An automatic method for generating chemistry review based on citation network

Abstract

The application provides an automatic academic review generation method based on a citation network, and relates to the technical field of artificial intelligence. First, seed literatures are automatically retrieved around a target research theme, a literature set is expanded based on co-citation relationships between the seed literatures, and a representative candidate literature set is obtained; subsequently, a mixed citation network is constructed by fusing real citation relationships, literature semantic similarity relationships and co-citation relationships, and explicit citation structures and implicit semantic correlations between literatures are described; the correlation of literatures and the research theme, the structural importance in the citation network and the publication time factor are comprehensively considered, the candidate literatures are filtered, and a high-quality reference literature set is obtained; a community detection algorithm is used to divide the literatures, high-contribution-degree communities and representative literatures in the high-contribution-degree communities are screened according to normalized community contribution degrees, and a hierarchical review outline is generated on this basis; finally, corresponding review main content is automatically generated under the constraints of the outline structure and the representative literatures.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an automated method for generating academic reviews based on citation networks. Background Technology

[0002] With the rapid development of information technology, especially the continuous evolution of cutting-edge fields such as artificial intelligence and machine learning, related academic research results have shown an explosive growth trend. A large number of papers have emerged, greatly promoting scientific and technological progress. However, at the same time, this has also brought significant challenges to researchers in acquiring, understanding, and systematically summarizing the overall development trajectory of a particular research field. Against this backdrop, review articles, as an important academic form for systematically summarizing and generalizing specific research fields, play a crucial role in summarizing the current research status, outlining technological evolution paths, analyzing key issues, and projecting future development directions. They have become an indispensable foundational resource in scientific research activities.

[0003] However, traditional review writing primarily relies on field experts manually searching, screening, reading, and summarizing literature. This process is typically time-consuming, labor-intensive, and demands a high level of academic expertise and domain knowledge from the authors. Given the rapid increase in the number of publications and the increasingly specialized nature of research directions, the cost of producing high-quality reviews entirely manually is constantly rising, making it difficult to meet the demands for improved research efficiency. Therefore, how to leverage automation to assist or even achieve review generation has gradually become a research focus in academia.

[0004] In recent years, large-scale language models have demonstrated strong capabilities in natural language understanding and generation tasks, providing new technical pathways for automated review article generation. Existing research typically utilizes large language models to summarize retrieved paper abstracts or main texts, and then generates relatively complete review texts through multi-stage generation strategies or retrieval enhancement mechanisms. Some methods further introduce multi-agent collaboration, heuristic outline generation, or memory mechanisms to improve the coherence and structural rationality of the generated content. These technologies, to some extent, alleviate the burden of manually writing reviews and promote the development of automated review article generation technology.

[0005] Existing automated review generation methods still have several shortcomings. First, most methods treat papers as independent text samples, primarily relying on semantic similarity or keyword matching for processing. This fails to effectively model citation relationships between papers, making it difficult to reflect knowledge inheritance and methodological evolution among research findings, thus affecting the review's portrayal of the field's development trajectory. Second, existing literature retrieval and screening strategies often lack systematic structural constraints, making it difficult to ensure both topic relevance and the representativeness and academic impact of the literature, easily introducing redundant literature or omitting key foundational work. Third, in the review outline and main text generation stages, existing methods have limited ability to constrain the content structure, often resulting in unclear chapter hierarchy, overlapping and mixed themes, and insufficient logical connections, failing to meet the structural and systematic requirements of high-quality academic reviews. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide an automated academic review generation method based on citation networks to address the shortcomings of the prior art. This method can fully model the structural relationships between documents, take into account semantic relevance and academic influence, and use structured information to guide the organization and content generation of the review.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0008] On the one hand, this invention provides an automated academic review generation method based on citation networks, comprising the following steps:

[0009] Step 1: Retrieve seed literature. Based on the research topic, keywords or search conditions entered by the user, use the academic search engine to perform literature retrieval operations, obtain an initial set of literatures semantically related to the target research field, and perform preliminary filtering based on similarity to serve as the seed literature set.

[0010] The seed literature collection is used to provide preliminary semantic coverage of the target research topic, and its size and content are dynamically adjusted according to the search criteria.

[0011] Step 2: Based on the seed literature set obtained in Step 1, analyze the reference list of each seed literature and count the number of times different literatures are cited by multiple seed literatures; for literatures whose co-citation counts meet the preset co-citation threshold condition, add them to the candidate literature set; the preset co-citation threshold is set according to the number of seed literatures or the total size of literatures, and is used to control the quantity and quality of expanded literatures; through co-citation expansion processing, basic or classic literatures that are repeatedly cited in the research field are introduced;

[0012] Step 3: Using the documents in the document set obtained in Step 2 as network nodes, construct a basic citation network structure based on the real citation relationships between documents; on this basis, further construct supplementary association edges according to the semantic similarity relationship of document content and the co-citation relationship between documents, and merge different types of relationships to form a hybrid citation network;

[0013] Step 4: Based on the hybrid citation network obtained in Step 3, perform comprehensive filtering and sorting on the candidate literature set; evaluate the relevance of each literature to the target research topic, the importance of the literature in the citation network structure, and the publication time factor of the literature, and calculate the comprehensive score of the literature by comprehensively calculating the factors; according to the results of the comprehensive score, screen the candidate literature, retain the literature that meets the preset conditions, and form the final high-quality reference set.

[0014] Step 5: Based on the final high-quality reference set obtained in Step 4, the citation network is reconstructed using the hybrid citation network construction method described in Step 3. Then, a community detection algorithm is used to divide the literature into communities, grouping literature with similar or closely related research directions into the same community. Based on the community division, the contribution of each community to the overall literature set is further evaluated, and communities with high contributions are selected as core research directions. For each core community, a representative set of literature is selected, and a corresponding first-level chapter outline is generated using a large language model. Simultaneously, fine-grained clustering of literature is performed within each community, and a corresponding second-level chapter outline is generated using a large language model, thus forming a hierarchical and thematically clear overall review outline structure.

[0015] Step 6: Under the constraints of the outline structure obtained in Step 5, and combined with the literature collection corresponding to each chapter, extract the titles and abstracts of the literature as input, and use the large language model to generate the main body of the review text chapter by chapter; during the generation of the main body, ensure that the generated content is consistent with the theme of the corresponding chapter and maintain the logical connection between each chapter, and finally form a review text with complete structure and coherent content.

[0016] Furthermore, the preliminary filtering based on similarity specifically involves:

[0017] For each document in the initial document set, its title and abstract text are extracted, and the text is vectorized using a pre-defined text embedding model to obtain a semantic vector representation. At the same time, the query text is vectorized in the same way to obtain a query vector representation. Based on the above vector representation, cosine similarity is used as a semantic relevance measure to calculate the semantic relevance between each document and the query topic. The documents in the initial document set are sorted according to the semantic relevance score, and the top A documents in the sorting results are selected as the seed document set according to a pre-defined ranking threshold A.

[0018] Furthermore, the method for constructing the basic citation network structure in step 3 is as follows:

[0019] Let the set of candidate documents obtained through step 2 be:

[0020] ;

[0021] in, Indicates the first in the candidate document set This document, Indicates the total number of candidate documents;

[0022] For each document Extract its reference list, denoted as Based on the reference list, a set of true citation relationships is constructed, defined as follows:

[0023] ;

[0024] Among them, directed edges ( ) indicates literature For the literature Citation relationships are used to characterize the knowledge inheritance path and the direction of research dissemination among documents.

[0025] Furthermore, in step 3, the specific method for forming the hybrid citation network is as follows:

[0026] First, construct semantically enhanced relations; then, for the candidate document set... Each document in The title and abstract are vectorized using a pre-defined text embedding model to obtain the semantic representation vector of the document:

[0027] ;

[0028] Based on the semantic representation vectors mentioned above, cosine similarity is used to calculate the similarity between any two documents. and Semantic similarity between them:

[0029] ;

[0030] For each document Only retain the top-ranked results based on semantic similarity. And the similarity is greater than the preset threshold. The literature is used as a candidate neighbor; wherein, the parameter It is adaptively determined based on the size of the literature, and its definition is as follows:

[0031] ;

[0032] in, is a proportionality coefficient used to control the relative retention ratio of semantic edges in the overall candidate space, thereby ensuring that the sparsity of the graph is consistent under document sets of different sizes; This is a smoothing compensation term used to provide necessary smoothing compensation when the document size is small, to avoid excessively sparse graph structure due to insufficient candidate quantity; This is the strength parameter used to adjust the smoothing compensation term; and Top- The lower and upper bounds are used to prevent degenerate or overly dense graph structures in extremely small or extremely large-scale scenarios.

[0033] When the literature With literature When a genuine citation relationship already exists between two sources, the corresponding semantically enhanced relationship will not be constructed again.

[0034] Secondly, construct co-citation enhancement relationships; assuming the literature With literature The reference sets are respectively and The co-citation count between the two is defined as:

[0035] ;

[0036] Based on this, the co-citation similarity between documents is defined as:

[0037] ;

[0038] When the literature and If a genuine citation relationship already exists between two sources, no further co-citation enhancement relationship will be constructed.

[0039] Finally, semantic enhancement relations and co-citation enhancement relations are fused to construct hybrid enhancement edges; for any two documents and Its hybrid enhancement edge is defined as:

[0040] ;

[0041] in, , which are weighting coefficients used to adjust the relative contributions of semantic similarity and co-citation similarity in the mixed relation; when one type of enhancement relation does not exist, its corresponding weight term takes a value of zero;

[0042] Based on the hybrid enhanced edges, construct an enhanced edge set. This is then merged with the set of real citation relationships to obtain a hybrid citation network:

[0043] ;

[0044] in, It is the set of real citation relationships constructed based on real citation relationships in the previous steps. It is the set of candidate documents obtained in step 2;

[0045] In this context, network nodes represent documents, real citation edges are used to characterize explicit academic citation relationships, and hybrid augmented edges are used to supplement the potential semantic and research-based associations between documents.

[0046] Furthermore, the specific method for the integrated filtering and sorting process in step 4 is as follows:

[0047] Step 4.1: Calculation of the relevance of the document topic;

[0048] Let the hybrid citation network constructed in step 3 be:

[0049] ;

[0050] Where E represents the set of citation and enhancement relationships between documents;

[0051] For each document Based on its title and abstract text, a semantic vector representation of the document is obtained using a pre-defined text embedding model. ;

[0052] At the same time, the query text for the target research topic The query vector is obtained by vectorization. ;

[0053] Based on the above vector representation, calculate the literature The semantic relevance between a research topic and the target research topic is defined as:

[0054] ;

[0055] in, This is used to measure the degree of match between the content of the literature and the target research topic;

[0056] Step 4.2: Calculation of the importance of document structure.

[0057] Based on the hybrid citation network The PageRank algorithm is used to calculate the importance of document nodes in order to evaluate the academic influence of documents in the overall citation network structure.

[0058] Document PageRank score in citation networks is This represents structural importance. Its calculation method is based on iteratively obtaining the connections between nodes in the network, and the result is normalized to ensure its value falls within a certain range. Structural importance Used to reflect the centrality of a document on the network and its citation impact;

[0059] Step 4.3: Calculation of the literature time compensation factor;

[0060] To balance classic and recent literature in the literature filtering process, a time compensation factor is introduced to give appropriate weight to recently published literature without excessively weakening the influence of classic literature.

[0061] Document The time difference between the publication time and the current time is Then its time compensation factor is defined as:

[0062] ;

[0063] in, This is a preset time decay parameter used to control the degree of influence of time factors on the overall score;

[0064] Step 4.4: Comprehensive scoring and literature filtering;

[0065] Based on the above calculation results, for each document Calculate its overall score, defined as follows:

[0066] ;

[0067] in, This is a relevance adjustment parameter used to enhance the impact of topic relevance on the final score; This is a time compensation weight parameter used to adjust the contribution ratio of time factors in the overall score;

[0068] Step 4.5: Based on the comprehensive score... The candidate literature set is sorted, and a filtering strategy based on percentiles or number thresholds is used to retain literature with scores in the top 80% percentile, forming the final high-quality reference set. , .

[0069] Furthermore, the specific method of step 5 is as follows:

[0070] Step 5.1: Construct the citation network;

[0071] Let the set of high-quality documents obtained through step 4 be:

[0072] ;

[0073] in, This indicates the number of documents used to generate the review;

[0074] Using the hybrid citation network construction method from step 3, based on Construct a new citation network ;

[0075] Step 5.2: Divide communities based on network structure;

[0076] For the new citation network The Leiden community discovery algorithm is used to perform community partitioning, dividing document nodes into several non-overlapping communities, denoted as:

[0077] ;

[0078] in, Indicates the first A literature community Indicates the number of communities;

[0079] Step 5.3: Calculate the contribution of each literature community to characterize the overall contribution of literature within that community to the target research field under the combined effects of topic relevance, structural importance, and time factors;

[0080] Based on the literature obtained in step 4 Overall rating Then the first The contribution of each community is defined as follows:

[0081] ;

[0082] The community contribution score is normalized to obtain the normalized contribution score:

[0083] ;

[0084] Step 5.4: Community screening and generation of first-level headings for the methodology section. After obtaining the normalized contribution of each literature community, an adaptive screening strategy based on the cumulative contribution threshold is used to determine the research sub-directions covered by the review.

[0085] First, all communities are sorted from highest to lowest according to their normalized community contribution, forming an ordered community sequence. Then, the normalized contribution scores of each community are accumulated sequentially, and the smallest set of communities that meets the following conditions is selected:

[0086] ;

[0087] in, This represents the selected set of communities. A preset community contribution threshold is used to control the completeness of the research directions covered by the review at the macro level;

[0088] For each selected community Furthermore, based on the contribution of the literature at the literature level, the literature within the community is screened to determine the representative set of literature used to generate the first-level headings;

[0089] For the community Each document in Based on the comprehensive score defined in step 4 The relative contribution of each document within the community is calculated, and the documents are sorted from highest to lowest relative contribution. Then, using the same cumulative threshold strategy as the community selection, the smallest set of documents meeting the following criteria is selected:

[0090] ;

[0091] in, Indicates community A collection of representative literature, The preset threshold for literature contribution;

[0092] The representative literature collection Used to characterize the main research content and technological context of the corresponding research direction of the community;

[0093] Based on representative literature collections The title, abstract, and keywords of Chinese literature are automatically used to generate corresponding first-level headings, which summarize the overall research theme of the research sub-direction represented by the community.

[0094] Step 5.5: Semantic Clustering within the Community and Generation of Second-Level Headings for the Methodology Section; After determining the first-level headings, generate second-level headings for each literature community. The internal literature further performs fine-grained clustering based on semantic similarity, dividing the community into several semantically consistent sub-clusters, denoted as:

[0095] ;

[0096] in, Indicates community The first in Subclusters;

[0097] Based on the semantic features of the documents in each sub-cluster, corresponding secondary headings are automatically generated to describe the specific research questions, technical routes, or methodological categories in that research sub-direction.

[0098] The outline corresponding to the method chapters is obtained through the above method, and its output is expressed as follows:

[0099] ;

[0100] in, The outline corresponding to the chapter on representation methods is at page number 1. One first-level heading, This indicates the corresponding second-level heading. Indicates the number of communities selected;

[0101] Step 5.6: Generate the outlines for the remaining chapters;

[0102] For the outlines of non-method sections, the same method as that used for generating method sections is employed. The difference lies in that the literature information used for generating non-method sections is a collection of representative papers from each community. Finally, the outlines of method sections and non-method sections are merged to obtain a complete review outline.

[0103] Furthermore, the specific method for step 6 is as follows:

[0104] Step 6.1: Input structure for generating the main text. Let the outline structure obtained in Step 5 be as follows:

[0105] ;

[0106] in, This indicates the first section of the review outline. One first-level heading, This indicates the corresponding second-level heading;

[0107] For each first-level heading In step 5, the corresponding representative literature set has been determined, denoted as:

[0108] ;

[0109] For each second-level heading From representative literature collections Further, a subset of documents most relevant to the semantics of the second-level heading is selected as the reference input for that subsection;

[0110] Step 6.2: Segmentation generation strategy based on chapter structure. To ensure the structural consistency and logical coherence of the generated text, a hierarchical and segmented text generation strategy is adopted.

[0111] Following the hierarchical structure of the review outline, proceed sequentially to address each first-level heading. and its subordinate second-level headings Execute the text generation process;

[0112] When generating a text paragraph corresponding to a certain second-level heading, the second-level heading is used as a subject constraint, and the corresponding subset of literature is used as the content basis, so that the generated text revolves around a clear research question or method category.

[0113] Step 6.3: Generate text based on the content constraints of representative literature. The main text is generated using a conditional text generation method based on a large language model. During the generation process, the key information of the representative literature corresponding to each chapter is used as the input of the generation conditions, so that the generated text is semantically constrained by the content of the literature. For a certain chapter, the title and abstract information of each literature in its representative literature set are integrated and used as reference context input into the large language model, so that the generated text can summarize the main research content, technical route and development context of the research work involved in the chapter.

[0114] Using the above-mentioned text generation method based on structural and literal constraints, a complete academic review text is finally output. Its content is organized according to the preset first-level and second-level headings, and the content of each chapter is consistent in terms of theme, logic and expression style.

[0115] On the other hand, this application proposes an electronic device comprising: one or more processors, and a memory for storing instructions that, when executed by the one or more processors, cause the one or more processors to perform the aforementioned citation network-based automated academic review generation method.

[0116] Thirdly, this application proposes a computer-readable storage medium storing executable instructions that, when executed, cause a processor to perform the described automated academic review generation method based on citation networks.

[0117] Fourthly, this application proposes a computer program product, including a computer program or instructions that, when executed by a processor, implement the aforementioned method for generating automated academic reviews based on citation networks.

[0118] The beneficial effects of adopting the above technical solution are as follows: The automated academic review generation method based on citation networks provided by this invention introduces a co-citation expansion mechanism in the literature acquisition stage to supplement basic and representative literature that may be missed during topic retrieval; in the literature modeling stage, it constructs a hybrid citation network by integrating real citation relationships, semantic similarity relationships, and co-citation relationships, thereby simultaneously characterizing the explicit citation structure and implicit semantic connections between documents; in the literature screening stage, it comprehensively considers the relevance of documents to the target topic, their structural importance in the citation network, and time factors, and designs a filtering mechanism based on relevance-importance-time compensation, effectively improving the quality of the literature set used for review generation. This invention can automatically generate academic review texts with clear structure and reliable content without manual intervention, providing an efficient, stable, and scalable technical solution for large-scale literature analysis and academic knowledge organization, with good practical application value and promising prospects for promotion. Attached Figure Description

[0119] Figure 1 A flowchart illustrating the automated academic review generation method based on citation networks provided in the first embodiment of the present invention;

[0120] Figure 2 A flowchart for generating a first-level chapter outline provided in the first embodiment of the present invention;

[0121] Figure 3 The flowchart for generating a second-level chapter outline is provided for the first embodiment of the present invention. Detailed Implementation

[0122] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0123] Example 1:

[0124] An automated academic review generation method based on citation networks, such as Figure 1 As shown, the method of this embodiment is described below.

[0125] Step 1: Retrieve seed literature. Based on the research topic, keywords, or search criteria entered by the user, use an academic search engine to perform a literature search operation, obtain an initial set of literatures semantically related to the target research field, and perform preliminary filtering based on similarity to form a seed literature set.

[0126] This embodiment takes receiving research topic information input by a user as an example. The research topic information may include one or more keywords, phrases, or a topic description in natural language form, used to characterize the research area of ​​interest to the target review. Let the research topic be represented as query text. .

[0127] Based on the query text An academic search engine is used to perform a literature search operation to obtain a set of documents that are semantically related to the query text. The initial set of documents obtained through the search operation is denoted as:

[0128] ;

[0129] in, Represents the first in the initial document set This is a search of literature. This indicates the total number of documents retrieved.

[0130] For each document The title and abstract text are extracted, and the text is vectorized using a pre-defined text embedding model. Let the document... The semantic vector representation is as follows:

[0131] ;

[0132] in, This represents a text embedding model.

[0133] At the same time, for the query text Performing the same vectorization process, the query vector representation is obtained as follows:

[0134] .

[0135] Based on the above vector representation, the semantic relevance between each document and the query topic is calculated. In this embodiment, cosine similarity is used as the semantic relevance metric, and its calculation method is as follows:

[0136] ;

[0137] in, Documents The semantic relevance score between the document and the target research topic indicates the degree of relevance between the document and the research topic. The higher the score, the more relevant the document is to the research topic.

[0138] Based on the semantic relevance score For the initial set of documents The documents in the list are sorted, and based on a preset quantity threshold A, the top A documents are selected from the sorting results as a seed document set.

[0139] ;

[0140] in, , and set The literature in the literature can semantically cover the main content of the target research topic.

[0141] The seed literature set obtained through the above steps This step serves as the input foundation for subsequent co-citation expansion, citation network construction, and literature filtering. It ensures that subsequent processing revolves around the target research topic at the semantic level, thereby reducing the interference of irrelevant literature on the generated review results.

[0142] Step 2: After completing the seed literature search in Step 1, the literature set is further expanded based on co-citation relationships to introduce fundamental and representative literature in the target research field. The specific implementation method is as follows:

[0143] For each seed document Extract its reference list, denoted as:

[0144] ;

[0145] in, Documents The collection of cited references.

[0146] Based on the above reference list, a statistical analysis was performed on all references cited by the seed references. For any cited reference... Its cocitation frequency is defined as:

[0147] ;

[0148] in, As an indicator function, when the literature Appeared in seed literature The value is 1 if it is in the reference list, and 0 otherwise.

[0149] Cocitation frequency Documents The number of times a document is cited by different seed documents is used to characterize the fundamental nature and importance of the document in the target research field.

[0150] Cited literature was screened based on co-citation frequency statistics. When literature... When the co-citation frequency of a document meets the following preset co-citation threshold conditions, the document will be included in the list of references. Inclusion of candidate literature:

[0151] ;

[0152] in, This is the preset co-citation threshold. It can be set according to the number of seed documents or the total size of the documents to control the quantity and quality of the extended documents.

[0153] The candidate document set obtained through the above co-citation expansion process is represented as follows:

[0154] ;

[0155] Among them, set It includes the original seed documents as well as documents obtained through co-citation relationships.

[0156] This step introduces literature cited by multiple seed documents, effectively supplementing classic literature or foundational work that might be missed by topic-based retrieval alone. This enhances the representativeness and completeness of the candidate literature set while maintaining topic relevance. This will serve as input for subsequent construction of the hybrid citation network and execution of document filtering.

[0157] Step 3: Based on the candidate literature set obtained in Step 2 A hybrid citation network is constructed to uniformly model various relationships between documents. The specific implementation method is as follows:

[0158] Step 3.1: Constructing the relationship between bibliographic nodes and actual citations:

[0159] Let the set of candidate documents obtained through step 2 be:

[0160] ;

[0161] in, Indicates the first in the candidate document set This document, This indicates the total number of candidate documents.

[0162] For each document Extract its reference list, denoted as Based on the aforementioned reference list, a set of true citation relationships is constructed, defined as follows:

[0163] ;

[0164] Among them, directed edges ( ) indicates literature For the literature Citation relationships are used to characterize the knowledge inheritance path and the direction of research dissemination among documents.

[0165] Step 3.2: Construction of enhanced relations based on semantic similarity: To alleviate the sparsity problem that may exist in real citation networks, enhanced relations based on semantic similarity are further introduced.

[0166] First, the candidate literature set Each document in By using a pre-defined text embedding model to vectorize the title and abstract, a semantic vector representation of the document is obtained:

[0167] ;

[0168] Based on the semantic vector representation described above, calculate the semantic vector representation of any two documents. and The semantic similarity between them, which is defined in the form of cosine similarity:

[0169] ;

[0170] To control the number of semantically enhanced relations, for each document Only retain the top-ranked results based on semantic similarity. And the similarity is greater than the preset threshold. The literature is used as a candidate neighbor. Wherein, the parameters... It is adaptively determined based on the size of the literature, and its definition is as follows:

[0171] .

[0172] in, is a proportionality coefficient used to control the relative retention ratio of semantic edges in the overall candidate space, thereby ensuring that the sparsity of the graph is consistent under document sets of different sizes; This is a smoothing compensation term used to provide necessary smoothing compensation when the document size is small, to avoid excessively sparse graph structure due to insufficient candidate quantity; This is the strength parameter used to adjust the smoothing compensation term; and Top- The lower and upper bounds are used to prevent degenerate or overly dense graph structures in extremely small or extremely large scenarios.

[0173] When the literature With literature When a genuine citation relationship already exists between two sources, the corresponding semantic enhancement relationship will not be reconstructed, thus avoiding redundant connections.

[0174] Step 3.3: Constructing Enhanced Connections Based on Co-citation Relationships. To characterize the similarity between documents at the research foundation level, enhanced connections are constructed based on co-citation relationships.

[0175] Document With literature The reference sets are respectively and The co-citation count between the two is defined as:

[0176] .

[0177] Based on this, the co-citation similarity between documents is defined as:

[0178] .

[0179] Co-citation similarity considers both the number of shared references and the differences in reference size, making the similarity between different documents comparable. When documents... and If a genuine citation relationship already exists between two sources, a corresponding co-citation enhancement relationship will not be constructed.

[0180] Step 3.4: Fusion of Hybrid Enhancement Relationships and Network Construction.

[0181] To unify semantic similarity information and co-citation similarity information, the two types of enhanced relations are fused to construct a hybrid enhanced edge. For any two documents... and Its hybrid edge enhancement edge is defined as:

[0182] ;

[0183] in, is a weighting coefficient used to adjust the relative contribution of semantic similarity and co-citation similarity in the mixed relation; when one type of enhancement relation does not exist, its corresponding weight term takes a value of zero.

[0184] Based on the above hybrid reinforcement edges, construct a set of reinforcement edges. This is then merged with the set of real citation relationships to obtain a hybrid citation network:

[0185] .

[0186] In this context, network nodes represent documents, real citation edges are used to characterize explicit academic citation relationships, and hybrid augmented edges are used to supplement the potential semantic and research-based associations between documents.

[0187] The hybrid citation network constructed through the above steps maintains the original citation structure while improving the network's connectivity and semantic consistency, providing a reliable structured foundation for subsequent literature importance analysis, filtering, and review structure generation.

[0188] Step 4: After constructing the hybrid citation network, a comprehensive filtering process based on relevance, structural importance, and time factors is performed on the candidate reference set to obtain a high-quality reference set for review generation. The specific implementation method is as follows:

[0189] Step 4.1: Calculation of the relevance of the literature topic.

[0190] Let the hybrid citation network constructed in step 3 be:

[0191] ;

[0192] Where E represents the set of citation and enhancement relationships between documents, i.e. .

[0193] For each document Based on its title and abstract text, a semantic vector representation of the document is obtained using a pre-defined text embedding model. .

[0194] At the same time, the query text for the target research topic The query vector is obtained by vectorization. .

[0195] Based on the above vector representation, calculate the literature The semantic relevance between a research topic and the target research topic is defined as:

[0196] ;

[0197] in, It is used to measure the degree of match between the content of the literature and the target research topic.

[0198] Step 4.2: Calculation of the importance of document structure.

[0199] Based on the hybrid citation network The PageRank algorithm is used to calculate the importance of document nodes in order to evaluate the academic influence of documents in the overall citation network structure.

[0200] Document PageRank score in citation networks is This represents structural importance. Its calculation method is based on iteratively obtaining the connections between nodes in the network, and the result is normalized to ensure its value falls within a certain range. .

[0201] Structural importance It is used to reflect the centrality of a document in the network and its citation influence.

[0202] Step 4.3: Calculation of the literature time compensation factor.

[0203] To balance classic and recent literature in the literature filtering process, a time compensation factor is further introduced.

[0204] Document The time difference between the publication time and the current time is Then its time compensation factor is defined as:

[0205] ;

[0206] in, The preset time decay parameter is used to control the degree of influence of time factors on the overall score. The time compensation factor is used to give appropriate weight to recently published literature without excessively weakening the influence of classic literature.

[0207] Step 4.4: Comprehensive scoring and literature filtering.

[0208] Based on the above calculation results, for each document Calculate its overall score, defined as follows:

[0209] ;

[0210] in, This is a relevance adjustment parameter used to enhance the impact of topic relevance on the final score; This is a time compensation weight parameter used to adjust the contribution ratio of time factors in the overall score.

[0211] By using an index for relevance items, topic relevance plays a soft constraint role in the scoring process, thereby preventing literature with low relevance to the target topic from receiving excessively high scores.

[0212] Step 4.5: Based on the comprehensive score... The candidate literature set is sorted, and a filtering strategy based on percentiles or number thresholds is used to retain literature with scores in the top 80% percentile, forming the final high-quality reference set. .

[0213] By employing the comprehensive filtering method based on relevance, structural importance, and time factors, we can ensure the consistency of the literature's themes while also taking into account its academic influence and timeliness, thus providing reliable and representative literature input for the subsequent generation of review outlines and main texts.

[0214] Step 5: After completing the literature filtering, the literature collection is divided into communities based on the structural relationships and semantic similarities between the literature, and an outline structure of the academic review is automatically generated accordingly. The specific implementation method is as follows.

[0215] Step 5.1: Construction of the citation network.

[0216] Let the set of high-quality documents obtained through step 4 be:

[0217] ;

[0218] in, This indicates the number of documents used to generate the review.

[0219] Using the hybrid citation network construction algorithm described in step 3, based on Construct a new citation network .

[0220] Step 5.2: Community partitioning based on network structure. This involves analyzing the document subnetwork. The Leiden community discovery algorithm is used to perform community partitioning, dividing document nodes into several non-overlapping communities, denoted as:

[0221] ;

[0222] in, Indicates the first A literature community Indicates the number of communities.

[0223] The literature within each literature community has a high degree of consistency in citation structure and semantic content, corresponding to a relatively independent research sub-direction in the target research field;

[0224] Step 5.3: Calculation of Community Contribution. To measure the importance of different research sub-directions within the overall research field, this invention further calculates the contribution of each literature community.

[0225] Based on the comprehensive literature score obtained in step 4, let the literature... The overall score is Then the first The contribution of each community is defined as follows:

[0226] .

[0227] The community contribution score is used to characterize the overall contribution of literature within the community to the target research field under the combined effects of topic relevance, structural importance, and time factors.

[0228] Furthermore, to facilitate comparisons between different communities, the community contribution scores are normalized to obtain normalized contribution scores:

[0229] .

[0230] Step 5.4: Community Screening and Generation of First-Level Headings for Methodology Chapters. After obtaining the normalized contribution of each literature community, an adaptive screening strategy based on a cumulative contribution threshold is used to determine the research sub-directions covered by the review, such as... Figure 2 As shown.

[0231] First, all communities are sorted from highest to lowest according to their normalized community contribution, forming an ordered community sequence. Subsequently, the normalized contribution scores of the aforementioned communities are summed sequentially, and the smallest set of communities satisfying the following conditions is selected:

[0232] ;

[0233] in, This represents the selected set of communities. The preset community contribution threshold is used to control the completeness of the research directions covered by the review at the macro level.

[0234] By selecting the smallest set of communities that meet the above conditions, it is possible to ensure that the review covers the main research directions while avoiding the introduction of research sub-directions with low contribution and insufficient representativeness.

[0235] For each selected community Furthermore, based on the contribution of the literature at the literature level, the literature within the community is screened to determine the representative set of literature used to generate the first-level headings.

[0236] Specifically, for the community Each document in Based on the comprehensive score defined in step 4 The relative contribution of each document within the community is calculated, and they are sorted from highest to lowest contribution. Then, using the same cumulative threshold strategy as the community selection, the smallest set of documents meeting the following criteria is selected:

[0237] ;

[0238] in, Indicates community A collection of representative literature, This is the preset threshold for literature contribution.

[0239] The representative literature collection It includes core literature that has high thematic relevance and structural importance in this research sub-direction, and is used to characterize the main research content and technical context of the corresponding research direction of this community.

[0240] Based on the aforementioned representative literature set The titles, abstracts, and keywords of Chinese literature are used to automatically generate corresponding first-level headings, which summarize the overall research theme of the sub-direction represented by that community. By introducing a two-layer screening mechanism of communities and literature based on contribution thresholds, the generated first-level headings can reflect the overall structure of the research field while also possessing good representativeness and stability.

[0241] Step 5.5: Generation of second-level headings for the "Semantic Clustering and Methodology" section within the community. For example... Figure 3 As shown, after the first-level headings are determined, each literature community... The internal literature further performs fine-grained clustering based on semantic similarity, dividing the community into several semantically consistent sub-clusters, denoted as:

[0242] ;

[0243] in, Indicates community The first in Subclusters.

[0244] Based on the semantic features of the documents within each subcluster, corresponding secondary headings are automatically generated to describe the specific research questions, technical routes, or methodologies within that research sub-direction.

[0245] The outline corresponding to the method chapters was obtained through the above method, and its output can be expressed as follows:

[0246] ;

[0247] in, The outline corresponding to the chapter on representation methods is at page number 1. One first-level heading, This indicates the corresponding second-level heading. This indicates the number of communities selected.

[0248] The methodology outline generated through this step can cover the main research directions in the research field at the macro level and reflect the research structure within each sub-direction at the micro level, providing clear structural constraints for the automatic generation of the subsequent review text.

[0249] Step 5.6: Generating outlines for the remaining chapters. For the outlines of non-methodology chapters, the same algorithm as for methodology chapters is used. The difference is that for non-methodology chapters, the literature information used is a collection of representative papers from each community. Finally, the outlines of methodology chapters and non-methodology chapters are merged to obtain the complete review outline.

[0250] Step 6: After the review outline is generated in Step 5, the main text of the academic review is automatically generated based on the selected representative literature set under the constraints of the outline structure. The specific implementation method is as follows.

[0251] Step 6.1: Input structure for generating the main text. Let the outline structure obtained in Step 5 be as follows:

[0252] ;

[0253] in, This indicates the first section of the review outline. One first-level heading, This indicates the corresponding second-level heading.

[0254] For each first-level heading In step 5, the corresponding representative literature set has been determined, denoted as:

[0255] .

[0256] For each second-level heading From the aforementioned representative literature collection The next step is to select the subset of documents that are most relevant to the semantics of the second-level heading and use it as the reference input for that subsection.

[0257] Step 6.2: Segmentation generation strategy based on chapter structure. To ensure the structural consistency and logical coherence of the generated text, a hierarchical and segmented text generation strategy is adopted.

[0258] Specifically, following the hierarchical structure of the review outline, each first-level heading is addressed sequentially. and its subordinate second-level headings Perform the text generation process.

[0259] When generating a text paragraph corresponding to a certain second-level heading, the second-level heading is used as a subject constraint, and the corresponding subset of literature is used as the content basis, so that the generated text revolves around a clear research question or method category.

[0260] By limiting the text generation process to a preset chapter structure, the problems of content duplication and topic drift between different chapters can be effectively avoided.

[0261] Step 6.3: Content Constraint Generation Based on Representative Literature. The main text generation adopts a conditional text generation method based on a large language model. During the generation process, the key information of the representative literature corresponding to each chapter is used as the generation condition input, so that the generated text is semantically constrained by the content of the literature. Specifically, for a certain chapter, the titles and abstracts of each literature in its representative literature set are integrated and used as reference context input into the large language model, so that the generated text can summarize the main research content, technical route and development context of the research work involved in the chapter.

[0262] By introducing representative literature as content constraints, the problem of factual bias or deviation from the context of the literature in the generated text can be significantly reduced.

[0263] Using the above-mentioned text generation method based on structural and literal constraints, a complete academic review text is finally output. Its content is organized according to the preset first-level and second-level headings, and the content of each chapter is consistent in terms of theme, logic and expression style.

[0264] This embodiment proposes an automated academic review generation method based on citation networks. This method integrates literature retrieval, citation relationship modeling, literature quality assessment, and structured text generation to form an end-to-end automated review generation technology solution. The core innovations of this embodiment are: in the literature acquisition stage, a co-citation expansion mechanism is introduced to supplement fundamental and representative literature that may be missed during topic retrieval; in the literature modeling stage, a hybrid citation network is constructed by integrating real citation relationships, semantic similarity relationships, and co-citation relationships, thereby simultaneously characterizing the explicit citation structure and implicit semantic connections between documents; in the literature screening stage, a filtering mechanism based on relevance-importance-time compensation is designed, comprehensively considering the relevance of documents to the target topic, their structural importance in the citation network, and time factors, effectively improving the quality of the literature set used for review generation.

[0265] Furthermore, this embodiment automates the division of research fields based on the community structure of a hybrid citation network, and adaptively selects research sub-directions and representative literature through normalized contribution measurement. This ensures that the generated review outline covers the main research directions within the field while avoiding structural redundancy. Building upon this, under the constraints of a clear outline structure and representative literature, this embodiment employs a hierarchical and segmented text generation strategy, achieving consistency in thematic content, logical coherence, and information completeness. Through these methods, this embodiment can automatically generate well-structured and reliable academic review texts without manual intervention, providing an efficient, stable, and scalable technical solution for large-scale literature analysis and academic knowledge organization, demonstrating significant practical application value and promising prospects for wider adoption.

[0266] Example 2:

[0267] This embodiment proposes an electronic device, including: one or more processors, and a memory for storing instructions, which, when executed by the one or more processors, cause the one or more processors to perform the automated academic review generation method based on citation networks.

[0268] The electronic device may be a mobile phone, computer, or tablet computer, etc., and includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, implements the automated academic review generation method based on citation networks as described in the embodiments. It is understood that the electronic device may also include input / output (I / O) interfaces and communication components.

[0269] The processor is used to execute all or part of the steps in the citation network-based automated academic review generation method described in the above embodiments. The memory is used to store various types of data, which may include, for example, instructions for any application or method in an electronic device, as well as application-related data.

[0270] The processor can be implemented as an Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), controller, microcontroller, microprocessor, or other electronic components, and is used to execute the automated academic review generation method based on citation networks described in the above embodiments.

[0271] Example 3:

[0272] This embodiment proposes a computer-readable storage medium that stores executable instructions. When these instructions are executed, if they are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0273] The computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the citation network-based automated academic review generation method described in the various embodiments of this application.

[0274] The aforementioned storage media include: flash memory, hard disks, multimedia cards, card-type memory (e.g., SD (Secure Digital Memory Card) or DX (Memory Data Register, MDR) memory), random access memory (RAM), static random-access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic storage, disks, optical discs, servers, APP (Application) application stores, and other media capable of storing program verification codes. These media store computer programs, which, when executed by a processor, can implement the various steps of the aforementioned automated academic review generation method based on citation networks.

[0275] Example 4:

[0276] This embodiment proposes a computer program product, including a computer program or instructions, which, when executed by a processor, implements the aforementioned automated academic review generation method based on citation networks.

[0277] Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a computer program product.

[0278] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the present invention.

Claims

1. A method for generating automated academic reviews based on citation networks, characterized in that: Includes the following steps: Step 1: Retrieve seed literature. Based on the research topic, keywords or search conditions entered by the user, use the academic search engine to perform literature retrieval operations, obtain an initial set of literatures semantically related to the target research field, and perform preliminary filtering based on similarity to serve as the seed literature set. The seed literature collection is used to provide preliminary semantic coverage of the target research topic, and its size and content are dynamically adjusted according to the search criteria. Step 2: Based on the seed literature set obtained in Step 1, analyze the reference list of each seed literature and count the number of times different literatures are cited by multiple seed literatures; for literatures whose co-citation counts meet the preset co-citation threshold condition, add them to the candidate literature set; the preset co-citation threshold is set according to the number of seed literatures or the total size of literatures, and is used to control the quantity and quality of expanded literatures; through co-citation expansion processing, basic or classic literatures that are repeatedly cited in the research field are introduced; Step 3: Using the documents in the document set obtained in Step 2 as network nodes, construct a basic citation network structure based on the real citation relationships between documents; on this basis, further construct supplementary association edges according to the semantic similarity relationship of document content and the co-citation relationship between documents, and merge different types of relationships to form a hybrid citation network; Step 4: Based on the hybrid citation network obtained in Step 3, perform comprehensive filtering and ranking processing on the candidate document set; The relevance of each document to the target research topic, the importance of the document in the citation network structure, and the publication time of the document are evaluated separately. The comprehensive score of the document is obtained by comprehensively calculating the various factors. Based on the comprehensive score, candidate documents are screened, and those that meet the preset conditions are retained to form a final high-quality reference set. Step 5: Based on the final high-quality reference set obtained in Step 4, the citation network is reconstructed using the hybrid citation network construction method described in Step 3. Then, the community detection algorithm is used to divide the literature into communities, and literature with similar or closely related research directions are grouped into the same community. Based on the community division, we further evaluated the contribution of each community to the overall literature collection and selected communities with high contribution as core research directions. For each core community, a representative collection of literature is selected, and a corresponding first-level chapter outline is generated using a large language model. At the same time, the literature is clustered in a fine-grained manner within the community, and a corresponding second-level chapter outline is generated using a large language model, thus forming a clear hierarchical and thematic overall review outline structure. Step 6: Under the constraints of the outline structure obtained in Step 5, and combined with the literature collection corresponding to each chapter, extract the titles and abstracts of the literature as input, and use the large language model to generate the main text of the review article chapter by chapter; During the text generation process, ensure that the generated content is consistent with the theme of the corresponding chapter and maintain the logical connection between each chapter, so as to form a complete and coherent review text.

2. The automated academic review generation method based on citation networks according to claim 1, characterized in that: The preliminary filtering based on similarity specifically involves: For each document in the initial document set, its title and abstract text are extracted, and the text is vectorized using a pre-defined text embedding model to obtain a semantic vector representation. At the same time, the query text is vectorized in the same way to obtain a query vector representation. Based on the above vector representation, cosine similarity is used as a semantic relevance measure to calculate the semantic relevance between each document and the query topic. The documents in the initial document set are sorted according to the semantic relevance score, and the top A documents in the sorting results are selected as the seed document set according to a pre-defined ranking threshold A.

3. The automated academic review generation method based on citation networks according to claim 1, characterized in that: The method for constructing the basic citation network structure in step 3 is as follows: Let the set of candidate documents obtained through step 2 be: ； in, Indicates the first in the candidate document set This document, Indicates the total number of candidate documents; For each document Extract its reference list, denoted as Based on the reference list, a set of true citation relationships is constructed, defined as follows: ； Among them, directed edges ( ) indicates literature For the literature Citation relationships are used to characterize the knowledge inheritance path and the direction of research dissemination among documents.

4. The automated academic review generation method based on citation networks according to claim 3, characterized in that: In step 3, the specific method for forming the hybrid citation network is as follows: First, construct semantically enhanced relations; then, for the candidate document set... Each document in The title and abstract are vectorized using a pre-defined text embedding model to obtain the semantic representation vector of the document: ； Based on the semantic representation vectors mentioned above, cosine similarity is used to calculate the similarity between any two documents. and Semantic similarity between them: ； For each document Only retain the top-ranked results based on semantic similarity. And the similarity is greater than the preset threshold. The literature is used as a candidate neighbor; wherein, the parameter It is adaptively determined based on the size of the literature, and its definition is as follows: ； in, is a proportionality coefficient used to control the relative retention ratio of semantic edges in the overall candidate space, thereby ensuring that the sparsity of the graph is consistent under document sets of different sizes; This is a smoothing compensation term used to provide necessary smoothing compensation when the document size is small, to avoid excessively sparse graph structure due to insufficient candidate quantity; This is the strength parameter used to adjust the smoothing compensation term; and Top- The lower and upper bounds are used to prevent degenerate or overly dense graph structures in extremely small or extremely large-scale scenarios. When the literature With literature When a genuine citation relationship already exists between two sources, the corresponding semantically enhanced relationship will not be constructed again. Secondly, construct co-citation enhancement relationships; assuming the literature With literature The reference sets are respectively and The co-citation count between the two is defined as: ； Based on this, the co-citation similarity between documents is defined as: ； When the literature and If a genuine citation relationship already exists between two sources, no further co-citation enhancement relationship will be constructed. Finally, semantic enhancement relations and co-citation enhancement relations are fused to construct hybrid enhancement edges; for any two documents and Its hybrid enhancement edge is defined as: ； in, , which are weighting coefficients used to adjust the relative contributions of semantic similarity and co-citation similarity in the mixed relation; when one type of enhancement relation does not exist, its corresponding weight term takes a value of zero; Based on the hybrid enhanced edges, construct an enhanced edge set. This is then merged with the set of real citation relationships to obtain a hybrid citation network: ； in, It is the set of real citation relationships constructed based on real citation relationships in the previous steps. It is the set of candidate documents obtained in step 2; In this context, network nodes represent documents, real citation edges are used to characterize explicit academic citation relationships, and hybrid augmented edges are used to supplement the potential semantic and research-based associations between documents.

5. The automated academic review generation method based on citation networks according to claim 4, characterized in that: The specific method for the comprehensive filtering and sorting process in step 4 is as follows: Step 4.1: Calculation of the relevance of the document topic; Let the hybrid citation network constructed in step 3 be: ； Where E represents the set of citation and enhancement relationships between documents; For each document Based on its title and abstract text, a semantic vector representation of the document is obtained using a pre-defined text embedding model. ; At the same time, the query text for the target research topic The query vector is obtained by vectorization. ; Based on the above vector representation, calculate the literature The semantic relevance between a research topic and the target research topic is defined as: ； in, This is used to measure the degree of match between the content of the literature and the target research topic; Step 4.2: Calculation of the importance of document structure; Based on the hybrid citation network We use a webpage ranking algorithm to calculate the importance of document nodes in order to evaluate the academic influence of documents in the overall citation network structure. Document PageRank score in citation networks is This represents structural importance. Its calculation method is based on iteratively obtaining the connections between nodes in the network, and the result is normalized to ensure its value falls within a certain range. Structural importance Used to reflect the centrality of a document on the network and its citation impact; Step 4.3: Calculation of the literature time compensation factor; To balance classic and recent literature in the literature filtering process, a time compensation factor is introduced to give appropriate weight to recently published literature without excessively weakening the influence of classic literature. Document The time difference between the publication time and the current time is Then its time compensation factor is defined as: ； in, This is a preset time decay parameter used to control the degree of influence of time factors on the overall score; Step 4.4: Comprehensive scoring and literature filtering; Based on the above calculation results, for each document Calculate its overall score, defined as follows: ； in, This is a relevance adjustment parameter used to enhance the impact of topic relevance on the final score; This is a time compensation weight parameter used to adjust the contribution ratio of time factors in the overall score; Step 4.5: Based on the comprehensive score... The candidate literature set is sorted, and a filtering strategy based on percentiles or number thresholds is used to retain literature with scores in the top 80% percentile, forming the final high-quality reference set. , .

6. The automated academic review generation method based on citation networks according to claim 5, characterized in that: The specific method for step 5 is as follows: Step 5.1: Construct the citation network; Let the set of high-quality documents obtained through step 4 be: ； in, This indicates the number of documents used to generate the review; Using the hybrid citation network construction method from step 3, based on Construct a new citation network ; Step 5.2: Divide communities based on network structure; For the new citation network The Leiden community discovery algorithm is used to perform community partitioning, dividing document nodes into several non-overlapping communities, denoted as: ； in, Indicates the first A literature community Indicates the number of communities; Step 5.3: Calculate the contribution of each literature community to characterize the overall contribution of literature within that community to the target research field under the combined effects of topic relevance, structural importance, and time factors; Based on the literature obtained in step 4 Overall rating Then the first The contribution of each community is defined as follows: ； The community contribution score is normalized to obtain the normalized contribution score: ； Step 5.4: Community screening and generation of first-level headings for the methodology section. After obtaining the normalized contribution of each literature community, an adaptive screening strategy based on the cumulative contribution threshold is used to determine the research sub-directions covered by the review. First, all communities are sorted from highest to lowest according to their normalized community contribution, forming an ordered community sequence. Then, the normalized contribution scores of each community are accumulated sequentially, and the smallest set of communities that meets the following conditions is selected: ； in, This represents the selected set of communities. A preset community contribution threshold is used to control the completeness of the research directions covered by the review at the macro level; For each selected community Furthermore, based on the contribution of the literature at the literature level, the literature within the community is screened to determine the representative set of literature used to generate the first-level headings; For the community Each document in Based on the comprehensive score defined in step 4 The relative contribution of each document within the community is calculated, and the documents are sorted from highest to lowest relative contribution. Then, using the same cumulative threshold strategy as the community selection, the smallest set of documents meeting the following criteria is selected: ； in, Indicates community A collection of representative literature, The preset threshold for literature contribution; The representative literature collection Used to characterize the main research content and technological context of the corresponding research direction of the community; Based on representative literature collections The title, abstract, and keywords of Chinese literature are automatically used to generate corresponding first-level headings, which summarize the overall research theme of the research sub-direction represented by the community. Step 5.5: Semantic Clustering within the Community and Generation of Second-Level Headings for the Methodology Section; After determining the first-level headings, generate second-level headings for each literature community. The internal literature further performs fine-grained clustering based on semantic similarity, dividing the community into several semantically consistent sub-clusters, denoted as: ； in, Indicates community The first in Subclusters; Based on the semantic features of the documents in each sub-cluster, corresponding secondary headings are automatically generated to describe the specific research questions, technical routes, or methodological categories in that research sub-direction. The outline corresponding to the method chapters is obtained through the above method, and its output is expressed as follows: ； in, The outline corresponding to the chapter on representation methods is at page number 1. One first-level heading, This indicates the corresponding second-level heading. Indicates the number of communities selected; Step 5.6: Generate the outlines for the remaining chapters; For the outlines of non-method sections, the same method as that used for generating method sections is employed. The difference lies in that the literature information used for generating non-method sections is a collection of representative papers from each community. Finally, the outlines of method sections and non-method sections are merged to obtain a complete review outline.

7. The automated academic review generation method based on citation networks according to claim 6, characterized in that: The specific method for step 6 is as follows: Step 6.1: Input structure for generating the main text. Let the outline structure obtained in Step 5 be as follows: ； in, This indicates the first section of the review outline. One first-level heading, This indicates the corresponding second-level heading; For each first-level heading In step 5, the corresponding representative literature set has been determined, denoted as: ； For each second-level heading From representative literature collections Further, a subset of documents most relevant to the semantics of the second-level heading is selected as the reference input for that subsection; Step 6.2: Segmentation generation strategy based on chapter structure. To ensure the structural consistency and logical coherence of the generated text, a hierarchical and segmented text generation strategy is adopted. Following the hierarchical structure of the review outline, proceed sequentially to address each first-level heading. and its subordinate second-level headings Execute the text generation process; When generating a text paragraph corresponding to a certain second-level heading, the second-level heading is used as a subject constraint, and the corresponding subset of literature is used as the content basis, so that the generated text revolves around a clear research question or method category. Step 6.3: Generate text based on the content constraints of representative literature. The main text is generated using a conditional text generation method based on a large language model. During the generation process, the key information of the representative literature corresponding to each chapter is used as the input of the generation conditions, so that the generated text is semantically constrained by the content of the literature. For a certain chapter, the title and abstract information of each literature in its representative literature set are integrated and used as reference context input into the large language model, so that the generated text can summarize the main research content, technical route and development context of the research work involved in the chapter. Using the above-mentioned text generation method based on structural and literal constraints, a complete academic review text is finally output. Its content is organized according to the preset first-level and second-level headings, and the content of each chapter is consistent in terms of theme, logic and expression style.

8. An electronic device, characterized in that: The method includes one or more processors and a memory for storing instructions that, when executed by the one or more processors, cause the one or more processors to perform the automated academic review generation method based on citation networks as described in any one of claims 1-7.

9. A computer-readable storage medium, characterized in that: It stores executable instructions that, when executed, cause the processor to perform the automated academic review generation method based on citation networks as described in any one of claims 1-7.

10. A computer program product, characterized in that: It includes a computer program or instructions that, when executed by a processor, implement the automated academic review generation method based on citation networks as described in any one of claims 1-7.

Images