A data standard intelligent reference and automatic review closed loop method fusing large model implicit rule mining

Abstract

The application discloses to the technical field of artificial intelligence auxiliary examination, specifically to a data standard intelligent reference and automatic examination closed loop method fusing large model implicit rule mining, first, obtaining explicit standards and historical sample data, constructing an intelligent rule library, then performing semantic recognition and standard intelligent reference on a model design object, generating a semantic vector, automatically binding rules and rules to be confirmed; subsequently, performing multi-dimensional automatic examination according to the bound rules, generating hierarchical problem evidence of naming, code, field definition, architecture compliance and context early warning; finally, according to the examination problems, rectification results and version difference contraction review range, calculating the access control risk score to obtain a pre-release pass or block conclusion; based on this, the application can update the rule credibility through examination, rectification, manual review and online quality feedback, so that the rule library can adapt to the changes in model design practice.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence-assisted review technology, specifically to a data standard intelligent citation and automatic review closed-loop method that integrates large-scale model implicit rule mining. Background Technology

[0002] In the process of enterprise-level data model design, the model typically includes objects such as entities, fields, relationships, code values, named results, and value domain definitions. These objects will be reused in subsequent data processing, API calls, report analysis, access control, and data asset management processes. Therefore, the standardization and compliance of the model design phase directly affects the stability of subsequent data processing processes.

[0003] Current data standard reviews typically rely on explicit standard documents, manual approval experience, or fixed rule verification. Explicit standards can constrain field naming, code values, field types, and architectural hierarchy, but in long-term modeling practice, a large number of empirical rules are formed that are not formally written into standard documents. If these implicit rules cannot be identified and transformed into enforceable review provisions, the standard review will still suffer from insufficient coverage.

[0004] A common technical approach is to perform post-submission validation based on manually entered rules. This method can cover explicitly registered standards, but it doesn't fully utilize the implicit modeling experience accumulated in historical compliant and rejected models. As the model size increases, if the rule base relies solely on manual maintenance, it is prone to issues such as delayed rule updates, insufficient rule coverage, and inconsistent review results.

[0005] Another technical approach is to perform hard matching on model object attributes, such as matching by object type, field name, or field type rules. This method is simple to implement, but it lacks support for objects with similar semantics, significantly different names, or those that require contextual relationships for judgment. For example, if a field name does not directly contain a sensitive meaning, but the field description, the entity it belongs to, and historical similar fields all point to sensitive information, relying solely on character-level matching is prone to missing detections.

[0006] In addition, existing designs mostly stay at the process level of problem discovery, rectification, and review. The review results, rectification records, manual review opinions, and data quality feedback after going live are not stably fed back to the rule base, which makes it impossible to dynamically adjust the rule confidence based on subsequent verification results, and the review capability is also difficult to continuously update with the model design practice.

[0007] The existing technical challenges mainly lie in three aspects. First, implicit modeling rules are scattered across historical models, rejection records, and manual review opinions, lacking a mechanism for automatic extraction, evaluation, and storage from samples. Second, standard citations need to consider object type, semantic meaning, subject domain, and contextual relationships simultaneously; relying solely on hard matching is insufficient to reliably locate the most suitable rule. Third, the review loop needs to reflect the problem results, rectification process, review conclusions, and post-launch feedback in reverse to the rule base; otherwise, it's merely a process loop and cannot achieve continuous iteration of rule capabilities.

[0008] To address the aforementioned issues, this invention provides a closed-loop method for intelligent data standard referencing and automatic review, integrating implicit rule mining from large-scale models. This method merges explicit standard rules and implicit empirical rules mined from historical models into an intelligent rule base. When a model design object is created, modified, saved, or submitted, it combines basic matching, semantic matching, and contextual matching to complete intelligent standard referencing. Subsequently, multi-dimensional automatic review is performed, generating hierarchical problem evidence. During the rectification and re-review stages, the scope of re-review is narrowed based on version differences, object associations, and problem status. Finally, the review, rectification, verification, and deployment feedback are fed back to the rule base, achieving rule confidence updates and self-evolution of review capabilities. Summary of the Invention

[0009] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0010] A data standard intelligent citation and automatic review closed-loop method that integrates implicit rule mining from large models includes the following steps:

[0011] S1, Explicit standards and implicit rules are integrated to build an intelligent rule base: Obtain explicit standards and historical sample data to build an intelligent rule base. The data includes explicit standard documents, historical models, review opinions and online quality feedback.

[0012] S2, Semantic Recognition and Standard Intelligent Reference of Model Design Objects: Perform semantic recognition and standard intelligent reference on model design objects to generate semantic vectors, automatic binding rules and rules to be confirmed;

[0013] S3, Multi-dimensional Automatic Review and Layered Issue Localization: Performs multi-dimensional automatic review based on bound rules, generating layered issue evidence including naming, code, field definition, architecture compliance, and contextual alerts;

[0014] S4, Rectification Review and Pre-Release Closed-Loop Control: Based on the review issues, rectification results and version differences, narrow the scope of the review, calculate the access control risk score, and obtain the pre-release approval or blocking conclusion;

[0015] S5, Review Data Feedback and Rule Base Self-Evolution: Update rule credibility, matching threshold and access control threshold based on feedback from review, rectification, verification and launch, forming a closed loop of rule base self-evolution.

[0016] As a preferred embodiment of the data standard intelligent referencing and automatic review closed-loop method for integrating large-model implicit rule mining described in this invention, the specific steps of S1 are as follows:

[0017] Extract explicit rules from explicit standard documents, code standards, naming conventions, field definition conventions, and architectural constraints;

[0018] Training samples were constructed from historical compliance models, historical rejection models, and manual review opinions, with historical compliance models serving as positive samples and historical rejection models or repeated rectification models serving as negative samples.

[0019] The large model is used to extract candidate implicit rules from field naming, entity relationships, value domain binding, code set references and rectification opinions, and the output of the large model is transformed into rule number, rule content, applicable object, applicable stage, positive and negative sample hit status and example object;

[0020] The credibility of candidate implicit rules is calculated. When the credibility of a candidate implicit rule reaches or exceeds the threshold, it is marked as an implicit empirical rule and added to the intelligent rule base. When the credibility is between the threshold, it is included in the observation rules or recommended confirmation rules. When the credibility is below the threshold, it is not used for automatic review. At the same time, for rules that have already been added to the base, they are downgraded or deactivated.

[0021] Explicit rules and implicit empirical rules filtered by credibility are merged into an intelligent rule base, and each rule is recorded with information such as rule source, credibility, applicable objects, applicable stage, start / stop status and version.

[0022] As a preferred embodiment of the data standard intelligent referencing and automatic review closed-loop method for integrating large-model implicit rule mining described in this invention, the specific steps of S2 are as follows:

[0023] Collect information on the name, type, subject domain, entity, field description, data type, code set, and value range of entities, fields, relationships, code values, and value ranges for the definition objects; then, semantically represent the object name, description text, and associated entities to form an object semantic vector, and form a rule semantic vector for the rule content, applicable objects, and rule examples;

[0024] A three-layer matching process, comprising basic matching, semantic matching, and contextual matching, is performed to calculate a comprehensive matching score between the model object and the rule. The basic matching filters candidate rules based on object type, applicable stage, and rule start / stop status. The semantic matching determines the semantic closeness based on the similarity between the object's semantic vector and the rule's semantic vector. The contextual matching corrects the matching results based on the subject domain, associated entities, code set, and value domain contextual information.

[0025] The referencing results are determined based on the matching score: rules that reach the automatic binding threshold are automatically bound to objects and participate in subsequent reviews; rules in the recommended range enter the confirmation queue, where modelers or standards administrators confirm their applicability; rules below the filtering threshold do not participate in the review.

[0026] As a preferred embodiment of the data standard intelligent referencing and automatic review closed-loop method for integrating large-model implicit rule mining as described in this invention, the specific steps of S3 are as follows:

[0027] Hard review is performed on rules that have been automatically bound or confirmed; for explicit blocking rules, if the actual design value does not meet the expected value after the rule is hit, a hard review issue is generated directly; for non-blocking rules, rectification suggestions are generated first, and it is decided whether to enter the access control calculation based on the severity.

[0028] For designs that formally meet explicit rules but are clearly inconsistent with similar historical objects, a large model is used to assist in outputting soft review warnings, calculate context anomaly scores, and generate ordinary warnings, high-risk warnings, or defaults to not entering the approval report based on the scores.

[0029] Issues are organized at the model level, object level, and rule item level. The model level is used to present an overview of the issue; the object level is used to locate entities, fields, relationships, code values, or value range definitions; and the rule item level is used to record rule number, rule content, actual value, expected value, matching score, and warning information. The severity of the issue is determined based on the default severity of the rule, the scope of the object's impact, historical rectification costs, and soft review risks.

[0030] As a preferred embodiment of the data standard intelligent referencing and automatic review closed-loop method for integrating large-model implicit rule mining as described in this invention, the specific steps of S4 are as follows:

[0031] Generate rectification suggestions for the review issues to provide directions for modification, and record the rule number, matching score, expected design value and reason for rectification.

[0032] After rectification, compare the differences between the model versions before and after rectification, identify the modified objects, related objects, objects affected by rule version changes, and unclosed problem objects, and the review will only cover the above objects, while the valid conclusions of the previous round will be used for unaffected objects.

[0033] Before release, an access control risk score is calculated based on unclosed issues and historical risks. If the access control risk score is higher than the threshold, the current model version will not enter the pre-release approval state. If the access control risk score is between the threshold, manual review or supplementary approval comments are required. If the access control risk score is lower than the threshold and the approval conclusion is passed, the current model version is marked as a compliant version. At the same time, if there are unclosed issues with a severity level of blocking, release will be directly blocked even if the comprehensive score does not exceed the blocking threshold.

[0034] As a preferred embodiment of the data standard intelligent referencing and automatic review closed-loop method for integrating large-model implicit rule mining as described in this invention, the specific steps of S6 are as follows:

[0035] The review results, rectification records, manual review opinions, review conclusions, and post-launch quality feedback will be fed back as rule samples; positive feedback will be added to the corresponding rules for issues that are confirmed to be valid, and negative feedback will be added to the rules that are manually rejected or have been falsely reported multiple times.

[0036] The credibility of implicit rules is periodically recalculated, and based on the updated credibility, the rules are either retained, downgraded, deactivated, or re-mined.

[0037] Adjust the matching threshold and access control threshold based on feedback;

[0038] After the rule base is iterated, the new rule version enters the next round of model design object review, thus forming a self-evolving process of review, rectification, re-review, online feedback and rule update.

[0039] Compared with existing technologies:

[0040] 1. This invention expands data standard review from simple explicit rule verification to a review that integrates explicit standards and implicit rules, and forms a self-evolving closed loop through semantic matching, soft review warning, rectification review and feedback iteration; its technical effect comes from the continuous role of rule credibility, semantic matching score and access control risk score in the main process, rather than simply introducing a large model concept.

[0041] 2. Regarding rule coverage, this invention mines implicit rules through historical compliance models and rejection models, and uses C(r) to evaluate rule credibility, which can solve the problem of insufficient explicit standard coverage. Regarding citation matching, this invention integrates object type, semantic similarity, contextual relationship, and topic domain matching through M(o,r), which can reduce missed or false judgments caused by character-level hard matching. Regarding review positioning, this invention combines hard and soft review to locate problems at the model level, object level, and rule item level, enabling review opinions to drill down to specific objects and rule evidence.

[0042] 3. Regarding access control review, this invention narrows the scope of review by version differences and uses G(v) to calculate the risk before release, which can avoid repeated full review of irrelevant objects after rectification and make the blocking conditions before release have a clear calculation basis. Regarding rule iteration, this invention updates the credibility of rules through review, rectification, manual review and online quality feedback, so that the rule base can adapt to changes in model design practice. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the overall self-evolutionary review closed loop of the present invention;

[0044] Figure 2 This is a flowchart of the implicit rule mining and standard reference process of this invention;

[0045] Figure 3 This is a flowchart of the closed-loop process for review, rectification, and self-evolution of this invention. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0047] To address the problems of insufficient utilization of implicit rules, coarse standard reference matching, difficulty in drilling down into review issues, and inability to update the rule base in reverse feedback in existing technologies, this invention provides a data standard intelligent reference and automatic review closed-loop method that integrates implicit rule mining of large models.

[0048] Specifically, the process involves using historical compliance and rejection models to uncover implicit rules, filtering executable rules based on rule credibility, calculating matching scores between model objects and rules using semantic representations and contextual relationships, generating tiered evidence of issues using hard and soft reviews, controlling the review and pre-release status using version differences and access control risk scoring, and updating rule confidence levels using review feedback and deployment feedback. Its core processing objects include explicit standard rules, implicit empirical rules, model design objects, object semantic representations, contextual relationships, review issues, rectification records, review conclusions, deployment feedback, and rule confidence levels.

[0049] The overall processing flow consists of five steps: S1 Integrating explicit standards and implicit rules to construct an intelligent rule base; S2 Model design object semantic recognition and intelligent standard referencing; S3 Multi-dimensional automatic review and hierarchical problem localization; S4 Rectification review and pre-release closed-loop control; S5 Review data feedback and rule base self-evolution. Their overall relationship is as follows: Figure 1 As shown.

[0050] The data foundation utilized in this invention comprises two categories: one is explicit standards that can be directly executed, such as data standards, code standards, naming conventions, field definition conventions, and architectural constraints; the other is sample records that reflect historical modeling experience, such as historical compliance models, historical rejection models, manual review opinions, rectification records, and post-deployment quality feedback. Explicit standards are used to provide deterministic constraints, while historical samples are used to discover implicit modeling rules that are not explicitly written in standard documents.

[0051] During the model design process, the system collects object information when entities, fields, relationships, code values, and value domain definitions are created, modified, saved, or submitted. The collected information is not required to be limited to a specific database structure, but it should be able to describe the object's name, object type, subject domain, field type, code set, value domain, and the entity or relationship endpoint it belongs to, as well as its contextual relationships. This allows subsequent rule matching to consider both object attributes and the model environment in which the object exists.

[0052] For rule-side data, each rule should at least be able to represent its content, applicable objects, applicable stage, rule source, credibility status, and start / stop status. For model object-side data, each object should at least be able to be associated with its model version and adjacent context objects. The above information collectively forms the processing basis for implicit rule mining, semantic-level standard citation, automatic review, review access control, and rule base self-evolution in this invention. To achieve the above objectives, this invention adopts the following technical solution:

[0053] S1: Integrating explicit standards and implicit rules to construct an intelligent rule base

[0054] This step is used to integrate the explicit standards maintained manually with the implicit empirical rules accumulated in the historical model to form an executable, evaluable, and iterative intelligent rule base.

[0055] The system first extracts explicit rules from explicit standard documents, code standards, naming conventions, field definition conventions, and architectural constraints. Then, it constructs training samples from historical compliance models, historical rejection models, and manual review comments. Historical compliance models can serve as positive samples, while historical rejection models or repeated rectification models can serve as negative samples. A large model is used to assist in extracting candidate implicit rules. This can be achieved through prompt word extraction, retrieval-enhanced extraction, or lightweight fine-tuning, enabling the model to summarize empirical rules from field naming, entity relationships, value domain binding, code set references, and rectification comments. The output of the large model is not directly stored in the database but is first transformed into rule number, rule content, applicable objects, applicable stages, positive and negative sample hit rates, and example objects.

[0056] For candidate implicit rules, the system calculates the rule's credibility. This can be done using the following formula:

[0057]

[0058] Where C(r) represents the credibility of rule r; Cov pos (r) represents the coverage ratio of rule r in the positive samples, which can usually be calculated by dividing the number of positive samples that match the rule by the total number of positive samples; Cov neg H(r) represents the proportion of rule r in negative samples, used to measure whether this rule also frequently appears in rejected or repeatedly revised models; H(r) represents the degree of confirmation of rule r by manual review or online feedback, usually calculated by combining the number of times manual confirmation is valid, the number of times the review is passed, and the number of times there are no quality anomalies after going online. The weighting coefficients can use the default configuration of α = 0.5, β = 0.3, and γ = 0.2, so that the positive sample coverage rate dominates the credibility evaluation, while taking into account the exclusion of negative samples and manual feedback.

[0059] In the default implementation, it is recommended to set the rule entry threshold to 0.75. When the confidence level of a candidate implicit rule reaches 0.75 or higher, the system marks it as an implicit empirical rule and adds it to the intelligent rule base; when the confidence level is between 0.60 and 0.75, the system can include it in the observation rules or recommended confirmation rules; when the confidence level is below 0.45, the system will not use it for automatic review. For rules that have already been entered into the database, if subsequent feedback causes their confidence level to remain below 0.45, the system can downgrade or disable the rule; if the false alarm rate is higher than 8% for two consecutive statistical periods, the system can appropriately increase the entry threshold or require manual review before reactivation.

[0060] After the above processing, explicit rules and implicit empirical rules are merged into an intelligent rule base, and each rule is recorded with its source, credibility, applicable objects, applicable stage, start / stop status, and version information. This intelligent rule base serves as the rule foundation for subsequent standard references and automatic review of model objects.

[0061] S2: Semantic Recognition of Model Design Objects and Standard Intelligent Reference

[0062] This step is used to identify object semantics when model design objects are generated or changed, and to match them with appropriate standard rules.

[0063] The system collects information such as name, type, subject domain, entity, field description, data type, code set, and value range for objects including entities, fields, relationships, code values, and value range definitions. It then semantically represents the object name, description text, and associated entities, forming an object semantic vector. Similarly, it generates rule semantic vectors for rule content, applicable objects, and rule examples. Here, the semantic vector encodes field names, Chinese descriptions, business terms, entity affiliations, and rule descriptions into a computable numerical vector. The system can generate this vector using an enterprise domain dictionary, large-model text embedding, or a combination of both. Subsequently, it uses cosine similarity to determine whether the object and the rule express similar meanings.

[0064] The standard reference employs a three-layer matching process. The first layer is basic matching, filtering candidate rules based on object type, applicable stage, and rule start / stop status. The second layer is semantic matching, determining semantic closeness based on the similarity between the object's semantic vector and the rule's semantic vector. The third layer is contextual matching, refining the matching results based on contextual information such as the subject domain, associated entities, code sets, and value domains. Contextual matching measures the object's actual business environment within the model structure, such as whether the entity to which the field belongs, adjacent fields, bound code sets, data types, relationship endpoints, and upstream / downstream references are consistent with the rule's applicable conditions. It is typically calculated by weighting four factors: entity relationships, code and value domains, data types, and adjacent objects. Subject domain matching determines whether the object's business domain is consistent with the rule's applicable business domain. For example, a customer domain rule that is completely consistent with the customer entity field can be recorded as 1.00; adjacent subject domains in parent / child subject domains or within the same business line can be recorded as around 0.70; and when there is a significant cross-domain issue, it can drop below 0.30. The matching degree between object o and rule r can be expressed using the following formula:

[0065]

[0066] Where M(o, r) represents the overall matching score between object o and rule r; T(o, r) represents whether the object type and applicable stage match, taking 1 for a complete match, 0 for no match, and 0.5 for a partial match; cos(E o E r ) represents the semantic vector E of the object. o With the regular semantic vector E rThe system uses cosine similarity; Ctx(o, r) represents the contextual relationship matching degree; and D(o, r) represents the matching degree of the subject domain or business domain. The weighting coefficients can use the default configuration of w1 = 0.25, w2 = 0.35, w3 = 0.25, and w4 = 0.15, giving semantic similarity a higher weight while preserving the constraints of object type, context, and subject domain. For rules with strong semantic expression but inconsistent context, the system will not directly bind them based solely on text similarity, thereby reducing the risk of cross-domain mismatch.

[0067] The system determines the citation results based on the matching score. The automatic binding threshold is set to 0.80 by default; rules with a matching score of 0.80 or higher are automatically bound to objects and participate in subsequent review. The recommended range is set to 0.60 to 0.80 by default; rules within this range enter the confirmation queue for modelers or standards administrators to confirm their applicability. The filtering threshold is set to 0.50 by default; rules below 0.50 are not reviewed. If the false positive rate is high, the system can gradually increase the automatic binding threshold to 0.85; if the false negative rate is high, the system can lower the lower limit of the recommended range from 0.60 to 0.55 to allow more candidate rules to enter manual confirmation. This generates object semantic recognition results, standard citation candidates, automatic binding rules, rules to be confirmed, and filtering records, which then proceed to the subsequent automatic review steps. The connection between implicit rule mining, credibility screening, and standard intelligent citation is as follows: Figure 2 As shown.

[0068] S3: Multi-dimensional automated review and hierarchical problem localization

[0069] This step is used to perform a multi-dimensional review of the bound rules and pinpoint issues at the model, object, and rule item levels.

[0070] The system performs hard reviews on rules that have been automatically bound or confirmed. Hard reviews include naming checks, code matching checks, field definition checks, and schema compliance checks. For example, naming checks determine if field names conform to prefix, root, suffix, and length rules; code matching checks determine if fields are bound to valid code sets; field definition checks determine if field types, descriptions, and value ranges are complete; and schema compliance checks determine if entity hierarchy, relationship direction, and cross-domain references meet schema constraints. For explicit blocking rules, if the actual design value does not meet the expected value after the rule is hit, the system directly generates a hard review issue; for non-blocking rules, the system can first generate rectification suggestions and decide whether to proceed to access control calculations based on the severity.

[0071] For designs that formally meet explicit rules but are significantly inconsistent with historical similar objects, the large model assists in outputting soft review warnings. Soft review does not replace hard rule conclusions but is included in the report as a contextual risk indication. The system can calculate a contextual anomaly score between 0 and 1 for soft review warnings. A score of 0.65 generates a normal warning, 0.85 generates a high-risk warning, and scores below 0.50 are not included in the approval report by default. This score is derived by comprehensively considering consistency with historical similar objects, differences in field descriptions and naming meanings, sensitive attribute identification results, and reasons for manual rejection in the past.

[0072] Review issues are organized at the model, object, and rule item levels. The model level presents an overview of the issue; the object level identifies entities, fields, relationships, code values, or value range definitions; and the rule item level records rule numbers, rule content, actual values, expected values, matching scores, and warning information. Issue severity is determined by the rule's default severity level, the scope of object impact, historical rectification costs, and soft review risks. Severity is categorized into four levels: 1 to 4, where 1 indicates a warning, 2 indicates a general issue, 3 indicates a significant issue, and 4 indicates a blocking issue. Issues related to primary key fields, sensitive fields, core code sets, or publication blocking rules can be elevated one level. High-risk soft review warnings typically do not directly block publication before manual confirmation but can be used as a weighting factor in access control risk scoring. The resulting hard review issues, soft review warnings, issue severity levels, and tiered approval report evidence are used for drill-down positioning in approval reports, subsequent rectification reviews, and access control judgments.

[0073] S4: Rectification Review and Pre-Release Closed-Loop Control

[0074] This step is used to generate rectification suggestions based on the review issues, narrow the scope of the review, and determine whether the current version poses a risk of blocking before release.

[0075] The system generates rectification suggestions based on review issues. For naming-related issues, it provides candidate names that conform to the rules; for code set or value range issues, it suggests the code set or value range that should be bound; for field definition issues, it suggests missing field descriptions, types, or constraints; for architectural deviation issues, it suggests the violated hierarchical or relational constraints. The rectification suggestions not only provide directions for modification but also record the rule number, matching score, expected design value, and corresponding rectification reasons, facilitating reuse during the review phase.

[0076] After rectification, the system compares the differences between the model versions before and after rectification, identifying modified objects, related objects, objects affected by rule version changes, and objects with unresolved issues. The review only covers these objects; unaffected objects retain the valid conclusions from the previous round. Before release, the system calculates the access control risk score based on unresolved issues and historical risks. The following formula can be used:

[0077]

[0078] Wherein, G(v) represents the access control risk score of model version v; i represents the review issue in version v; Sev(i) represents the severity of the issue, which can be assigned a value from 1 to 4; Impact(i) represents the scope of the issue's impact, which can typically be between 0.2 and 1.0, with field-level local impact taking a value of 0.3 to 0.6, and core entity, master data, or cross-system interface impact taking a value of 0.8 to 1.0; Hist(i) represents the degree of correlation between this type of issue and historical online quality risks, i.e., the historical risk factor. The historical risk factor can be obtained by combining the historical recurrence rate of similar issues, the online quality anomaly correlation rate, and the frequency of manual rejection. The default value can be set to 1.0, and if similar issues have caused quality anomalies multiple times in the past three months, it can be increased to 1.3 to 1.5; Open(i) indicates whether the issue is not closed, with 1 for not closed and 0 for closed.

[0079] The default blocking threshold for access control risk scoring can be set to 6.0. When G(v) is higher than 6.0, the current model version will not enter the pre-release approval state; when G(v) is between 3.0 and 6.0, the system can request manual review or supplementary approval comments; when G(v) is lower than 3.0 and the approval conclusion is passed, the current model version can be marked as a compliant version. If there is an unclosed blocking issue with a severity of 4, the system can directly block the release even if the overall score does not exceed 6.0. The access control threshold can be dynamically adjusted based on feedback after launch. If a certain type of issue is frequently associated with quality anomalies, the historical risk factor of that type of issue will be increased or the blocking threshold will be tightened to 5.0; if a certain type of rule is confirmed to be invalid for a long time, its impact weight will be reduced or a rule review will be triggered. This forms rectification suggestions, review scope, review conclusion, access control risk score, and pre-release approval or blocking conclusion, which serve as the basis for rule base feedback and version status updates.

[0080] S5: Reviewing Data Feedback and Rule Base Self-Evolution

[0081] This step is used to transform review and deployment feedback into the basis for rule base updates, enabling continuous iteration of the rule base and review capabilities.

[0082] The system feeds back review results, rectification records, manual review comments, review conclusions, and post-launch quality feedback as rule samples. For issues confirmed as valid, positive feedback is added to the corresponding rules; for rules rejected manually or with multiple false alarms, negative feedback is added. The system periodically recalculates the implicit rule credibility C(r). Credibility is increased when a rule consistently covers positive samples, appears infrequently in negative samples, and receives good manual feedback; credibility is decreased when a rule is rejected multiple times or inconsistent with post-launch quality feedback.

[0083] For rules in the rule base, the system performs processing based on the updated credibility level, including retention, demotion, disabling, or re-mining. Rules with high and stable credibility continue to participate in automatic binding; rules with credibility between 0.60 and 0.75 enter the recommendation confirmation stage; rules with credibility below 0.45 are suspended from automatic review. The system can also adjust matching thresholds and access control thresholds based on feedback. For example, if the manual confirmation rate of a rule after 100 consecutive automatic bindings is higher than 95%, its automatic binding threshold can be lowered by 0.02; if the false alarm rate of a rule is higher than 8%, the automatic binding threshold can be raised by 0.03, with a maximum of 0.88. After the rule base is iterated, the new rule version enters the next round of model design object review, thus forming a self-evolving process of review, rectification, re-review, online feedback, and rule updates.

[0084] The data flow of this invention begins with historical models and explicit standards entering S1, forming an intelligent rule base; model objects and the rule base enter S2, forming standard references; reference results enter S3, forming review issues and tiered evidence; issues and version differences enter S4, forming rectification review and access control conclusions; review, rectification, verification, and online feedback enter S5, updating the rule base. Specifically, when a rule's credibility meets the entry criteria, it enters reference matching; when the matching score reaches the automatic binding threshold, it enters review; if the review fails, it enters rectification review; when the access control risk is below the threshold and approval is granted, a compliant version is formed, and online feedback flows back to the rule base. This closed loop not only records the process status but also applies the feedback results to rule credibility, matching threshold, and access control threshold, updating the review capability for the next round. The closed-loop relationship between automatic review, rectification review, access control scoring, and the self-evolution of the rule base is as follows: Figure 3 As shown.

[0085] Based on the above, the present invention includes, but is not limited to, the following embodiments:

[0086] This embodiment applies to an enterprise-level data modeling platform that needs to perform standard referencing, automatic review, rectification review, and rule base iteration on newly added and modified models. The implementation objects include historical compliant models, historical rejected models, currently pending review models, field objects, code sets, and value domain definitions. The implementation scenario can be a power grid business data modeling platform, where more than 20 new or modified models are added or modified daily. Existing standard rules are mainly maintained manually, with a rule coverage rate of approximately 28% and an initial model submission pass rate of approximately 31%. The sample size, rule size, and effect data in the following embodiments are used to illustrate the specific implementation of this invention; actual deployment can be adjusted according to the enterprise model scale and statistical cycle.

[0087] S1: Integrating explicit standards and implicit rules to construct an intelligent rule base

[0088] The data input in this phase includes 487 historical compliance models that have been deployed without quality issues and reused at least 5 times, 112 historical problem models that have been rejected at least 3 times, and 231 manually entered explicit standard rules. Explicit standard rules include field naming rules R01, code set binding rules, field definition integrity rules, and architecture hierarchy constraint rules; historical samples include field names, field descriptions, subject areas, bound code sets, reasons for approval rejections, and rectification records.

[0089] The system first uses 487 historical compliance models as positive samples and 112 historical problem models as negative samples. The large model then extracts candidate implicit rules from field naming, sensitive field binding, statistical indicator suffixes, and value range reference records. For example, the system obtains rule R23 for de-identifying and binding mobile phone numbers, which states that fields identified as having the semantic meaning of mobile phone numbers should be bound to the mobile phone number de-identification rule. The system also obtains rule R45 for encrypting sensitive fields in the customer domain, which states that when sensitive fields in the customer domain are stored encrypted, the field name should reflect an encrypted suffix or be configured with an encrypted identifier.

[0090] The system then calculates the credibility of the candidate rules. Taking R23 as an example, in the positive samples, 443 similar mobile phone number fields meet the de-identification binding requirements, resulting in a positive sample coverage rate of 0.91. In the negative samples, the proportion of anomalies related to this rule is 0.08, therefore the negative sample exclusion item is 0.92; the degree of confirmation through manual review and online feedback is 0.95. After calculation using the default weights of α = 0.5, β = 0.3, and γ = 0.2, the credibility of R23 is approximately 0.921, which is higher than the entry threshold of 0.75, thus it is included in the intelligent rule base. Taking R45 as an example, its positive sample coverage rate is 0.78, the negative sample exclusion item is 0.82, and the degree of manual confirmation is 0.82, resulting in a calculated credibility of approximately 0.800. It also meets the entry conditions, but because it is close to the threshold, the system can first set it as an implicit rule that can be automatically referenced but requires key tracking and feedback.

[0091] After this stage of processing, the system merges 127 implicit empirical rules with 231 explicit rules, forming a smart rule base of 358 rules. Each rule records the rule number, rule content, applicable objects, applicable stage, credibility, start / stop status, and version number. Among them, R01 is the naming rule for explicit fields, R23 is the de-identification and binding rule for mobile phone number fields, and R45 is the encryption suffix rule for sensitive fields in the customer domain. This result serves as the input for the S2 standard smart reference.

[0092] S2: Semantic Recognition of Model Design Objects and Standard Intelligent Reference

[0093] This stage inputs the field object F_CUST_MOBILE from the model currently under review. The entity to which this field belongs is E_CUSTOMER, the subject domain is Customer, the field description is "Customer Contact Number," the data type is VARCHAR, the length is 11, no de-identification rules are currently bound, and the current field name does not contain an encrypted suffix. The system also inputs the intelligent rule base formed in S1, focusing on R01, R23, and R45.

[0094] The system first performs basic matching, identifying the object type of F_CUST_MOBILE as a field, applicable to the model saving and submission stages. Therefore, R01, R23, and R45 are all filtered by object type and stage. The system then encodes the field name F_CUST_MOBILE, the field description "customer contact number," the entity E_CUSTOMER, and the customer domain context into object semantic vectors, and calculates similarity with the rule semantic vectors. For R23, the field semantics are highly similar to rule examples such as mobile phone number, contact number, and mobile phone number, with a semantic similarity of 0.92; the context matching score is calculated to be 0.88 based on the customer entity, sensitive field attributes, and field length of 11; and the topic domain matching score is 1.00. After calculation using w1 = 0.25, w2 = 0.35, w3 = 0.25, and w4 = 0.15, the matching score for R23 is approximately 0.942, higher than the automatic binding threshold of 0.80, therefore it is automatically bound to F_CUST_MOBILE.

[0095] For R01, the field name begins with F_, uses uppercase letters and underscores, and the name length does not exceed the default limit. The basic matching result is 1.00, and the comprehensive matching score is 0.86, which is higher than the automatic binding threshold. The system automatically binds this naming rule. For R45, the field is located in the customer domain and has sensitive field characteristics. However, whether to use encrypted storage in the current model still needs to be confirmed in conjunction with the design configuration. Therefore, its comprehensive matching score is 0.76, falling into the recommended range of 0.60 to 0.80. The system puts R45 into the confirmation queue and prompts the modeler to confirm whether the field belongs to the sensitive field of the customer domain that needs to be encrypted. After the modeler confirms, R45 enters the subsequent review scope. Thus, the output obtained in this stage includes the object semantic recognition result of F_CUST_MOBILE, the automatic binding rules R01 and R23, the rule R45 recommended for review after confirmation, and the filtered records of low matching rules.

[0096] S3: Multi-dimensional automated review and hierarchical problem localization

[0097] This stage inputs the rule binding results, current field design values, and expected rule values ​​for S2. For F_CUST_MOBILE, the current field name is F_CUST_MOBILE, the field description is "Customer Contact Number," the data type is VARCHAR(11), the desensitization configuration is not configured, and the encryption identifier is not configured. The system performs automatic review based on R01, R23, and the confirmed applicable R45.

[0098] The system first performs a hard review. For R01, the field name meets the rules of F_ prefix, uppercase English letters, and underscores, and the length is within the default allowed range, so the naming check passes. For R23, the rule requires that the semantic field of the mobile number be bound to the MASK_MOBILE masking rule or an equivalent masking strategy by default. However, the current field has not been configured with a masking rule, so the system generates object-level issue P20260515-001, with rule item R23, actual value of not configured masking, expected value of binding to MASK_MOBILE, and severity set to 3. For R45, the rule requires that sensitive fields in the customer domain reflect the _ENC suffix or be configured with an encryption identifier when stored in encrypted form. However, the current field has neither an encryption identifier nor a reflected encryption suffix, so the system generates object-level issue P20260515-002, with rule item R45, actual value of not configured encryption suffix and encryption identifier, expected value of adjusting the field name to F_CUST_MOBILE_ENC or configuring an equivalent encryption identifier, and severity set to 3.

[0099] The system also issued a soft review alert. Since over 95% of the customer mobile phone number fields in historically similar fields were bound to anonymization rules, and the field description and length both pointed to the semantics of the mobile phone number, the large model output a context anomaly alert with an anomaly score of 0.88, reaching the high-risk alert threshold of 0.85. This alert will not replace the hard review conclusions of R23 and R45, but it will be included in the approval report to explain the deviation of this field from the consistency model of historical compliance.

[0100] This phase outputs tiered review evidence. The model-level report shows two significant issues and one high-risk warning in the current model; the object-level report pinpoints the field F_CUST_MOBILE; and the rule-level reports record the rule content, matching score, actual value, expected value, severity, and remediation suggestion entry points for R23 and R45, respectively. Based on these results, approvers can drill down from the model overview to the field objects and further examine the specific rule-level evidence.

[0101] S4: Rectification Review and Pre-Release Closed-Loop Control

[0102] This stage inputs issues P20260515-001 and P20260515-002 generated by S3, the current version of the field object F_CUST_MOBILE, and model version difference records. The system generates rectification suggestions for R23, prompting the user to change the anonymization configuration from unconfigured to MASK_MOBILE; for R45, it generates rectification suggestions, prompting the user to change the field name to F_CUST_MOBILE_ENC, or, if enterprise standards allow, retain the original field name but add an equivalent encryption identifier. In this embodiment, the modeler chooses to change the field name to F_CUST_MOBILE_ENC and bind it to MASK_MOBILE, while also adding the encryption algorithm identifier AES256.

[0103] After rectification, the system compared the model versions before and after rectification and only identified changes in the name, masking configuration, encryption identifier, and two interface mappings dependent on the field F_CUST_MOBILE. Based on this, the system narrowed the audit scope to this field, related masking configurations, encryption configurations, and the two affected interface mappings, no longer performing full audits on unmodified entities and irrelevant fields. The audit results showed that the expected value of R23 was met, and the status of P20260515-001 was updated to closed; the expected value of R45 was met, and the status of P20260515-002 was updated to closed.

[0104] Before release, the system calculates the access control risk score. Before rectification, P20260515-001 had a severity of 3, an impact range of 0.80, a historical risk factor of 1.40, an unclosed status of 1, and a risk contribution of 3.36; P20260515-002 had a severity of 3, an impact range of 0.70, a historical risk factor of 1.30, an unclosed status of 1, and a risk contribution of 2.73. The model version's total score was 6.09, higher than the default blocking threshold of 6.0, therefore it could not be released directly. After the rectification review was passed, the unclosed status of both issues became 0, the access control risk score dropped to 0, below the passing reference line of 3.0, and the approval conclusion was passed. Therefore, the current model version is marked as a compliant version.

[0105] The outputs of this phase include remedial recommendations for R23 and R45, a narrowed review scope, issue closure records, access control risk scores, and a pre-release approval conclusion. This approach ensures that the review focuses on the objects to be modified and the affected rules, avoiding redundant full reviews of irrelevant objects.

[0106] S5: Reviewing Data Feedback and Rule Base Self-Evolution

[0107] This phase inputs the rectification records, review conclusions, manual confirmation opinions, and quality feedback after the model went live for S4. After the model went live, the system found no data quality anomalies related to F_CUST_MOBILE_ENC within the 30-day observation period, nor did it experience any interface call failures or anomalies in the de-identified display. The system uses the hit records, rectification, review approval, and no anomaly records for R23 and R45 in this review as positive feedback samples.

[0108] The system updates the credibility of relevant rules accordingly. For R23, since the issue reviewed was confirmed to be effective and the system was stable after rectification, the manual confirmation level increased from 0.95 to 0.96, and the overall credibility increased from approximately 0.921 to approximately 0.923, maintaining automatic binding status. For R45, since the recommendation was proven effective after confirmation, the manual confirmation level increased from 0.82 to 0.86, and the overall credibility increased from approximately 0.800 to approximately 0.808; simultaneously, the system records confirmation samples in sensitive field scenarios within the customer domain, making it easier for fields with the same subject domain and the same sensitive attributes to enter automatic binding or high-priority recommendations in the future.

[0109] After several statistical cycles, the platform's rule coverage increased from approximately 28% before implementation to approximately 91%, the initial model submission pass rate increased from approximately 31% to approximately 77%, the workload for manual rule maintenance decreased by approximately 83%, and the incidence of data quality issues in online models decreased by approximately 79%. These results illustrate that, under the same statistical caliber, this invention, through implicit rule mining, semantic-level referencing, multi-dimensional review, rectification and review access control, and feedback loop, can form a continuously updated review loop.

[0110] This embodiment covers the complete process from S1 to S5, demonstrating the specific implementation methods from historical sample input, candidate rule extraction, credibility filtering, standard reference of field F_CUST_MOBILE, automatic review of R01 / R23 / R45, rectification review, access control scoring to rule base self-evolution.

[0111] Although the present invention has been described above with reference to embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, as long as there is no structural conflict, the features in the disclosed embodiments can be combined with each other in any manner. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A data standard intelligent referencing and automatic review closed-loop method integrating large-scale model implicit rule mining, characterized in that, Includes the following steps: S1, Explicit standards and implicit rules are integrated to build an intelligent rule base: Obtain explicit standards and historical sample data to build an intelligent rule base. The data includes explicit standard documents, historical models, review opinions and online quality feedback. S2, Semantic Recognition and Standard Intelligent Reference of Model Design Objects: Perform semantic recognition and standard intelligent reference on model design objects to generate semantic vectors, automatic binding rules and rules to be confirmed; S3, Multi-dimensional Automatic Review and Layered Issue Localization: Performs multi-dimensional automatic review based on bound rules, generating layered issue evidence including naming, code, field definition, architecture compliance, and contextual alerts; S4, Rectification Review and Pre-Release Closed-Loop Control: Based on the review issues, rectification results and version differences, narrow the scope of the review, calculate the access control risk score, and obtain the pre-release approval or blocking conclusion; S5, Review Data Feedback and Rule Base Self-Evolution: Update rule credibility, matching threshold and access control threshold based on feedback from review, rectification, verification and launch, forming a closed loop of rule base self-evolution.

2. The data standard intelligent referencing and automatic review closed-loop method for integrating large-scale model implicit rule mining as described in claim 1, characterized in that, The specific steps of S1 are as follows: Extract explicit rules from explicit standard documents, code standards, naming conventions, field definition conventions, and architectural constraints; Training samples were constructed from historical compliance models, historical rejection models, and manual review opinions, with historical compliance models serving as positive samples and historical rejection models or repeated rectification models serving as negative samples. The large model is used to extract candidate implicit rules from field naming, entity relationships, value domain binding, code set references and rectification opinions, and the output of the large model is transformed into rule number, rule content, applicable object, applicable stage, positive and negative sample hit status and example object; The credibility of candidate implicit rules is calculated. When the credibility of a candidate implicit rule reaches or exceeds the threshold, it is marked as an implicit empirical rule and added to the intelligent rule base. When the credibility is between the threshold, it is included in the observation rules or recommended confirmation rules. When the credibility is below the threshold, it is not used for automatic review. At the same time, for rules that have already been added to the base, they are downgraded or deactivated. Explicit rules and implicit empirical rules filtered by credibility are merged into an intelligent rule base, and each rule is recorded with information such as rule source, credibility, applicable objects, applicable stage, start / stop status and version.

3. The data standard intelligent referencing and automatic review closed-loop method for integrating large-scale model implicit rule mining as described in claim 1, characterized in that, The specific steps of S2 are as follows: Collect information on the name, type, subject domain, entity, field description, data type, code set, and value range of entities, fields, relationships, code values, and value ranges for the definition objects; then, semantically represent the object name, description text, and associated entities to form an object semantic vector, and form a rule semantic vector for the rule content, applicable objects, and rule examples; A three-layer matching process, comprising basic matching, semantic matching, and contextual matching, is performed to calculate a comprehensive matching score between the model object and the rule. The basic matching filters candidate rules based on object type, applicable stage, and rule start / stop status. The semantic matching determines the semantic closeness based on the similarity between the object's semantic vector and the rule's semantic vector. The contextual matching corrects the matching results based on the subject domain, associated entities, code set, and value domain contextual information. The referencing results are determined based on the matching score: rules that reach the automatic binding threshold are automatically bound to objects and participate in subsequent reviews; rules in the recommended range enter the confirmation queue, where modelers or standards administrators confirm their applicability; rules below the filtering threshold do not participate in the review.

4. The data standard intelligent referencing and automatic review closed-loop method for integrating large-scale model implicit rule mining as described in claim 1, characterized in that, The specific steps of S3 are as follows: Hard review is performed on rules that have been automatically bound or confirmed; for explicit blocking rules, if the actual design value does not meet the expected value after the rule is hit, a hard review issue is generated directly; for non-blocking rules, rectification suggestions are generated first, and it is decided whether to enter the access control calculation based on the severity. For designs that formally meet explicit rules but are clearly inconsistent with similar historical objects, a large model is used to assist in outputting soft review warnings, calculate context anomaly scores, and generate ordinary warnings, high-risk warnings, or defaults to not entering the approval report based on the scores. Issues are organized at the model level, object level, and rule item level. The model level is used to present an overview of the issue; the object level is used to locate entities, fields, relationships, code values, or value range definitions; and the rule item level is used to record rule number, rule content, actual value, expected value, matching score, and warning information. The severity of the issue is determined based on the default severity of the rule, the scope of the object's impact, historical rectification costs, and soft review risks.

5. The data standard intelligent referencing and automatic review closed-loop method for integrating large-scale model implicit rule mining as described in claim 1, characterized in that, The specific steps of S4 are as follows: Generate rectification suggestions for the review issues to provide directions for modification, and record the rule number, matching score, expected design value and reason for rectification. After rectification, compare the differences between the model versions before and after rectification, identify the modified objects, related objects, objects affected by rule version changes, and unclosed problem objects, and the review will only cover the above objects, while the valid conclusions of the previous round will be used for unaffected objects. Before release, an access control risk score is calculated based on unclosed issues and historical risks. If the access control risk score is higher than the threshold, the current model version will not enter the pre-release approval state. If the access control risk score is between the threshold, manual review or supplementary approval comments are required. If the access control risk score is lower than the threshold and the approval conclusion is passed, the current model version is marked as a compliant version. At the same time, if there are unclosed issues with a severity level of blocking, release will be directly blocked even if the comprehensive score does not exceed the blocking threshold.

6. The data standard intelligent referencing and automatic review closed-loop method for integrating large-scale model implicit rule mining as described in claim 1, characterized in that, The specific steps of S6 are as follows: The review results, rectification records, manual review opinions, review conclusions, and post-launch quality feedback will be fed back as rule samples; positive feedback will be added to the corresponding rules for issues that are confirmed to be valid, and negative feedback will be added to the rules that are manually rejected or have been falsely reported multiple times. The credibility of implicit rules is periodically recalculated, and based on the updated credibility, the rules are either retained, downgraded, deactivated, or re-mined. Adjust the matching threshold and access control threshold based on feedback; After the rule base is iterated, the new rule version enters the next round of model design object review, thus forming a self-evolving process of review, rectification, re-review, online feedback and rule update.

Images