Ontology-constrained knowledge graph pipeline adaptive arrangement and feedback method and system

Abstract

The application discloses a body constraint knowledge graph pipeline adaptive arrangement and feedback method and system, and belongs to the technical field of knowledge engineering and data quality management. A five-tuple body constraint descriptor is constructed; a pipeline topology is automatically generated through semantic pattern alignment; a quality gate is set before warehousing, a quality score vector Q and a defect detail D are calculated; when the quality score vector Q and the defect detail D do not reach the standard, the main cause of the defect is determined according to the ratio method and the pipeline is dynamically adjusted, and re-inspection is performed until the standard is reached; after reaching the standard, cross-dimension consistency verification is performed, and an incremental triple carrying a version signature Sigma is pushed; a repair strategy is reversely deposited.Ψ, forming a bidirectional closed loop. The body driven arrangement, quantitative evaluation and adaptive repair before warehousing, and version incremental release are realized, the operation and maintenance cost is reduced, the rectification period is shortened, and the graph consistency and version traceability are guaranteed.

Description

Technical Field

[0001] This invention relates to the fields of knowledge engineering, knowledge graph construction, big data ETL pipeline orchestration, and data quality governance, specifically to a method and system for adaptive orchestration and closed-loop quality feedback of knowledge graph data production pipeline based on domain ontology constraints. Background Technology

[0002] Knowledge graphs, as structured knowledge carriers of entities, relationships, and attributes, have been widely applied in core business scenarios such as intelligent retrieval, financial risk control, public opinion analysis, and intelligent decision-making. Enterprise-level knowledge graphs rely on multi-source heterogeneous data access and complete iterative construction and continuous updates through a full pipeline process including ETL collection, data cleaning, entity relationship extraction, knowledge fusion, and graph storage. Existing knowledge graph data production and quality governance technologies have the following prominent shortcomings: 1. Ontology modeling and pipeline orchestration are disconnected: Traditional ETL pipelines rely solely on manual configuration of node processes based on data table structures. When the categories, attributes, relation cardinality, value range constraints, and business rules of the domain ontology undergo version iterations, they cannot be automatically mapped to pipeline node operators, extracted model parameters, and cleaned rule configurations. This requires manual secondary modifications, which can easily lead to schema heterogeneity inconsistencies, resulting in a large workload for pipeline operation and maintenance configuration and low iteration efficiency.

[0003] 2. Quality management adopts a post-event offline sampling inspection mode: Existing solutions mostly conduct offline quality verification after data is entered into the database. They lack a pre-entry quality gate and cannot quantitatively identify typical quality problems such as missing attributes, duplicate entities, out-of-bounds value ranges, and invalid foreign key associations before data is entered into the database. Furthermore, quality anomalies cannot be accurately traced to specific execution nodes in the pipeline, resulting in long troubleshooting and rectification cycles.

[0004] 3. The annotation system, model rules and ontology version are not bound and linked: the manual annotation specifications, entity extraction model, cleaning and verification rules and domain ontology version are independent of each other. The annotation results cannot be backflowed to iterate ontology constraints, the model parameters cannot be updated adaptively with ontology rules, and the corrected knowledge data is difficult to pass ontology consistency verification, which hinders the knowledge base backflow.

[0005] 4. Inefficient data product release model: The common practice is to export data in its entire lifecycle, which cannot generate incremental triples based on ontology category and relation dimension. It lacks a version signature traceability mechanism, making it difficult for downstream businesses to ensure schema version consistency. Full transmission also leads to a waste of network bandwidth and storage resources.

[0006] 5. The production chain is a one-way linear process: The traditional process only realizes the one-way flow of "modeling - orchestration - production - warehousing - release". The experience of quality defect rectification, the results of annotation, and the feedback from downstream business use cannot be used to iteratively optimize the ontology constraints and pipeline scheduling strategies. The system does not have the ability to adapt and evolve.

[0007] To address the aforementioned technical issues, existing technical document CN115619104A discloses a knowledge graph construction method, but its ontology model and ETL pipeline are manually configured and cannot achieve automatic synchronous updates; document CN114492653A proposes a data quality detection scheme, but it adopts a post-event offline sampling mode and lacks a quality gate before data entry. Summary of the Invention

[0008] The purpose of this invention is to overcome the above-mentioned defects of the prior art and provide a method and system for adaptive orchestration and closed-loop quality feedback of knowledge graph data production pipeline based on domain ontology constraints. This method and system realizes automatic generation and dynamic updating of versioned ontology-driven pipelines, multi-dimensional quality quantification assessment before data entry, intelligent determination of the main cause using the quality defect ratio method, adaptive closed-loop repair in multiple scenarios, reverse sedimentation of repair results into iterative ontology, incremental release of ontology subsets with version, and bidirectional self-optimization closed loop across the entire link.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: This invention discloses an adaptive orchestration and closed-loop quality feedback method and system for knowledge graph data production pipelines based on domain ontology constraints, comprising: S100: Construct a version-manageable domain knowledge ontology and generate an ontology constraint descriptor with a 5-tuple structure. The version signature Σ performs a hash operation on the serialized content of C, A, R, and Ψ to ensure... Uniquely bound to version. Ψ includes rules for required attributes, value range validation, foreign key references, and quality thresholds. Executed by the knowledge ontology modeling module, the output is as follows: Figure 4 The ontology constraint descriptor node is shown in Figure 206.

[0010] S200: Receives external data source metadata (corresponding to node 207 in Figure 4) and target knowledge base identifier, completes field-attribute matching through the semantic pattern alignment engine, automatically generates pipeline topology and configures processing operators; A mapping is established with the pipeline topology and operators. The class inheritance relationship of C determines the node hierarchy, the mandatory constraints of A drive the selection of the 201 source access node field, the cardinality constraint of R determines the batch parameters of the 203 entity relationship extraction node, and the verification entries of Ψ generate the 209 quality gate detection template. The five core operators correspond to 201 (source access), 202 (data cleaning), 203 (entity relationship extraction), 204 (knowledge fusion), and 205 (graph lake entry) in Figure 4. This is executed by the pipeline topology generation module, and semantic pattern alignment uses a three-dimensional similarity model. , α+β+γ=1 and α>β>γ.

[0011] S300: Press Figure 4 The topology shown executes schedules 201-204 sequentially, setting a quality gate 209 before warehousing at 205. Based on... .Ψ Calculate the quality score vector Q=( , , The system consists of an ontology and a detailed defect set D. Higher scores for each component indicate better quality. Each record in D is a quadruple of (ontology category, entity identifier, defect type, pipeline failure node). The calculation logic is detailed below. Figure 5 Node 310 on the left; Figure 5 In node 310, D and Q are shown together, representing the quality assessment results of the synchronous output. Threshold comparison is only performed on... , , The three components are compared with T item by item.

[0012] S400: From .Ψ Read the quality threshold vector T=( , , Q satisfies T, meaning ≥ and ≥ and ≥ If Q already satisfies T, proceed to S500; otherwise, when any component of Q is below the threshold, proceed as follows. The root cause is identified and dynamic adjustments are triggered. Only one root cause is selected at a time to execute the corresponding control strategy; 122 / 123 / 124 and 301 / 302 / 303 are mutually exclusive, not executed in parallel. After control is completed, the adaptive closed-loop feedback module triggers a return to S300 to recalculate Q, forming a quality re-inspection inner loop (see...). Figure 1The process continues until Q satisfies T; if the preset maximum number of re-checks is reached but T is still not satisfied, then 205 is blocked from entering the database and an alarm is output, while the unclosed defect is written to D. The decision-making process is shown in Figure 3, Node 120 (rough judgment, corresponding to Node 312 in Figure 5). Figure 5 Nodes 313-303; Missing attributes → 122 / 301 (adjust θ + annotation), Duplicate attributes → 123 / 302 (disambiguation + incremental fusion), Consistent attributes → 124 / 303 (rule switching + blocking entry into the lake). If determined to be an attribute missing attribute class, lower the entity relation extraction confidence threshold and generate an attribute completion and annotation task, adjusting the formula. =max( θ−η・( - The system uses both intersection-union ratio and sample consistency rate to determine the pre-labeling results. If the result is determined to be an entity duplication class, entity disambiguation processing is initiated and the incremental fusion sub-pipeline is triggered. If the result is determined to be a data inconsistency class, the cleaning rule version is switched and abnormal data is blocked from entering the database.

[0013] S500: After Q satisfies T, perform a cross-dimensional consistency check on the instance after pipeline processing and S400 dynamic adjustment—that is, based on... In the C / A / R / Ψ framework, instances undergo joint validation of class attribution, attribute range, relation cardinality, and foreign key references. Upon successful validation, the data is written to the knowledge base, and an incremental triple carrying Σ is generated based on the snapshot difference constrained by the C class subset and the R relation subset. Figure 3 121 enters the lake / release node and Figure 5 The 314 lake entry point meets the standards.

[0014] S600: After S500 is completed, the repair strategy parameters generated by the dynamic adjustment of S400 are written in reverse. Ψ, incrementing v and recalculating Σ, triggers topology synchronization update. This is achieved by the knowledge ontology modeling module and... Figure 3 The 125 closed-loop feedback link executes collaboratively, returning to S100 / S1 via 125, forming a solidified outer loop relative to the inner loop of S400→S300, and... Figure 1 The steps in S600 are consistent and together they form a two-way closed loop of body constraints, pipeline arrangement, and quality feedback.

[0015] The present invention also provides a system, such as Figure 2As shown, the system includes a knowledge ontology modeling module, a pipeline topology generation module, a pipeline execution and quality inspection module, an adaptive closed-loop feedback module, and a closed-loop re-feedback and product release module. These modules are interconnected via a unified API service layer and a data bus. The knowledge ontology modeling module executes steps S100 and S600; the pipeline topology generation module executes step S200; the pipeline execution and quality inspection module executes step S300; the adaptive closed-loop feedback module executes step S400; and the closed-loop re-feedback and product release module executes step S500. Each module is implemented through an API service layer and a data bus. The asynchronous transmission of the quality score vector Q forms a complete closed loop of ontology—pipeline—quality inspection—repair—warehousing—release—reverse iteration.

[0016] Beneficial effects 1. Versioned ontology drives automatic pipeline iteration, significantly reducing operation and maintenance costs: By building... It also establishes a strict mapping between itself and pipeline operators, and automatically updates pipeline topology and operator configuration when the ontology version is iterated, thus solving the problem of the disconnect between ontology modeling and pipeline orchestration.

[0017] 2. Pre-warehousing quality gate and precise traceability significantly shorten the quality repair cycle: Set up a quality gate before warehousing, calculate Q and defect details D, and achieve precise traceability of quality defects to specific pipeline execution nodes.

[0018] 3. Adaptive closed-loop feedback and rule self-iteration significantly improve data quality: based on The minimum ratio criterion is used to determine the root cause of the defect, and differentiated adjustments are implemented for three types of scenarios. The remediation strategies are then documented and refined. Ψ.

[0019] 4. Incremental triples and version retrospective release: Efficiently save resources and ensure consistency: Incremental triples are generated based on snapshot differences and pushed with version signature Σ to ensure downstream schema consistency.

[0020] 5. End-to-end bidirectional self-optimizing closed loop, adapting to long-term iterative evolution needs: through positive driving ( The system achieves adaptive evolution through pipeline orchestration and reverse iteration (quality feedback-driven ontology updates). Attached Figure Description

[0021] Figure 1 is a schematic diagram of the system method of the knowledge graph data production pipeline adaptive orchestration and closed-loop quality feedback method and system of the present invention; wherein, the inner loop of quality re-examination is formed between step S400 and step S300 by returning to recalculate the quality score vector Q; the outer loop of ontology sedimentation is formed between step S600 and step S100 by ontology version update, and the two together realize the bidirectional closed loop of ontology constraint-pipeline orchestration-quality feedback; Figure 2 shows the overall system architecture of the adaptive orchestration and closed-loop quality feedback method and system for knowledge graph data production pipeline of the present invention; it includes a knowledge ontology modeling module, a pipeline topology generation module, a pipeline execution and quality inspection module, an adaptive closed-loop feedback module, a closed-loop refeedback and product release module, as well as a unified API service layer and data bus; the execution order of each module is as follows: Figure 1 The S100-S600 series correspond to each other. Figure 3 This is a schematic diagram of the adaptive closed-loop feedback process of the knowledge graph data production pipeline adaptive orchestration and closed-loop quality feedback method and system of the present invention; where node 120 corresponds to Figure 5 Threshold determination at node 312; node 125 → S1 corresponds to the bulk precipitation outer ring in step S600. Nodes 122, 123, and 124 are mutually exclusive branches, based on... The minimum ratio criterion triggers one of three control mechanisms: missing attributes, duplicate entities, and inconsistent data. See the detailed logic below. Figure 5 ; Figure 4 This is a schematic diagram of the data production pipeline topology of the knowledge graph data production pipeline adaptive orchestration and closed-loop quality feedback method and system of the present invention; where nodes 201-205 are, in order, source access, data cleaning, entity relation extraction, knowledge fusion, and graph entry into the database; node 209 is the quality gate before data entry. Node 208 is an optional relation completion sub-node, which is selected when the minimum cardinality of the relation is greater than zero and the fill rate is lower than the lower limit of relation fill. When the control flow is less than 204, node 210 is an optional entity alignment child node, inserted between nodes 203 and 204; At that time, the control flow is inserted between nodes 204 and 209.

[0022] Figure 5 This is a schematic diagram illustrating the quality score vector calculation and decision-making logic of the adaptive orchestration and closed-loop quality feedback method and system for knowledge graph data production pipeline of the present invention; node 310 synchronously outputs the quality score vector Q and defect details D; nodes 312 and 313 complete threshold comparison and... The main cause is determined; nodes 301, 302, and 303 are mutually exclusive control branches. Executing one of them will return to step S300 to recalculate Q, thus forming... Figure 1 The inner loop of quality re-inspection is shown; node 314 indicates that the lake meets the standard after Q meets the threshold. Detailed Implementation

[0023] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0024] Preferred implementation parameters (all fixed parameters are listed here and are not included in the claims) In a preferred engineering embodiment of the present invention, the values ​​of each parameter are as follows: Semantic similarity weights: α=0.52, β=0.23, γ=0.25; High similarity threshold =0.83, low similarity threshold =0.58; Ambiguity threshold δ = 0.07; Relationship filling lower limit =0.90, lower limit of entity alignment rate =0.83; Consistency weight: =0.23、 =0.42、 =0.35; Quality threshold: =0.895, Repeat Score Threshold =0.965, Consistency score threshold =0.900; Confidence parameters: The initial threshold θ is 0.74, θ_min=0.58, and η is a preset adjustment coefficient (preferably η=0.62. Experiments have verified that η can converge stably in the range of 0.5-0.8. Among them, when η=0.5, the convergence speed is slower but the accuracy is higher, and when η=0.8, the convergence speed is faster but there is a risk of over-adjustment. η=0.62 is the preferred value to balance convergence speed and stability). The above are only preferred configurations. Those skilled in the art can make adaptive fine-tuning according to business scenarios, which does not limit the scope of protection of this invention.

[0025] Implementation steps: The present invention provides an adaptive orchestration and feedback method for ontology-constrained knowledge graph pipelines. In this embodiment, steps S1-S6 correspond to steps S100-S600 in claim 1, and the specific correspondence is as follows: S1↔S100 (ontology modeling), S2↔S200 (pipeline topology generation), S3↔S300 (pipeline execution and quality inspection). S4↔S400 (Adaptive Closed-Loop Feedback), S5↔S500 (Closed-Loop Recharge and Product Release), S6↔S600 (Closed-Loop Feedback and System Update).

[0026] Combined with appendix Figure 1-5 Specifically, it includes the following steps: S1, Ontology Modeling (S100): Corresponding to Figure 3 S1 and Figure 4 206 nodes in the middle. The knowledge ontology modeling module defines C, A, R, Ψ, and Σ, and generates... This serves as a unified basis for subsequent pipeline arrangement and quality verification.

[0027] S2, Pipeline Topology Generation (S200): Corresponding Figure 3 S2 and Figure 4 The module receives data source metadata (207) and ontology constraint descriptors (206), and automatically generates quality gate topology nodes 201-205 and 209; 208 is an optional relation completion child node, used when the minimum cardinality constraint of the relation is greater than 0 and the fill rate is less than 0. ( Figure 4 middle Insert between 203 and 204; 210 is an optional entity alignment child node, inserted when the entity alignment rate is lower than 100%. The control flow is inserted between 204 and 209.

[0028] S3. Pipeline Execution and Quality Inspection (S300): Before entering the lake at 205, calculate Q and D at the 209 quality gate. The calculation logic is as follows: Figure 5 Node 310.

[0029] S4, Adaptive Closed-Loop Feedback (S400): Figure 3 Node 120 and Figure 5 Node 312 performs a threshold comparison. Figure 5 Node 313 execution The main cause is determined, and one of the following is triggered: 122 / 301, 123 / 302, or 124 / 303. After the adjustment is completed, return to S3 / S300 to recalculate Q (quality re-inspection inner loop).

[0030] S5, Closed-loop backfeeding and product release (S500): After verification, write to the knowledge base and push the incremental triplet carrying Σ.

[0031] S6, Closed-loop feedback and ontology update (S600): via Figure 3 S125 returns to S100, and the repair strategy is then stored in... .Ψ.

[0032] Example 1: Application of Financial Public Opinion Knowledge Graph Production This embodiment is deployed in a container cluster environment, with orchestration and scheduling based on Kubernetes. The underlying graph database is used as the knowledge base storage, and the entity reasoning model and sentiment classification model are deployed as an online reasoning service. The entity fingerprint index and task queue are implemented based on Redis, and alarms are pushed via the web.

[0033] Construct a financial public opinion graph knowledge base with approximately 860,000 entities. Define the Company class and the Event class for public opinion events. Configure the Event class with required attributes for type (event type), time (occurrence time), and source (information source). Define the "occurred at" relationship between Company and Event and set cardinality constraints.

[0034] S1. Construct and publish ontology constraint descriptors (correspond Figure 4 (Node 206 in the middle), bind version signature, and configure the required field identifier, value range and foreign key rules for each attribute.

[0035] S2. Connect to a JSON format log data source (corresponding to...) Figure 4 (207 nodes in the middle), field similarity meets the threshold for automatic mapping, generating a standard five-node pipeline topology (corresponding to) Figure 4 (Nodes 201-205 and 209 in the middle).

[0036] S3, processing 86,000 entities per day, a quality assessment was triggered when an attribute missing score was slightly below the threshold. Figure 4 209 nodes in Figure 5 The quality determination is triggered at nodes 310-312 in the middle.

[0037] S4 Determined as a missing class (corresponding) Figure 5 313→301 Figure 3 (122), automatically lower the extraction threshold, and return to S3 to recalculate Q until the target is met; this batch only performs missing class control according to the one-choice principle, and does not trigger duplicate class disambiguation in parallel.

[0038] S5. After successful verification, the knowledge base is re-injected, generating an incremental triplet with a version signature for push (corresponding to...). Figure 3 The 121 lake entry / publishing nodes in Figure 5 are shown as 314.

[0039] S6: The repair strategy is refined through 125 closed-loop feedback. .Ψ( Figure 3 (125 nodes in the middle).

[0040] Example 2: Adaptive Pipeline Update for Ontology Version Iteration The ontology has been upgraded from v1.2 to v1.3, adding a required attribute "sentiment" to the Event class; the system automatically generates new version ontology constraint descriptors (corresponding appendix). Figure 4 After updating node 206 in the entity relationship and identifying the newly added required attributes, the node is automatically extracted without manual intervention. Figure 4Insert a sentiment classification sub-node after node 203 in the middle, and bind it to the SA-v2 sentiment analysis model; automatically generate attribute completion tasks for historical event entities, and at the lake inflow quality gate (attached) Figure 4 In node 209, a missing sentiment attribute statistic was added, and the triggering conditions for optional nodes 208 and 210 were updated simultaneously to verify the automatic iteration effect of the versioned ontology-driven pipeline.

[0041] This embodiment demonstrates Figure 4 When the 206-node version is updated, the system automatically adjusts the operators of node 203, the optional child nodes of node 208, and the quality detection rules of node 209. The technical effectiveness of the "versioned ontology-driven pipeline automatic iteration" of the present invention has been verified.

[0042] Example 3 Comparative Experiment Experimental environment: Detailed description of the experimental procedure (combined with...) Figure 1-5 ): 1. Data preprocessing stage (corresponding to...) Figure 4 (Nodes 201-202 in the middle) Control group: Data flow was manually configured using Apache NiFi, financial sentiment log data was imported, and data cleaning rules were configured.

[0043] Experimental group: The system of this invention is used to automatically generate pipeline topology based on ontology constraint descriptors. Automatically configure operators for each node.

[0044] 2. Model binding stage (corresponding to) Figure 4 (Node 203 NER-24Q3 in the middle) Control group: The NER model was trained manually, and the entity relation extraction parameters were configured.

[0045] Experimental group: The system automatically binds the NER-24Q3 model, according to The cardinality constraint of the relational subset in R automatically adjusts the extraction parameters.

[0046] 3. Quality assessment phase (see attached document) Figure 4 209 nodes in Figure 5 (Nodes 310-312 in the middle) Control group: After the data was entered into the database, Great Expectations was used for offline quality testing to generate a quality report.

[0047] Experimental group: Before entering the lake, a multidimensional quality quantification assessment was performed on the quality gate to generate a quality defect vector Q (corresponding to...). Figure 5(310 nodes in the middle).

[0048] 4. Defect Repair Phase (corresponding to) Figure 5 (Nodes 313-303 in the middle) Control group: Quality reports were manually analyzed to pinpoint problem areas and adjust pipeline configuration manually.

[0049] Experimental group: The system automatically determines the root cause of the defect (corresponding to...) Figure 5 (Node 313 in the middle), triggering the adaptive control strategy (corresponding to) Figure 5 (Nodes 301-303 in the pipeline) automatically adjust the pipeline topology.

[0050] 5. Iterative optimization phase (corresponding to) Figure 3 (125 closed-loop feedback nodes in the middle) Control group: Experience was manually summarized, and ontology constraints and pipeline configurations were manually updated. Experimental group: System automatic sedimentation repair strategy to Ψ subset (corresponding appendix) Figure 3 (125 nodes in the system) to achieve rule self-iteration.

[0051] Evaluation indicators and results: The above indicators were tested repeatedly for 90 days, with daily data recording. A paired-samples t-test was used to assess statistical significance, with degrees of freedom (df) = 89 and a significance level (α) = 0.05. The test was two-tailed. The difference between the experimental and control groups was determined by a paired-samples t-test, with a t-statistic of 15.32 and a p-value << 0.01 (specifically, p = 2.3 × 10⁻¹). 6 The results were statistically significant. Cohen's effect size was 2.81, which is considered a large effect size. The 95% confidence interval was [2.15, 2.89], indicating that the experimental results were stable and reliable. The standard deviations were all less than 5%, further verifying the reproducibility of the experimental results.

[0052] This experiment uses a hardware configuration of an 8-core CPU, 32GB of memory, and 1TB of SSD storage, which can meet the daily batch processing requirements of the financial public opinion knowledge graph.

[0053] Experimental results show that, under this configuration, the system of the present invention can operate stably, and its performance indicators are significantly better than those of the control group.

[0054] Experimental data verified Figure 5 The complete process of Q generation, threshold comparison, defect* determination, and 301-303 classification and control.

Claims

1. An adaptive orchestration and feedback method for ontology-constrained knowledge graph pipelines, characterized in that, Includes the following steps: S100: Construct a version-manageable domain knowledge ontology and generate an ontology constraint descriptor with a 5-tuple structure. Where C is the class set, A is the attribute set, R is the relation set with cardinality constraints, Ψ is the set of structured validation rules including mandatory attributes, value range, foreign key references and quality threshold rules, and Σ is the version signature of the serialized content of C, A, R and Ψ. S200: Receives external data source metadata and target knowledge base identifier, based on... Semantic pattern alignment is performed between C, A, and data source pattern features. Based on the alignment results, a data production pipeline topology is automatically generated, and each node is configured with [the appropriate configuration]. The processing operators associated with C, A, R, and Ψ; A mapping is established with the pipeline topology and processing operators, where the class inheritance relationship of C determines the topology node level, the mandatory attribute constraint of A drives the field selection of the data extraction operator, the relation cardinality constraint of R determines the batch parameters of the entity relation extraction operator, and the verification entries of Ψ generate the quality gate detection template. S300: Schedule each node to execute according to the pipeline topology, and set a quality gate before the knowledge graph is entered into the database; based on The entity completeness score is calculated for each of the Ψ items. Deduplication quality score and data consistency score The composite quality score vector Q=( , , The system generates a defect detail set D; higher scores indicate better quality; each record in D contains a quadruple location information of (ontology category, entity identifier, defect type, pipeline fault node); S400: From The Ψ subset reads the quality threshold vector T=( , , The condition Q has been satisfied; T refers to... ≥ and ≥ and ≥ If Q already satisfies T, then execute S500; otherwise, when any component of Q is lower than the corresponding threshold, the minimum ratio criterion applies. Determine the root cause of the defect and trigger dynamic adjustment of the pipeline topology based on the root cause of the defect; the dynamic adjustment includes at least one of the following: lowering the confidence threshold of entity relationship extraction, starting entity disambiguation fusion sub-pipeline, switching the cleaning rule version and blocking abnormal data from entering the database; after triggering the dynamic adjustment, return to execute S300 to recalculate Q until Q satisfies T; if the preset maximum number of re-inspections is reached but T is still not satisfied, then block data from entering the database and output alarm information; S500: After Q satisfies the quality threshold vector T, the knowledge instance after pipeline processing and dynamic adjustment is based on... Perform cross-dimensional consistency checks. Once the checks pass, write the data to the target knowledge base and use it as the subject of the check. The difference between the current and historical snapshots of the graph constrained by the ontology class subset of C and the relation subset of R generates an incremental triplet carrying Σ. S600: After S500 is completed, the repair strategy parameters generated by the dynamic adjustment triggered by S400 during the current pipeline operation are written in reverse. The Ψ subset is incremented, the ontology version identifier v is recalculated, and the pipeline topology and operator configuration are updated synchronously to form a two-way closed loop of ontology constraints, pipeline orchestration and quality feedback.

2. The method according to claim 1, characterized in that, The ontology constraint descriptor Defined as: Wherein: C is a set of ontology class hierarchies defined using the OWL standard, which includes class identifiers, class tags, and class inheritance relationships; A is a set of attribute definitions that are bound to various types in C. Each attribute definition includes an attribute identifier, data type, required field identifier, and value range constraint expression. R is a set of relations with cardinality constraints. Each relation definition includes a relation identifier, domain class, range class, minimum cardinality, and maximum cardinality. The Ψ is a set of structured validation rules, including rules for mandatory attributes, rules for value range validation, rules for foreign key references, and rules for quality thresholds. Each rule is encoded in JSON-LD format. The Σ is a version signature generated by hashing the serialized content of C, A, R, and Ψ using the SHA-256 hash algorithm, ensuring... Content and version are uniquely linked; The 'v' is the ontology version identifier, which adopts a semantic versioning format, MAJOR.MINOR.PATCH, where MAJOR represents incompatible ontology structure changes. MINOR indicates the addition of backward compatibility features, and PATCH indicates the fixing of backward compatibility issues.

3. The method according to claim 1, characterized in that, The semantic pattern alignment adopts a method based on A 3D semantic similarity model driven by subset A: Calculate the semantic similarity between the data source field f and the ontology attribute a. α + β + γ = 1, and α > β > γ; The This is the string similarity between field names and attribute labels based on edit distance; The To determine the compatibility between field data types and attribute data types based on a data type compatibility matrix, the data type compatibility matrix is ​​predefined in... The Ψ subset; The This refers to the percentage overlap between the field value range sampling and the attribute value range constraint expression. Set a high similarity threshold Low similarity threshold When sim(f,a)≥ Automatic binding at any time ≤sim(f,a)< The time marker is for manual confirmation, sim(f,a)< The time stamp is Expand the candidates; when the difference between the best and second-best attribute similarity is less than the ambiguity threshold δ, it is marked as ambiguous and pending confirmation.

4. The method according to claim 1, characterized in that, The quality score vector Q=( , , ),in ,in The number of instances of category c. This represents the total number of instances. This represents the number of missing required attributes for category c. This represents the total number of required attributes for category c. =( ); The number of duplicate entities detected using SimHash or MinHash fingerprinting based on the identifier attributes of subset A; = + + The weighting coefficients satisfy the normalization constraint + + =1; For field specification compliance rate, For the value range validation pass rate, This represents the foreign key resolvability rate.

5. The method according to claim 1 or 4, characterized in that, In step S400: when multiple When all values ​​are the minimum, the primary cause of the defect is determined based on priority: missing, duplicate, or consistent. If the maximum number of re-checks is reached but T is still not met, the unclosed defect is written into the defect detail set D and an operation and maintenance alarm is triggered. If the defect is determined to be missing an attribute, the confidence threshold for entity relationship extraction is lowered. =max( θ−η・( - And generate attribute completion annotation tasks; When an entity is determined to be duplicated, the entity disambiguation and incremental fusion sub-pipeline is started; when data is determined to be inconsistent, the cleaning rule version in Ψ is switched and abnormal data is blocked from entering the database; when the minimum cardinality of a relation in R is greater than zero and the fill rate is lower than the lower limit, a relation completion child node is inserted after the entity relation extraction node; when the entity alignment rate is lower than the lower limit, an entity alignment child node is inserted after the knowledge fusion node.

6. The method according to claim 1, characterized in that, The repair strategy parameters written in reverse in step S600 include at least one of the following: confidence adjustment coefficient, cleaning rule version identifier, disambiguation fusion parameter, and quality threshold adjustment amount.

7. An adaptive orchestration and feedback system for ontology-constrained knowledge graph pipelines, characterized in that, include: The knowledge ontology modeling module is used to execute step S100 and step S600 after step S500 is completed; The pipeline topology generation module is used to execute S200 and listen. Version changes are synchronized with topology and operators; The pipeline execution and quality inspection module is used to execute S300; The adaptive closed-loop feedback module is used to execute S400. When Q meets the quality threshold vector T, it triggers the closed-loop backfeed and product release module to execute S500. When T is not met, it triggers the dynamic adjustment and control of the pipeline execution and quality inspection module to re-execute S300. When the preset maximum number of re-inspections is reached and T is still not met, it blocks data entry into the database. The closed-loop power-on and product release module is used to execute S500; Each module is interconnected with a data bus via a unified API service layer, and the data bus transmits data asynchronously via message queues. And the quality score vector Q.

8. The system according to claim 7, characterized in that, The knowledge ontology modeling module generates The Ψ in the middle uses JSON-LD encoding, and the Σ uses SHA-256 to form a version signature chain; the pipeline topology generation module uses the three-dimensional semantic similarity model defined in claim 3 for semantic pattern alignment; the adaptive closed-loop feedback module uses the defect main cause judgment criterion and pipeline dynamic adjustment strategy defined in claim 5; the closed-loop backfeeding and product release module generates an incremental triple carrying the version signature Σ and pushes it to the target knowledge base.

9. A computer device comprising a processor and a memory, the memory storing a computer program, wherein the processor, when executing the computer program, implements the method of any one of claims 1 to 6.

10. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method of any one of claims 1 to 6.

Images