Academic knowledge retrieval enhancement generation method and system based on multi-dimensional citation relationship graph

Abstract

The application provides an academic knowledge retrieval enhancement generation method and system based on a multidimensional citation relationship graph, relates to the technical field of retrieval enhancement, and comprises the following steps: reading academic literature metadata, and constructing a sparse metadata graph with data-semantic-temporal triple enhancement; when performing online retrieval, based on multidimensional retrieval technology of citation coupling and Pareto frontier, recalling academic literature candidate nodes from the sparse metadata graph, and establishing a recall result set; relatively sorting and adjusting the candidate nodes in the recall result set; performing enhancement verification on the relatively sorted and adjusted result based on a dynamic graph state machine and a semantic topology double-space probability fusion logic alignment generation mechanism, and establishing a retrieval enhancement answer. The technical problem of insufficient accuracy and reliability of the retrieval content in the prior art is solved. High-precision retrieval and constraint generation under a multidimensional graph are achieved, and the technical effects of improving the accuracy and reliability of academic knowledge retrieval are achieved.

Description

Technical Field

[0001] This invention relates to the field of retrieval enhancement technology, specifically to a method and system for enhancing academic knowledge retrieval based on multidimensional citation graphs. Background Technology

[0002] Existing Retrieval Augmentation (RAG) techniques in academic literature scenarios typically treat papers as independent text units, achieving semantic similarity matching and fragment retrieval through vector encoding. This ignores the objectively existing citation relationships, data dependencies, and temporal evolution structures between papers, lacking the ability to model the topological structure of academic knowledge networks. Traditional methods primarily rely on static semantic vector retrieval, failing to effectively integrate structural information such as citation coupling, co-citation structures, and scientific data associations. Consequently, search results only reflect surface-level semantic similarity, failing to reveal the deep structural connections within the research context.

[0003] Furthermore, in multi-hop reasoning or domain evolution analysis tasks, traditional RAG lacks structural constraint mechanisms, which can easily lead to cross-domain jumps, path breaks, or information drift. The generated content lacks traceable graph structure support, reducing the credibility and academic interpretability of the results, resulting in low accuracy and reliability of academic knowledge retrieval results.

[0004] Existing technologies suffer from the problem of treating academic knowledge as isolated text entities for retrieval, while lacking constraints, resulting in insufficient accuracy and reliability of the retrieved content. Summary of the Invention

[0005] The purpose of this application is to provide an enhanced academic knowledge retrieval generation method and system based on multidimensional citation graphs, which addresses the technical problem that existing technologies treat academic knowledge as isolated text entities for retrieval, and lack constraints, resulting in insufficient accuracy and reliability of the retrieved content.

[0006] In view of the above problems, this application provides an academic knowledge retrieval enhancement generation method and system based on multidimensional citation relationship graphs.

[0007] The first aspect of this application provides an academic knowledge retrieval enhancement generation method based on a multidimensional citation graph. The method includes: after reading academic literature metadata, constructing a sparse metadata graph enhanced by a data-semantic-temporal framework; when performing an online retrieval, recalling candidate nodes of academic literature from the sparse metadata graph based on citation coupling and Pareto front multidimensional retrieval techniques to establish a recall result set; adjusting the relative ranking of candidate nodes within the recall result set based on a progressive implicit reasoning mechanism; and performing enhanced verification of the relative ranking adjustment results based on a logical alignment generation mechanism that fuses dynamic graph state machines and semantic topology dual-space probabilistics to establish an enhanced retrieval answer.

[0008] The metadata of academic documents is parsed into paper entities, and author associations and journal affiliations are constructed based on these paper entities, so that paper nodes form a first-layer academic structure network through author edges and journal edges respectively. On the basis of the first-layer academic structure network, the scientific data set associated with the paper is abstracted into independent scientific data nodes, and data association edges are established between paper nodes and corresponding scientific data nodes. For multiple paper nodes pointing to the same scientific dataset, bidirectional bridging edges are formed by sharing scientific data nodes. In the graph structure, a data homology identifier attribute is set for the bidirectional bridging edges, and implicit association channels across citation chains are formed by introducing scientific data nodes.

[0009] Optionally, data association edges are constructed by calculating edge weights, as follows: ;in, Representing edge weights, and These are the weighting coefficients. , Characterizing semantic cosine similarity, This is a domain-adaptive parameter used to represent the average citation half-life of journals in the corresponding subject area over 5 years. Characterizes the time difference.

[0010] Optionally, semantic similarity between isolated nodes is calculated based on Sentence-BERT vectors of keywords and summaries for all isolated nodes; node pairs with no reference path and semantic distance less than a preset threshold are identified based on the semantic similarity; implicit association edges marked as virtual type are inserted between the node pairs, and the confidence weight of the implicit association edges is configured to construct implicit association channels.

[0011] Optionally, during the multi-hop traversal, the document coupling fingerprint between the current node and its neighboring nodes is calculated, and the document coupling fingerprint is obtained based on the Jaccard coefficient of the reference set; during the path expansion process, the rate of change of the coupling strength sequence is calculated in real time, and a structural entropy index is constructed based on the calculation results; if the structural entropy index exceeds a preset threshold, a path soft truncation mechanism is triggered, and the traversal strategy is switched from depth-first to breadth-first.

[0012] Optionally, the retrieval task is decomposed into three mutually orthogonal objective functions using a multi-objective non-dominated ranking algorithm. These objective functions include a semantic relevance objective function, a structural centrality objective function, and a temporal novelty objective function, thereby establishing a recall result set.

[0013] Optionally, a special identifier token for triggering ranking inference is inserted into the input sequence; multiple rounds of pairwise comparison inference of papers are performed in the hidden representation layer, and an intermediate ranking score vector is generated after each round of comparison; a list-based ranking loss function is used to iteratively optimize the intermediate scores, and the Softmax temperature parameter is gradually reduced during the ranking process.

[0014] Optionally, when the generated content is detected to contain anchor entities in the graph, the K-hop neighbor subgraph of the corresponding anchor entity is activated in real time, and the K-hop neighbor subgraph is converted into a finite state automaton structure; during the decoding stage, a topological mask is applied to the output probability distribution, allowing only entities or relations existing in the K-hop neighbor subgraph to be generated, and assigning negative infinite weights to entity candidates that do not exist in the K-hop neighbor subgraph.

[0015] Optionally, the semantic space output probability and the topological constraint mask are fused at the Logits level; a probability boosting factor is applied to the legitimate path based on the citation strength or the number of co-citations; a look-ahead verification module is introduced to verify the subsequent reachability in the graph before selecting the next entity. If the predicted path enters a graph island, the cluster search backtracking mechanism is automatically triggered to re-plan and generate the path.

[0016] A second aspect of this application provides an academic knowledge retrieval enhancement generation system based on a multidimensional citation graph. The system includes: a data graph construction module, used to construct a sparse metadata graph with data-semantic-temporal enhancement after reading academic literature metadata; a candidate node recall module, used to recall academic literature candidate nodes from the sparse metadata graph and establish a recall result set based on multidimensional retrieval techniques of citation coupling and Pareto fronts when performing online retrieval; a ranking adjustment module, used to perform relative ranking adjustment on candidate nodes within the recall result set based on a progressive implicit reasoning mechanism; and an enhancement verification module, used to perform enhancement verification of the relative ranking adjustment results based on a logical alignment generation mechanism that fuses dynamic graph state machines and semantic topology dual-space probabilistics, and establish an enhanced retrieval answer.

[0017] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0018] The method provided in this application constructs a sparse metadata graph enhanced by data, semantics, and temporal sequence after reading the metadata of academic documents. During online retrieval, based on citation coupling and Pareto front multidimensional retrieval techniques, candidate nodes of academic documents are recalled from the sparse metadata graph to establish a recall result set. A progressive implicit reasoning mechanism is used to adjust the relative ranking of candidate nodes within the recall result set. An enhanced verification of the relative ranking adjustment results is performed based on a logical alignment generation mechanism that fuses dynamic graph state machines and semantic topology probabilistic spaces, thus establishing an enhanced retrieval answer. This achieves high-precision retrieval and constraint generation under multidimensional graph constraints, improving the accuracy and reliability of academic knowledge retrieval.

[0019] The above description is merely an overview of the technical solution of this application. To better understand the technical means of this application and to facilitate its implementation according to the description, and to make the above and other objects, features, and advantages of this application more apparent, specific embodiments of this application are described below. It should be understood that the content described in this section is not intended to identify key or important features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent through the following description. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0021] Figure 1 A flowchart illustrating the academic knowledge retrieval enhancement generation method based on multidimensional citation relationship graphs provided in this application.

[0022] Figure 2 A schematic diagram of the structure of the academic knowledge retrieval enhancement generation system based on multidimensional citation relationship graph provided in this application.

[0023] Figure labeling: Data graph construction module 11, candidate node recall module 12, sorting adjustment module 13, enhanced verification module 14. Detailed Implementation

[0024] This application provides an enhanced academic knowledge retrieval generation method and system based on a multidimensional citation graph. It addresses the technical problem that existing technologies treat academic knowledge as isolated text entities for retrieval, lacking constraints and resulting in insufficient accuracy and reliability of the retrieved content. The method achieves high-precision retrieval and constraint generation under multidimensional graph constraints, thereby improving the accuracy and reliability of academic knowledge retrieval.

[0025] The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. It should be understood that the present invention is not limited to the exemplary embodiments described herein. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. It should also be noted that, for ease of description, only the parts related to the present invention are shown in the accompanying drawings, not all of them.

[0026] Example 1, as Figure 1 As shown, this application provides an academic knowledge retrieval enhancement generation method based on a multidimensional citation graph. The method includes: After reading the metadata of academic literature, a sparse metadata graph with data-semantic-temporal enhancement is constructed.

[0027] Furthermore, a ternary augmented sparse metadata graph based on data, semantics, and temporal sequence is constructed, including: parsing academic literature metadata into paper entities, and constructing author associations and journal affiliations based on the paper entities, so that paper nodes form a first-layer academic structure network through author edges and journal edges respectively; on the basis of the first-layer academic structure network, the scientific data set associated with the paper is abstracted into independent scientific data nodes, and data association edges are established between paper nodes and corresponding scientific data nodes; for multiple paper nodes pointing to the same scientific dataset, bidirectional bridging edges are formed by sharing scientific data nodes; in the graph structure, a data homology identifier attribute is set for the bidirectional bridging edges, and implicit association channels across citation chains are formed by introducing scientific data nodes.

[0028] Specifically, academic literature metadata is a structured data set describing the basic characteristics and attributes of academic literature, covering key information such as title, author, abstract, keywords, publication date, journal name, volume number, page number, and reference list. Data in JSON / XML format is obtained by calling database API interfaces, parsing locally exported BibTeX / RIS files, or parsing the webpage structure to read the academic literature metadata. After obtaining the academic literature metadata, it undergoes structured parsing. This involves mapping fields such as title, author, abstract, publication date, and references to paper entities using entity recognition and field mapping methods. These paper entities are the basic units for constructing the entire data graph, containing basic information about the paper, such as title, author, abstract, publication date, and references, used to uniformly express the semantic and structural information of the literature.

[0029] Based on paper entities, author associations and journal affiliations are constructed. Paper nodes and author nodes are linked through author edges to form author associations, reflecting collaboration between authors of different papers. Simultaneously, journal edges are established between papers and their respective journals to construct journal affiliations, clarifying the journal information in which the paper was published. Through these two relationships, paper nodes are interconnected via author edges and journal edges, forming the first layer of the academic structure network. This first layer is a many-to-many connection structure centered on papers.

[0030] After the first layer of the academic structure network is constructed, the scientific data sets associated with the papers are extracted and parsed. Based on CSCD (China Standard Data Center), the scientific data and literature data are semantically linked. This means that some papers can find directly related or relevant scientific data sets, and the same scientific data may be associated with multiple papers. Scientific data resources, such as experimental datasets, simulation data, and observational data, are treated as independent scientific data nodes. These scientific data nodes are graph node entities that provide a structured representation of the data resources that the papers rely on or generate, aiming to achieve traceability and relevance of data resources. Data association edges are established between paper nodes and their corresponding scientific data nodes to express the direct dependencies between the papers and data resources.

[0031] For scenarios where multiple paper nodes point to the same scientific dataset, a paper-data-paper bridging structure is formed by sharing the same scientific data node. This paper-data-paper bridging structure is a bidirectional bridging relationship, meaning different papers form an implicit coupling relationship through commonly dependent data entities. This bridging edge is defined as a paper-data-paper bidirectional bridging edge, used to express the association between papers due to data homology. It can uncover homologous studies using the same dataset even in the absence of direct citations, significantly improving the retrieval accuracy for reproducibility and methodological questions. To enhance the structural expressiveness, a data homology identifier attribute is further set for the bidirectional bridging edge in the graph structure. This data homology identifier attribute uniquely identifies the shared data source, such as assigning a data source ID, version number, or source signature. This explicitly marks whether two papers originate from the same data foundation in the data-semantic-temporal ternary enhanced sparse metadata graph, enhancing the interpretability and traceability of the graph. By introducing scientific data nodes as intermediate bridging entities, the data-semantic-temporal ternary enhanced sparse metadata graph forms new implicit association channels outside the original citation relationship chains. These implicit association channels across citation chains refer to the ability of two papers to be indirectly connected through the same scientific data node, even if there is no direct citation relationship or indirect citation path between them. This breaks the local isolation structure of traditional citation networks and enhances the connectivity and information dissemination capabilities of the graph. The data-semantic-temporal ternary enhanced sparse metadata graph is stored using a graph database or adjacency list structure. Node type tags distinguish between paper nodes, author nodes, journal nodes, and scientific data nodes; edge type encoding distinguishes between author edges, journal edges, and data edges; and attribute fields are added to bridging edges to achieve structured storage and scalable management.

[0032] For example, consider three academic papers: Paper A, Paper B, and Paper C. Paper A, authored by Author 1 and Author 2, is published in Journal X and relies on scientific dataset D1. Paper B, authored by Author 2 and Author 3, is published in Journal Y and also relies on scientific dataset D1. Paper C, authored by Author 4, is published in Journal Z and has no direct citation relationship with the first two papers. However, Sentence-BERT vector calculations using keywords and abstracts reveal a semantic similarity to Papers A and B, with the similarity falling below a preset threshold. When constructing the sparse metadata graph, Papers A, B, and C are first parsed as paper entities. Then, author relationships are built based on these entities, such as Paper A and Paper B being linked through Author 2. Journal affiliation relationships are also constructed, such as Paper A belonging to Journal X. Papers A, B, and C form the first layer of the academic structure network through author edges and journal edges, respectively.

[0033] The scientific dataset D1 is abstracted as a scientific data node D1. Data association edges are established between paper A and scientific data node D1, and between paper B and scientific data node D1, with edge weights calculated according to the edge weight calculation formula. Since both paper A and paper B point to scientific data node D1, a bidirectional bridge edge is formed by sharing scientific data node D1, and this edge is assigned a data origin identification attribute. For paper C, although it has no direct citation path to paper A and paper B, but is semantically similar, implicit association edges marked as virtual types are inserted between paper C and paper A, and between paper C and paper B, based on the calculated semantic similarity, thus constructing implicit association channels. In this way, a sparse metadata graph with data-semantic-temporal ternary enhancement, containing paper entities, scientific data nodes, and various association edges, is constructed.

[0034] By constructing a ternary enhanced sparse metadata graph of data, semantics, and time series, various information of academic literature, including paper content, authors, journals, scientific data, and their relationships, is comprehensively and meticulously integrated in the form of a graph structure. The integration process enhances the global connectivity of the graph by introducing scientific data nodes and a bidirectional bridging mechanism to open up non-citation association paths between papers. The interpretability and traceability of the graph are improved by using data homology identification attributes, thereby enabling more comprehensive and accurate processing of academic knowledge retrieval and generation tasks.

[0035] Furthermore, data association edges are constructed through edge weight calculation, as follows: ;in, Representing edge weights, and These are the weighting coefficients. , Characterizing semantic cosine similarity, This is a domain-adaptive parameter used to represent the average citation half-life of journals in the corresponding subject area over 5 years. Characterizes the time difference.

[0036] Specifically, in a sparse metadata graph, data association edges represent dependencies or mutual influences between two nodes. To evaluate the strength of this relationship, each data association edge is assigned a weight value. The edge weight calculation for data association edges is expressed as follows: ;in, The weights of the edges between nodes i and j are represented. and These are weighting coefficients, used to adjust the relative importance of semantic similarity and time decay, respectively. They can be increased for review generation tasks. Emphasizing the evolution over time, if it's a conceptual explanation task, the time frame can be increased. Emphasis is placed on semantic accuracy and minimizing the impact of time. , The semantic cosine similarity between node i and node j is represented by cosine similarity calculation. This is a domain-adaptive parameter used to represent the average citation half-life of journals in a corresponding subject area over 5 years. The citation half-life refers to the time elapsed between the publication of the newest half of all currently used literature in a subject area, reflecting the rate of aging of literature in that field. The citation half-life varies across different subject areas. By introducing... This can make edge weight calculation more in line with the characteristics of different disciplines. It represents the difference between the current time and the time the paper was published.

[0037] By comprehensively considering the semantic content similarity and temporal relevance between papers, the data association edges in the sparse metadata graph are ensured to reasonably reflect the strength of relationships between academic documents. Through a weighted mechanism, not only can papers with directly related content be obtained, but time-based citation weights can also be dynamically adjusted, giving newer research results higher weight and priority in the sparse metadata graph, thereby improving the accuracy and efficiency of the entire retrieval and generation process.

[0038] Furthermore, by introducing scientific data nodes to form implicit association channels across citation chains, the following steps are taken: calculating the semantic similarity between isolated nodes based on the Sentence-BERT vectors of keywords and summaries for all isolated nodes; identifying node pairs that do not have citation paths and whose semantic distance is less than a preset threshold based on the semantic similarity; inserting implicit association edges marked as virtual types between the node pairs and configuring the confidence weights of the implicit association edges to construct implicit association channels.

[0039] Specifically, after constructing the basic citation relationships, to address the structural isolation problem in the sparse metadata graph caused by the lack of explicit citations, scientific data nodes are introduced and implicit connection edges are constructed to form implicit connection channels across citation chains. First, isolated nodes in the sparse metadata graph are identified. Isolated nodes refer to paper nodes that do not have direct or multi-hop citation paths in the current citation network or data bridging network and whose structural connectivity is below the structural connectivity threshold. The structural connectivity threshold is adaptively set based on the statistical characteristics of the node degree distribution in the sparse metadata graph. For example, by analyzing the distribution of global node degree or connected component size, the lower quantile of the degree distribution is selected, such as the top 10% quantile, or an interval using the average degree plus or minus the standard deviation is used as a reference range. Keywords and abstract text of all isolated nodes are extracted, and semantic vector representations are generated using the Sentence-BERT model. Sentence-BERT is a sentence vector encoding model based on a dual-tower structure that maps text to dense vectors of fixed dimensions, making semantically similar texts closer together in the vector space. The semantic similarity value between nodes is obtained by calculating the cosine similarity between any two isolated node vectors.

[0040] The system determines whether there is a direct referencing path between node pairs based on semantic similarity values. Node pairs with a semantic distance less than a preset threshold are selected. The semantic distance is calculated as the angular distance between vectors minus the cosine similarity, used to measure the semantic closeness of the text. The preset threshold can be determined statistically using a validation set based on the domain corpus distribution, for example, by taking the upper quantile of the similarity distribution to avoid false connections. For node pairs that meet the conditions of no direct referencing path and semantic closeness, an implicit association edge marked as "Virtual" is inserted between them. This "Virtual" attribute indicates that the implicit association edge does not represent a real referencing behavior but is a structural compensation edge generated by semantic inference, used to distinguish it from real referencing edges or data association edges. Simultaneously, a confidence weight is assigned to each implicit association edge. The confidence weight can be calculated by combining the semantic similarity value and a stability index, such as W. Virtual =γ·Sim content γ is the confidence scaling factor, used to control the influence of implicit edges in graph computation. By constructing implicit association channels across citation chains in the sparse metadata graph, semantically similar papers that were originally separated due to lack of citations can form computable connection paths, thereby improving the connectivity and information dissemination capabilities of the overall graph structure.

[0041] For example, papers A and B have no citation relationship and are not connected in the current graph, belonging to isolated nodes. After Sentence-BERT encoding, vectors VA and VB are obtained, and their cosine similarity is calculated to be 0.87. Setting the semantic similarity threshold θ to 0.80, 0.87 is greater than the threshold, and graph traversal confirms that there is no direct citation path between them. Therefore, a virtual type implicit association edge is inserted between A and B, and the confidence weight is set to 0.87 × γ. For example, if γ = 0.9, the weight is 0.783. In this way, during subsequent retrieval or generation processes, structural reachability from paper A to paper B can be achieved through this implicit association channel, enhancing the knowledge connection capability across citation chains.

[0042] By introducing scientific data nodes to form implicit association channels across citation chains, sparse citation trees are transformed into dense semantic networks, achieving structural compensation for potential academic associations. This solves the problem that unpopular fields or newly published papers cannot be recalled due to zero citations, thereby improving recall coverage and the continuity of knowledge reasoning, while avoiding detachment from structural constraints by simply relying on text vector retrieval.

[0043] When performing online retrieval, multidimensional retrieval techniques based on citation coupling and Pareto fronts are used to recall candidate nodes of academic literature from the sparse metadata graph and establish a recall result set.

[0044] Furthermore, the process of recalling candidate nodes of academic documents from the sparse metadata graph and establishing a recall result set includes: calculating the document coupling fingerprint between the current node and its neighboring nodes during multi-hop traversal, wherein the document coupling fingerprint is obtained based on the Jaccard coefficient of the reference set; calculating the rate of change of the coupling strength sequence in real time during path expansion and constructing a structural entropy index based on the calculation results; and triggering a path soft truncation mechanism when the structural entropy index exceeds a preset threshold, switching the traversal strategy from depth-first to breadth-first.

[0045] Furthermore, the retrieval task is decomposed into three mutually orthogonal objective functions using a multi-objective non-dominated ranking algorithm. These objective functions include a semantic relevance objective function, a structural centrality objective function, and a temporal novelty objective function, thereby establishing a recall result set.

[0046] Specifically, when performing online academic knowledge retrieval, the user query is used as the starting point. The query statement is encoded into a semantic vector, and several anchor paper nodes with semantic similarity to the query are located in the sparse metadata graph as initial seed nodes. Based on this, a multi-dimensional retrieval technique using citation coupling and Pareto fronts is employed. First, entropy-controlled path expansion based on citation coupling is used. Citation coupling refers to the potential relevance of two papers on the same academic topic if they share the same references. The coupling strength can be calculated using the Jaccard coefficient of the reference set. Therefore, when performing multi-hop traversal starting from the recall node, text similarity is not relied upon. Instead, citation network features are used to calculate the document coupling fingerprint between the current node and its neighboring nodes. The document coupling fingerprint is obtained based on the Jaccard coefficient of the reference set, which is the ratio of the intersection to the union of two reference sets. The document coupling fingerprint is essentially a quantitative representation of the structural similarity between nodes.

[0047] During multi-hop traversal of the sparse metadata graph, as the path expands from the current node to adjacent nodes, the changes in coupling strength along the path are monitored in real time, and the rate of change of the coupling strength sequence is calculated. This rate of change refers to the magnitude of the change in coupling value between adjacent hops as the path expands, reflecting whether the path relevance remains stable. A structural entropy index is constructed based on the rate of change of the coupling strength sequence. Structural entropy measures the degree of uncertainty in structural information during path expansion. When the coupling strength decreases rapidly or fluctuations intensify, the structural entropy value increases, indicating that the path may have entered a semantic divergence region. The structural entropy index is compared with a preset threshold. If the structural entropy index is greater than the preset threshold, it indicates a jump to an irrelevant domain, which is judged as semantic drift, triggering a path soft truncation mechanism. The preset threshold is determined by statistically analyzing the distribution of structural entropy during multi-hop traversal on a historical retrieval sample set, combined with manually annotated relevant paths and semantic divergence paths as supervision signals. The threshold is selected using ROC curves or the F1 maximization criterion to best distinguish the inflection point between relevant expansion and domain drift, and can be adaptively calibrated according to different disciplines. The path soft truncation mechanism refers to not immediately terminating the traversal, but instead switching the traversal strategy from depth-first to breadth-first search. Depth-first is conducive to in-depth mining along strongly related paths, while breadth-first helps to re-expand potential highly related nodes in a local range. The path adaptive control is achieved through a dynamic switching mechanism. In scenarios without full text, this mechanism makes full use of the characteristics of citation behavior and accurately replaces the consistency of text semantics, thereby avoiding retrieval from the topic or getting caught in local structural noise.

[0048] After generating an initial candidate set through structural traversal, the results are filtered using a multi-objective collaborative ranking algorithm based on Pareto fronts. This algorithm abandons the traditional linear weighted summation formula and employs a multi-objective non-dominated ranking algorithm to optimize and filter candidate nodes. This multi-objective non-dominated ranking refers to finding a set of Pareto front nodes that are not completely superior to other candidates across all objective dimensions, rather than ranking them based on a single metric under multiple conflicting or independent objective functions. The multi-objective non-dominated ranking algorithm decomposes the retrieval objective into three mutually orthogonal objective functions: semantic relevance, structural centrality, and temporal novelty. The semantic relevance objective function is calculated using the cosine similarity between the query vector and the paper vector. The structural centrality objective function measures the structural influence of a node in the academic network using its PageRank value, degree centrality, or betweenness centrality. The temporal novelty objective function calculates a time decay score based on the difference between the publication time and the current time, reflecting the cutting-edge nature of the research. By performing non-dominated sorting on candidate nodes, a set of papers located on the three-dimensional Pareto front is output, that is, nodes that have high semantic matching, high structural influence and novelty are preferentially retained, forming the final recall result set.

[0049] By constructing a high-quality recall mechanism that combines structural constraints and multidimensional optimization, compared to traditional retrieval methods based solely on vector similarity, this mechanism strengthens structural consistency through citation coupling, controls path divergence risk through structural entropy, and balances semantic, structural, and temporal dimensions through multi-objective Pareto optimization. This enhances the topic concentration, academic authority, and cutting-edge nature of the recall results, fundamentally solving the Matthew effect problem in traditional ranking where highly cited older papers suppress less cited new discoveries. It ensures a balance between the diversity and authority of the result set, achieving high-quality, multidimensional recall driven by the synergy of structure and semantics.

[0050] The relative ranking of candidate nodes in the recall result set is adjusted based on the progressive implicit reasoning mechanism.

[0051] Furthermore, the progressive implicit inference mechanism includes: inserting a special token into the input sequence to trigger ranking inference; performing multiple rounds of pairwise comparison inference of papers in the hidden representation layer and generating an intermediate ranking score vector after each round of comparison; iteratively optimizing the intermediate scores using a list-based ranking loss function and gradually reducing the Softmax temperature parameter during the ranking process.

[0052] Specifically, after completing multidimensional recall and forming a candidate node set, in order to further improve the ranking accuracy and structural consistency of the results, a relative ranking adjustment process based on a progressive implicit reasoning mechanism is introduced to solve the problem that large models cannot directly handle the ranking of large-scale academic knowledge and improve the accuracy of relative ranking.

[0053] The progressive implicit inference mechanism includes encoding the paper titles, abstract vector representations, and key structural features of candidate nodes in the recall result set into a unified input sequence, and inserting a special token at the beginning of the sequence or a specified position to trigger ranking inference. This special token is a predefined functional control symbol in the retrieval model's vocabulary, which explicitly activates the ranking inference subnetwork within the retrieval model, switching it from a generation mode to a comparative inference mode. This guides the hidden layer attention mechanism to focus on modeling the relative merits of candidate nodes. The special token forces the retrieval model to perform multi-step academic influence comparative inference in the hidden representation layer. Each step generates an intermediate representation and a prediction score. As the number of inference steps increases, the retrieval model's understanding of the paper's motivation, methodology, and its position at the forefront of research becomes deeper and more accurate, addressing the problem of LLM's lack of awareness of new knowledge.

[0054] During the forward propagation process, the retrieval model performs multiple rounds of pairwise comparison reasoning of papers in the hidden representation layer. This involves constructing pairwise interactive representations between candidate papers and calculating the relative advantages of any two papers in dimensions such as semantic matching, structural importance, and temporal novelty using cross-attention or dual-tower differential encoding. After each round of comparison, an intermediate ranking score vector is output. This intermediate ranking score vector represents the relative score distribution of each candidate paper in the current round, reflecting the stage-wise ranking results. A list-based ranking loss function is then used for overall iterative optimization of the intermediate ranking score vector. Unlike point-based or pairwise ranking methods, the list-based ranking loss function optimizes the entire candidate list as the unit of optimization. By maximizing the likelihood probability of the true ranking order, it ensures that the score sequence generated by the retrieval model is consistent with the true order annotated by experts, thus avoiding the accumulation of local comparison errors. During the multiple rounds of iterative optimization, the Softmax temperature parameter is gradually reduced. This parameter is an adjustment factor that controls the smoothness of the probability distribution. A higher temperature results in a smoother output distribution, which is beneficial for exploring different ranking possibilities. As the temperature gradually decreases, the probability distribution becomes sharper, causing the ranking results to gradually converge and increasing determinism.

[0055] Through an exploratory and convergent progressive strategy, the retrieval model maintains ranking flexibility in the early stages, enhances the discriminative power of optimal candidates in the later stages, and ultimately outputs stable relative ranking results. The progressive implicit reasoning mechanism compensates for the fine-grained priority differences that still exist within the candidate optimal set formed by the preceding multi-objective non-dominated ranking. Multiple rounds of implicit reasoning strengthen the relative comparison ability between candidate nodes, enabling ranking to be dynamically optimized not only by static index calculations but also through semantic and structural comprehensive reasoning within the retrieval model. This shifts the optimization objective from absolute scores to precise alignment of the list order, placing more relevant and important documents first and improving the orderliness and usability of the retrieval results.

[0056] A logical alignment generation mechanism based on the probabilistic fusion of dynamic graph state machine and semantic topology is used to perform enhanced verification of relative ranking adjustment results and establish enhanced retrieval answers.

[0057] Furthermore, an enhanced verification of the relative ranking adjustment results is performed based on the logical alignment generation mechanism of dynamic graph state machine and semantic topology dual-space probability fusion to establish enhanced retrieval answers, including: when the generated content is detected to contain anchor entities in the graph, the K-hop neighbor subgraph of the corresponding anchor entity is activated in real time, and the K-hop neighbor subgraph is converted into a finite state automaton structure; during the decoding stage, a topological mask is applied to the output probability distribution, allowing only the generation of entities or relations that exist in the K-hop neighbor subgraph, and assigning negative infinite weights to entity candidates that do not exist in the K-hop neighbor subgraph.

[0058] Furthermore, the semantic-topological dual-space probabilistic fusion includes: performing Logits-level fusion of semantic space output probabilities and topological constraint masks; applying probability enhancement factors to legitimate paths based on citation strength or co-citation counts; and introducing a look-ahead verification module to verify subsequent reachability in the graph before selecting the next entity. If the predicted path enters a graph island, a cluster search backtracking mechanism is automatically triggered to re-plan and generate the path.

[0059] Specifically, after completing the relative ranking optimization, in order to further ensure that the generated content maintains logical consistency with the graph structure and ranking results, a logical alignment generation mechanism based on dynamic graph state machine and semantic topology dual-space probability fusion is introduced to enhance the verification of the ranking adjustment results and generate the final enhanced retrieval answer.

[0060] The dynamic graph state machine monitors the generated text stream in real time. When the decoding stage detects that the output sequence contains anchor entities from the graph, it identifies the position of these anchor entities in the graph in real time. An anchor entity refers to a specific paper ID or author name. Centered on the anchor entity, its corresponding K-hop neighbor subgraph is activated. The K-hop neighbor subgraph is a local subgraph structure formed by expanding outwards through K layers of reachable nodes from the anchor entity. It is used to limit the generation range and provide structural constraints. K is a positive integer and can be set according to actual needs. The text representation of the K-hop neighbor subgraph is converted into a finite state automaton (FSA) structure, where the entity states in the subgraph are considered as state nodes, and the relationship edges between entities are considered as state transition conditions, thus forming a dynamically traversable state transition machine. This transformation changes the graph structure from a static network to a dynamic graph state machine that can be updated in real time during the generation process, achieving structure-guided generation.

[0061] During the decoding process, a topological mask is applied to the output probability distribution. This topological mask is a binary mask matrix constructed based on the set of legal transitions in the current state machine. It only allows the generation of entities or relations existing in the K-hop neighbor subgraph, assigning negative infinite weights to candidate entities or relations not belonging to the K-hop neighbor subgraph, causing their probabilities to approach zero after Softmax computation. This forces the generated path to fall within the allowed topological range of the graph, thus physically blocking the possibility of generating non-existent reference relationships. Through this process, the generated content is no longer freely sampled but rather probabilistically decided under structural constraints.

[0062] Semantic-topological dual-space probabilistic fusion includes: performing Logits-level fusion of the semantic space output probability and the topological constraint mask. This Logits-level fusion refers to directly weighting and summing the semantic prediction score and the topological constraint score of the retrieval model before Softmax, rather than simply multiplying them at the probability layer, thus avoiding information loss during probability normalization. Simultaneously, probability enhancement factors are applied to legitimate paths based on citation strength or co-citation counts. Citation strength typically represents the frequency or weight of the citation relationship between two entities, while co-citation counts represent the number of times two papers are jointly cited by third-party literature. These structural statistical indicators are used as path credibility enhancement factors, and high-frequency citation or high-co-occurrence paths are probabilistically weighted to improve the academic rationality of the generated paths.

[0063] To balance the fluency (semantic space) and realism (topological space) of the generated data, the final output is corrected using the following formula: ,in, This represents the original output score of the retrieval model before the introduction of graph constraints. It is the unnormalized probability value (the score before Softmax) calculated by the retrieval model based on the semantic context, and is used to reflect the generation preferences of the retrieval model itself. This represents the graph topological constraint mask. In hard constraint mode, it assigns a value of 0 to legal entities or relational paths existing in the graph, and assigns a value of negative infinity to candidate paths that do not exist or are illegal in the graph. Its function is to directly shield illegal generated paths at the Logits level, thereby achieving structural mandatory constraints. In open generation tasks such as academic evaluation, it automatically switches to soft constraint mode and no longer uses negative infinity hard truncation. The relevance score of a path or candidate token can be calculated based on citation strength, such as co-citation count, citation frequency, or semantic matching degree. It measures the academic influence or structural credibility of the current candidate within the graph structure. β represents the adjustment coefficient of the graph structure constraint, used to control the influence of the topological mask on the final output Logits. When β is large, the structural constraint has a stronger control over the generated results; when β approaches 0, the retrieval model relies more on its own semantic generation. γ represents the relevance enhancement factor or citation strength scaling coefficient, used to adjust the weighted influence of the structure score on Logits. In soft constraint mode, γ is usually positively correlated with citation strength and is used to amplify the probability of highly co-cited or highly influential paths. This modified formula achieves the fusion of three spaces: semantic generation score, graph structure constraint, and academic influence score. By dynamically adjusting the weights of different signals through parameters β and γ, it achieves strong control over path legitimacy under hard constraint mode and structural preference guidance under soft constraint mode. This ensures that the generated enhanced search results not only conform to the graph topology but also take into account academic relevance and content rationality. It realizes the enhanced verification of relative ranking adjustment results by the logical alignment generation mechanism based on the probabilistic fusion of dynamic graph state machine and semantic topology.

[0064] Furthermore, a look-ahead verification module is introduced. Before selecting the next entity, this module predicts and analyzes the reachability of subsequent steps, specifically by predicting whether the current path will enter a graph island with no subsequent connections through graph traversal. If the prediction result indicates that the path will enter an isolated subgraph or the set of reachable nodes is empty, a beam search backtracking mechanism is automatically triggered. This mechanism involves retaining multiple candidate path branches based on Beam Search. When a path is about to fail, it reverts to the alternative paths to replan the generation direction, ensuring that the generated argument is logically complete and closed, thereby improving generation stability and global connectivity, avoiding entering graph islands, and improving the rationality and fluency of the generated answer. Through a logical alignment generation mechanism that fuses dynamic graph state machines with semantic topology probabilistics, a dynamic balance is achieved between structural constraints and semantic generation, ultimately outputting an enhanced retrieval answer after structural verification and probabilistic optimization.

[0065] By embedding graph structure constraints deep into the generation stage, consistency verification and dynamic correction between the sorting results and the graph topology are achieved, thereby preventing the retrieval generation from being freely generated without regard to the graph facts. Through state machine control, topology masking and probability fusion mechanisms, the traceability and structural legality of the generation path are jointly guaranteed, further improving the accuracy and effectiveness of the retrieval enhancement response, and thus improving the quality of academic knowledge retrieval results.

[0066] Example 2 is based on the same inventive concept as the academic knowledge retrieval enhancement generation method based on multidimensional citation relationship graphs in the previous examples, such as... Figure 2 As shown, this application provides an academic knowledge retrieval enhancement generation system based on a multidimensional citation graph, wherein the academic knowledge retrieval enhancement generation system based on a multidimensional citation graph includes: The data graph construction module 11 is used to construct a sparse metadata graph with data-semantic-temporal enhancement after reading the metadata of academic literature; the candidate node recall module 12 is used to recall candidate nodes of academic literature from the sparse metadata graph and establish a recall result set when performing online retrieval, based on citation coupling and Pareto front multidimensional retrieval technology; the ranking adjustment module 13 is used to perform relative ranking adjustment of candidate nodes in the recall result set based on a progressive implicit reasoning mechanism; and the enhancement verification module 14 is used to perform enhancement verification of the relative ranking adjustment results based on a logical alignment generation mechanism that fuses dynamic graph state machine and semantic topology dual-space probabilistics, and establish an enhanced retrieval answer.

[0067] Furthermore, the data graph construction module 11 is also used to: parse academic literature metadata into paper entities, and construct author association relationships and journal affiliation relationships based on the paper entities, so that paper nodes form a first-layer academic structure network through author edges and journal edges respectively; on the basis of the first-layer academic structure network, abstract the scientific data set associated with the paper into independent scientific data nodes, and establish data association edges between the paper nodes and the corresponding scientific data nodes; for multiple paper nodes pointing to the same scientific dataset, form a bidirectional bridging edge between paper and data by sharing scientific data nodes; set a data homology identifier attribute for the bidirectional bridging edge in the graph structure, and form an implicit association channel across the citation chain by introducing scientific data nodes.

[0068] Furthermore, the data graph construction module 11 is also used to construct data association edges through edge weight calculation, as follows: ;in, Representing edge weights, and These are the weighting coefficients. , Characterizing semantic cosine similarity, This is a domain-adaptive parameter used to represent the average citation half-life of journals in the corresponding subject area over 5 years. Characterizes the time difference.

[0069] Furthermore, the data graph construction module 11 is also used to: calculate the semantic similarity between isolated nodes based on the Sentence-BERT vectors of keywords and summaries for all isolated nodes; identify node pairs that do not have a reference path and whose semantic distance is less than a preset threshold based on the semantic similarity; insert implicit association edges marked as virtual types between the node pairs, and configure the confidence weight of the implicit association edges to construct implicit association channels.

[0070] Furthermore, the candidate node recall module 12 is also used to: calculate the document coupling fingerprint between the current node and its neighboring nodes during the multi-hop traversal process, wherein the document coupling fingerprint is obtained based on the Jaccard coefficient of the reference set; calculate the rate of change of the coupling strength sequence in real time during the path expansion process, and construct a structural entropy index based on the calculation results; and trigger a path soft truncation mechanism when the structural entropy index exceeds a preset threshold, switching the traversal strategy from depth-first to breadth-first.

[0071] Furthermore, the candidate node recall module 12 is also used to: decompose the retrieval task into three mutually orthogonal objective functions through a multi-objective non-dominated ranking algorithm, wherein the objective functions include a semantic relevance objective function, a structural centrality objective function, and a temporal novelty objective function, and establish a recall result set.

[0072] Furthermore, the sorting adjustment module 13 is also used to: insert a special identifier Token for triggering sorting inference into the input sequence; perform multiple rounds of pairwise comparison inference of papers in the hidden representation layer, and generate an intermediate sorting score vector after each round of comparison; iteratively optimize the intermediate score using a list-style sorting loss function, and gradually reduce the Softmax temperature parameter during the sorting process.

[0073] Furthermore, the enhanced verification module 14 is also used to: when the generated content is detected to contain anchor entities in the graph, activate the K-hop neighbor subgraph of the corresponding anchor entity in real time, and convert the K-hop neighbor subgraph into a finite state automaton structure; apply a topological mask to the output probability distribution during the decoding stage, allowing only the generation of entities or relations that exist in the K-hop neighbor subgraph, and assigning negative infinite weights to entity candidates that do not exist in the K-hop neighbor subgraph.

[0074] Furthermore, the enhanced verification module 14 is also used to: perform Logits-level fusion of semantic space output probability and topological constraint mask; apply probability enhancement factor to legitimate paths based on citation strength or co-citation count; introduce a look-ahead verification module to verify subsequent reachability in the graph before selecting the next entity; if the predicted path enters a graph island, automatically trigger the cluster search backtracking mechanism to re-plan and generate the path.

[0075] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The academic knowledge retrieval enhancement generation method and specific examples based on multidimensional citation relationship graphs in the foregoing embodiment one are also applicable to the academic knowledge retrieval enhancement generation system based on multidimensional citation relationship graphs in this embodiment. Through the foregoing detailed description of the academic knowledge retrieval enhancement generation method based on multidimensional citation relationship graphs, those skilled in the art can clearly understand the academic knowledge retrieval enhancement generation system based on multidimensional citation relationship graphs in this embodiment. Therefore, for the sake of brevity, it will not be described in detail here.

[0076] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0077] Obviously, those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. An academic knowledge retrieval enhancement generation method based on multidimensional citation relationship graphs, characterized in that, The method includes: After reading the metadata of academic literature, a sparse metadata graph with data-semantic-temporal enhancement is constructed; When performing online retrieval, multidimensional retrieval technology based on citation coupling and Pareto front is used to recall candidate nodes of academic literature from the sparse metadata graph and establish a recall result set. The relative ranking of candidate nodes in the recall result set is adjusted based on the progressive implicit reasoning mechanism. A logical alignment generation mechanism based on the probabilistic fusion of dynamic graph state machine and semantic topology is used to perform enhanced verification of relative ranking adjustment results and establish enhanced retrieval answers.

2. The academic knowledge retrieval enhancement generation method based on multidimensional citation graph as described in claim 1, characterized in that, Constructing a ternary augmented sparse metadata graph of data, semantics, and time series, including: The metadata of academic documents is parsed into paper entities, and author association and journal affiliation relationships are constructed based on the paper entities, so that the paper nodes form the first layer of academic structure network through author edges and journal edges respectively. Based on the first-layer academic structure network, the scientific data set associated with the paper is abstracted into independent scientific data nodes, and data association edges are established between the paper node and the corresponding scientific data node. For multiple paper nodes pointing to the same scientific dataset, a bidirectional bridge edge is formed between paper and data nodes by sharing scientific data nodes; In the graph structure, a data origin identification attribute is set for the bidirectional bridging edge, and an implicit association channel across the reference chain is formed by introducing scientific data nodes.

3. The academic knowledge retrieval enhancement generation method based on multidimensional citation relationship graph as described in claim 2, characterized in that, Data association edges are constructed by calculating edge weights, as follows: ； in, Representing edge weights, and These are the weighting coefficients. , Characterizing semantic cosine similarity, This is a domain-adaptive parameter used to represent the average citation half-life of journals in the corresponding subject area over 5 years. Characterizes the time difference.

4. The academic knowledge retrieval enhancement generation method based on multidimensional citation relationship graph as described in claim 2, characterized in that, Implicit links across citation chains are formed by introducing scientific data nodes, including: For all isolated nodes, semantic similarity between isolated nodes is calculated based on Sentence-BERT vectors of keywords and summaries; Based on the semantic similarity, node pairs that do not have a reference path and whose semantic distance is less than a preset threshold are identified. Implicit association edges labeled as virtual type are inserted between the node pairs, and the confidence weights of the implicit association edges are configured to construct implicit association channels.

5. The academic knowledge retrieval enhancement generation method based on multidimensional citation relationship graph as described in claim 1, characterized in that, From the sparse metadata graph, candidate nodes of academic documents are recalled, and a recall result set is established, including: During the multi-hop traversal, the document coupling fingerprint between the current node and its neighboring nodes is calculated, and the document coupling fingerprint is obtained based on the Jaccard coefficient of the reference set. The rate of change of coupling strength sequence is calculated in real time during the path expansion process, and a structural entropy index is constructed based on the calculation results; If the structural entropy index exceeds a preset threshold, a soft path truncation mechanism is triggered, switching the traversal strategy from depth-first to breadth-first.

6. The academic knowledge retrieval enhancement generation method based on multidimensional citation relationship graph as described in claim 5, characterized in that, The retrieval task is decomposed into three mutually orthogonal objective functions using a multi-objective non-dominated ranking algorithm. These objective functions include a semantic relevance objective function, a structural centrality objective function, and a temporal novelty objective function, thereby establishing a recall result set.

7. The academic knowledge retrieval enhancement generation method based on multidimensional citation relationship graph as described in claim 1, characterized in that, Progressive implicit reasoning mechanisms include: Insert a special token into the input sequence to trigger sorting inference; Perform multiple rounds of pairwise comparison reasoning of papers in the hidden representation layer, and generate an intermediate sorted score vector after each round of comparison; A list-based sorting loss function is used to iteratively optimize the intermediate scores, and the Softmax temperature parameter is gradually reduced during the sorting process.

8. The academic knowledge retrieval enhancement generation method based on multidimensional citation relationship graph as described in claim 1, characterized in that, A logical alignment generation mechanism based on the probabilistic fusion of dynamic graph state machine and semantic topology is used to perform enhanced verification of relative ranking adjustment results, and to establish enhanced retrieval answers, including: When the generated content is detected to contain anchor entities in the graph, the K-hop neighbor subgraph of the corresponding anchor entity is activated in real time, and the K-hop neighbor subgraph is converted into a finite state automaton structure. During the decoding phase, a topological mask is applied to the output probability distribution, allowing only entities or relations that exist in the K-hop neighbor subgraph to be generated, and assigning negative infinite weights to entity candidates that do not exist in the K-hop neighbor subgraph.

9. The academic knowledge retrieval enhancement generation method based on multidimensional citation relationship graph as described in claim 7, characterized in that, Semantic topological dual-space probabilistic fusion includes: The semantic space output probability is fused with the topological constraint mask at the Logits level. Apply a probability boosting factor to legitimate paths based on citation strength or co-citation count; A look-ahead verification module is introduced to verify the subsequent reachability in the graph before selecting the next entity. If the predicted path enters a graph island, the cluster search backtracking mechanism is automatically triggered to replan and generate a new path.

10. An academic knowledge retrieval enhancement generation system based on multidimensional citation relationship graphs, characterized in that: The steps for implementing the academic knowledge retrieval enhancement generation method based on multidimensional citation graphs as described in any one of claims 1 to 9 include: The data graph construction module is used to construct a sparse metadata graph with data-semantic-temporal enhancement after reading the metadata of academic literature; The candidate node recall module is used to recall academic literature candidate nodes from the sparse metadata map and establish a recall result set when performing online retrieval, based on citation coupling and Pareto front multidimensional retrieval technology. The sorting adjustment module is used to perform relative sorting adjustment on candidate nodes in the recall result set based on a progressive implicit reasoning mechanism. The enhanced verification module is used to perform enhanced verification of the relative ranking adjustment results based on the logical alignment generation mechanism of dynamic graph state machine and semantic topology dual-space probability fusion, and to establish enhanced retrieval answers.

Images