated using precision, recall, and F1 score.

Images