A rule-based semantic fusion large model data verification system and method

Abstract

The application discloses a rule semantic fusion-based large model data verification system and method, and relates to the technical field of data processing and artificial intelligence. The system comprises a data receiving and preprocessing module, an intelligent verification engine module, an abnormality detection and risk assessment module, an execution and linkage module, and a feedback iteration optimization module. The system analyzes multi-source heterogeneous data, outputs verification results through rule compliance and semantic rationality double checking, fuses a four-fold detection mechanism in the abnormality detection module and dynamically weights risk scores, outputs a graded report according to risk levels in the execution module, and continuously optimizes the model and rules based on a time decay mechanism in the feedback module to form a closed loop. The application solves the technical problems of insufficient verification accuracy, difficult automatic arbitration of rule conflicts, poor risk assessment scene adaptability, and abnormality missed detection in the prior art, and improves the intelligentization and self-adaptation level of data verification.

Description

Technical Field

[0001] This invention relates to the fields of data processing and artificial intelligence technology, specifically to a large-scale model data verification system and method based on rule-based semantic fusion. Background Technology

[0002] As enterprises and various institutions continue to increase their digitalization, the data generated by business systems is characterized by diverse sources, heterogeneous formats, and massive volumes. Data verification has become a crucial link in ensuring data quality and meeting internal risk control management and external regulatory compliance requirements. However, existing data verification technologies generally have the following shortcomings in practical applications:

[0003] First, existing verification methods mostly adopt verification logic based on fixed rules, which can only identify explicit anomalies that violate plain text rules. They are unable to identify implicit anomaly patterns that are not exhaustively covered by the rules and require semantic understanding to identify, thus limiting the coverage and accuracy of verification results.

[0004] Secondly, existing verification systems are typically designed only for structured data. They lack unified and effective parsing and verification methods for unstructured data such as financial statements, contracts, and invoices in PDF or scanned image formats, as well as semi-structured data such as JSON. This makes it difficult to cope with complex verification scenarios with heterogeneous data sources and diverse formats.

[0005] Third, when the data to be verified involves multiple rule sources such as regulatory policies, industry standards and internal corporate systems, the reasoning conclusions of different rule bases may contradict each other. Existing technology lacks an automated rule conflict arbitration mechanism and usually relies on manual adjudication one by one, which is inefficient and difficult to guarantee consistency.

[0006] Fourth, most existing risk assessment methods use fixed weights to calculate the weights of each risk dimension, failing to dynamically adjust the weights of each dimension according to the characteristics of different business scenarios. This results in a disconnect between the risk assessment results and the risk composition characteristics of actual business scenarios, and insufficient applicability in different scenarios such as financial risk control and government review.

[0007] Fifth, existing anomaly detection technologies typically only make isolated judgments on single data items and fail to combine the relationships between data to detect the propagation of anomalies. When there are causal anomalies among related data items, it is easy for anomalies to be discovered only at the trigger point and related data items to be missed.

[0008] Sixth, when faced with new business scenarios, existing verification systems usually require manual re-sorting and configuration of verification rules. This results in long cold start adaptation cycles and high manual configuration costs due to the lack of historical data for new scenarios. At the same time, the feedback data from manual review is not effectively used for the continuous optimization of verification models and rules. The verification system is unable to adapt to changes in business rules and data distribution, causing the verification accuracy to decline over time.

[0009] In summary, there is an urgent need for an intelligent data verification system and method that can integrate multi-source heterogeneous data processing, rule-based and semantic dual verification, dynamic risk assessment, cascade anomaly detection, and continuous feedback optimization capabilities to solve the technical problems of insufficient verification accuracy, poor scenario adaptability, and high dependence on manual labor in existing technologies. Summary of the Invention

[0010] To address the aforementioned issues, this invention discloses a large-scale model data verification system based on rule-based semantic fusion, comprising a data receiving and preprocessing module, an intelligent verification engine module, an anomaly detection and risk assessment module, an execution and linkage module, and a feedback iterative optimization module.

[0011] The data receiving and preprocessing module is used to receive structured, semi-structured and unstructured data to be verified in multiple forms. It uses OCR and NLP technologies to intelligently parse unstructured documents. After data cleaning and format normalization, it extracts data structure features, data semantic features and data quality features to generate standardized data information and outputs it to the intelligent verification engine module and the anomaly detection and risk assessment module.

[0012] The intelligent verification engine module is equipped with the WxDataVerifyLM data verification big model. Through a three-layer architecture of data semantic understanding layer, rule fusion reasoning layer and verification decision generation layer, it integrates the business rule base and historical verification cases in the data warehouse, performs four-dimensional verification reasoning of completeness, consistency, accuracy and timeliness, and outputs the verification result after a dual verification mechanism of rule compliance verification and semantic rationality verification.

[0013] The anomaly detection and risk assessment module constructs a four-fold detection mechanism: statistical analysis detection, pattern recognition detection, semantic reasoning detection, and rule verification detection. It uses a weighted voting mechanism to integrate the results of the four-fold detection to calculate a comprehensive anomaly score. The WxDataVerifyLM model dynamically allocates weights for four core dimensions—anomaly severity, business impact scope, compliance requirements, and historical risk cases—based on the scenario feature vector to calculate a comprehensive risk score and classify risk levels.

[0014] The execution and linkage module uses the data batch ID and the unique identifier of the data record as the association primary key to accurately link and match the verification results with the abnormal information. After sorting by risk level, it outputs a standardized verification report through a multi-terminal adaptation interface and builds a hierarchical alarm system and a manual review workbench.

[0015] The feedback iterative optimization module performs hierarchical diagnosis of system anomaly information, processes manual review feedback data based on a time decay weighting mechanism, and performs full-link iterative optimization of the WxDataVerifyLM model, verification engine verification rules, and anomaly detection parameters. The optimized parameters and rules are dynamically synchronized to each core module, forming a closed-loop technical link of "verification - detection - execution - feedback - optimization".

[0016] In the intelligent verification engine module, the data semantic understanding layer employs a data semantic mask matrix. The improved Transformer encoder structure uses the following formula for calculating attention weights:

[0017] ;

[0018] in For the data semantic mask matrix, higher attention weights are assigned to key fields, core business indicators and highly sensitive data, and a verification context vector with 768 dimensions is constructed by integrating historical verification behavior data;

[0019] When verification rules are added, changed, or abolished in the business rule base of the data warehouse, they are automatically updated. The mask weight values ​​of the corresponding rule fields are set, and the incremental re-verification process is triggered for historical and existing datasets affected by rule changes. This realizes the automatic synchronization between dynamic rule iteration and semantic representation inside the model, and eliminates the deviation problem of the model using old semantic weights for incorrect verification due to the lag in rule updates.

[0020] In the intelligent verification engine module, the rule fusion reasoning layer adopts a knowledge embedding attention mechanism to vectorize and embed the business rule base in the data warehouse in the form of a knowledge graph. The rule embedding formula is as follows:

[0021] ;

[0022] in For data feature vectors, To verify the context vector, This is the rule weight matrix;

[0023] By introducing a rule priority arbitration matrix, when the reasoning conclusions from multiple rule bases such as regulatory policies, industry standards and internal systems are contradictory, the rule priority arbitration matrix automatically resolves the rule conflict according to the preset priority level, determines the final applicable rule, and solves the pain point of enterprises being unable to automatically arbitrate rule conflicts and having high manual intervention costs in scenarios where multiple regulatory frameworks coexist.

[0024] The verification decision generation layer is based on an improved Transformer decoder and configures a dedicated decoding template matrix for different business scenarios. :

[0025] ;

[0026] in For data-rule fusion vectors, It is a scenario bias term; and a verification self-verification mechanism is introduced to verify the logical consistency of the decision in real time through a consistency verifier during the generation of verification decisions.

[0027] In the dual verification mechanism, semantic reasonableness verification is achieved by calculating a comprehensive score based on the accuracy of the verification. To achieve this, the calculation formula is:

[0028] ;

[0029] in For the first Explicit rule validation pass rate The number of explicit rule validation items. For the first Coverage of latent anomaly detection This represents the number of dimensions for detecting latent anomalies.

[0030] Based on data semantic vectors With abnormal pattern vector Cosine similarity calculation:

[0031] ;

[0032] Similarity hour ,otherwise ;when When semantic rationality is checked, The system automatically adjusts the verification strategy and triggers secondary verification. The threshold of 0.9 is determined through A / B testing of no less than 5,000 real verification cases to balance verification accuracy and processing efficiency. The verification result can only be output to subsequent modules when both rule compliance verification and semantic rationality verification are passed simultaneously.

[0033] In the anomaly detection and risk assessment module, the formula for determining the dual pass condition of the quadruple detection mechanism is as follows:

[0034] ;

[0035] in The threshold is the cosine similarity between the semantic vector of the data to be verified and the scenario vector of the business rule. ; The ratio of the number of data items that meet the rule validation conditions to the total number of rule validation conditions is required, which requires that all rule sub-conditions be met simultaneously; when The test is deemed passed upon completion; the overall anomaly score is considered accurate. The data items are marked as anomaly candidates and output to the risk assessment layer.

[0036] In the anomaly detection and risk assessment module, the formula for calculating the overall risk score is as follows:

[0037] ;

[0038] in The feature value represents the severity of the anomaly. For the feature values ​​of the business impact scope dimension, For compliance requirements dimension feature values, For historical risk case feature values, all are standardized to interval; This is a personalized correction value that the WxDataVerifyLM model dynamically adjusts based on historical risk cases and current data characteristics.

[0039] Dynamic weights to The WxDataVerifyLM model is based on scene feature vectors Automatic calculation:

[0040] ;

[0041] Automatically increase the weight of compliance requirements in financial risk control scenarios. Automatically increase the weight of the business impact scope dimension for government affairs review scenarios. Automatically increase the weight of historical risk cases for scenarios with frequent historical risks. Risk levels are classified as follows: High risk, Medium risk. Low risk.

[0042] The data receiving and preprocessing module is also used to calculate the Data Integrity Index (DCI):

[0043] ;

[0044] Among them, the first The integrity scores for each field are as follows:

[0045] ;

[0046] Field Business Importance Weight Based on the rule dependency graph, the frequency and dependency depth of each field in the verification rules are calculated. Fields with higher frequency and deeper dependency depth are assigned greater weights. The weight calculation formula is as follows:

[0047] ;

[0048] in For the first Semantic vectors of each field, For scene feature vectors, This indicates vector concatenation; It was deemed qualified. Trigger supplementary verification The system directly marks data as high-risk and blocks all subsequent verification processes, sending a data completion request to the data provider to address the engineering pain point of severely incomplete data entering the verification pipeline and causing the spread of misjudgments.

[0049] In the feedback iterative optimization module, a verification misjudgment feedback vector is constructed based on a time decay mechanism. : ;

[0050] in For the first The feature vectors returned by the second manual review This is a time decay factor, which gives more weight to recent review conclusions. The total number of feedback responses within the past 90 days is taken;

[0051] Will With weighting coefficients Verification reasoning logic integrated into the WxDataVerifyLM model: ;

[0052] This addresses the engineering pain points of semantic mismatch between old manual review feedback and current rules due to frequent iterations of business rules, and model verification bias caused by excessively high weighting of long-term feedback; simultaneously... Parameters are used for anomaly detection model correction, enabling continuous optimization through multi-module collaboration.

[0053] In the feedback iterative optimization module, the formula for calculating the priority weight of system exception handling is: ;

[0054] in The abnormal level is (Level 1 = 3, Level 2 = 2, Level 3 = 1). The scope of the impact of the anomaly (affecting full verification = 3, affecting partial verification = 2, affecting a single data item = 1). , This combination of coefficients was determined by the business loss cost model. The emergency optimization process is triggered in time. It is then included in the regular iterative optimization queue.

[0055] The intelligent verification engine module also has the following dynamic adaptation capabilities:

[0056] Maintain a mapping relationship library between verification scenarios and verification strategies. When a new business verification scenario appears, match the same type of scenario from the mapping relationship library through scenario semantic similarity retrieval. Use the verification strategy parameters of the same type of scenario as the initialization parameters of the new scenario through transfer learning to automatically complete the strategy adaptation without manual reconfiguration of verification rules. This overcomes the cold start problem caused by the scarcity of historical data for new scenarios. After accumulating a sufficient number of verification cases in the new scenario, the adaptation strategy is continuously fine-tuned through an online learning mechanism.

[0057] Regularly monitor the mean, variance, and distribution of the data to be verified, and automatically trigger the model parameter adjustment mechanism when a significant drift is detected.

[0058] When business rules or compliance requirements change, the system automatically identifies the affected existing datasets and initiates an incremental re-verification process based on the updated rules, ensuring the timeliness of historical data verification conclusions in rule iteration scenarios.

[0059] The intelligent verification engine module supports verification result output at four granular levels: field level, record level, table level, and batch level. When field-level anomalies are aggregated to the record level, a weighted counting logic is used. When the proportion of abnormal fields in a record exceeds a preset threshold or any high-weight field anomaly exists, the record-level conclusion is upgraded to "fail". When record-level conclusions are aggregated to the table and batch levels, a hierarchical statistical strategy is used. When the proportion of high-risk records exceeds a threshold, a high-risk warning is triggered at the corresponding level. Verification conclusions and anomaly information at each granular level are bidirectionally associated and stored, supporting full-link traceability queries from the batch level to the field level, solving the engineering pain points of inconsistent aggregation of multi-granularity verification conclusions and broken traceability chains across granularities.

[0060] And a data verification method for large-scale models based on rule-based semantic fusion, including the following steps:

[0061] S1. Data Reception and Standardization Processing: Receives structured, semi-structured, and unstructured data in various forms to be verified. Uses OCR and NLP technologies to intelligently parse unstructured documents. After data cleaning, format normalization, and multi-dimensional feature extraction, standardized data information containing data structure features, data semantic features, and data quality features is generated.

[0062] S2, Intelligent Verification Engine Reasoning: Calls the WxDataVerifyLM data verification big model, and through a three-layer architecture of data semantic understanding layer, rule fusion reasoning layer and verification decision generation layer, it integrates the business rule base and historical verification cases in the data warehouse, performs four-dimensional verification reasoning of completeness, consistency, accuracy and timeliness, and outputs the verification result after double verification of rule compliance and semantic rationality.

[0063] S3. Multi-dimensional anomaly detection and risk assessment: The verification results are subjected to four-fold anomaly detection: statistical analysis detection, pattern recognition detection, semantic reasoning detection, and rule verification detection. A weighted voting mechanism is used to integrate the four-fold detection results to calculate a comprehensive anomaly score. The WxDataVerifyLM model dynamically assigns weights based on scene feature vectors, calculates a comprehensive risk score, and classifies the risk into three levels: high, medium, and low.

[0064] S4. Linked output of verification results and abnormal information: Using the data batch ID and the unique identifier of the data record as the association primary key, the verification results and abnormal information are accurately linked and matched. After being sorted by risk level, a standardized verification report is output through a multi-terminal adaptation interface, and a graded alarm notification is triggered simultaneously.

[0065] S5. Real-time monitoring and manual review: Real-time monitoring of the entire system execution status and capture of system anomaly information; triggering manual review process for high-risk verification results, and collecting manual review conclusions in structured data format and storing them in the data warehouse;

[0066] S6. Feedback Iterative Optimization: Classify and diagnose system anomalies and formulate targeted adjustment strategies. Process manual review feedback data based on time decay weighting mechanism, construct verification misjudgment feedback vector, implement full-link iterative optimization of WxDataVerifyLM model, verification engine verification rules and anomaly detection parameters, and dynamically synchronize the optimized parameters and rules to each core module.

[0067] In step S2, the specific implementation of the dual verification mechanism is as follows:

[0068] Rule compliance verification: The business rule base based on the data warehouse performs rule compliance checks on the verification conclusions. When the inference conclusions from multiple rule bases are contradictory, the rule priority arbitration matrix is ​​called to automatically resolve the conflict according to the priority level.

[0069] Semantic rationality verification: Calculate the overall accuracy score using the following formula. ,when When the verification is passed, Automatically adjust verification strategies and trigger secondary verification:

[0070] ;

[0071] Only when the verification result passes both the rule compliance check and the semantic reasonableness check can it be output to step S3 for further processing.

[0072] In step S3, the method for calculating the overall risk score is as follows:

[0073] ;

[0074] Dynamic weights to The WxDataVerifyLM model is based on scene feature vectors pass Automatic calculation to automatically meet the requirements of financial risk control scenarios. Automatically improve for government review scenarios Automatically improve for scenarios with frequent historical risks Personalized correction value The model is dynamically adjusted based on historical risk cases and current data characteristics.

[0075] Step S1 also includes the following data integrity pre-interception steps:

[0076] According to the formula Calculate the overall data integrity index, with field importance weights. Dynamic allocation based on rule-based dependency graph;

[0077] when When this happens, the data batch is immediately marked as high-risk and all subsequent verification processes are blocked; a data completion request is sent to the data provider. When this occurs, a supplementary verification process is triggered to conduct targeted verification of incomplete fields; when If the data is deemed complete, it will proceed to the next verification process.

[0078] In step S6, the specific steps for constructing and fusing the time-decay weighted feedback vector are as follows:

[0079] Collect manual review feedback from the past 90 days, and analyze the feature vector of each feedback. Apply time decay coefficient Construct the feedback vector:

[0080] ;

[0081] With weighting coefficients Integrating WxDataVerifyLM model inference logic: System anomalies are handled according to priority weights. Perform hierarchical processing. Trigger emergency optimization. Included in regular iterations.

[0082] Step S2 also includes a cold start adaptation step for new business verification scenarios:

[0083] Maintain a mapping database of verification scenarios and verification strategies, storing accumulated verification strategy parameters corresponding to different business scenarios. When a new business verification scenario is identified, the system uses semantic similarity retrieval to match the most similar scenario from the mapping database. The verification strategy parameters of the similar scenario are then used as the initialization parameters for the new scenario through transfer learning, automatically completing the strategy adaptation without manual intervention to reconfigure the verification rules. After accumulating a sufficient number of verification cases for the new scenario, the system continuously fine-tunes the adaptation strategy through an online learning mechanism, feeding back the adaptation effect to the mapping database.

[0084] In step S3, semantic reasoning detection also includes a cascading propagation detection step for anomalies in relational data:

[0085] Based on the primary and foreign key relationship graph of the data tables stored in the data warehouse, when a high-risk anomaly is detected in a certain data item, all related data items that reference this data item through primary and foreign key relationships are automatically tracked, and cascading anomaly detection is triggered for the related data items. The cascading detection results are incorporated into the risk comprehensive score calculation of the related data items to ensure the complete propagation coverage of anomalies in relational data, and solve the engineering pain point of missed detection due to the failure of primary key anomalies to spread to foreign key related data.

[0086] In step S5, the real-time monitoring of the system includes continuous monitoring of core indicators such as verification throughput, verification accuracy, anomaly detection recall rate and interface response speed. When any indicator exceeds the preset threshold, the system will trigger an anomaly alarm and synchronize the anomaly information to step S6.

[0087] Step S6 also includes a data distribution drift adaptive adjustment step: periodically perform statistical analysis on the mean, variance and distribution pattern of the data to be verified. When a significant distribution drift is detected, the adaptive adjustment of the WxDataVerifyLM model parameters is automatically triggered. This can maintain the continuous adaptability of the verification engine to the dynamic changes in data distribution without manual intervention, thus solving the engineering pain point of verification accuracy decay caused by the evolution of data distribution after the model is statically deployed.

[0088] And a computer-readable storage medium storing a computer program, characterized in that, when the computer program is executed by a processor, it implements any of the rule-based semantic fusion large model data verification methods described in the present invention.

[0089] Compared with the prior art, the present invention has the following advantages:

[0090] The system unifies the access and parsing of structured, semi-structured, and unstructured data through a data receiving and preprocessing module. Combined with OCR and NLP technologies, it achieves intelligent parsing of unstructured documents, enabling the system to adapt to complex verification scenarios with heterogeneous data sources and diverse formats, thereby improving the system's data compatibility and scenario adaptability.

[0091] The intelligent verification engine module employs a dual verification mechanism, combining rule compliance and semantic rationality checks. This integrates explicit verification based on plaintext rules with implicit anomaly identification based on semantic similarity reasoning, overcoming the limitations of single-rule verification methods in identifying implicit anomaly patterns and improving the accuracy and coverage of verification results. Furthermore, a rule priority arbitration matrix automatically resolves conflicts between contradictory reasoning conclusions from multiple rule bases, avoiding the inefficiency and inconsistent adjudication standards associated with manual rule conflict adjudication. This reduces the cost of manual intervention in scenarios with multiple regulatory frameworks.

[0092] By dynamically calculating the weights of each dimension of risk assessment based on scenario feature vectors using the WxDataVerifyLM model, the calculation of the comprehensive risk score can adapt to the risk composition characteristics of different business scenarios, improving the matching degree between risk assessment results and actual business scenarios. Through a cascading propagation detection mechanism based on the primary and foreign key relationship graph of data tables, the scope of anomaly detection is expanded from a single data item to all data items with related relationships, avoiding engineering problems such as anomalies being processed only at the trigger point and missing related data, thus improving the completeness of anomaly detection.

[0093] By leveraging a mapping library between verification scenarios and verification strategies, along with a transfer learning mechanism, rapid cold-start adaptation for new business verification scenarios is achieved, eliminating the need for manual reconfiguration of verification rules. This shortens the deployment cycle for new scenarios and reduces manual configuration costs. Furthermore, by constructing a verification misjudgment feedback vector based on a time-decay weighted mechanism, manual review feedback data is dynamically integrated into the model's verification inference logic. The verification engine rules and anomaly detection parameters undergo end-to-end iterative optimization, forming a closed-loop technical chain of "verification—detection—execution—feedback—optimization." This allows the system's verification capabilities to continuously and adaptively optimize with changes in business rules and data distribution, preventing the degradation of verification accuracy over time. Attached Figure Description

[0094] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative effort. In the drawings:

[0095] Figure 1 This is an architecture diagram of a large-scale model data verification system based on rule-based semantic fusion.

[0096] Figure 2 This is a flowchart of a large-scale model data verification method based on rule-based semantic fusion. Detailed Implementation

[0097] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0098] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0099] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0100] Example 1

[0101] This example uses the verification of loan business data from a commercial bank as an application scenario. The bank needs to conduct a comprehensive verification of its core loan business data every quarter to meet internal risk control management requirements and regulatory compliance reporting requirements. The data to be verified totals approximately 1.2 million records, sourced from the core banking system (CBS), loan business system, borrower-submitted financial statement PDF files, and external credit data interfaces. Data formats include structured database records, CSV format repayment detail files, JSON format external interface data, and unstructured financial statement PDF documents. The overall data volume is large, the sources are heterogeneous, the formats are diverse, and the compliance requirements are stringent, making it a typical complex and challenging data verification scenario.

[0102] The data receiving and preprocessing module first receives and parses multi-source heterogeneous data to be verified. For structured data from the core banking system and loan business system, the module accesses the data through a batch import interface, receiving approximately 1.2 million loan business records. For the 32,000 PDF financial statements submitted by borrowers, the module's integrated OCR engine performs image recognition on the PDF pages, extracting key financial data items such as total assets, total liabilities, and debt-to-asset ratio from the balance sheet, and operating revenue, operating costs, and net profit from the profit and loss statement. In the NLP processing flow, the Named Entity Recognition (NER) model performs semantic annotation on the extracted financial fields, automatically recognizing and annotating the original field name "amt_rcvbl" as "accounts receivable amount" and "tot_asset" as "total assets," completing the semantic alignment of heterogeneous field names. The data cleaning and standardization process identified and deleted 312 duplicate loan records. Based on business rules, 256 empty values ​​in the "guarantee method" field were inferred and filled (when the loan amount exceeded 5 million yuan and there was no explicit guarantee method record, "guarantee" was inferred and filled based on business rules). The date field was uniformly converted to YYYY-MM-DD format, the amount field was uniformly converted to ten thousand yuan unit, and the enterprise identifier was uniformly converted to the 18-digit unified social credit code standard format.

[0103] In the data integrity pre-interception stage, the module calculates the Data Integrity Index (DCI) for each batch to be processed. This stage serves as a mandatory pre-gate for subsequent verification processes, preventing severely incomplete data from entering the verification pipeline and spreading erroneous conclusions. Taking a batch of 3500 loan records as an example, this batch contains 48 verification fields (…). The WxDataVerifyLM model analyzes the frequency and dependency depth of each field in the verification rules based on a rule dependency graph: frequently referenced fields (such as "loan amount", "lender's unified social credit code", and "loan term") are assigned higher weights. Approximately 0.04 to 0.06), low-frequency fields (such as "loan purpose remarks") have lower weights. (Approximately 0.005), the weights of each field are determined by... Dynamic allocation, satisfying The completeness score for each field was calculated. The "Loan Guarantee Information" field contained 87 blank values ​​and 23 invalid values. After weighted aggregation, this batch , in Within a given timeframe, a supplementary verification process is triggered, involving targeted checks on the guarantee information field and other incomplete fields. This does not disrupt subsequent verification processes. Simultaneously, a data completion request is sent to the loan business system, requiring the missing guarantee information records to be supplemented within 24 hours. After multi-dimensional feature extraction, the module integrates data structure features (types, constraints, and relationships of 48 fields), data semantic features (field business meanings and business logic relationships), and data quality features (completeness, consistency, and accuracy indicators) into standardized data information in JSON format. This information includes a batch ID (BID-20240915-001), data source tags, and feature dimension annotations. This data is then simultaneously output to the intelligent verification engine module and the anomaly detection and risk assessment module, and the original data and standardized data information are stored in the data warehouse.

[0104] The intelligent verification engine module calls the WxDataVerifyLM model to load the verification rule base for this batch of loan business (containing 247 rules including financial accounting rules, internal bank risk control rules, and CBRC compliance requirements), the historical verification case base (12,000 historical cases of similar loan data verification in the past two years), and industry reference data for similar businesses from the data warehouse. At the data semantic understanding layer, the model introduces a data semantic mask matrix. The improved Transformer encoder encodes standardized data information, and the attention weights are calculated as follows: ,in High-risk core fields such as "loan amount", "asset-liability ratio" and "lender's unified social credit code" are given higher attention weights, while low-sensitivity fields such as "loan purpose remarks" are given relatively lower weights. A verification context vector with 768 dimensions is constructed by integrating historical verification behavior data.

[0105] It is worth noting that a week before this verification began, the China Banking Regulatory Commission (CBRC) issued new regulations regarding the early warning threshold for the asset-liability ratio of real estate loans, lowering the threshold from 65% to 60%. The rule base was subsequently updated accordingly. The system automatically detected the rule base update event, triggering... The dynamic update mechanism increases the mask weight of rule fields related to "debt-to-asset ratio" from 0.031 to 0.048, giving the model greater attention to this indicator in subsequent inference. Simultaneously, the system automatically identifies historical datasets affected by this rule change, includes them in the incremental re-verification queue, and re-verifies historical batches according to the new threshold rules. This automatic synchronization of rule changes and the model's internal semantic representation is achieved without any manual intervention, completely eliminating the verification lag bias caused by the model using old semantic weights after rule updates.

[0106] In the rule fusion reasoning layer, the model embeds 247 verification rules into a knowledge graph format using a knowledge embedding attention mechanism. The rule embedding formula is as follows: ,in For the regular weight matrix, To verify the context vector. During the verification inference process of this batch, the model detected a conflict between the inference conclusions of two rules: the bank's internal risk control rules stipulate that "the asset-liability ratio of real estate enterprise loans shall not exceed 65%", while the new regulations of the China Banking Regulatory Commission (CBRC) require that "the asset-liability ratio shall not exceed 60%". Both rules apply to the same verified data item, forming a typical scenario of conflicting rules across multiple regulatory frameworks. The rule priority arbitration matrix automatically determines that the new CBRC regulations have a higher priority based on the preset priority hierarchy (regulatory policies > industry standards > internal systems), using 60% as the final applicable threshold to resolve the conflict and ensure the compliance and uniqueness of the verification conclusion. Without this arbitration mechanism, manually identifying conflicting rules would be time-consuming and prone to erroneous conclusions due to misunderstandings. This module achieves automated and accurate resolution of rule conflicts in scenarios with multiple regulatory frameworks coexisting.

[0107] The verification decision generation layer is based on data-rule fusion feature vectors. A dedicated decoding template matrix is ​​used for loan risk control scenarios. Execution output The model performs real-time verification of the logical consistency of decisions using a self-verification mechanism. It performs four-dimensional reasoning on the 3500 records in this batch: completeness verification (filling of required fields), consistency verification (consistency between loan contract amount and loan disbursement record, consistency between lender information and credit data), accuracy verification (correctness of debt-to-asset ratio calculation, consistency between operating revenue data and industry trends), and timeliness verification (whether the financial statements are valid data from the most recent year). The model then outputs the verification decision for each record.

[0108] In the dual verification of accuracy, the verification data of a loan enterprise (hereinafter referred to as "Enterprise A") in this batch is used as the specific example. A total of 12 explicit rule verifications were performed on Enterprise A. ), of which 11 passed (corresponding to One item failed (the debt-to-asset ratio of 63.7% exceeds the new regulation's 60% threshold, corresponding to...). ); Perform 4D latent anomaly detection ( Based on cosine similarity Calculate each dimension The cosine similarity between the semantic vector of company A's operating revenue and the abnormal pattern vector of "abnormal revenue growth" is 0.83. Therefore The semantic similarity of the guarantee information consistency is 0.71. , The semantic similarity of the related-party transaction's reasonableness is 0.76. The semantic similarity of financial trend rationality is 0.85 ( Substitute the above values ​​into the formula for the comprehensive score of verification accuracy:

[0109] ;

[0110] If the semantic reasonableness check fails, the system automatically adjusts the verification strategy, reloads recent similar real estate industry loan verification cases from the historical case library, adjusts the weight of implicit anomaly detection, and performs a second verification. The cosine similarity of the related transaction reasonableness dimension is recalculated to 0.82. The cosine similarity of the consistency dimension of guarantee information was recalculated to be 0.86. Substituting back into the formula, we get... Through double verification, the final verification conclusion is output: Company A's asset-liability ratio (verified value 63.7%) exceeds the new regulatory threshold of 60%, the verification conclusion is "not passed", the abnormality type is marked as "compliance violation - asset-liability ratio exceeds threshold", the confidence level is 0.94, and the verification basis field includes structured information such as rule name, violation value, and applicable regulatory documents.

[0111] The anomaly detection and risk assessment module performed four-fold anomaly detection on the verification results of this batch. Statistical analysis detection, based on the Z-Score method, performed distribution analysis on the operating revenue data of 3,500 records, identifying that Company A's year-on-year operating revenue growth rate (+187%) exceeded the mean by 3.2 standard deviations, marking it as an explicit statistical anomaly. Pattern recognition detection, based on the WxDataVerifyLM model, learned anomaly patterns from historical cases, matching 23 records in this batch that conformed to the known anomaly pattern of "abnormal fluctuations in operating revenue but no synchronous change in accounts receivable". Semantic reasoning detection, through deep semantic reasoning of the business logic relationships between data, identified a significant logical contradiction in Company A's "operating revenue increased by 187% but operating costs only increased by 12%", and "related party transaction amounts are seriously mismatched with the company's operating scale", marking it as a deep semantic anomaly. Rule verification detection, based on the rule base, identified 17 explicit rule violation records with asset-liability ratios exceeding the threshold. For Company A, all four detection results triggered anomalies. A weighted voting mechanism was used to calculate the overall anomaly score: The model dynamically allocated the weights of the four detections for the loan risk control scenario (statistical analysis 0.20, pattern recognition 0.25, semantic reasoning 0.35, rule validation 0.20). Company A's overall anomaly score was:

[0112] These are marked as abnormal candidates and output to the risk assessment layer.

[0113] Semantic reasoning detection also simultaneously triggers a cascading anomaly detection process for relational data. Based on the primary and foreign key relationship graph of the data tables stored in the data warehouse, after the system detects a high-risk anomaly in Company A (the lender's main table record), it automatically tracks all related data items that reference Company A's primary key through foreign key relationships. This includes 3 related guarantee records, 2 related credit records, and 1 related corporate loan record for Company A, totaling 6 related data items. The system triggers cascading anomaly detection for these 6 related data items, propagating the high-risk anomaly status of Company A to the risk comprehensive score calculation process of each related data item. This prevents missed detections due to the primary key anomaly not spreading to related data, thus solving the engineering pain point that single-point anomaly detection cannot cover the scope of influence of relationships in relational data scenarios.

[0114] At the risk assessment layer, the WxDataVerifyLM model is based on feature vectors from loan risk control scenarios. pass Calculate the dynamic weights of the four dimensions and output the results. (Severity of abnormality) (Scope of business impact) (Compliance requirements are automatically met in financial risk control scenarios) ), (Historical risk cases) meet the requirements The feature values ​​of each dimension are quantized and standardized to... Interval: (The debt-to-asset ratio exceeds the threshold, coupled with a logical contradiction in operating revenue, indicating a high degree of abnormal severity.) (Involves 6 related records, with a moderate impact) (This directly violates the regulatory policies of the China Banking Regulatory Commission and has extremely high compliance requirements.) (A high percentage of similar high-risk real estate loan cases in the past). Personalized adjustment value. (The model is dynamically adjusted upwards based on the current overall upward trend of risk in the real estate industry.)

[0115] Substituting into the risk comprehensive score calculation formula, ;

[0116] Company A's corresponding verification data item was classified as high-risk, triggering an immediate alarm and manual review process. Of the 3,500 records in this batch, 37 were ultimately classified as high-risk (including Company A and its 6 cascading related data items), 82 as medium-risk, and 3,381 as low-risk. The risk distribution results were synchronized to the execution and linkage module.

[0117] The execution and linkage module uses the batch ID (BID-20240915-001) and the unified social credit code of enterprise A as the primary key of association. It performs precise one-to-one linkage matching between the verification conclusion (failure) and abnormal information (asset-liability ratio exceeding the threshold of 63.7%, contradiction between the year-on-year growth of operating revenue and cost logic, and abnormal scale of related-party transactions). After sorting by risk level (high-risk items are placed at the top), a structured PDF verification report is generated. The content includes the verification conclusion, abnormal description, risk comprehensive score (0.855), cascading impact scope (a list of 6 related data items), and suggested handling measures, which are output through the risk control management platform interface. The multi-granularity verification conclusions are aggregated from the field level to the record level (for example, if three fields—asset-liability ratio, operating revenue, and related-party transactions—are abnormal in Enterprise A's record, exceeding the threshold for triggering abnormalities in high-weight fields, the record-level conclusion is upgraded to "fail"). Then, from the record level, they are aggregated to the batch level (if 1.06% of records are high-risk, a high-risk warning is issued at the batch level). Verification conclusions and abnormal information at each granularity level are bidirectionally associated and stored, supporting full-chain traceability queries from the batch level down to the field level. Based on the high-risk level, the tiered alarm system simultaneously pushes three-level alarm notifications via WeChat to the loan business department head, risk control and compliance department, and credit approval personnel. The alarm content includes an anomaly description, a Risk_Score value, and a suggested processing timeframe of 24 hours for manual review. The system's full-process execution status monitoring system synchronously records and sends this three-level anomaly information to the feedback and iteration optimization module.

[0118] A total of 37 high-risk verification conclusions were identified, triggering a manual review process. The person in charge reviewed the verification details for Company A, examining the original financial statement PDF, the results returned from the credit data interface, and data from related companies. The review conclusion confirmed that the debt-to-asset ratio exceeded the threshold, and that the investigation revealed the revenue data was inflated by related parties. The "high-risk, not approved" conclusion was maintained, with the additional review comment: "It is recommended to include it in the key risk monitoring list and suspend new credit approvals." The manual review conclusions were collected in a structured data format (including fields such as reviewer ID, review timestamp, review conclusion, whether the original verification conclusion needs modification, and additional comments) and stored in the data warehouse, providing labeled data support for the feedback and iterative optimization module.

[0119] After receiving the execution data for this batch, the feedback iteration optimization module performs time-decay weighted processing on all manually reviewed feedback from the past 90 days. This module has currently accumulated 60 such instances. Manual review and feedback, most recent feedback ( The weights after applying time decay are: (including the feature vector of "revenue of related parties falsely inflated by company A") The earliest feedback ( Weight decay to The weight of recent feedback is approximately 20.2 times that of the earliest feedback. This time decay mechanism effectively avoids the following engineering pain point: Before the release of the new regulations by the China Banking Regulatory Commission, historical manual review feedback was based on the old threshold of 65%. Without time decay, this batch of old feedback would influence model inference with the same weight as the review feedback under the new rules, causing the model to still favor the verification logic of the old rule semantics during the implementation of the new rules, resulting in a systematic verification bias. After attenuation, the old feedback weights decreased significantly, and recent model optimization has focused on verification feedback based on the new regulations. Based on this, a verification misjudgment feedback vector is constructed. With weighting coefficients Integrating WxDataVerifyLM model inference logic This improves the model's sensitivity to semantic reasoning in identifying latent anomalies such as "inflated revenue from related parties".

[0120] For the Level 3 system anomalies captured in this batch, based on the priority weight formula... ,because The data was included in the regular iterative optimization queue, and the alarm notification link response mechanism was optimized in the next iteration cycle. In-depth mining of the manually reviewed data in this batch revealed two false negatives in the implicit anomaly detection of the "revenue related party inflation" anomaly pattern (i.e., abnormal data was not initially detected). Corresponding misjudgment feature vectors were extracted and included as incremental samples in the next round of small-batch fine-tuning training dataset for the WxDataVerifyLM model, specifically optimizing the model's semantic reasoning ability for logical contradictions in related-party transaction revenue. After optimization, the updated model parameters, verification engine verification rules (the rule expression for the new asset-liability ratio threshold of 0.60), and anomaly detection model parameters (the identification weights for the "revenue related party inflation" pattern) were synchronously updated to each core module and the data warehouse. A comparison of the verification accuracy before and after optimization was stored in the data warehouse, showing that the false negative rate was reduced from 2 / 37 in this batch to the expected 0 / 37. This completed the full-link closed-loop iteration of "verification—detection—execution—feedback—optimization" for this loan data verification cycle.

[0121] Example 2

[0122] This embodiment uses the annual corporate income tax declaration data verification of a provincial tax bureau as an application scenario. Each year, the provincial tax bureau needs to conduct compliance verification on all corporate income tax declaration data within its jurisdiction. The verification objectives cover the authenticity of declared income, the compliance of cost and expense listings, the validity of tax incentive qualifications, and the reasonableness of related-party transaction pricing, in order to ensure the quality of tax collection and administration and meet the data reporting requirements of the State Taxation Administration. This verification involves a total of 492,000 declaration records, covering all existing enterprises within the jurisdiction. Data sources include: the provincial electronic declaration system (structured declaration form data), financial statements and related supporting materials uploaded by enterprises (PDF attachments, totaling 216,000 copies), business registration data interface (providing basic enterprise information, equity structure, and related party information), social security payment data interface (used for cross-verification of the scale and proportion of enterprise R&D personnel), and financial institution transaction summary data (used to verify the flow of funds for large amounts of income and costs). The diversity of data formats and sources places high demands on the verification system's multimodal processing capabilities and multi-dimensional joint verification capabilities.

[0123] The data receiving and preprocessing module completed the full data reception through a combination of batch import and real-time streaming access. The OCR engine performed layout analysis and text recognition on each of the 216,000 PDF financial attachments, extracting core data fields from the balance sheet, profit and loss statement, cash flow statement, and R&D expense special statement. The recognition results were semantically annotated using the named entity recognition model in the NLP process, labeling "R&Dexp" as "research and development expenses" and "tech_dev_fund" as "technology development funding investment". This mapped synonymous field names in different companies' financial reports to a unified standard field system, solving the engineering pain point of extremely fragmented field naming and the inability to directly compare cross-enterprises in multi-entity declaration scenarios. The data cleaning step identified and cleaned 538 duplicate declaration records, corrected 2,341 abnormal coded fields, and performed unit normalization on monetary fields. After format normalization, a clean dataset conforming to the standard field mapping table was output.

[0124] In the calculation of the Data Integrity Index (DCI), taking a batch of 8,000 declaration records (batch ID: BID-TAX-20241115-003, including high-tech enterprise declaration records) as an example, the parameter composition and calculation process of DCI are explained in detail. This batch contains a total of 62 fields to be verified ( The formula for calculating DCI is:

[0125] ;

[0126] The meanings of each parameter are as follows: This refers to the total number of fields covered in this batch of application forms, including four categories: basic application fields (company name, unified social credit code, application period), tax base fields (taxable income, operating revenue, operating costs), preferential fields (R&D expense deduction, high-tech enterprise preferential tax rate application mark, technology transfer income reduction and exemption), and voucher fields (high-tech qualification certificate number, R&D project establishment document number). For the first The business importance weight of each field is determined by... Dynamic allocation, where The first digit obtained by encoding the WxDataVerifyLM model Semantic vectors of each field, This is the scenario feature vector for the current verification scenario (corporate income tax compliance verification). This represents a vector concatenation operation. MLP stands for Multilayer Perceptron. Softmax normalizes the MLP output to ensure that the weights of each field satisfy the specified conditions. Under the rule dependency graph analysis, tax base fields (such as "taxable income") are cited most frequently in the verification rules, with an average of 12 rules directly referencing them, and a corresponding weight of approximately 0.038 to 0.052. Voucher fields are cited less frequently, with an average of 3 rules referencing them, and a corresponding weight of approximately 0.008 to 0.015. For the first The completeness score for each field is calculated using the following formula:

[0127] ;

[0128] in This represents the number of empty values ​​for this field in this batch of 8000 records, i.e., the number of records where the field value is empty or not filled in. This represents the number of invalid values ​​for the field, i.e., the number of records whose field values ​​do not conform to the preset format or value range (e.g., the high-tech qualification certificate number does not conform to the 18-digit national standard encoding format). This is the total number of records for this field; in this batch, it is 8000. Taking the "R&D Expense Additional Deduction" field as an example, this field has 341 blank values ​​(some companies that did not declare R&D expenses did not fill in this field) and 89 invalid values ​​(the amount filled in exceeds the upper limit of the total R&D expenditure for the year). Taking the "High-tech Enterprise Qualification Certificate Number" field as an example, there are 217 blank values ​​(the system identifies that such enterprises should have corresponding certificate numbers based on the high-tech preferential mark, but they are not actually filled in) and 134 invalid values. After weighted aggregation, this batch , in During the specified period, the supplementary verification process was triggered. The system initiated targeted verification of fields with low completeness, such as the R&D expense deduction amount and the high-tech qualification certificate number. At the same time, a data completion notice was sent to the application system, requiring the relevant enterprises to complete the missing application items within the specified time limit. This did not block the subsequent main verification process.

[0129] The multi-dimensional feature extraction process extracts three core features from standardized data: data structure features (types, constraints, and cross-table relationships of 62 fields, including one-to-many relationships between the main declaration table and the R&D expense details appendix, and foreign key relationships between declaration data and business registration data), data semantic features (business meanings of fields and logical relationships across fields, such as the logical constraint semantics of "taxable income = operating revenue − operating costs − period expenses − other items deductible before tax"), and data quality features (quantitative indicators of completeness, consistency, and accuracy). These features are integrated into standardized data information in JSON format, synchronously output to the intelligent verification engine module and the anomaly detection and risk assessment module, and stored in the data warehouse.

[0130] The intelligent verification engine module calls the WxDataVerifyLM model, loading the verification rule library for this scenario, the historical verification case library, and the declaration data of enterprises in the same province and industry from the data warehouse as a reference benchmark. The data semantic understanding layer uses a data semantic mask matrix. Differentiated attention weights are assigned to key fields in the verification scenario. Fields with the most direct impact on tax collection and administration, such as "taxable income", "R&D expense deduction", and "related transaction amount", are given higher mask weights, while low-related fields such as "company contact number" have relatively lower weights. A verification context vector with a dimension of 768 is constructed by integrating historical verification behavior data (verification conclusions, manual review results, and tax audit cases) from the past three years, providing a semantic basis for subsequent reasoning.

[0131] In this embodiment, the "New Carbon Deduction Regulations" (hereinafter referred to as the "Carbon Deduction New Regulations") have been issued, which have added a "Carbon Deduction Tax Deduction Amount" declaration item to the corporate income tax return. This declaration item is a completely new business type, and there are no verification records of similar declaration data in the historical verification case library. The intelligent verification engine module starts the cold start adaptation process for the new scenario as follows: In the verification scenario and verification strategy mapping relationship library, the "Carbon Deduction Tax Deduction Verification" scenario is compared with existing scenarios one by one through scenario semantic similarity calculation. The scenario with the highest semantic similarity is "R&D Expense Additional Deduction Verification" (semantic similarity score 0.83). The two are highly similar in verification logic structure, both involving three core logics: preferential qualification determination, amount compliance verification, and supporting document integrity verification. The model uses transfer learning to apply the verification strategy parameters from the "R&D expense deduction verification" scenario as the initialization parameters for the "carbon emission reduction tax deduction verification" scenario. It then adjusts the initialization parameters based on the specific provisions of the new carbon deduction regulations (e.g., changing the applicable object of the qualification verification rules from "certified high-tech enterprises" to "carbon emission reduction certified enterprises," and changing the upper limit rule from "not exceeding 120% of actual R&D expenditure" to "not exceeding the upper limit of the deduction amount converted from certified emission reductions"). This automatically adapts the verification strategy to the new scenario without requiring manual intervention to reconfigure the verification rules. This cold-start adaptation solution was successfully applied to the first batch of 1247 verification tasks involving carbon deduction declaration records within the jurisdiction. Subsequently, as verification cases accumulate, the system fine-tunes the adaptation strategy through an online learning mechanism, updating the optimized parameters to the mapping database for future retrieval and utilization in similar new scenarios.

[0132] The rule fusion reasoning layer detected a rule conflict during the verification process for this batch: The rule states that "the upper limit of the additional deduction ratio for R&D expenses of high-tech enterprises is 100% of the actual R&D expenditure," while the local tax incentive policy issued by the province stipulates that "the upper limit of the additional deduction ratio for R&D expenses of key high-tech enterprises within the jurisdiction can be increased to 120%." The two rules contradict each other in calculating the additional deduction amount for R&D expenses declared by high-tech enterprises within the same jurisdiction. Based on the preset priority hierarchy, the rule priority arbitration matrix should have determined that the 100% upper limit takes precedence; however, the arbitration matrix also embeds a "legal exception" rule: when a local preferential policy is explicitly approved and filed by the State Taxation Administration, the local policy has priority over the national policy within the jurisdiction. The system retrieves approval and filing records from the rule base and confirms that the province's local policy has been filed with the State Taxation Administration. Based on this, the 120% cap is identified as the final applicable rule for key high-tech enterprises in the jurisdiction, resolving rule conflicts and ensuring that the verification conclusion is consistent with local regulatory practices. At the same time, the arbitration decision logic and basis are stored in the conflict resolution case records in the rule base for direct reference in subsequent similar conflicts, avoiding the time loss of manual repeated judgment.

[0133] The verification decision generation layer is based on data-rule fusion feature vectors. Calling the dedicated decoding template matrix for government affairs review scenarios ,implement The completeness, consistency, accuracy, and timeliness of the application records in this batch are verified and reasoned across four dimensions, with a self-verification mechanism verifying the logical consistency of the decisions in real time. Taking a specific applicant company (hereinafter referred to as "Company B") in this batch as an example, the parameter composition of the dual verification process for accuracy is explained in detail. Company B applied for a deduction of 5.8 million yuan for R&D expenses, but the actual R&D expenditure listed in the financial statement appendix was only 2.1 million yuan, and social security data showed that Company B had only 12 R&D personnel, which is seriously mismatched with the scale of the applied R&D expenses.

[0134] The compliance verification of Company B involved 16 explicit rule verification items. , This represents the total number of explicit rule checks performed on Enterprise B's data item according to the verification strategy, including rules on R&D expense caps, validity of high-tech enterprise qualification certificates, and matching during the application period, etc., of which 15 passed. ), 1 item failed (R&D expense deduction exceeded the 120% limit), ), Simultaneously perform 4D latent anomaly detection ( , (Representing the total number of dimensions for semantic reasoning detection performed by the WxDataVerifyLM model on the enterprise B's declaration data), each dimension Based on the semantic vector of the data to be verified With the corresponding anomaly pattern vector cosine similarity Determine similarity hour ,otherwise In terms of the compatibility between R&D personnel size and R&D expenditure, Company B's average R&D expenditure per employee is approximately 483,000 yuan, far exceeding the industry average of 92,000 yuan, with a corresponding cosine similarity of 0.91. In terms of the rationality of the fund flow reasoning, there is a logical discrepancy between the RMB 1.48 million outflow of R&D-related funds in the cash flow statement and the RMB 2.1 million declared R&D expenditure, with a cosine similarity of 0.88. In terms of the rationality reasoning of R&D output, Company B has no patent application records in the past three years but has continuously declared large amounts of R&D expenses. The cosine similarity is 0.82. In terms of the reasonableness of related-party transactions, there are records of R&D service procurement between Company B and its holding subsidiary. Semantic reasoning has not yet fully linked to the abnormal pattern, and the cosine similarity is 0.74. ), , Substitute each parameter into the comprehensive accuracy score formula:

[0135] ;

[0136] The meanings of the coefficients in the formula are explained as follows: In the numerator, 0.7 represents the contribution weight of the explicit rule verification item to the overall score, reflecting the dominant role of deterministic rule verification in the accuracy assessment; 0.3 represents the contribution weight of the latent anomaly detection coverage rate to the overall score, reflecting the necessary supplementary role of semantic reasoning to explicit rules; the sum of the two is 1.0 to maintain consistency in molecular dimensions; the normalization coefficient in the denominator is taken as... Instead This is because of the number of dimensions for latent anomaly detection. The confidence level of the test results is lower than the number of explicit rule validation items. (Semantic reasoning has a probability of misjudgment, while explicit rule violations are deterministic judgments.) A reduction factor of 0.5 is applied to prevent the number of detection dimensions from excessively increasing the overall score, thus achieving a reasonable balance between the contributions of the two types of verification. The overall score threshold of 0.9 was determined through A / B testing of no less than 5,000 real verification cases. When the threshold is higher than 0.9, the verification accuracy can be stably maintained above 97%. When the threshold is lower than 0.9, the misjudgment rate increases significantly. This threshold is the optimal critical value for balancing verification accuracy and processing efficiency.

[0137] Calculation results If the semantic reasonableness check fails, the system automatically triggers the verification strategy adjustment mechanism. It retrieves 81 verification cases related to "false reporting of R&D expense deduction" from the historical case database within the past three years and adds them to the current verification reasoning context. The weights of the implicit anomaly detection dimension are readjusted, and the cosine similarity of the related-party transaction reasonableness dimension is recalculated to 0.86. Substituting into the formula for a second verification yields... The value remained below 0.9, triggering another strategy adjustment. The model further introduced cross-source consistency verification rules for social security data and bank transaction data, resulting in a third verification. Through dual verification, the final verification conclusion is: Enterprise B's R&D expense deduction amount (declared value of 5.8 million yuan) is suspected of being falsely reported. The actual deductible limit is 2.52 million yuan, and the over-declaration is 3.28 million yuan. The verification conclusion is "not passed" with a confidence level of 0.96. The verification basis fields include structured information such as rule name, numerical deviation amount and supporting abnormal evidence (mismatch in R&D personnel scale, deviation in fund flow, and missing R&D results).

[0138] The anomaly detection and risk assessment module performed quadruple anomaly detection on the verification results of this batch. Statistical analysis, based on the IQR interquartile range method, analyzed the distribution of the R&D expense deduction rate (deduction amount / actual R&D expenditure) for all companies in the batch, identifying Company B with a deduction rate of 276% (580 / 210), far exceeding the average for the entire batch. The upper boundary value formed is marked as an explicit statistical outlier; the pattern recognition detection is based on the "overstating of R&D expenses" anomaly pattern learned from the historical case library by the WxDataVerifyLM model. The matching degree between the data features of Enterprise B and this pattern reaches 0.91, triggering the pattern recognition anomaly labeling; the semantic reasoning detection, based on the identification of the contradiction between the scale of R&D personnel and the logical of expenses, further infers the implicit related-party transaction contradiction of "after Enterprise B pays R&D service fees to its holding subsidiary, the subsidiary repeatedly lists it as its own R&D expenses in the declaration", and marks it as a deep semantic anomaly; the rule verification detection identifies the explicit rule violation of the R&D expense deduction amount exceeding the upper limit.

[0139] The dual pass conditions of the quadruple test results further supplement and verify the findings: taking Enterprise B's data as an example. ,in The semantic vector of the data submitted by Company B (encoding features such as abnormal R&D expense submissions, mismatch in personnel size, and deviation in fund flow). This is a scenario vector for the business rules of "compliance verification of corporate income tax R&D expenses" (encoding the semantic representation of R&D expense recognition rules, personnel requirement rules, and voucher requirement rules). ; To determine the ratio of the number of data items meeting the rule verification conditions to the total number of rule verification conditions, Company B's verification rule verification conditions cover four sub-conditions: "Additional deduction amount does not exceed the upper limit," "Reasonable personnel size," "Completeness of vouchers," and "Reasonableness of fund flow." Company B only meets the "Completeness of vouchers" and "Completeness of voucher materials" sub-conditions within "Completeness of vouchers" and "Additional deduction amount does not exceed the upper limit." Substituting into the double pass condition formula:

[0140] ;

[0141] The two 0.5 weight coefficients in the formula reflect the equal contribution of semantic similarity and rule matching degree to the comprehensive detection judgment, embodying the design principle of giving equal importance to semantic reasoning and rule determinism testing; The test did not meet the passing criteria, and the overall abnormal score of Company B after weighted voting for the four tests was 0.87. The abnormal candidate threshold is marked as an abnormal candidate and output to the risk assessment layer.

[0142] Semantic reasoning detection synchronously triggers cascading anomaly detection in relational data: Based on the primary and foreign key relationship graph stored in the data warehouse, after detecting a high-risk anomaly in Company B (the parent company's main table record), the system automatically tracks the declaration records of two holding subsidiaries (Company B-1 and Company B-2) that reference Company B's primary key through equity relationship foreign keys. Cascading anomaly detection is triggered for each subsidiary, focusing on verifying whether there are instances of duplicate listings of R&D service expenses purchased from Company B. The detection results show that Company B-1 included RMB 1.37 million of R&D service purchase expenses from Company B in its own R&D expense declaration, suspected of double deduction; no double listing was found in Company B-2. The cascading anomaly detection result for Company B-1 is included in its overall risk score calculation, extending the impact of Company B's relational anomalies to the subsidiary level, preventing the risk of missed detection due to primary key anomalies not spreading to related data.

[0143] At the risk assessment layer, the WxDataVerifyLM model is based on feature vectors from government affairs review scenarios. ,pass The model automatically calculates dynamic weights for four core dimensions. The MLP consists of two fully connected layers (with ReLU activation function), and Softmax ensures the sum of the output weights is 1. For government approval scenarios, the model automatically outputs... , , , ,satisfy ;in The weight of the business impact scope dimension is automatically increased in government review scenarios, compared to financial risk control scenarios. () is significantly higher.

[0144] The quantization process for each dimension of feature values ​​is as follows: In terms of the severity of the anomalies, the over-reporting amount of 3.28 million yuan is relatively large, and the mismatch between personnel size and actual size exceeds 300% (the per capita cost exceeds the industry average by 5.2 times). The sub-features are quantified to 0.82 after being weighted according to the preset weights. In terms of the scope of business impact, in addition to the main entity B, it also involves the abnormal association of subsidiary B-1, and the relevant declaration data will be summarized and reported to the State Taxation Administration, affecting the quality of national tax data, and the quantitative value is too high. In terms of compliance requirements, it directly violates the tax law's rules for recognizing R&D expenses, and there are signs of subjective violations such as suspected double deduction by related parties, resulting in a high compliance quantification value. Corresponding to the historical risk case dimension, the average risk level of similar historical cases of overstated R&D expenses is medium to high risk, and the proportion of such overstated expenses ultimately verified by audits in historical cases is 73%. Personalized adjustment value. , The range of values ​​is The WxDataVerifyLM model dynamically calculates the final value after comprehensively considering the current tax audit focus and the historical credit record of Company B (no violations in the past two years, which slightly lowers the risk). Substitute into the risk comprehensive score calculation formula:

[0145] ;

[0146] Enterprise B's reported anomalies were classified as high-risk, triggering an immediate alert and manual review process. Of the 8,000 reports submitted in this batch, 51 were ultimately classified as high-risk (including cascading anomalies from Enterprise B and its subsidiary B-1), 203 as medium-risk, and 7,746 as low-risk. The risk distribution results were synchronized to the execution and linkage module.

[0147] The execution and linkage module uses the batch ID (BID-TAX-20241115-003) and the unified social credit code of enterprise B as the primary key of association. It precisely links and matches the verification conclusion (failure), anomaly description (excessive R&D expense deduction of 3.28 million yuan, mismatch in R&D personnel size, suspected related-party duplicate deduction of 1.37 million yuan), comprehensive risk score (0.827), and cascading impact scope (list of suspected duplicate deductions by subsidiary B-1). This generates a structured PDF verification report containing verification conclusions at three granular levels: field-level, declaration item-level, and enterprise-level, which is then output to the tax risk control platform. When field-level anomalies (R&D expense deduction field) are aggregated to the declaration item level, because this field is a high-weight field (…),… Furthermore, if there are definitive rule violations, the verification conclusion at the application item level is directly upgraded to "fail." When the application item level conclusion is aggregated to the enterprise level, a weighted counting logic is used. Although the proportion of failed applications is relatively low among all applications, it triggers the high-weight field anomaly priority rule, and the enterprise-level verification conclusion is upgraded to "fail." When the enterprise-level verification conclusion is aggregated to the batch level, the proportion of high-risk enterprises is 0.64% (51 / 8000), which does not exceed the batch-level high-risk warning threshold (3%). The batch level issues a medium-risk warning. Verification conclusions and anomaly information at each granularity level are bidirectionally associated and stored, supporting full-link traceability queries from the batch level to the field level. The hierarchical alarm system pushes three-level alarm notifications to the supervising tax officer, risk control and audit department, and data governance manager through internal messages and emails on the tax risk control platform. The alarm content includes structured information such as Risk_Score, excess declaration amount, and suggested handling measures. The system's full-process execution status monitoring showed that the total verification time for this batch of 492,000 records was 37.6 minutes, with an average interface response time of 287ms. The monitoring system identified two level-two anomalies, and the anomaly information was simultaneously sent to the feedback iteration and optimization module. After the manual review workbench reviewed the 51 high-risk verification conclusions, the tax experts maintained the high-risk failure conclusion for Company B, confirming that the over-declaration was true, and recommended initiating tax audit procedures for Company B and its subsidiary B-1 to recover taxes and late payment fees. The review conclusions were stored in the data warehouse in a structured data format.

[0148] The feedback iteration optimization module processes the feedback data generated in this batch. For the two secondary system anomalies (business registration data interface call timeout), they are processed according to a priority weight formula. The processing priority is calculated, and the meanings of each parameter are as follows: The anomaly level coefficient reflects the weight of the anomaly level on the urgency of handling. This value is determined by the business loss cost model (if a severe anomaly is not handled in time, the cumulative loss it causes is far greater than the loss caused by the expansion of the impact range, so the weight of the level coefficient should be higher than the impact range coefficient). For the influence range coefficient, Ensure that the contributions of both are completely consistent; This is the level quantification value corresponding to this Level 2 anomaly (Level 1 = 3 indicates that it seriously affects the execution of full verification, Level 2 = 2 indicates that it affects some verification functions, and Level 3 = 1 indicates that it is a minor anomaly). This is the quantified value for the impact range corresponding to the approximately 1200 records affected by this interface timeout. Substitute into the formula for calculation: ,because The timeout retry mechanism for the business registration data interface will be optimized in the next iteration cycle to prevent the instability of the interface from causing data loading delays that will have a lasting impact on the timeliness of verification.

[0149] For manual review feedback, a time-decay weighted verification misjudgment feedback vector is constructed from all manual review feedback accumulated within the past 90 days. The meanings of each parameter are as follows: This is the time decay coefficient. The logic behind determining this value is as follows: Although tax verification rules are adjusted annually with policy changes, the validity period of the core rules is usually between 1 and 3 years. Both recent and longer-term review feedback have certain reference value. The decay rate is ensured to be within a feedback window of approximately 90 days (actually approximately...). (times), most recent feedback weight () ) is approximately the earliest feedback weight ( ) is 9.6 times that of the previous tax policy, which retains historical experience while making recent feedback dominate the optimization direction, effectively avoiding the interference of historical feedback generated under the old tax policy with excessive weight to the inference semantics of the model under the new policy rules; For the first The feature vector of the second manual review feedback encodes the data characteristics of the review object (such as the semantic features of "excessive declaration of R&D expenses"), whether the original verification conclusion was overturned (all 51 high-risk reviews in this batch maintained the original conclusion, and no overturning record was found) and the reviewer's handling suggestions (such as "initiating tax audit" and other decision semantics). Use the total number of feedback responses within the last 90 days to ensure the timeliness of the feedback data and avoid diluting recent optimization directions with overly distant feedback information; Subscript The order of arrangement is from most recent to oldest time, that is... In response to the most recent feedback, The time decay coefficient corresponds to the earliest feedback within the 90-day window. Follow Increase and decrease accordingly to ensure that the impact of recent review conclusions on model optimization has a higher weight. With weighting coefficients Integrating WxDataVerifyLM model inference logic: ,in The value of ensures that the influence of human feedback on model inference is moderately controllable, so that the model fully absorbs human review experience while avoiding overfitting to a small number of cases, which would reduce the model's ability to generalize detection of new types of anomalies. The value has been verified as the optimal coefficient for the current scenario through multiple rounds of ablation experiments.

[0150] In-depth analysis of the manually reviewed data in this batch revealed that the verification engine's error patterns were insufficient in its initial identification of the complex anomaly of "overstated R&D expenses coupled with repeated deductions by related parties." This was found to be due to inadequate semantic reasoning regarding related-party transactions. (To identify the judgment bias), the misjudgment feature vector is extracted as an incremental sample and incorporated into the next round of mini-batch fine-tuning training dataset for the WxDataVerifyLM model. This optimizes the model's deep semantic reasoning capabilities for scenarios involving the combined superposition of related-party transaction anomalies and entity declaration anomalies. The optimized model parameters, related-party transaction semantic reasoning rule weights, and online learning update parameters for the "carbon emission reduction tax deduction verification" cold start adaptation strategy are synchronized to each core module and data warehouse, completing the full-link closed-loop optimization of this corporate income tax declaration data verification cycle.

[0151] Example 3

[0152] This embodiment uses a multi-source data verification platform within an enterprise as an application scenario. This enterprise encompasses multiple business lines, including financial reimbursement, human resources payroll, supply chain procurement, and contract management. Each business line has its own independent business system, with different data formats, structures, and update frequencies. The enterprise needs to conduct unified verification of the core data generated by each business line regularly to support internal auditing, business decision-making, and compliance management needs. The data to be verified totals approximately 387,000 records, sourced from the financial system (structured accounting records), the human resources system (structured payroll and attendance records), the procurement system (semi-structured purchase orders and approval flow JSON data), supplier-submitted contracts and invoices (PDF files totaling 43,000), and unstructured scanned image files such as bank receipts. The overall data exhibits characteristics of dispersed business lines, heterogeneous data structures, and the continuous emergence of new business scenarios.

[0153] The data reception and standardization process is executed by the data reception and preprocessing module in step S1. This module uniformly accesses structured accounting records from the financial system, structured payroll records from the human resources system, JSON-formatted approval flow data from the procurement system, and scanned images of contracts, invoices (PDF files), and bank receipts submitted by suppliers. For PDF contracts and invoices, the module uses OCR technology to recognize the text content in the images and combines this with NLP technology to perform semantic analysis of the recognition results, automatically extracting key fields such as contract amount, contract period, invoice tax number, and invoice date. For structured and semi-structured data, the module performs field alignment, null value marking, and format normalization. After these processes, the module extracts the data structure features, semantic features, and quality features of various data types, generating standardized data information in a unified format. This information is then output to the intelligent verification engine module and the anomaly detection and risk assessment module, completing the processing flow of step S1.

[0154] In this step, the module synchronously calculates the comprehensive integrity index of data from each business line: ;

[0155] The integrity scores for each field are as follows: ;

[0156] In and These are the actual engineering quantity values ​​obtained by the data receiving and preprocessing module during the field parsing and format normalization process, i.e., the [number]th [item]. The number of records with null values ​​in each field in this batch of data, and the number of records judged as invalid values ​​(such as field type mismatch or exceeding the preset value range) after format validation, are both derived from a field-by-field statistical analysis of real data, and are not subjectively set abstract values. Essentially, it's an objective measure of the actual completeness of a field within this batch of data. Field weight. Instead of manually preset fixed coefficients, the data receiving and preprocessing module calculates the frequency and dependency depth of this field across all current business rules of the enterprise based on the rule dependency graph. The calculation is dynamic, meaning that the more frequently a field is referenced and the more profound its impact in the verification rule system, the higher its weight. This weight is automatically recalculated as business rules are added, changed, or abolished, without the need for manual intervention. This calculation process relies on the engineered structure of the rule dependency graph stored in the data warehouse and the real-time reasoning of the neural network model, rather than a simple mathematical weighted summation of fixed weights, thus distinguishing it from purely artificial intelligence rules.

[0157] Taking the procurement system data as an example, the "Supplier Taxpayer Identification Number" field is frequently referenced and has a deep dependency in historical verification rules, resulting in a high weight value obtained through the aforementioned weight calculation formula. Calculation shows that the DCI of this batch of procurement data is 0.91, falling within the range of [0.85, 0.95). The system automatically triggers a supplementary verification process, initiating a targeted supplementary verification for the 312 procurement order records missing this field, rather than directly blocking the entire data line. This avoids the engineering problem of a large amount of compliant data being intercepted due to the missing partial field.

[0158] Regarding the DCI thresholds, these two thresholds are not arbitrarily set, but determined based on statistical patterns in the company's historical verification data: When the DCI is below 0.85, historical data indicates that the number of key fields with missing or invalid values ​​in this batch of data has significantly affected the reliable input for subsequent verification inference. Continuing verification will produce a large number of misjudgments, so it is directly blocked and data completion is required; when the DCI is in the range of [0.85, 0.95), historical data indicates that the missing values ​​are mainly concentrated in a few specific fields. Targeted verification of these fields can restore data availability without blocking the overall process; when the DCI reaches above 0.95, historical statistics indicate that data integrity no longer has a substantial impact on the reliability of the verification conclusion. The technical problem solved by this threshold setting is to avoid the verification system adopting a "one-size-fits-all" approach to all incomplete data, and to minimize the scope of verification process interruptions caused by data integrity issues while ensuring the reliability of verification conclusions.

[0159] After standardized data information is generated, step S2 is executed by the intelligent verification engine module. This module, equipped with the WxDataVerifyLM data verification model, sequentially encodes the semantics of data from different business lines such as finance, human resources, and procurement through a data semantic understanding layer, combining this with the semantic mask matrix of each business line. The system assigns higher attention weight to core business indicators such as contract amount and salary base; subsequently, through a rule fusion and reasoning layer, multiple business rule bases, including the company's internal financial system, human resource compensation management methods, and procurement management regulations, are vectorized and embedded in the form of a knowledge graph, resulting in:

[0160] ;

[0161] The rule vector and data vector are fused; finally, the decision generation layer verifies the data and uses a dedicated decoding template matrix configured for different business lines. Generate verification decisions.

[0162] and These are the current data feature vector output by the data semantic understanding layer and the verification context vector generated after fusing historical verification behavior. Both are intermediate calculation results actually generated by this module in the preceding processing stage, rather than symbolic variables assumed out of thin air. This is a rule weight matrix, whose values ​​are determined through learning from historical verification cases during model training. It is retrained or incrementally updated as the business rule base changes. Therefore, this fusion formula describes a learnable computational process that dynamically evolves with business rules, rather than a fixed mathematical mapping. The specific technical problem this formula solves is that enterprise business rules typically exist in the form of natural language clauses, which cannot be directly computed with data feature vectors processed by neural networks. This formula achieves a technical fusion between unstructured business rules and structured data features through knowledge graph vectorization and matrix weighting operations. This allows verification decisions to have both data-driven and rule-driven basis, resolving the technical contradictions of purely data-driven models lacking interpretability of business rules and purely rule-driven systems struggling to handle complex semantic business scenarios.

[0163] In this embodiment, a certain order in the procurement system simultaneously triggers both the clauses regarding approval authority for single procurement amounts in the "Procurement Management Regulations" and the clauses regarding invoice compliance in the "Financial Reimbursement System," resulting in a contradiction in the approval level determination between the two rules. The intelligent verification engine module, based on the rule priority arbitration matrix and according to the preset priority level of "specialized systems take precedence over general systems," automatically resolves the conflict, determining that the conclusion of the procurement management regulations is the final applicable rule, without requiring manual intervention.

[0164] Subsequently, the module follows the formula: ;

[0165] Calculate the overall score for verification accuracy. (Among them...) Cosine similarity determination based on data semantic vectors and abnormal pattern vectors:

[0166] ;

[0167] in, For the first The pass rate of explicit rule verification corresponds to the verification items in the business rule base that can be explicitly clause-based and rule-based for verification (such as whether the approval authority matches or whether the invoice tax number format is compliant). Its verification results are relatively certain. For the first The coverage of latent anomaly detection corresponds to latent anomaly patterns that cannot be directly enumerated by the rules and require semantic similarity reasoning for identification (such as transaction behavior patterns that deviate from the historical normal range but do not violate any explicit rules).

[0168] Because the determinism and verifiability of explicit rule verification results are higher than those of implicit detection results derived from semantic similarity reasoning, explicit rule verification items are assigned a higher weight coefficient of 0.7 in the comprehensive score calculation, while implicit anomaly detection items are assigned 0.3. This ratio was determined through A / B testing of no less than 5,000 real verification cases, and is not set based on subjective experience. The judgment threshold of 0.9 was also determined through the aforementioned A / B testing, optimizing between accuracy and false positive rate: a threshold that is too high will cause a large amount of normal data to be misjudged as requiring secondary verification due to minor semantic deviations, increasing unnecessary manual review costs; a threshold that is too low will allow data that is substantially abnormal but not covered by explicit rules to pass.

[0169] This batch of data has been calculated. If the threshold of 0.9 is reached, the semantic rationality check is deemed to have passed. After all the compliance checks of the superimposed rules have passed, the verification result is output by this module to step S3, and handed over to the anomaly detection and risk assessment module for further processing.

[0170] After the verification results are generated, the anomaly detection and risk assessment module executes step S3. This module performs statistical analysis detection, pattern recognition detection, semantic reasoning detection, and rule verification detection in parallel on the verification results, specifically:

[0171] ;

[0172] The system determines whether the test passes and uses a weighted voting mechanism to integrate the four test results to calculate a comprehensive anomaly score. Data items with a comprehensive anomaly score of 0.7 are marked as anomaly candidates and enter the risk assessment layer.

[0173] and The system measures the semantic and rule compliance levels of the data to be verified and the business rule scenarios, respectively. These two measures reflect different attribute dimensions of the detected objects and do not have inherent differences in importance. Therefore, in general scenarios lacking specific business scenario bias indicators, equal weights of 0.5 and 0.5 are used for fusion to avoid artificially introducing unverified dimensional bias. When the system identifies evidence of dimensional bias in a specific business scenario, adjustments can be made through a dynamic weighting mechanism driven by scenario feature vectors (such as the method used in risk comprehensive score calculation). The two are not contradictory. The thresholds of 0.8 and 0.7 are both derived from the score distribution statistics of confirmed abnormal and normal cases in the enterprise's historical verification data: a value of 0.8 corresponds to the statistical boundary point where semantic similarity and rule compliance can effectively distinguish abnormal and normal scenarios, while a value of 0.7 corresponds to the statistical boundary point in the comprehensive abnormal score that can balance the coverage of abnormal identification and the false alarm rate.

[0174] In this embodiment, a supplier payment record in the financial system is marked as a high-risk candidate due to an abnormal amount. The anomaly detection and risk assessment module, based on the primary and foreign key relationship diagram of the data tables stored in the enterprise data warehouse, automatically tracks all data items that reference this payment record through the supplier's primary key. This includes all unsettled orders related to the supplier in the procurement system, and existing contract records signed with the supplier in the contract management system. Cascading anomaly detection is triggered for these related data items, and the cascading detection results are incorporated into the risk score calculation for each of these related data items. Through this cascading detection step, six related contracts and eleven related orders under the supplier that were not initially directly marked as anomaly are included in the risk review scope, avoiding the problem of missed detection caused by anomalies being handled only at the trigger point without spreading to related data for investigation.

[0175] The overall risk score is calculated using the following formula:

[0176] ;

[0177] The dynamic weights are automatically calculated by the WxDataVerifyLM model based on the scene feature vectors.

[0178] ;

[0179] in, to The feature values ​​for the four dimensions of anomaly severity, business impact scope, compliance requirements, and historical risk cases are all standardized and mapped to the [0,1] interval. This gives the four types of original indicators with different dimensions (such as anomaly severity which may be derived from detection scores and the number of historical risk cases which may be derived from statistical counts) a mathematical basis for weighted summation. This standardization process is an actual data processing step executed within the module, rather than an abstract assumption.

[0180] Dynamic weights to The reason for not using fixed values ​​is that the importance of the four dimensions to the final risk assessment varies objectively across different business scenarios (e.g., the importance of compliance requirements is significantly higher in financial regulatory scenarios than in other scenarios). Using fixed weights would cause the model's risk assessment conclusions to become disconnected from the actual importance of the business when the scenario is transferred. This formula dynamically calculates weights using a neural network driven by scenario feature vectors, making weight allocation a technical process that can automatically learn and adjust according to the scenario. (Correction value) The value range is limited to a small interval of [-0.05, 0.05] to ensure that personalized adjustments reflect the subtle correlation between historical risk cases and current data characteristics, while avoiding excessive deviation from the main conclusion of the four-dimensional weighted calculation, and to prevent the adjustment item from overshadowing the main conclusion and causing the risk level determination to lose stability.

[0181] Since this scenario involves internal corporate compliance management rather than external financial regulation, the model does not forcibly increase the weight of the compliance requirements dimension. Instead, it automatically increases the weight of the business impact scope dimension based on the distribution of historical risk cases. This module has the ability to dynamically adapt to different business scenarios.

[0182] The Risk_Score of the aforementioned supplier payment record is calculated to be 0.84, which is considered high risk, and proceeds to step S4.

[0183] After the risk assessment is completed, the execution and linkage module performs step S4. This module uses the data batch ID and the unique identifier of the data record as the association primary key to accurately link and match the verification results and abnormal information of the aforementioned high-risk payment records and their cascaded associated contract and order records.

[0184] In this embodiment, the module supports verification result output and aggregation at four granular levels: field level, record level, table level, and batch level. At the field level, the "Payment Amount" and "Receiving Account" fields in the aforementioned payment records are marked as abnormal. During record-level aggregation, since "Receiving Account" is a preset high-weight field, even if the proportion of abnormal fields in the record does not exceed the preset threshold, the module still determines the record-level conclusion as "fail" based on the weighted counting logic that "the record-level conclusion is upgraded to failure when any high-weight field is abnormal." During table-level aggregation, the proportion of high-risk records in this batch in the financial payment table is 1.2%, which does not exceed the preset table-level warning threshold, so no overall table-level warning is triggered. However, during batch-level aggregation, since the high-risk record and its associated multiple order and contract records are included in the total number of high-risk records in this verification batch, the proportion of high-risk records in the batch reaches the preset threshold, triggering a batch-level high-risk warning. Verification conclusions and anomaly information at each granularity level are bidirectionally associated and stored. Business personnel can drill down from batch-level alerts to the specific fields that triggered the alerts, completing a full-chain traceability query.

[0185] After sorting by risk level, the module outputs a standardized verification report to the enterprise's risk control department through a multi-terminal adaptation interface. Based on the preset hierarchical alarm system, it pushes the batch of high-risk records to the manual review workbench and proceeds to step S5.

[0186] During the entire data verification process, the execution and linkage module synchronously undertakes some functions of step S5, and monitors key indicators such as verification throughput, verification accuracy, anomaly detection recall rate, and interface response speed in real time. Reviewers review the aforementioned high-risk supplier payment records and their cascading related records in the manual review workbench. The review conclusions are collected in structured data format and stored in the enterprise data warehouse as input data for subsequent feedback and iterative optimization.

[0187] After the data is verified and consolidated, the feedback iterative optimization module executes step S6. This module performs a graded diagnosis of the abnormal information generated by the system in this round of verification: ;

[0188] The system calculates the priority weights for anomaly handling. Level and Impact represent the severity level of the anomaly itself (Level 1 = 3, Level 2 = 2, Level 3 = 1) and the business scope of the anomaly's impact (affecting full verification = 3, affecting partial verification = 2, affecting a single data item = 1), respectively. Both are discretized values ​​derived from actual anomalies generated by the system according to preset rules, rather than arbitrarily assumed values. Weight coefficients of 0.6 and 0.4 assign a slightly higher weight to the anomaly level than to the impact scope. This is because the anomaly level itself implicitly reflects the severity of potential business losses, while the impact scope reflects more the breadth of the anomaly's spread. Based on historical anomaly handling cost statistics and business loss cost model calculations, the anomaly level has a slightly greater determinant effect on the final handling priority than the impact scope alone; therefore, this weight combination is determined. The threshold 2.2 is also determined based on this business loss cost model: when... When the threshold of 2.2 is reached, the business loss cost caused by the corresponding abnormal event in historical statistics has exceeded the loss caused by the processing latency that the regular iterative optimization queue can withstand. Therefore, it is necessary to trigger the emergency optimization process for priority handling.

[0189] In this embodiment, because the aforementioned supplier payment anomaly, after being detected through cascading propagation, affected the verification conclusions of multiple related tables, it was identified as a Level 2 anomaly (Level=2) and its impact scope was determined to be "affecting partial verification" (Impact=2). The calculation yields... The emergency optimization trigger threshold of 2.2 was not reached, so it was included in the regular iterative optimization queue. However, within the same verification cycle, another event, caused by an anomaly in the human resources system data interface leading to a complete interruption of payroll verification, was identified as a Level 1 anomaly (Level=3) with an impact scope of "affecting full verification" (Impact=3). The calculation... If the threshold of 2.2 is exceeded, an emergency optimization process is triggered. The feedback iterative optimization module then prioritizes the repair and parameter adjustment of the data receiving and preprocessing module configuration corresponding to that interface. The comparison of the actual calculation results shows that the weight and threshold combination used in this formula can effectively distinguish the processing priority between secondary local anomalies and primary global anomalies. This is not a formal mathematical calculation, but directly drives the actual engineering decision-making of subsequent module resource scheduling.

[0190] For anomalies in the regular iterative optimization queue, this module processes the manual review feedback data from the past 90 days based on a time decay weighted mechanism, specifically:

[0191] ;

[0192] Construct a verification misjudgment feedback vector and use weight coefficients Integrating model verification reasoning logic:

[0193] ;

[0194] in For the first The feature vector actually generated during the manual review process, and the time decay coefficient. The purpose is to give higher weight to human review feedback that is more recent in time in the feedback vector construction, while the weight of earlier feedback decays exponentially over time. The value of this coefficient is determined by statistical analysis of the frequency of historical business rule changes and the failure cycle of human review feedback; fusion weight Limiting the extent to which manual review feedback corrects the model's verification reasoning logic enables the feedback mechanism to continuously correct model biases, while preventing feedback vectors from dominating the verification reasoning logic and weakening the model's verification ability based on rules and data itself.

[0195] The entire process of iterative optimization of the WxDataVerifyLM model, verification engine verification rules, and anomaly detection parameters was completed. The optimized parameters and rules were then dynamically synchronized to the data receiving and preprocessing module, the intelligent verification engine module, and the anomaly detection and risk assessment module, forming a complete closed-loop technology chain.

[0196] In this embodiment, the enterprise adds an electronic contract signing business line, generating electronic signature log data that the verification platform has not previously encountered, thus forming a new business verification scenario. In step S2, the intelligent verification engine module identifies that this new scenario lacks historical verification cases. It then calls its maintained verification scenario and verification strategy mapping relationship library, performing a similarity search based on the semantic feature vector of the new scenario and existing scenarios in the library, matching the scenario with the highest semantic similarity to the "Contract Management" business line. Using transfer learning, the module uses the verification strategy parameters accumulated in the "Contract Management" scenario as the initialization parameters for the electronic contract signing scenario, automatically adapting the strategy for this new scenario. Enterprise managers can verify the electronic signature log data without manually configuring any verification rules. After the new scenario runs continuously and accumulates more than two thousand verification cases, the intelligent verification engine module continuously fine-tunes the adapted strategy parameters through an online learning mechanism, gradually aligning the verification strategy with the characteristic distribution of the electronic signature log data itself. This optimization effect is simultaneously fed back to the mapping relationship library for further reuse when the enterprise adds new business lines in the future.

Claims

1. A large-scale model data verification system based on rule-based semantic fusion, characterized in that, It includes a data receiving and preprocessing module, an intelligent verification engine module, an anomaly detection and risk assessment module, an execution and linkage module, and a feedback and iterative optimization module; The data receiving and preprocessing module is used to receive structured, semi-structured and unstructured data to be verified in multiple forms. It uses OCR and NLP technologies to intelligently parse unstructured documents. After data cleaning and format normalization, it extracts data structure features, data semantic features and data quality features to generate standardized data information and outputs it to the intelligent verification engine module and the anomaly detection and risk assessment module. The intelligent verification engine module is equipped with a large data verification model. It adopts a three-layer architecture of data semantic understanding layer, rule fusion reasoning layer and verification decision generation layer. It integrates business rule base and historical verification cases in data warehouse, performs verification reasoning in four dimensions of completeness, consistency, accuracy and timeliness, and outputs verification results after passing the dual verification mechanism of rule compliance verification and semantic rationality verification. The anomaly detection and risk assessment module constructs a four-fold detection mechanism: statistical analysis detection, pattern recognition detection, semantic reasoning detection, and rule verification detection. It uses a weighted voting mechanism to integrate the results of the four-fold detection to calculate a comprehensive anomaly score. The data verification model dynamically allocates weights for four core dimensions—anomaly severity, business impact scope, compliance requirements, and historical risk cases—based on the scenario feature vector, calculates a comprehensive risk score, and classifies risk levels. The execution and linkage module uses the data batch ID and the unique identifier of the data record as the association primary key to accurately link and match the verification results with the abnormal information. After sorting by risk level, it outputs a standardized verification report through a multi-terminal adaptation interface and builds a hierarchical alarm system and a manual review workbench. The feedback iterative optimization module performs hierarchical diagnosis of system anomaly information, processes manual review feedback data based on a time decay weighting mechanism, performs full-link iterative optimization of the data verification big model, verification engine verification rules and anomaly detection parameters, and dynamically synchronizes the optimized parameters and rules to each core module.

2. The rule-based semantic fusion-based large model data verification system according to claim 1, characterized in that, In the intelligent verification engine module, the data semantic understanding layer employs a data semantic mask matrix. The improved Transformer encoder structure uses the following formula for calculating attention weights: ； in The data semantic mask matrix assigns higher attention weights to key fields, core business metrics, and highly sensitive data; it is automatically updated when verification rules in the data warehouse's business rule base are added, changed, or abolished. The mask weight value of the corresponding rule field is determined, and an incremental re-core process is triggered for historical and existing datasets affected by rule changes.

3. The rule-based semantic fusion-based large model data verification system according to claim 2, characterized in that, In the intelligent verification engine module, the rule fusion reasoning layer adopts a knowledge embedding attention mechanism to vectorize and embed the business rule base in the data warehouse in the form of a knowledge graph. The rule embedding formula is as follows: ; in For data feature vectors, To verify the context vector, This is the rule weight matrix; A rule priority arbitration matrix is ​​introduced. When the reasoning conclusions from the rule base are contradictory, the rule priority arbitration matrix automatically resolves the rule conflict according to the preset priority level and determines the final applicable rule. The verification decision generation layer is based on an improved Transformer decoder and configures a dedicated decoding template matrix for different business scenarios. : ; in For data-rule fusion vectors, For scene bias; A self-verification mechanism is introduced to verify the logical consistency of decisions in real time during the generation of verification decisions through a consistency verifier.

4. The rule-based semantic fusion-based large model data verification system according to claim 1, characterized in that, In the aforementioned dual verification mechanism, semantic rationality verification is achieved by calculating a comprehensive score based on the accuracy of the verification. To achieve this, the calculation formula is: ； in For the first Explicit rule validation pass rate The number of explicit rule validation items. For the first Coverage of latent anomaly detection This represents the number of dimensions for detecting latent anomalies. Based on data semantic vectors With abnormal pattern vector Cosine similarity calculation: ； Similarity hour ,otherwise ;when When semantic rationality is checked, The system automatically adjusts the verification strategy and triggers secondary verification. The threshold of 0.9 is determined through A / B testing of no less than 5,000 real verification cases to balance verification accuracy and processing efficiency. The verification result is output to the subsequent modules only when both rule compliance verification and semantic rationality verification are passed simultaneously.

5. The rule-based semantic fusion-based large model data verification system according to claim 1, characterized in that, In the anomaly detection and risk assessment module, the formula for determining the dual pass condition of the quadruple detection mechanism is as follows: ； in The threshold is the cosine similarity between the semantic vector of the data to be verified and the scenario vector of the business rule. ; The ratio of the number of data items that meet the rule validation conditions to the total number of rule validation conditions is required, which requires that all rule sub-conditions be met simultaneously; when The test is deemed passed upon completion; the overall anomaly score is considered accurate. The data items are marked as anomaly candidates and output to the risk assessment layer.

6. The large-scale model data verification system based on rule-based semantic fusion according to claim 1, characterized in that, In the anomaly detection and risk assessment module, the formula for calculating the comprehensive risk score is as follows: ; in The feature value represents the severity of the anomaly. For the feature values ​​of the business impact scope dimension, For compliance requirements dimension feature values, For historical risk case feature values, all are standardized to interval; Personalized correction values ​​are dynamically adjusted by the large data verification model based on historical risk cases and current data characteristics; Dynamic weights to The large model is verified by data based on scene feature vectors Automatic calculation: 。 7. The rule-based semantic fusion-based large model data verification system according to claim 1, characterized in that, The data receiving and preprocessing module is also used to calculate the Data Integrity Index (DCI): ; Among them, the first The integrity scores for each field are as follows: ; Field Business Importance Weight Based on the rule dependency graph, the frequency and dependency depth of each field in the verification rules are calculated. Fields with higher frequency and deeper dependency depth are assigned greater weights. The weight calculation formula is as follows: ; in For the first Semantic vectors of each field, For scene feature vectors, This indicates vector concatenation; It was deemed qualified. Trigger supplementary verification It is directly marked as high risk and all subsequent verification processes are blocked, and a data completion request is sent to the data provider.

8. The rule-based semantic fusion-based large model data verification system according to claim 1, characterized in that, In the feedback iterative optimization module, a verification misjudgment feedback vector is constructed based on a time decay mechanism. : ; in For the first The feature vectors returned by the second manual review This is a time decay factor, which gives more weight to recent review conclusions. The total number of feedback responses within the past 90 days is taken; Will With weighting coefficients Verification reasoning logic integrated into a large-scale data verification model: ; Will Parameters are used for anomaly detection model correction to achieve continuous optimization through multi-module collaboration.

9. A large-scale model data verification system based on rule-based semantic fusion according to claim 1, characterized in that, In the feedback iterative optimization module, the formula for calculating the priority weight of system anomaly handling is as follows: ; in It is an abnormal level. The scope of the abnormal impact. , This combination of coefficients was determined by the business loss cost model. The emergency optimization process is triggered in time. It is then included in the regular iterative optimization queue.

10. A method for verifying large-scale model data based on rule-based semantic fusion, characterized in that, Includes the following steps: S1. Data reception and standardization processing: Receives structured, semi-structured and unstructured data in various forms to be verified. Intelligently parses unstructured documents using OCR and NLP technologies. After data cleaning, format normalization and multi-dimensional feature extraction, standardized data information containing data structure features, data semantic features and data quality features is generated. S2, the intelligent verification engine reasoning, calls the large data verification model, and adopts a three-layer architecture of data semantic understanding layer, rule fusion reasoning layer and verification decision generation layer. It integrates the business rule base in the data warehouse and historical verification cases, and performs four-dimensional verification reasoning of completeness, consistency, accuracy and timeliness. After double verification of rule compliance and semantic rationality, the verification result is output. The specific implementation of the double verification mechanism is as follows: rule compliance verification: the verification conclusion is checked for rule compliance based on the business rule base in the data warehouse. When the reasoning conclusions from multiple rule bases are contradictory, the rule priority arbitration matrix is ​​called to automatically resolve the conflict according to the priority level. Perform semantic reasonableness verification and calculate the comprehensive score of verification accuracy using the following formula. ,when When the verification is passed, Automatically adjust verification strategies and trigger secondary verification: ； The output is sent to step S3 only when the verification result passes both the rule compliance check and the semantic reasonableness check. S3. Perform four-fold anomaly detection on the verification results: statistical analysis detection, pattern recognition detection, semantic reasoning detection, and rule verification detection. Use a weighted voting mechanism to integrate the four-fold detection results to calculate a comprehensive anomaly score. The data verification big model dynamically assigns weights based on scene feature vectors to calculate a comprehensive risk score and classify it into high, medium, and low risk levels. The method for calculating the comprehensive risk score is as follows: ； Among them, dynamic weights to The large-scale data verification model is based on scene feature vectors. pass Automatic calculation to automatically meet the requirements of financial risk control scenarios. Personalized correction value The model is dynamically adjusted based on historical risk cases and current data characteristics; S4. Linked output of verification results and abnormal information: Using the data batch ID and the unique identifier of the data record as the association primary key, the verification results and abnormal information are accurately linked and matched. After being sorted by risk level, a standardized verification report is output through a multi-terminal adaptation interface, and a graded alarm notification is triggered simultaneously. S5. Real-time monitoring and manual review: Monitor the execution status of the entire system process in real time and capture abnormal information; trigger the manual review process for high-risk verification results, and collect the manual review conclusions in structured data format and store them in the data warehouse. S6. Classify and diagnose abnormal information in the system and formulate targeted adjustment strategies. Process manual review feedback data based on time decay weighting mechanism, construct verification misjudgment feedback vector, implement full-link iterative optimization of data verification big model, verification engine verification rules and abnormal detection parameters, and dynamically synchronize the optimized parameters and rules to each core module. The specific steps for constructing and fusing the time-decay weighted feedback vector are as follows: collecting manual review feedback, and processing the feature vector for each feedback. Apply time decay coefficient Construct the feedback vector: ; With weighting coefficients Integrating data verification into the reasoning logic of the large model: System anomalies are handled according to priority weights. Perform hierarchical processing. Trigger emergency optimization. Included in regular iterations.

Images