Literature explainable recommendation method and system based on knowledge graph and large model

Abstract

The application relates to the technical field of literature recommendation, and discloses a literature explainable recommendation method and system based on a knowledge graph and a large model. A fine-grained field knowledge graph is constructed by fusing paper metadata, technical entities, citation contexts and multi-modal content, the paper is refined from a coarse-grained description by a unit of an article into a structured expression of "technical entity-relation", deep semantic analysis and computable measurement of the technical contribution and citation motivation of the paper are realized, a time decay correction factor based on publication time and cited information fitting is further introduced, the cross-year calibration of the influence potential score is carried out to weaken the path dependence and "Matthew effect" caused by citation accumulation. Therefore, the influence evaluation is more traceable in the semantic evidence level, the attribution is more objective, and the influence evaluation is more comparable in the time dimension, so that the objectivity and accuracy of the paper recommendation and influence evaluation are improved as a whole.

Description

Technical Field

[0001] This invention relates to the field of literature recommendation technology, specifically to a method and system for interpretable literature recommendation based on knowledge graphs and large models. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] With the deepening of interdisciplinary integration, related research results have experienced explosive growth; for example, core journals in the textile field now publish tens of thousands of articles annually. This surge in output makes cutting-edge and high-value innovations more susceptible to being drowned out by information noise, placing a heavy burden on researchers, reviewers, and journal editors in terms of topic selection, research outcome assessment, and research decision-making. Therefore, there is an urgent need for evaluation and recommendation methods that are more closely aligned with the characteristics of the textile discipline and can more accurately identify academic contributions and innovative value, in order to improve the efficiency of research decision-making and academic dissemination.

[0004] Existing paper recommendation and impact assessment mainly rely on two paths: one is bibliometric indicators, and the other is general deep learning or graph model recommendation algorithms, but both have key technical problems that affect the identification of true value. First, the core of econometric indicators, such as citation frequency and impact factor, stems from a lack of understanding of the paper's content and citation context. It fails to distinguish citation motivations (support, comparison, negation, methodological borrowing, etc.) and their specific technical contributions within the textile field, resulting in coarse evaluation granularity and weak value analysis. Second, cross-year comparisons are prone to time unfairness due to the significant lag and path-dependent effect of citation accumulation. Earlier papers tend to consistently achieve higher indicators due to accumulated advantages, creating a "Matthew effect" that systematically underestimates recent high-quality achievements. Third, general recommendation models are limited in vertical disciplines because they rely heavily on general metadata and relational modeling, making it difficult to characterize the unique technical entities, process chains, and disciplinary knowledge structures of the textile field. They also struggle to fully utilize the numerous charts, structural diagrams, and other multimodal content in papers that carry key information. Furthermore, deep models often lack interpretability because their decision-making process is difficult to trace back to clear semantic evidence and reasoning, leaving users without explanatory details and limiting their application in demanding scenarios such as review and editorial decisions. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes a document-interpretable recommendation method and system based on knowledge graphs and large-scale models. By constructing a fine-grained domain knowledge graph integrating paper metadata, technical entities, citation context, and multimodal content, it refines the coarse-grained description of papers into a structured expression of "technical entities-relationships," enabling deep semantic analysis and computable measurement of the technical contributions and citation motivations of papers. Furthermore, it introduces a time decay correction factor based on a fit between publication time and citation information to calibrate the impact potential score across years, mitigating path dependence and the "Matthew effect" caused by citation accumulation. This makes impact evaluation more traceable at the semantic evidence level, more objective in attribution, and more comparable over time, thereby improving the overall objectivity and accuracy of paper recommendation and impact evaluation.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: One or more embodiments provide a document-interpretable recommendation method based on knowledge graphs and large models, comprising the following steps: Acquire paper data in the target domain, extract semantic features from the unstructured text in the paper data, extract paper entities, paper technical entities and relationships, and construct a fine-grained multimodal domain knowledge graph; For the paper data in the target field and the constructed domain knowledge graph, extract the paper feature vectors, and calculate the influence potential score of each paper based on the paper feature vectors; For historical papers, a decay factor λ is fitted and calculated based on the paper's publication time and citation information. An exponential decay term is then calculated based on the decay factor λ to calibrate the impact potential score of historical papers, resulting in a time-faired impact potential score. Based on the impact potential scores of the obtained papers, a ranking learning model is used to rank the papers and output a set of candidate papers of a predetermined number as the recommendation result.

[0007] One or more embodiments provide a document-interpretable recommendation system based on knowledge graphs and large models, including: The fine-grained multimodal domain knowledge graph construction module is configured to acquire paper data in the target domain, extract semantic features from the unstructured text in the paper data, extract paper entities, paper technical entities and relationships, and construct a fine-grained multimodal domain knowledge graph. The scoring module is configured to extract feature vectors from papers in the target domain and the constructed domain knowledge graph, and calculate the impact potential score of each paper based on the feature vectors. The score correction module is configured to calculate a decay factor λ based on the publication time and citation information of historical papers, calculate an exponential decay term based on the decay factor λ, and calibrate the influence potential score of historical papers to obtain a time-faired influence potential score. The recommendation module is configured to rank the papers based on their impact potential scores using a ranking learning model and output a preset set of candidate papers as the recommendation result.

[0008] An electronic device includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the steps in the above-described document-interpretable recommendation method based on knowledge graphs and large models.

[0009] A computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the steps in the above-described document-interpretable recommendation method based on knowledge graphs and large models.

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention integrates metadata, textual semantic structure, and citation network structure within the same knowledge graph, and further refines the paper content into a semantically structured expression of "technical entities-relationships." This allows recommendations and evaluations to move beyond coarse-grained citation counts and differentiate between various technical elements, process chains, and citation contexts within a domain at a fine-grained level, thereby enhancing the ability to analyze and measure the deep contributions of papers. By introducing a decay factor λ based on publication time and citation information, and applying exponential decay calibration to historical paper scores, path dependence caused by citation accumulation can be suppressed when ranking across years, giving recent high-quality results a fairer opportunity to be displayed, thus mitigating the "Matthew effect" in cross-period evaluation. By integrating text, image, and graph structural features and employing graph attention networks for domain-adaptive modeling, the unique technological connections and innovative value of the target domain can be captured more accurately.

[0011] The advantages of the present invention, as well as its additional advantages, will be described in detail in the following specific embodiments. Attached Figure Description

[0012] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute a limitation thereof.

[0013] Figure 1 This is the first flowchart of the document-interpretable recommendation method of Embodiment 1 of the present invention; Figure 2This is a schematic diagram of the historical document influence potential score calibration process in Embodiment 1 of the present invention; Figure 3 This is the second flowchart of the document-interpretable recommendation method of Embodiment 1 of the present invention. Detailed Implementation

[0014] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0015] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0016] It should be noted that the terminology used herein is for describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof. It should be noted that, without conflict, the various embodiments and features within those embodiments can be combined with each other. The embodiments will now be described in detail with reference to the accompanying drawings.

[0017] Explanation of technical terms: (1) Heterogeneous academic data sources: refers to collections of academic data with different sources and structures, including paper metadata, abstract text, citation relationships, author information, etc.

[0018] (2) Metadata: refers to the set of data fields that describe the basic information of a paper, such as title, author, keywords, publication time, etc.

[0019] (3) Knowledge Graph (KG) / Academic Knowledge Graph: Graph structure data consisting of entities and relations, used to represent papers, authors, institutions, technical entities and their associations.

[0020] (4) Entities: refers to objects that can be identified and standardized in academic texts and metadata, such as papers, authors, institutions, technical terms, etc.

[0021] (5) Relationship: refers to the type of association between entities, such as citation, co-citation, collaboration (co-authorship), etc.

[0022] Knowledge extraction: refers to the process of identifying entities and relationships from unstructured academic texts and outputting them in a structured manner.

[0023] (6) Graph database (Neo4j): A database system used for storing and querying knowledge graphs; this invention uses Neo4j as an example.

[0024] (7) Graph Neural Network (GNN): A neural network framework for processing graph structure data; this invention is used in the influence quantification process.

[0025] (8) Graph Attention Network (GAT): A graph neural network based on attention mechanism, used to aggregate neighbor information and learn paper influence representation.

[0026] (9) Semantic embedding features: refers to the feature representation of encoding text such as paper abstracts into vectors.

[0027] (10) Learning to Rank: A supervised learning method aimed at ranking quality, used to rank candidate papers by influence score.

[0028] (11) XGBoost Ranking: Ranking implementation based on gradient boosting tree. This invention is used for annual paper candidate ranking.

[0029] (12) Structured Prompt: The candidate paper information and evidence chain are organized into input prompts according to preset fields to constrain the generation of large models and the basis of conclusions.

[0030] Example 1 In one or more of the technical solutions disclosed in the embodiments, such as Figures 1 to 3 As shown, a document-interpretable recommendation method based on knowledge graphs and large models includes the following steps: S1. Obtain paper data in the target domain, extract semantic features from the unstructured text in the paper data, extract paper entities, paper technical entities and relationships, and construct a fine-grained multimodal domain knowledge graph. S2. Extract feature vectors from the paper data and the constructed domain knowledge graph for the target domain, and calculate the influence potential score of each paper based on the feature vectors. S3. For historical papers, the decay factor λ is calculated based on the publication time and citation information of the paper. The exponential decay term is calculated based on the decay factor λ to calibrate the influence potential score of historical papers and obtain the time-faired influence potential score. S4. Based on the obtained impact potential scores of the papers, the papers are ranked using a ranking learning model, and a set of candidate papers in a preset number of candidate paper lists or recommended ranking lists is output as the recommendation results. Explainable information is obtained by reasoning based on a large language model based on the recommendation results.

[0031] In the above implementation, by integrating metadata, textual semantic structure, and citation network structure within the same knowledge graph, and further refining the paper content into a semantically structured expression of "technical entities-relationships," recommendations and evaluations no longer rely solely on coarse-grained citation counts. Instead, they can distinguish different technical elements, process chains, and citation contexts within the domain at a fine-grained level, thereby enhancing the ability to analyze and measure the deep contributions of papers. By introducing a decay factor λ based on publication time and citation information, and applying exponential decay calibration to historical paper scores, path dependence caused by citation accumulation can be suppressed when ranking across years, giving recent high-quality results a fairer opportunity to be displayed, thus mitigating the "Matthew effect" in cross-period evaluation. By integrating text, image, and graph structural features and employing graph attention networks for domain-adaptive modeling, the unique technological connections and innovative value of the target domain can be captured more accurately.

[0032] S1. Obtain paper data in the target domain, extract semantic features from the unstructured text in the paper data, extract paper entities, technical entities and relationships, and construct a fine-grained multimodal domain knowledge graph, including: S11. Obtain paper data in the target field, wherein the paper data includes at least the structured metadata, unstructured text and citation information of each paper; Optionally, incremental Python crawlers based on Selenium can be used to obtain paper data in the target field from academic platforms such as CNKI. This paper data includes at least: structured metadata for each paper, unstructured text for each paper, and citation information for each paper. It can also record incremental information such as update time, batch number, and source identifier to support continuous updates and backtracking.

[0033] The structured metadata is the basic information of the target paper, which may include title, author, author affiliation, journal or conference name, year, volume and issue, page number, keywords, identifiers or numbers, such as unique identifiers like DOI / document number / URL, etc. Unstructured text refers to the continuous text in a paper that exists as natural language, without a fixed field structure, and requires NLP / large models to understand before information can be extracted. It can include the abstract, introduction, main body paragraphs, and paragraphs under chapter titles. Unstructured text consists of free text segments that require entity extraction, relation extraction, and numerical unit extraction to be converted into structured "entity-relationship" structures.

[0034] Citation information includes the number of times a paper has been cited, a list of references, a list of cited papers, and citation edges. Citation edges specifically refer to the citation relationships between papers, such as the relationship of "Paper A - Citation - Paper B". Optionally, data quality can also be ensured by using rule-based cleaning and deduplication mechanisms to clean the obtained papers in the target field; S12. Use a large language model to extract semantic information from unstructured text, extract entities and relations from academic text based on prompt templates, and output the technical entities and relations at the technical level of the paper content. Optionally, this embodiment employs a two-step approach: pre-training a large model and fine-tuning using prompting engineering. First, leveraging the semantic understanding capabilities of open-source large models such as ChatGLM and BGE, preliminary entity and relation extraction is performed from the academic text of the paper using prompt templates. Then, based on the sampling evaluation results, a small-scale, precisely labeled dataset is constructed for domain-specific terms and complex relationships. The model is then fine-tuned using lightweight methods to iteratively improve extraction accuracy.

[0035] S13. Based on technical entities and relationships, structured metadata and citation information, extract entities and relationships, construct a domain knowledge graph containing paper nodes, technical entity nodes and their associated relationship edges, obtain a fine-grained multimodal domain knowledge graph, and store it in a graph database. Specifically, paper nodes generated through structured metadata can include attributes such as title, author, and institution; technical entity nodes can be generated by entity extraction from the semantic information of unstructured text; technical entity association edges can be generated through relation extraction; and citation relationship edges between papers can be generated through citation information.

[0036] Optionally, the constructed domain knowledge graph can be stored based on graph databases such as Neo4j, and efficient relational reasoning and path query can be achieved using native query languages, thereby supporting subsequent graph applications and analysis.

[0037] This step, by constructing a fine-grained multimodal knowledge graph in the textile field, enables precise analysis and measurement of the deep semantic content and citation motivations of academic papers. It addresses the lack of fine-grained semantic discrimination capabilities in traditional measurement methods. Specifically, it simultaneously expresses metadata, textual semantic structure, and citation network structure within the same knowledge graph. The textual semantic structure includes multimodal information such as images and text, enabling the fusion modeling of these three types of heterogeneous information. By extracting entities and relationships from the unstructured text and constructing a domain knowledge graph containing paper nodes, technical entity nodes, and their associated edges based on the extraction results, this embodiment refines the existing coarse-grained organization method based on papers into a semantically structured expression centered on technical entities and relationships extracted from unstructured text, achieving deep analysis and computable modeling of paper content. This not only explicitly depicts the key technical elements within the paper and their interrelationships in the graph, but also provides more fine-grained and traceable knowledge support for subsequent academic evaluation, recommendation, and analysis.

[0038] This embodiment overcomes the semantic limitations of traditional measurement indicators, significantly improving the accuracy and domain adaptability of identifying high-value papers in the textile field. It constructs a fine-grained, multimodal knowledge graph of the textile field, utilizes a large model to extract unstructured semantic information from papers, and combines a graph attention network (GAT) to aggregate the structural features of neighboring nodes. This transforms the evaluation system from a simple "counting" approach to a deep understanding of "content and relationships," effectively identifying key papers on the core evolutionary path of technology and solving the problem of poor adaptability of general models in vertical domains.

[0039] S2. Extract feature vectors from the paper data and the constructed domain knowledge graph for the target domain, and calculate the influence potential score of each paper based on the feature vectors. Optionally, the paper feature vector includes semantic embedding features obtained from the paper text, quantitative features obtained from the paper's metadata and citation information, and topological features obtained from the domain knowledge graph; the paper feature vector can be represented as: ; Where Xi represents the comprehensive feature vector of paper node i; E is the vector concatenation symbol. sem (di) is the semantic embedding of the summary extracted using Sentence-BERT; E metric (ni) is a normalized metric for extracting traditional paper metadata. It uses an incremental Python crawler based on Selenium to obtain information about target-domain paper data from academic platforms such as CNKI, such as citation count and download count per paper; E topo (gi) are topological features extracted from the knowledge graph, used to characterize the position of a paper in the entire academic network (graph).

[0040] One specific implementation method for calculating the impact potential score of each paper based on its feature vectors includes the following steps: S21. Construct a paper feature vector for each paper based on the knowledge graph; The paper feature vector includes at least: semantic embedding features obtained from the paper text, quantitative features obtained from the metadata and citation information, and topological structure features obtained from the knowledge graph; S22. Obtain the paper ranking expert annotation dataset obtained from expert annotations, and use the paper feature vectors and the labels constructed in the expert annotation dataset to train the graph attention network (GAT) model to output the influence potential score of each paper. Optionally, taking the textile field as an example, a dataset of expert-annotated papers from the 2015-2023 "Journal of Textile Research" can be constructed by collecting back-to-back reviews and group discussions from senior experts in the textile field based on multiple dimensions such as originality, technical depth, industry attention, and dissemination.

[0041] A multi-layer graph attention network (GAT) model is constructed. This model takes a knowledge graph as input and uses an attention mechanism to aggregate the structural information of node attributes and multi-hop neighbors. At the same time, it considers edge features to optimize attention weights. The attention weight allocation logic is shown in formula (2). The multi-layer graph attention network (GAT) model uses a list-style loss function, as shown in formula (3), to drive its output paper ranking to be highly consistent with the expert ranking. The parameter learning is completed through iterative optimization on the training set, validation set, and test set.

[0042] Where, α ij Let represent the influence contribution weight of neighboring paper j to target paper i; W is the learnable state transition matrix; a is the learnable attention weight vector, and the arrow above indicates the vector; a h represents the transpose of vector a; i h represents the node feature vector of the target paper i; j h represents the node feature vector of neighboring paper j; k Let represent the feature vector of the node with neighboring paper k; || represents the vector concatenation operation; LeakyReLU represents the nonlinear activation function; exp represents the exponential function with the natural constant e as the base; N i Let k represent the set of all first-order neighbor nodes of the target paper i; k∈N i This indicates traversing set N. i Each neighbor node k in the model. Through this mechanism, the model can automatically identify key citation relationships in complex academic networks.

[0043] The list-style loss function used for training the multi-layer graph attention network (GAT) model is as follows: (3); Among them, L(y, ) represents the list-style loss value between the model's predicted ranking and the actual expert ranking; y represents the actual expert ranking sequence; This represents the predicted paper ranking sequence output by the multi-layer graph attention network (GAT) model; n represents the total number of papers in the paper list participating in the ranking calculation; i is the index number of the paper in the list; s i P represents the i-th paper in the candidate list; y (s i ) represents the i-th paper s determined based on the ranking of real experts y.i The true probability distribution of P; (s i () indicates ranking based on model prediction The calculated i-th paper s i The predicted probability distribution; ∑ represents the summation of the loss values ​​of all n papers in the list; log represents the natural logarithm function.

[0044] This embodiment integrates expert knowledge with objective data to establish an influence quantification model that better aligns with disciplinary evaluation standards. A "gold standard" dataset annotated by experts was constructed to train a graph attention network (GAT). A list-based loss function was used to optimize the model parameters, enabling the model to not only learn the statistical patterns of the data but also internalize the implicit evaluation standards of textile experts regarding originality and technological breakthroughs. This ensures that the algorithm's ranking results are highly consistent with the perceptions of top experts, effectively eliminating evaluation noise caused by relying solely on citation counts.

[0045] A time decay mechanism is designed to balance long-term accumulation and recent performance. In step S3, for historical papers, a decay factor λ is calculated based on the paper's publication date and citation information. Figure 2 As shown, it includes the following steps: Step S31: Sample a predetermined number of papers from the paper collection corresponding to the constructed domain knowledge graph; Step S32: Obtain the annual citation sequence of each paper sample and count the historical peak citation count of the paper; Step S33: The peak number of citations is approximated by multiplying it by an exponential decay term to obtain the predicted number of citations. The goal is to minimize the sum of the squared differences between the actual number of citations and the predicted number of citations. The time decay factor λ is obtained by nonlinear least squares fitting. Furthermore, the exponential decay term is formulated as follows: ; in, As the attenuation factor, This is the time difference between the publication of the paper and the present. The objective function is to minimize the sum of the squared differences between the actual number of citations and the predicted number of citations. (4); Among them, C p,t This represents the actual number of times paper p was cited in year t. This is its historical peak number of citations.

[0046] Specifically, 2000–3000 papers are randomly selected from the knowledge graph, and their citation history is traced using knowledge graph and large model technology. The citation frequency of each paper is counted year by year and historical peak. Then, the optimal attenuation factor λ value is obtained by fitting the data using the nonlinear least squares method. Next, the original measurement standard score is calibrated according to the decay function. The dynamic time decay influence quantification formula is shown in (5), which realizes the dynamic calibration of the influence of historical papers.

[0047]

[0048] Among them, C p,t This represents the actual number of times paper p was cited in year t. This is its historical peak number of citations.

[0049] In step S3, the influence potential score of historical papers is calibrated to obtain the time-faired influence score I. adj :

[0050] Among them, I raw I represents the influence potential score obtained from the original metrics using the GAT model. adj t0 represents the impact potential score corresponding to the original indicator after time fairness, where t0 is the publication time of the paper and λ is the time decay factor.

[0051] In this embodiment, the inherent advantage of older papers is weakened by weighting the indicators through an exponential decay term. This achieves a dynamic and fair evaluation of the influence of papers across different time periods, effectively overcoming the "Matthew effect" in academic evaluation. By introducing an exponential decay function and using nonlinear least squares to fit historical data to determine the optimal decay factor, the static quantitative indicators of papers are dynamically and temporally weighted and calibrated. While preserving the value of historical classic papers, the weight of recently active results is significantly increased, thus achieving a fair comparison of papers published in different years.

[0052] In step S4, based on the influence potential score of each paper obtained in step S2 and the influence potential score of historical papers corrected in S3, the influence of the papers can be ranked. Specifically, based on the influence potential score of the obtained papers, the papers can be ranked by a ranking learning model to obtain the ranking order of the paper influence, and a preset number of candidate paper lists or recommended ranking lists can be selected.

[0053] Another possible approach is to sort the papers published this year and the papers published in previous years separately, thus obtaining recommended lists of influential papers published this year and influential papers published in previous years. Specifically, for the selection of annual papers, based on the influence potential score extracted by the trained GAT model, the papers published in the same year are ranked by the ranking learning model to obtain the paper influence ranking order, and a preset number of candidate paper lists or recommended ranking lists can be selected and output.

[0054] Optionally, the ranking learning model can be the XGBoost Ranking model; Specifically, regarding the selection of historical papers, the influence potential score is extracted based on the trained GAT model. A time decay term is constructed based on the fitted time decay factor to correct the influence potential score. Based on the corrected influence potential score, the historical papers are sorted through a ranking learning model to obtain the paper influence ranking order. A preset number of candidate paper lists or recommended ranking lists can be selected and output.

[0055] In step S4 of this embodiment, based on the obtained candidate paper recommendation results, a large language model is used to conduct an interpretability review of the candidate papers, outputting the strengths and weaknesses of each paper, and completing the final paper selection and generating a review report. A further technical solution, a method for obtaining interpretable information based on the recommendation results using a large language model, includes the following steps: Step S41: Receive the candidate paper set output by the recommendation module in the previous stage, and extract the subgraph information associated with the paper from the domain knowledge graph for each candidate paper. Preferably, the top 20 candidate papers are selected. For each candidate paper, subgraph information associated with the paper is extracted from the domain knowledge graph. The subgraph information includes at least: technical entity nodes, relational edges, and citation association information that are directly or multi-hop associated with the paper nodes. Unique identifiers are assigned to the nodes and edges in the subgraph for subsequent citation and tracing.

[0056] Step S42: Convert the key information of each candidate paper and its associated subgraph information into structured prompts and combine them to form the input to the large language model.

[0057] Optionally, structured hints preferably include the following fields: Basic information about candidate papers may include paper identifier, title, author / institution, publication date, etc. Recommended output abstract: the paper's ranking, score, or confidence level in the candidate set; The evidence chain information can be a list of key nodes and key relationships extracted from the subgraph, including the node / edge ID, name, type, relationship direction, and relationship type in the knowledge graph; A summary of citation information, including citation relationships or citation statistics related to this paper (if any); Output constraints, such as requiring the large language model to perform analysis only based on the provided fields, and to reference the nodes / edges corresponding to the evidence chain IDs in the output to support the conclusions.

[0058] Step S43: Input the structured prompts obtained in S42 into GPT-4 or the locally deployed large language model, and guide the model to complete the review and interpretation through system prompts, and output interpretable information.

[0059] Explainable information may include: explanation results for individual papers, cross-paper comparisons, and the final selected explanations. Specifically, for each candidate paper, the model generates a corresponding explanation result, which may include: Highlights analysis can include summarizing the paper's contributions and providing information on the chain of evidence. The deficiency analysis may include summarizing the limitations or points of contention, and providing information on the chain of evidence cited; The reasons for recommendation can include an explanation of why the paper was included in the shortlist and its relative advantages, along with the content of the evidence chain cited. Specifically, for each candidate paper, the cross-paper comparison and final selection interpretation results may include: A comprehensive comparison and ranking is conducted based on the sufficiency of highlights, the completeness of the chain of evidence, and candidate ranking / score information. Output the final list of the set number of papers, and provide a brief reason for selection and a chain of evidence for each selected paper; Output an abstract explaining the reasons for the rejection of papers that were not selected but had similar scores.

[0060] This step in this embodiment leverages a large language model to reason and generate explanations from the evidence subgraph and citation context, significantly enhancing the interpretability and traceability of the recommendation results. This meets the demanding requirements of review and editorial decision-making scenarios for interpretable explanations. By constructing a collaborative mechanism of ranking and large model explanation, while outputting the paper's influence potential score and recommendation ranking results, it further generates interpretable comments that correspond one-to-one with the ranking results based on the structured evidence chain in the knowledge graph. Under the constraint of controlled evidence input, the large language model reads information such as technical entity nodes, relationship paths, citation context fragments, and attention weights / contributions related to the target paper, forming a traceable evidence-conclusion attribution explanation and outputting natural language explanatory text. Since the explanation generation process uses verifiable facts from the knowledge graph and quantifiable contributions from the model output as input boundaries, it reduces the probability of the explanation deviating from objective data, avoids the uncertainty brought about by relying solely on a general generation model, and thus achieves transparency and auditability in the recommendation decision-making process. Therefore, the black-box outputs that are difficult to trace in traditional recommendation systems can be transformed into explanatory results supported by evidence, improving the consistency and acceptance of researchers and journal editors in their understanding of recommendation rankings, thereby enhancing the credibility and applicability of the recommendation results.

[0061] Example 2 Based on Example 1, this example provides a document-interpretable recommendation system based on knowledge graphs and large models, including: The fine-grained multimodal domain knowledge graph construction module is configured to acquire paper data in the target domain, extract semantic features from the unstructured text in the paper data, extract paper entities, paper technical entities and relationships, and construct a fine-grained multimodal domain knowledge graph. The scoring module is configured to extract feature vectors from papers in the target domain and the constructed domain knowledge graph, and calculate the impact potential score of each paper based on the feature vectors. The score correction module is configured to calculate a decay factor λ based on the publication time and citation information of historical papers, calculate an exponential decay term based on the decay factor λ, and calibrate the influence potential score of historical papers to obtain a time-faired influence potential score. The recommendation module is configured to rank the papers based on their impact potential scores using a ranking learning model and output a preset set of candidate papers as the recommendation result.

[0062] It should be noted that each module in this embodiment corresponds one-to-one with each step in embodiment 1, and their specific implementation process is the same, so it will not be repeated here.

[0063] Example 3 Based on Embodiment 1, this embodiment provides an electronic device, including a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When the processor executes the computer instructions, it completes the steps in the document-interpretable recommendation method based on knowledge graphs and large models described in Embodiment 1.

[0064] Example 4 Based on Embodiment 1, this embodiment provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the steps in the document interpretable recommendation method based on knowledge graphs and large models described in Embodiment 1.

[0065] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0066] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A document-interpretable recommendation method based on knowledge graphs and large models, characterized in that, Includes the following steps: Acquire paper data in the target domain, extract semantic features from the unstructured text in the paper data, extract paper entities, paper technical entities and relationships, and construct a fine-grained multimodal domain knowledge graph; For the paper data in the target field and the constructed domain knowledge graph, extract the paper feature vectors, and calculate the influence potential score of each paper based on the paper feature vectors; For historical papers, a decay factor λ is fitted and calculated based on the paper's publication time and citation information. An exponential decay term is then calculated based on the decay factor λ to calibrate the impact potential score of historical papers, resulting in a time-faired impact potential score. Based on the impact potential scores of the obtained papers, a ranking learning model is used to rank the papers and output a set of candidate papers of a predetermined number as the recommendation result.

2. The document-interpretable recommendation method based on knowledge graphs and large models as described in claim 1, characterized in that, Constructing fine-grained, multimodal domain knowledge graphs, including: Acquire paper data in the target field, wherein the paper data includes at least the structured metadata, unstructured text, and citation information of each paper; A large language model is used to extract semantic information from unstructured text, and entity and relation extraction is performed on academic text based on prompt templates to output technical entities and relations at the technical level of the paper content. Based on technical entities and relationships, structured metadata, and citation information, entity and relationship extraction is performed to construct a domain knowledge graph containing paper nodes, technical entity nodes, and their associated relationship edges, resulting in a fine-grained multimodal domain knowledge graph.

3. The document-interpretable recommendation method based on knowledge graphs and large models as described in claim 1, characterized in that, The paper feature vector includes semantic embedding features obtained from the paper text, quantitative features obtained from the paper metadata and citation information, and topological features obtained from the domain knowledge graph.

4. The document-interpretable recommendation method based on knowledge graphs and large models as described in claim 1, characterized in that, The method for calculating the impact potential score of each paper based on its feature vectors includes the following steps: Based on the knowledge graph, a feature vector is constructed for each paper; Obtain the expert-annotated dataset of papers for ranking, and use the feature vectors of the papers and the labels built from the expert-annotated dataset to train a graph attention network model to output the impact potential score of each paper.

5. The document-interpretable recommendation method based on knowledge graphs and large models as described in claim 1, characterized in that, For historical papers, the decay factor λ is calculated based on the paper's publication date and citation information, including the following steps: A predetermined number of paper samples are obtained from the collection of papers corresponding to the constructed domain knowledge graph. Obtain the annual citation count sequence for each paper sample and count the historical peak citation count of the paper; The predicted citation count is approximated by multiplying the peak citation count by an exponential decay term. The time decay factor λ is obtained by nonlinear least squares fitting with the goal of minimizing the sum of the squared differences between the actual citation count and the predicted citation count.

6. The document-interpretable recommendation method based on knowledge graphs and large models as described in claim 5, characterized in that, The exponential decay term is expressed in the following formula: ； in, As the attenuation factor, This is the time difference between the publication of the paper and the present. The objective function is to minimize the sum of the squared differences between the actual number of citations and the predicted number of citations. ； Among them, C p,t This represents the actual number of times paper p was cited in year t. This is its historical peak number of citations.

7. The document-interpretable recommendation method based on knowledge graphs and large models as described in claim 1, characterized in that, Interpretable information is obtained by reasoning based on a large language model based on the recommendation results, including... Receive a set of candidate papers, and for each candidate paper, extract the subgraph information associated with that paper from the domain knowledge graph; The key information of each candidate paper and its related subgraph information are converted into structured prompts; The obtained structured prompts are input into the large language model, and the model is guided by system prompts to complete the review and interpretation, and output interpretable information.

8. A document-interpretable recommendation system based on knowledge graphs and large models, characterized in that: include: The fine-grained multimodal domain knowledge graph construction module is configured to acquire paper data in the target domain, extract semantic features from the unstructured text in the paper data, extract paper entities, paper technical entities and relationships, and construct a fine-grained multimodal domain knowledge graph. The scoring module is configured to extract feature vectors from papers in the target domain and the constructed domain knowledge graph, and calculate the impact potential score of each paper based on the feature vectors. The score correction module is configured to calculate a decay factor λ based on the publication time and citation information of historical papers, calculate an exponential decay term based on the decay factor λ, and calibrate the influence potential score of historical papers to obtain a time-faired influence potential score. The recommendation module is configured to rank the papers based on their impact potential scores using a ranking learning model and output a preset set of candidate papers as the recommendation result.

9. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the steps in the document-interpretable recommendation method based on knowledge graphs and large models as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, complete the steps in the document interpretable recommendation method based on knowledge graphs and large models as described in any one of claims 1-7.

Images