Intelligent management system and method for quality evaluation and self-repair of knowledge graph

Abstract

The invention discloses an intelligent management system and method for knowledge graph quality evaluation and self-repairing, belongs to the technical field of knowledge graphs, and aims to solve the problems that in traditional knowledge graph management, manual auditing efficiency is low, an effective automatic repairing means is lacked, and data complexity and real-time changes are difficult to deal with. The system firstly collects multi-source heterogeneous data in a target field, cleans the data through a deep learning noise recognition model, extracts entities and relationships by using a natural language processing technology, and adds metadata to convert the entities and relationships into graph structure data; then, a graph framework is defined based on the ontology, entity semantic alignment is achieved in combination with a graph neural network, and a knowledge graph is constructed by complementing implicit relations with the help of a pre-training language model. Then, the quality of the atlas is quantitatively evaluated through a four-layer quality evaluation system, meanwhile, a repair scheme is generated based on vulnerability feature extraction, knowledge base matching and decision fusion, and intelligent self-repair is achieved; the map can be monitored in real time and evaluated regularly, a repair strategy and a knowledge base are optimized through reinforcement learning, it is ensured that the map is kept accurate and time-efficient for a long time, and the practical value is improved.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph technology, and in particular to an intelligent management system and method for knowledge graph quality assessment and self-repair. Background Technology

[0002] In the field of knowledge graph technology, traditional knowledge graph construction and management methods often neglect continuous quality monitoring and self-repair capabilities after the graph is built. As the scale of knowledge graphs continues to expand, the complexity and uncertainty of the data also increase significantly, resulting in a large amount of noise, errors, or missing information in the graph. These problems directly affect the effectiveness and reliability of the knowledge graph.

[0003] Traditionally, the quality assessment of knowledge graphs has relied primarily on manual review. This approach is not only time-consuming and labor-intensive but also inefficient and ill-suited to handling real-time changes in large-scale datasets. Furthermore, the lack of effective automated methods for correcting errors or inconsistencies in knowledge graphs leads to persistent problems that negatively impact their practical value and application effectiveness. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes an intelligent management system and method for knowledge graph quality assessment and self-repair. The technical solution adopted by the intelligent management method for knowledge graph quality assessment and self-repair includes the following steps: Data collection and preprocessing: Collect structured, semi-structured, and unstructured data from the target domain; clean the data using a deep learning-based noise recognition model, and extract entities and relationships between entities from the text using natural language processing techniques; transform the extracted data into a unified graph structure data, and attach hierarchical classification labels and context-related dynamic keywords as metadata to the data entities; Knowledge graph construction: Based on ontology, a diversified architecture for the graph is defined, including entity types, relation types, and attribute constraints; through entity linking technology based on graph neural networks, entities identified in the preprocessing stage are semantically disambiguated and aligned with a known entity database to ensure the global uniqueness and consistency of entities; a pre-trained language model is used to automatically infer and complete potential implicit relationships between entities, enhancing the connectivity of the graph; Multi-dimensional quality assessment: Constructing a four-layer assessment system of background quality, content quality, application quality, and knowledge value density; employing a dynamic weight allocation strategy that integrates the analytic hierarchy process (AHP) and entropy weighting to assign differentiated weights to each dimension; constructing a knowledge signal-to-noise ratio model to quantify the ratio of core knowledge elements to noisy knowledge elements; and calculating the comprehensive quality score of the knowledge graph in a specific domain through weighted fusion. Intelligent self-repair: Based on error detection algorithms that combine inconsistency detection and knowledge conflict resolution, it identifies logical conflicts, errors, or missing information in the knowledge graph; it integrates rule-based systems and transfer learning-based prediction and correction models to generate repair strategies; and it dynamically adjusts the strength and scope of the repair strategies according to the real-time update stream of the knowledge graph. Continuous monitoring and optimization: Real-time quality monitoring and periodic in-depth evaluation of the knowledge graph; when a new type of vulnerability is detected, a solution generation process based on case-based reasoning and decision fusion is triggered; feedback signals from the remediation process are recorded through reinforcement learning mechanisms to continuously optimize the vulnerability knowledge base and remediation strategies, thereby optimizing the knowledge graph.

[0005] As a further aspect of the present invention, the specific steps of multidimensional quality assessment include: Knowledge background quality assessment uses Jaccard similarity to measure the similarity of topological structures between entities, and designs attribute similarity calculation functions for three attribute types: discrete, exact numerical, and linear relational. It introduces semantic similarity calculation based on knowledge graph embedding to make up for the shortcomings of symbolic similarity. It combines entity similarity and attribute similarity through AHP-entropy weighting strategy to obtain knowledge background quality score. Knowledge content quality assessment divides the knowledge graph into a layered evaluation: the concept layer is evaluated based on its consistency with the domain ontology terminology library; the entity layer is evaluated based on the richness and credibility of its descriptive information; and the relation layer is evaluated based on the confidence and semantic consistency of relation paths. Knowledge application quality assessment is conducted by evaluating the query response, reasoning ability, and decision support effect of knowledge graphs in real-world scenarios; the graphs are tested using a predefined question set, and the F1 score or cosine similarity between their output results and the expected goals is calculated to quantify their application performance. Knowledge value density assessment calculates the knowledge signal-to-noise ratio to obtain the ratio of core knowledge elements to redundant or marginal knowledge elements in the knowledge graph, thus addressing the problem of uneven knowledge contribution value density; it also calculates a multi-factor fusion model based on the topological importance, source authority, and assertion confidence of entities in the graph. The comprehensive quality fusion aggregates the scores of background quality, content quality, application quality, and knowledge value density through a non-linear weighted fusion function. The weights are adjusted based on historical evaluation data to output a comprehensive quality score for the knowledge graph. During the fusion process, the knowledge value density score acts as a moderating factor to strengthen or suppress the first three basic scores, thereby optimizing the final score and making it more reflective of the actual utility of the graph.

[0006] As a further aspect of the present invention, the method for calculating the knowledge signal-to-noise ratio is as follows: ,in: The total strength of the core knowledge element is calculated by aggregating entity centrality, knowledge source authority score, and the frequency of citation of the knowledge. The intensity of the noisy knowledge element is calculated by the attribute redundancy of the aggregated entity, the low confidence of the relation assertion, and the low confidence of the source. The formula for calculating integrated quality is: Where: w1, w2, and w3 are weighting coefficients. The quality scores are for the background, content, and application, respectively. As a knowledge value density moderating factor, when When the value is positive, it adds to the total score, but the effect of the gain is diminishing marginal utility.

[0007] As a further aspect of the present invention, the specific steps of the intelligent self-repair method include: Vulnerability feature extraction and representation adopts the topic event extraction method of event graph to extract structured feature triples from new vulnerabilities, including: trigger words, vulnerability type, and affected objects, and transforms them into low-dimensional vector representations for retrieval; Vulnerability matching and remediation decisions are made using a knowledge graph-based vulnerability knowledge base for vectorized semantic retrieval. If a new vulnerability feature matches a record in the database, the corresponding remediation script is automatically invoked to perform the patch operation. If the match fails, a graph attention network-based associated vulnerability mining module is activated to aggregate similar vulnerability information and submit it to the decision fusion engine. In the solution generation and knowledge update process, the decision fusion engine uses case-based reasoning and Bayesian networks for probabilistic reasoning to generate a set of candidate remediation solutions. It then selects the optimal solution for remediation through credibility evaluation. After the remediation is completed, the characteristics of the new vulnerability and its solutions will be automatically structured and updated to the vulnerability knowledge base, achieving closed-loop accumulation of knowledge.

[0008] As a further aspect of the present invention, an intelligent management system for knowledge graph quality assessment and self-repair includes the following modules: The knowledge graph construction module is used to realize multi-source data collection, intelligent cleaning, entity relationship extraction and graph construction; The multi-dimensional quality assessment module is used to perform assessments and fusion calculations of background quality, content quality, application quality, and knowledge value density. This module integrates a configurable weight management sub-module, an assessment indicator library, and... Computing engine; The intelligent self-repair module includes a vulnerability feature extractor, a vulnerability knowledge base manager, a decision fusion engine, and a repair executor, enabling automatic vulnerability identification, matching, repair decision-making, and execution. The continuous optimization management module is responsible for the real-time monitoring, evaluation, scheduling, repair learning, and meta-knowledge management of the knowledge graph, ensuring that the system has continuous optimization capabilities.

[0009] As a further aspect of the present invention, the vulnerability knowledge base is a structured knowledge base built based on knowledge graph technology, in which vulnerability information is stored in the form of entity-relationship, and is accompanied by attributes such as remediation scheme, effective conditions, and historical effective records, supporting semantic similarity retrieval and association reasoning.

[0010] Beneficial Effects: This invention effectively addresses the pain points of traditional knowledge graph management, offering significant benefits. Leveraging a deep learning noise recognition model and differentiated cleaning strategies, combined with NLP technology to extract entity relationships, it overcomes the bottleneck of processing multi-source heterogeneous data, providing high-quality data support for graph construction. Based on an ontology architecture and graph neural network entity links, it ensures the global uniqueness of entities, and pre-trained models complete implicit relationships, improving graph connectivity and resolving entity ambiguity and missing relationship issues in traditional construction. Simultaneously, it innovates a four-layer evaluation system of "background-content-application-knowledge value density," using AHP-entropy weighting and K_SNR model to quantify value, replacing inefficient manual review and achieving automated evaluation and quantifiable results. Furthermore, the closed-loop mechanism of "feature extraction-matching repair-knowledge update," combined with a decision fusion engine, reduces manual intervention by over 70%, with real-time monitoring and reinforcement learning for continuous optimization, ensuring the long-term accuracy and effectiveness of the graph and significantly improving the reliability and efficiency of domain querying, reasoning, and decision support. Attached Figure Description

[0011] Figure 1 This is a flowchart illustrating the steps of the intelligent management method for knowledge graph quality assessment and self-repair in this invention. Figure 2 This is a schematic diagram of the system modules of the intelligent management system for knowledge graph quality assessment and self-repair in this invention. Detailed Implementation

[0012] The present invention will be further described in detail below with reference to the embodiments.

[0013] Please see Figure 1 As shown, the technical solution adopted by the intelligent management method for knowledge graph quality assessment and self-repair in this invention includes the following steps: S100 Data Collection and Preprocessing: Data collection and preprocessing are fundamental data purification and standardization steps in knowledge graph construction. They transform multi-source, heterogeneous raw data into high-quality structured data that meets the requirements of knowledge graph construction. The specific steps are as follows: S101 Multi-Source Data Collection: Collects data covering all dimensions of the target domain. Based on the knowledge boundaries of the target domain, three types of data are collected to ensure the comprehensiveness and relevance of the data.

[0014] Structured data collection: Collect data with clear format specifications that can be directly stored in tables or databases. During collection, the business meaning and value constraints of data fields must be clearly defined. Emphasis should be placed on field completeness and data format consistency (e.g., date format should be uniformly "YYYY-MM-DD", and numeric formats should uniformly retain two decimal places). Semi-structured data collection: Collect data with some structural characteristics (e.g., labels, key-value pairs) but less strict formatting. During collection, the structured portion must be extracted while preserving contextual relationships. For XML and JSON formats, the corresponding content must be extracted using parsing tools, and the hierarchical relationships of the data must be recorded.

[0015] Unstructured data collection: This refers to text-based data without a fixed format. When collecting such data, it is necessary to clearly define the authority of the data source and its relevance to the topic. Examples include domain literature, expert interview transcripts, and scanned copies of handwritten reports. The original text needs to be obtained through text extraction (such as literature database APIs) or OCR technology (such as scan-to-text conversion), while simultaneously filtering out content related to the core knowledge of the target domain.

[0016] S102 Deep Learning-Based Noise Recognition and Data Cleaning: This involves eliminating invalid and erroneous information from the data. A deep learning model is used to accurately identify noise, followed by targeted cleaning operations. Specifically, this includes: Noise Recognition Model Construction: A classification model based on fine-tuning a pre-trained language model is used to categorize noise types into five categories: "typos," "duplicate records," "semantic conflicts," "formatting errors," and "invalid values." Furthermore, during model training, a domain-specific noise-annotated corpus is constructed (annotating the noise type and location for each data point). By fine-tuning the model's output layer to a multi-classifier, accurate identification of different noise types is achieved. The core evaluation metrics are the noise recognition accuracy (≥90%) and recall (≥85%), ensuring minimal omissions and false positives.

[0017] Further targeted cleaning operations include: differentiated cleaning based on noise type: typo correction: combining domain thesaurus and contextual semantics, identified typos are replaced, and semantic consistency is verified after correction (e.g., the corrected words must exist in the domain terminology database); duplicate record deletion: duplicate records are identified by calculating data field similarity (structured fields complete match, unstructured fields semantic similarity ≥95%), retaining the most complete record (no missing core fields); semantic conflict resolution: for contradictory information about the same entity, the correct information is determined by associating with authoritative third-party data, and the incorrect information is deleted; handling of formatting issues and invalid values: non-standard formatted data is standardized to a standard format, and invalid values ​​are marked as needing to be supplemented, subsequently improved through a data completion mechanism.

[0018] S103 NLP-based Entity and Relation Extraction: Extracts core elements (entities and relations) of the geographic graph from text, achieving structured extraction through natural language processing techniques. Specifically, this includes: Entity Extraction: Employing a pre-trained language model and a named entity recognition model, first defining domain entity types, then using the model to identify and extract corresponding entities from the text. During model training, the domain entity corpus needs to be annotated (annotating entity boundaries and types), with a focus on optimizing the accuracy of entity boundaries and the correctness of type classification.

[0019] Furthermore, relation extraction adopts a relation extraction method based on graph neural networks. First, the domain relation types are defined, and then the relations are extracted: the unlabeled text is automatically labeled using known entity-relation pairs in the domain knowledge base; relation classification: the labeled text fragments are input into the classification model fine-tuned by the pre-trained language model, and the relation types are output. The core evaluation index is the F1 score of relation extraction (≥88%), ensuring that the extracted relations conform to the domain semantic logic.

[0020] S104 Data Transformation and Metadata Attachment: Transforms extracted features into graph-compatible formats and enhances data retrievability. Specifically, this includes: Graph Structure Data Transformation: Transforms extracted entity-relation pairs into the resource description framework triple format (i.e., subject-verb-object structure) of knowledge graph standards, or attribute graph format (entities with attributes, relations with confidence), ensuring that the data can be directly imported into graph databases.

[0021] Furthermore, metadata addition: Two types of metadata are added to each entity to enhance the semantic richness and retrieval efficiency of the entity: Hierarchical classification tags: A three-level classification system of domain-major category-minor category is adopted to ensure the logicality of entity classification and facilitate subsequent management of the graph by category; Context-related dynamic keywords: Based on the context of the text in which the entity is located, 3-5 core keywords are extracted through TF-IDF or TextRank algorithms for subsequent semantic retrieval and similarity matching of entities.

[0022] S200 Knowledge Graph Construction: This involves transforming standardized data into a structured knowledge network, defining the architecture based on ontology, ensuring entity consistency, and completing implicit relationships. The specific steps are as follows: S201 defines a diversified architecture based on ontology: clearly defining the core components of the graph to ensure the standardization of subsequent data entry, specifically including: Entity type definition: Based on the knowledge system of the target domain, divide entities into primary and secondary types, and clarify the core attributes of each type of entity.

[0023] Relationship type definition: Define the semantic relationship between entities, clarify the direction and constraints of the relationship, and avoid relationship redundancy; Attribute constraint definition: By using a Web ontology language or domain rule base, define the value constraints of attributes to ensure the validity of attribute values.

[0024] S202 Entity Linking Based on Graph Neural Networks: Eliminating entity ambiguity and ensuring global uniqueness, this approach uses graph neural networks to align entities to be linked with known authoritative entity databases. Specifically, it includes: Entity feature construction: Construct a multi-dimensional feature vector for each entity to be linked, covering three types of features: Text features: Word2Vec or BERT vector of the entity name (reflecting the semantic information of the entity); Attribute characteristics: Encoded vectors of the core attribute values ​​of an entity; Contextual features: The contextual fragment vector of the text in which the entity is located (reflecting the usage scenario of the entity); The three types of features are concatenated into a unified feature vector, which is then used as the input for entity linking.

[0025] Furthermore, the graph neural network model is trained using a graph attention network or graph convolutional network, with a domain authority entity database as the target database. The model calculates the semantic similarity between the entity to be linked and the authority entity. The model focuses on the semantically matching-related dimensions of the entity features through an attention mechanism, sets a similarity threshold (e.g., ≥85%), and aligns the entity to be linked to a unique ID in the authority entity database.

[0026] Furthermore, consistency verification: After alignment is completed, global entities are verified, many-to-one duplicate entities are deleted, one-to-many misalignments are corrected, and finally the global uniqueness of entities in the graph is ensured.

[0027] S203 Pre-trained Language Model Completion of Implicit Relations: Inferring unspoken relationships between entities using a domain-specific pre-trained language model, specifically including: Model selection and fine-tuning: Select a domain-specific pre-trained language model and fine-tune it for the relation completion task. During fine-tuning, use entity pairs and relation labels as training samples to optimize the model's ability to predict domain relations. The core evaluation metric is the accuracy of relation prediction (≥85%).

[0028] Furthermore, the implicit relationship inference logic is as follows: Based on existing explicit relationships, implicit relationships are inferred through the following two logics: Based on semantic association: If entity A and entity B have a relationship R1, entity B and entity C have a relationship R2, and R3 can be derived from R1 and R2 in the domain knowledge, then it is inferred that A and C have a relationship R3; Based on attribute similarity: If the core attribute similarity between entity A and entity B is ≥80%, and A and entity C have a relationship R, then it is inferred that B and C also have a relationship R.

[0029] Further connectivity optimization verification: After completion, the connectivity rate of the graph needs to be calculated to ensure that there are no isolated nodes between core entities, thereby improving the knowledge coverage and retrieval efficiency of the graph.

[0030] S300 Multi-Dimensional Quality Assessment: Constructs a four-layer system of background, content, application, and knowledge value density, and calculates comprehensive quality through dynamic weights and a quantitative model. The specific steps are as follows: S301 Knowledge Background Quality Assessment: Entity Topological Similarity Calculation: Jaccard similarity is used to measure the overlap of the neighbor node sets of two entities, and the formula is as follows: Where: A and B are two knowledge graph entities to be evaluated, which are the core objects for similarity calculation; and Let J(A,B) be the set of neighboring nodes of entities A and B, respectively. This index reflects the degree of association between entities in the graph structure. By comparing the intersection and union of their nodes, the similarity of their topological structures is converted into a quantitative value between 0 and 1. The higher the ratio, the stronger the structural association between the two entities in the graph. J(A,B)∈[0,1]: When J(A,B)=1, A and B have exactly the same neighboring nodes (completely similar structure); when J(A,B)=0, A and B have no common neighbors (completely unrelated structure); in general applications, if J(A,B)≥0.6, the two entities can be considered to have highly similar topological structures.

[0031] Furthermore, entity attribute similarity calculation: Differentiated functions are designed for the three attribute types to ensure accurate quantification of attribute similarity. Discrete attributes: A matching degree function is used. If the attribute values ​​are exactly the same, the similarity is 1; otherwise, it is 0. Precise numerical attributes: Utilize the inverse absolute difference function, the formula is: Where: x1, x2: two precise values ​​of the same attribute; |x1-x2|: the absolute difference between x1 and x2, reflecting the absolute difference in values; max(x1,x2): the maximum value between x1 and x2, used to normalize the absolute difference to a relative difference, avoiding bias caused by the magnitude of the values. This transforms the numerical difference into a similarity between 0 and 1, ensuring that numerical attributes at different scales can be compared fairly. The higher the ratio, the more similar the two numerical attributes are.

[0032] Linear relational attributes: Similarity is calculated using the Pearson correlation coefficient, reflecting the degree of linear association between attribute values; Semantic similarity supplementary calculation: Introducing semantic similarity calculation based on knowledge graph embedding, mapping entities and relations to a low-dimensional vector space, and measuring semantic association through cosine similarity; AHP-Entropy Weighting Combination: This method combines subjective weights (AHP) and objective weights (entropy weighting) to determine the weights of each similarity indicator. AHP (Analytic Hierarchy Process): 5-8 domain experts are invited to determine weights through pairwise comparisons using a judgment matrix (e.g., experts assign weights of 0.45, 0.35, and 0.2 respectively, indicating that entity similarity is more important than attribute similarity). Entropy weighting: Weights are calculated based on the dispersion of historical evaluation data (e.g., entity similarity data with high dispersion is assigned a weight of 0.4; attribute similarity data with moderate dispersion is assigned a weight of 0.4). Combined weights: Take the weighted average of the two values, and finally obtain the background quality score by weighted summation.

[0033] S302 Knowledge Content Quality Assessment, specifically includes: Conceptual layer evaluation: Based on the domain authoritative ontology terminology database, the consistency between the graph concepts and the definitions of standard terms is evaluated. Specifically, the core definition entries for each concept in the graph are extracted, and the percentage of entries matching the standard terminology definition is calculated (e.g., matching 3 core entries results in a consistency score of 1.0, matching 2 results in 0.67). Finally, the average score of all concepts is taken as the conceptual layer quality score. Entity layer evaluation: Evaluation from two dimensions: richness and credibility. Richness: For each type of entity, a preset number of core attributes is set, and the ratio of the actual number of core attributes contained in the entity to the preset number is calculated (e.g., if it contains 4 core attributes, the richness score is 0.8). The average score of all entities is then taken. Credibility: Assign values ​​based on the authority of the entity's attribute values, and calculate the average score of all attribute values ​​of the entity as the credibility score; Entity layer total score: A weighted average is used, i.e., entity layer total score = richness score × 0.6 + credibility score × 0.4.

[0034] Furthermore, the relation layer is evaluated from two dimensions: confidence and semantic consistency. Confidence is calculated based on the source data of the relation, taking the average confidence of all relations. Semantic consistency is verified by the domain rule base to see if the relation conforms to common sense. If the relation conforms to the rules, the consistency score is 1.0; otherwise, it is 0. The average score of all relations is taken. The total relation layer score is calculated using a weighted average, i.e., total relation layer score = confidence × 0.7 + semantic consistency × 0.3.

[0035] Furthermore, the total content quality score is calculated using a weighted average: Total Content Quality Score = Conceptual Layer Score × 0.4 + Entity Layer Score × 0.3 + Relationship Layer Score × 0.3 (maximum score 1 point).

[0036] The S303 knowledge application quality assessment is quantitatively evaluated through functional testing, specifically including: Query Response Evaluation: Common query cases in the design domain are evaluated using two core metrics: Response Time: The average response time of queries is tested using a graph database query engine, which must meet the domain's real-time requirements (e.g., ≤1 second). The score is calculated by reversing the actual response time to the threshold. Accuracy: The ratio of the number of correctly returned results to the actual number of results that should have been returned; Query response score: weighted average (response time weight 0.3, accuracy weight 0.7).

[0037] Further, reasoning ability assessment: Design domain reasoning cases, output results through the graph reasoning model, and calculate the F1 score (harmonic average of precision and recall): Precision = Number of correct inference results / Total number of inference results; Recall = Number of correct inference results / Actual possible results; As a score for reasoning ability, it combines precision and recall to give a single-dimensional quantitative score. The higher the score, the better the graph reasoning ability.

[0038] Furthermore, the decision support effectiveness is evaluated by selecting real-world business cases in the domain, comparing the decision results recommended by the graph with the authoritative results, and calculating the conformity rate (number of consistent cases / total number of cases) as the decision support score.

[0039] Furthermore, the overall application quality score is calculated as a weighted average (query response score weight 0.3, reasoning ability score weight 0.4, decision support score weight 0.3).

[0040] The S304 knowledge value density assessment uses a knowledge signal-to-noise ratio (K_SNR) and multi-factor fusion model for evaluation. Specifically, it includes: Knowledge Signal-to-Noise Ratio (K_SNR) calculation: based on the formula: ,in: The total strength of core knowledge elements refers to entities / relationships that directly support the core business of the domain. The strength is calculated by weighting entity centrality, source authority score, and citation frequency (each weight is 1 / 3), and the sum is obtained. The total intensity of noisy knowledge elements refers to redundant and low-value entity relationships. The intensity is calculated by weighting attribute redundancy (number of redundant attributes / total number of attributes) + relationship low confidence (1 - confidence) + source low credibility (1 - source authority score) (each weight is 1 / 3), and then summing them to obtain the total intensity.

[0041] The larger the K_SNR value, the higher the proportion of core knowledge and the less noise.

[0042] Furthermore, the multi-factor fusion model calculates: the topological importance of the fused entity + the authority of the source + the assertion confidence, and sums them with equal weights (1 / 3 each) to obtain the fusion value, which serves as an auxiliary evaluation indicator for knowledge value density.

[0043] Furthermore, the knowledge value density score is based on K_SNR and is adjusted by combining the fusion value (e.g., if K_SNR ≥ 6dB and fusion value ≥ 0.8, the value density score is 1.0; if K_SNR = 4-6dB and fusion value = 0.6-0.8, the score is 0.7).

[0044] S305 integrates the results of four levels of evaluation and outputs a final quality score, using the following formula: The weights w1, w2, and w3 are dynamically adjusted based on domain requirements and historical evaluation data, determined using analysis of variance (e.g., the largest historical variance in application quality indicates a greater impact from fluctuations in that dimension, thus assigning higher weights, e.g., w1=0.25, w2=0.35, w3=0.4); Q(kbg), Q(kc), and Q(ka) are the scores for background quality, content quality, and application quality, respectively (all ranging from 0 to 1); adjustment factors... When K_SNR is positive, the first three base scores are increased, and the marginal utility of the increase effect decreases. Overall quality score Q (kg): The final output score used to measure the overall quality level of the spectrum.

[0045] The S400 intelligent self-repair system achieves a closed loop through feature extraction, matching decision-making, solution generation, and knowledge updating. The specific steps are as follows: S401 vulnerability feature extraction and representation transforms unstructured vulnerabilities into searchable structured features, specifically including: Vulnerability detection trigger: Vulnerabilities are identified through rule detection and statistical anomaly detection. Rule detection is based on the domain rule base, and statistical anomaly detection is based on deviations from the historical data mean ± 2 standard deviations. After vulnerability detection is triggered, the entities, relationships or attributes involved in the vulnerability are located.

[0046] Furthermore, thematic event extraction: using the ACE or ERE event extraction model, structured feature triples are extracted from the vulnerability, including: trigger words: key actions or attributes that cause the vulnerability (such as "relational assertion semantic conflict" or "attribute value format error"); vulnerability type: classified according to the nature of the vulnerability (such as "relational error", "attribute error", "entity duplication", or "information missing"); affected objects: graph elements directly involved in the vulnerability; Low-dimensional vector representation: Each element in the feature triplet is transformed into a low-dimensional vector to facilitate subsequent semantic retrieval. Trigger words, vulnerability types, and affected objects are transformed into fixed-dimensional vectors. The three vectors are concatenated into a vulnerability feature vector to ensure that the vector reflects the semantic essence of the vulnerability.

[0047] S402 vulnerability matching and remediation, specifically including: Vulnerability knowledge base construction: Vulnerability information is stored using a knowledge graph structure to ensure structure and searchability. Entities: Vulnerability type entity, affected object entity; Relationship: Vulnerability type - Impact - Affected objects; Vulnerability type - Corresponding remediation script ID; Remediation script - Effective conditions; Attributes: Repair script content, activation conditions, and historical activation records; Vectorized semantic retrieval: Calculate the cosine similarity between the feature vector of a new vulnerability and the feature vectors of historical vulnerabilities in the vulnerability knowledge base, set a matching threshold, and if the match is successful (similarity ≥ threshold): it indicates that there is a mature remediation solution, and the corresponding remediation script is automatically called; if the match fails (similarity < threshold): it indicates that it is a new type of vulnerability, and the association vulnerability mining based on graph attention network is initiated: The GAT model is input with the feature vector of the new vulnerability and all vulnerability feature vectors in the knowledge base, and the association weight between the new vulnerability and historical vulnerabilities is calculated through the attention mechanism; similar vulnerabilities with association weight ≥ 0.6 are mined, the remediation solutions of these similar vulnerabilities are summarized, and submitted to the decision fusion engine; Furthermore, the script execution verification is performed: after a successful match, the script is executed and the repair effect needs to be verified. If the verification fails, the operation is rolled back and manual intervention is triggered.

[0048] S403 solution generation and knowledge update, updating the knowledge base to form a closed loop, specifically includes: Decision fusion engine working logic: using a combination of case-based reasoning and Bayesian network methods to generate candidate remediation solutions; Case-based reasoning: extracting remediation actions from remediation cases of similar vulnerabilities, assigning weights according to case similarity, and fusing to obtain preliminary candidate solutions; Furthermore, Bayesian network probabilistic reasoning: Construct a Bayesian network, where nodes include vulnerability type, attribute type, remediation scheme, and credibility. Input the vulnerability type of the new vulnerability and the attribute type of the affected object, and calculate the probability of successful remediation for each candidate scheme. Furthermore, the optimal solution selection: select the candidate solution with the highest probability of successful repair and perform the repair. After repair, it needs to pass quality verification to ensure that the repair is effective. Furthermore, the knowledge loop is updated: after structuring the new vulnerabilities and their remediation solutions, the information is updated to the vulnerability knowledge base; the vulnerability knowledge base includes: new entities: if it is a new vulnerability type, a new vulnerability type entity is added; New relationships: New vulnerability type - Impact - Affected objects; New vulnerability type - Corresponding fix script ID; Added attributes: Add content, effective conditions, and initial history of the new repair script; ensure that when encountering similar vulnerabilities in the future, the repair solution can be directly matched, improving repair efficiency.

[0049] S500 continuous monitoring and optimization, including the following steps: S501 real-time quality monitoring, specifically including: Monitoring System Deployment: Utilizing a stream processing framework deployed on a distributed server, enabling real-time processing of graph update data; Data Access: Connecting to the graph database's update logs to acquire new data in real time; Parallel Processing: Setting a reasonable parallelism to ensure processing latency ≤1 second; Core Monitoring Indicator Design: Setting three quantifiable monitoring indicators, each with configured thresholds: Relationship Assertion Confidence: Threshold ≥ 0.6 (below triggers an alarm, indicating low relationship confidence); Entity Attribute Integrity: Threshold ≥ 0.9 (below triggers an alarm, indicating many missing core entity attributes); Attribute Format Correctness: Threshold = 100% (below triggers an alarm, indicating non-standard attribute format); Alarm and Handling Mechanism: When an indicator triggers a threshold, the system automatically executes: Alarm Push: Pushing alarm information to the administrator (including vulnerability type, involved elements, indicator value, and occurrence time); Temporary Interception: Performing "temporary interception" on low-quality data (such as relationships with a confidence of 0.5), preventing it from being written to the graph master database; Automatic Repair Attempt: For simple vulnerabilities (such as incorrect attribute formats), automatically repairing them by calling existing repair scripts, with a repair success rate ≥ 90%. If yes, proceed; otherwise, wait for manual processing.

[0050] S502 periodic in-depth assessment, specifically including: Evaluation cycle and execution method: Set the evaluation cycle and use a distributed computing framework to process the full map data: Cluster configuration: Configure the number of nodes according to the map size; Assessment content: Perform the entire process of multi-dimensional quality assessment, and simultaneously calculate the vulnerability occurrence rate and remediation success rate; Time control: Ensure that the full assessment takes ≤8 hours; Assessment Report Generation: Generates a periodic quality assessment report, including: Quality Score Trend: Compares the overall quality score Q (kg) and scores of each dimension over the past 3 periods, and analyzes the reasons for the changes; Vulnerability analysis: Statistics on the distribution of vulnerability types and the location of frequently occurring vulnerabilities; Optimization suggestions: Propose optimization directions for the low-scoring dimensions.

[0051] S503 reinforcement learning optimization continuously refines the repair strategy and knowledge base, iteratively improving repair effectiveness through feedback signals, specifically including: Feedback signal collection: Three types of feedback indicators are set for each remediation strategy to quantify the remediation effect, the occurrence rate of similar vulnerabilities after remediation, and the number of new similar vulnerabilities in the three periods after remediation; Remediation plan execution efficiency: the average execution time of the remediation script; Improvement of graph quality after remediation: the improvement value of the quality score of the corresponding dimension after remediation.

[0052] Furthermore, reward values ​​are assigned to the remediation strategies based on the performance indicators; reinforcement learning model training: Q-learning algorithm is used, with vulnerability type as state (S), remediation strategy as action (A), reward value as feedback (R), and the Q value (action value function) of the strategy is updated: Q value update formula: Where Q(s,a): the action value function of performing action a in state s, representing the expected long-term benefit of the strategy; s: the current state, referring to the vulnerability type; a: the current action, referring to the remediation strategy; s': the next state after performing action a; a': the optimal action in the next state s', i.e., Q(s',a')=max {Q(s',a1),Q(s',a2),...}; Learning rate ( ∈[0,1]), controlling the magnitude of each Q-value update, The larger it is, the more radical the updates become; The smaller the size, the smoother the updates; Discount factor ( ∈[0,1]), controlling the weight of future rewards. The closer γ is to 1, the more emphasis is placed on long-term benefits; the closer γ is to 0, the more emphasis is placed on immediate benefits; r: the immediate reward after executing action a, the feedback signal of the repair strategy, quantified as: r = the occurrence rate of similar vulnerabilities after repair (negative correlation) + the execution efficiency of the repair plan (positive correlation) + the improvement of the graph quality after repair (positive correlation). The specific value needs to be combined with the actual situation of the domain. For example, when the occurrence rate of similar vulnerabilities after repair is 0, r=10, and when the occurrence rate is 50%, r=2. By iteratively updating Q(s,a) through immediate rewards and expected future rewards, the Q value of efficient repair strategies gradually increases, and the Q value of inefficient strategies gradually decreases, ultimately achieving autonomous optimization of the repair strategy.

[0053] Furthermore, knowledge base and strategy optimization: quarterly optimization based on Q-value, strategy screening: deleting inefficient strategies with Q-value < 0.5; knowledge base update: adding remediation scripts and vulnerability cases corresponding to efficient strategies with Q-value ≥ 0.9; ultimately achieving improved remediation success rate and remediation efficiency.

[0054] For further details, please refer to Figure 2 The diagram shown is a system module schematic of the intelligent management system for knowledge graph quality assessment and self-repair of this invention. The system modules include: The knowledge graph construction module is used to realize multi-source data collection, intelligent cleaning, entity relationship extraction and graph construction; The multi-dimensional quality assessment module is used to perform assessments and fusion calculations of background quality, content quality, application quality, and knowledge value density. This module integrates a configurable weight management sub-module, an assessment indicator library, and... Computing engine; The intelligent self-repair module includes a vulnerability feature extractor, a vulnerability knowledge base manager, a decision fusion engine, and a repair executor, enabling automatic vulnerability identification, matching, repair decision-making, and execution. The continuous optimization management module is responsible for the real-time monitoring, evaluation, scheduling, repair learning, and meta-knowledge management of the knowledge graph, ensuring that the system has continuous optimization capabilities.

[0055] Furthermore, the vulnerability knowledge base is a structured knowledge base built on knowledge graph technology, in which vulnerability information is stored in the form of entity-relationship, and is accompanied by attributes such as remediation solutions, effective conditions, and historical effective records, supporting semantic similarity retrieval and association reasoning.

[0056] The above description is merely an embodiment of this specification and is not intended to limit this specification. Various modifications and variations can be made to this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this specification should be included within the scope of the claims of this application.

Claims

1. An intelligent management method for knowledge graph quality assessment and self-repairing, characterized in that, Comprising the following steps: Data collection and preprocessing: Collect structured, semi-structured and unstructured data in the target field; use deep learning-based noise identification model to clean up the data, and use natural language processing technology to extract entities and relationships between entities from the text; convert the extracted data into unified graph structure data, and add hierarchical classification labels and context-related dynamic keywords as metadata to the data entities; Knowledge graph construction: Define the diversified architecture of the graph based on ontology, including entity type, relationship type and attribute constraint; use graph neural network-based entity linking technology to perform semantic disambiguation and alignment between the entities identified in the preprocessing stage and the known entity library, ensuring the global uniqueness and consistency of the entities; Use pre-trained language models to automatically infer and complete the potential implicit relationships between entities, enhancing the connectivity of the graph; Multi-dimensional quality evaluation: Construct a four-layer evaluation system for background quality, content quality, application quality and knowledge value density; use the dynamic weight distribution strategy of AHP and entropy weight method fusion to assign different weights to each dimension; Construct a knowledge signal-to-noise ratio model to quantify the ratio of core knowledge elements to noise knowledge elements; calculate the comprehensive quality score of the knowledge graph in a specific field by weighted fusion; Intelligent self-repair: Based on the error detection algorithm of inconsistency detection and knowledge conflict analysis, identify logical conflicts, errors or missing information in the knowledge graph; integrate rule-based systems and transfer learning-based prediction correction models to generate repair strategies; dynamically adjust the strength and scope of repair strategies based on the real-time update flow of the knowledge graph; Continuous monitoring and optimization: Real-time quality monitoring and periodic deep evaluation of the knowledge graph; When a new type of vulnerability is detected, trigger the solution generation process based on case-based reasoning and decision fusion; Record feedback signals from the repair process through reinforcement learning mechanism to continuously optimize the vulnerability knowledge base and repair strategies, and achieve optimization of the knowledge graph. 2.The intelligent management method of knowledge graph quality evaluation and self-repairing according to claim 1, characterized in that, The specific steps of the multi-dimensional quality evaluation include: Knowledge background quality evaluation: Use Jaccard similarity to measure the similarity of the topological structure between entities, and design attribute similarity calculation functions for discrete, precise numerical and linear relationship types; introduce semantic similarity calculation based on knowledge graph embedding to make up for the shortcomings of symbolic similarity; combine entity similarity and attribute similarity by AHP entropy weight method combination weighting strategy to obtain knowledge background quality score; Knowledge content quality evaluation: Divide the knowledge graph into concept layer, entity layer and relationship layer for hierarchical evaluation: the concept layer is evaluated by its consistency with the definition of the domain ontology term library; the entity layer is evaluated by its description information richness and credibility; the relationship layer is evaluated by the confidence and semantic consistency of the relationship path; Knowledge application quality evaluation: Evaluate the query response, reasoning ability and decision support effect of the knowledge graph in real scenarios; use a pre-defined question set to test the graph, calculate the F1 value or cosine similarity between the output result and the expected target, and quantify its application performance; Knowledge value density evaluation, calculate knowledge signal-to-noise ratio, get the ratio of core knowledge elements and redundant or marginal knowledge elements in knowledge graph; Calculate the multi-factor fusion model based on the topological importance of entity in the graph, the authority of the source and the assertion confidence; Comprehensive quality fusion, the background quality, content quality, application quality and knowledge value density score are aggregated by a nonlinear weighted fusion function, where the weight is adjusted by historical evaluation data, and the comprehensive quality score of the knowledge graph is output; In the fusion process, the knowledge value density score is the adjustment factor, which strengthens or inhibits the first three basic scores. 3.The intelligent management method of knowledge graph quality evaluation and self-repairing according to claim 2, characterized in that, The calculation method of the knowledge signal-to-noise ratio is as follows: wherein: is the total strength of core knowledge elements, computed by aggregating the entity centrality, the knowledge source authority score, and the frequency of citation of this knowledge; is the strength of noise knowledge elements, computed by aggregating the attribute redundancy of the entity, the low confidence of the relation assertion, and the low trustworthiness of the source; The calculation formula of the comprehensive quality fusion is: Wherein: w1, w2, w3 are weight coefficients, are quality scores of background, content and application respectively, is a knowledge value density adjustment factor, when is a positive value, the gain is made to the total score, and the gain effect is marginal utility diminishing. 4.The intelligent management method of knowledge graph quality assessment and self-repairing according to claim 1, characterized in that, The specific steps of the intelligent self-repairing method include: Vulnerability feature extraction and representation, using the topic event extraction method of event graph, structured feature triples are extracted from new vulnerabilities, including trigger words, vulnerability types, and impact objects, and are converted into low-dimensional vector representation for retrieval; Vulnerability matching and repair decision, use the knowledge graph-based vulnerability knowledge base for vectorized semantic retrieval; If the new vulnerability feature matches the record in the library successfully, the corresponding repair script is automatically called to perform patching operation; If the matching fails, the associated vulnerability mining module based on graph attention network is started, and similar vulnerability information is summarized and submitted to the decision fusion engine; Solution generation and knowledge update, the decision fusion engine uses case-based reasoning and Bayesian network for probability reasoning to generate a set of candidate repair solutions, and selects the optimal solution for repair through credibility evaluation; After repair, the features of the new vulnerability and its solutions are automatically structured and updated to the vulnerability knowledge base, realizing the closed-loop accumulation of knowledge.

5. The intelligent management system for knowledge graph quality assessment and self-repairing, characterized in that, The system is used to realize the intelligent management method of knowledge graph quality evaluation and self-repairing according to any one of claims 1-4, comprising the following modules: Knowledge graph construction module, used for realizing multi-source data acquisition, intelligent cleaning, entity relationship extraction and graph construction; A multi-dimensional quality evaluation module is used to perform evaluation and fusion calculation of background quality, content quality, application quality and knowledge value density. The module integrates a configurable weight management submodule, an evaluation index library and a computing engine; Intelligent self-repairing module, including vulnerability feature extractor, vulnerability knowledge base manager, decision fusion engine and repair executor, realizing automatic identification, matching, repair decision and execution of vulnerabilities; Continuous optimization management module, responsible for real-time monitoring, evaluation scheduling, repair learning and meta-knowledge management of knowledge graph, ensuring the continuous optimization ability of the system. 6.The intelligent management system of knowledge graph quality evaluation and self-repairing according to claim 5, characterized in that, The vulnerability knowledge base is a structured knowledge base based on knowledge graph technology, in which vulnerability information is stored in the form of entity-relation, with repair solutions, effective conditions and historical effective record attributes, supporting semantic-based similarity retrieval and associated reasoning.

Images