A data quality evaluation method based on multi-source heterogeneous material driving and a system thereof

Abstract

The application discloses a data quality evaluation method and system based on multi-source heterogeneous material driving. The application constructs a unified evaluation knowledge graph by fusing traditional basic quality dimensions and high-quality data set expansion evaluation dimensions, receives multi-source heterogeneous evaluation materials, and completes intelligent analysis and cross-document information fusion by type, automatically maps data fields and samples to a complete set of standardized indexes based on multi-dimensional matching scores, generates executable check rules in batches, and completes intelligent correction of the rules through multi-layer cross-validation, realizes differentiated weighting scoring of each dimension relying on a hierarchical weight distribution mechanism, completes distributed parallel batch evaluation in combination with rule-dependent topology optimization, aggregates evaluation results in layers and outputs a unified standardized quality evaluation report, realizes full-process automation from multi-source evaluation material input to standardized data quality evaluation report output, shortens the full-process evaluation cycle, and improves the efficiency and accuracy of data quality evaluation.

Description

Technical Field

[0001] This invention relates to the field of data quality assessment technology, specifically to a data quality assessment method and system driven by multi-source heterogeneous materials. Background Technology

[0002] With the continuous advancement of market-oriented reforms of data elements, data quality assessment has become a core pre-requisite technical step for data asset ownership confirmation, trading, and accounting. It is also an important technical support for third-party data quality certification and corporate data compliance governance. In the implementation of data quality assessment, multi-source heterogeneous materials, consisting of data dictionaries, system architecture descriptions, business rule documents, and regulatory compliance requirement documents, serve as the core input basis for the assessment work. The assessment output results must strictly correspond to the requirements of the standardized indicator system and possess the characteristics of process traceability, result reproducibility, and conclusion comparability.

[0003] Currently, data quality assessment technology has mainly formed three types of technical solutions in practical applications: The first type is the rule engine-driven automated assessment solution, which takes predefined verification rules as the core and realizes parallel verification of large-scale data through a distributed computing engine; however, in this type of solution, all assessment rules, indicator mapping, and parameter configuration rely on manual definition and input by domain experts, which only realizes the automation of the assessment execution process. The front-end assessment material parsing and rule design processes still require a lot of manual input, resulting in low adaptation efficiency and long delivery cycle when facing multi-source heterogeneous assessment materials;

[0004] The second category is large-scale model-based evaluation schemes, which rely on the semantic understanding capabilities of large language models to directly parse evaluation requirements, generate verification logic, and even output quality scores, possessing strong flexibility and versatility; however, this type of method does not have a fixed and unified standardized indicator system, making it difficult to maintain consistency in the rule definitions and judgment criteria for different evaluation tasks, resulting in a lack of horizontal comparability of evaluation results, and the rule generation process lacks systematic verification, resulting in insufficient stability and difficulty in meeting the seriousness requirements of standardized evaluation.

[0005] The third category is assessment solutions empowered by knowledge graphs. These solutions use knowledge graphs to structure and store data quality rules and metadata information, supporting the retrieval and reuse of rules. However, in most existing solutions, knowledge graphs are only used as rule bases and metadata directories. They do not establish a complete semantic association and reasoning link between standardized indicator systems, business scenarios, compliance requirements, and assessment rules, and cannot support the fully automated processing from original assessment materials to the final assessment report.

[0006] The applicant previously published a paper with publication number CN122388953B, which disclosed a multimodal data quality assessment method that integrates meta-models and large models. By constructing an assessment meta-model based on category theory, it solved the problems of unified semantic abstraction, dependency-aware scheduling optimization, and heterogeneous scoring standardization in multimodal data quality assessment, effectively improving the efficiency of assessment execution and the comparability of results. However, the input of this technical solution is a predefined structured multimodal data assessment task, which does not cover the front-end transformation from the client's original unstructured assessment materials to the standardized assessment task. It still requires domain experts to invest a lot of manual labor to complete unstructured document parsing, data relationship sorting, standardized indicator mapping, and adaptation rule writing, and cannot achieve full automation of the data quality assessment process from material reception to report output.

[0007] It is evident that existing data quality assessment methods suffer from poor consistency in indicator weights, incomparable assessment results across different batches, failure to consider rule dependencies during assessment execution, significant duplication of calculations, and low assessment efficiency. Furthermore, they generally lack multi-dimensional systematic verification mechanisms for automatically generated rules, resulting in a high error rate and failing to meet the stringent accuracy requirements of data quality assessment.

[0008] Therefore, it is necessary to invent a data quality assessment method and system based on multi-source heterogeneous materials to solve the above problems. Summary of the Invention

[0009] The purpose of this invention is to provide a data quality assessment method and system based on multi-source heterogeneous materials to address the aforementioned shortcomings in the technology.

[0010] To achieve the above objectives, the present invention provides the following technical solution: a data quality assessment method based on multi-source heterogeneous materials, comprising the following steps:

[0011] Step 1: Construct an evaluation knowledge graph with a standardized indicator system at its core;

[0012] Step 2: Receive multi-source heterogeneous evaluation input materials and output a structured evaluation input model;

[0013] Step 3: Based on the multi-dimensional association between the structured evaluation input model and the evaluation knowledge graph, automatically map data fields to the corresponding standardized indicator nodes;

[0014] Step 4: Based on the evaluation rule templates corresponding to the standardized indicator nodes obtained from the mapping, automatically generate an initial executable evaluation rule set that is compatible with multiple execution engines;

[0015] Step 5: Perform multi-dimensional cross-validation and intelligent correction on the initial executable evaluation rule set to obtain the target executable evaluation rule set;

[0016] Step 6: Based on the unified basic weights pre-fixed by experts, assign evaluation weights to each indicator in combination with business scenarios and compliance requirements, optimize the parallel execution of the target executable evaluation rule set through rule dependency optimization, and generate a standardized data quality evaluation report.

[0017] The aforementioned data quality assessment method based on multi-source heterogeneous materials involves constructing an assessment knowledge graph with a standardized indicator system at its core in step 1. The specific steps are as follows:

[0018] 1.1 Constructing a node set ;

[0019] 1.2 Construct a set of relationships to describe the semantic, logical, and dependency relationships between nodes. ;

[0020] 1.3 Generate text descriptions for all nodes. Dense vector representation, and constructing vector indexes based on the approximate nearest neighbor algorithm. ;

[0021] 1.4. Provides CRUD (Create, Read, Update, Delete) interfaces for evaluating knowledge graphs, supporting version tracking and incremental updates based on node sets. Relation sets Vector index Obtain the evaluation knowledge graph .

[0022] The aforementioned data quality assessment method based on multi-source heterogeneous materials, in step 1.1, constructs a node set. The specific formula is as follows:

[0023] ;

[0024] in, A standardized set of indicator nodes; To evaluate the set of rule template nodes; A collection of data type nodes; A collection of business scenario nodes; For the set of execution engine nodes;

[0025] The standardized indicator node set, evaluation rule template node set, data type node set, business scenario node set, and execution engine node set are constructed as follows:

[0026] 1.1.1 Constructing a standardized set of indicator nodes Specifically, the calculation formula for each standardized indicator node is as follows:

[0027] ;

[0028] in, For the first One standardized indicator node; For the first A globally unique identifier for each indicator; For the first The standard name of each indicator; For the first Standardized definition of each indicator; For the first Description of the calculation method for each indicator; For the first The pass / fail threshold for each indicator; For the first The compliance documents and clause numbers corresponding to each indicator; For the first Priority of each indicator; For the first The set of data type identifiers applicable to each indicator; For the first A set of business scenario identifiers applicable to each indicator; For the first A set of evaluation rule template identifiers associated with each indicator;

[0029] 1.1.2 Constructing a set of evaluation rule template nodes Specifically, the calculation formula for each evaluation rule template node is as follows:

[0030] ;

[0031] in, For the first One evaluation rule template node; For the first A globally unique identifier for each template; For the first The name of the template; For the first The rule content of a template with parameter placeholders; For the first Each template contains a list of parameter definitions, and each parameter includes a parameter name, parameter type, and valid parameter range. For the first The syntax rules corresponding to each template; For the first Each template supports a set of execution engine identifiers; For the first Automatic validation rules corresponding to each template;

[0032] 1.1.3 Constructing a collection of data type nodes Specifically, the calculation formula for each data type node is as follows;

[0033] ;

[0034] Among them, For the first Each data type node; For the first A globally unique identifier for each data type; For the first The name of each data type; For the first A format specification description for each data type; For the first Description of common quality issues for each data type; For the first A set of evaluation method identifiers corresponding to each data type;

[0035] 1.1.4 Constructing a subset of business scenario nodes Specifically, the calculation formula for each business scenario node is as follows:

[0036] ;

[0037] in, For the first Each business scenario node; For the first A globally unique identifier for each business scenario; For the first The name of each business scenario; For the first A detailed description of each business scenario; For the first A set of standardized indicators that are of core concern in each scenario; For the first Description of mandatory compliance requirements in various scenarios;

[0038] 1.1.5 Constructing the Execution Engine Node Set Specifically, the calculation formula for each execution engine node is as follows:

[0039] ;

[0040] in, For the first One execution engine node; For the first A globally unique identifier for each execution engine; For the first The name of the execution engine; For the first The set of rule types supported by each engine; For the first Performance parameters of each engine; For the first The default configuration parameters for each engine;

[0041] In step 1.2, a set of relationships is constructed to describe the semantic associations, logical associations, and dependencies between the nodes. The specific formula is as follows:

[0042] ;

[0043] in, This is a subset of the relationships between indicators and applicable data types. This refers to a subset of the relationships between metrics and applicable business scenarios. This is a subset of the relationships between indicators and their corresponding rule templates; This refers to a subset of the relationship between rule templates and supporting execution engines. This is a subset of the relationships between business scenarios and core metrics of interest.

[0044] The following methods are used to construct the subsets of relationships between metrics and applicable data types, metrics and applicable business scenarios, metrics and corresponding rule templates, rule templates and supporting execution engines, and business scenarios and core metrics of interest:

[0045] 1.2.1 Construct a subset of relationships between indicators and applicable data types. Each relationship represents a standardized indicator that is applicable to one or more data types. The specific formula is as follows;

[0046] ;

[0047] in, To start from standardized indicator nodes Pointer to data type node A directed edge, representing the first... The first indicator applies to the first Data types;

[0048] 1.2.2 Construct a subset of relationships between metrics and applicable business scenarios. Each relationship represents a standardized metric that is applicable to one or more business scenarios. The specific formula is as follows:

[0049] ;

[0050] in, To start from standardized indicator nodes Pointing to business scenario nodes A directed edge, representing the first... The first indicator applies to the first One business scenario;

[0051] 1.2.3 Constructing a subset of relationships between indicators and corresponding rule templates Each relation represents a standardized indicator corresponding to one or more evaluation rule templates, as shown in the following formula:

[0052] ;

[0053] in, To start from standardized indicator nodes Point to the evaluation rule template node A directed edge, representing the first... The first indicator applies to the first A rule template generates evaluation rules. For the first A set of rule template identifiers corresponding to each indicator;

[0054] 1.2.4. Relationship between rule templates and supporting execution engines (subset) Each relation represents an evaluation rule template that can be run on one or more execution engines, as shown in the following formula:

[0055] ;

[0056] in, To evaluate the rule template node Point to the execution engine node A directed edge, representing the first... The rule template supports the first one. It runs on an execution engine;

[0057] 1.2.5 Constructing a subset of relationships between business scenarios and core metrics of interest Each relation represents a core metric of interest in a business scenario, and the specific formula is as follows:

[0058] ;

[0059] in, To start from business scenario nodes Pointing to standardized indicator nodes A directed edge, representing the first... The first indicator is the Key metrics to focus on for each business scenario;

[0060] In step 1.3, text descriptions are generated for all nodes. Dense vector representation, and constructing vector indexes based on the approximate nearest neighbor algorithm. The specific steps are as follows:

[0061] 1.3.1. Preprocess the core text description of each node, including word segmentation, stop word removal, and normalization, to generate a standardized text representation;

[0062] 1.3.2. A pre-trained language model is used to convert the standardized text representation of each node into... The 3D dense vector representation is given by the following formula:

[0063] ;

[0064] in, For the first Each node The dimension is represented as a dense vector. For the first Standardized text representation of each node, For pre-trained language model encoders, , for 3D real vector space;

[0065] 1.3.3 Constructing a vector index using an approximate nearest neighbor algorithm , all nodes The dense vector representation is indexed to support millisecond-level semantic similarity retrieval, as shown in the following formula;

[0066] ;

[0067] in, Functions for constructing approximate nearest neighbor indices. To assess the total number of nodes in a knowledge graph;

[0068] In step 1.4, CRUD (Create, Read, Update, Delete) interfaces for evaluating the knowledge graph are provided, supporting version tracking and incremental updates based on node sets. Relation sets Vector index Obtain the evaluation knowledge graph The specific steps are as follows:

[0069] 1.4.1 Generate a unique version number for each evaluation of knowledge graph changes. Record the changes, the time of change, and the person making the change, for use in evaluating the version traceability of the knowledge graph;

[0070] 1.4.2 Supports incremental updates of the knowledge graph for evaluation. When a knowledge node or relationship is added or modified, the corresponding vector index is automatically updated. ;

[0071] 1.4.3. Provides a batch import / export interface for evaluation knowledge graphs, supporting batch import of evaluation knowledge graphs from external files, and also supporting export of evaluation knowledge graphs to a standardized format, thus obtaining a complete evaluation knowledge graph capable of machine reasoning. The specific formula is as follows:

[0072] .

[0073] The aforementioned data quality assessment method based on multi-source heterogeneous materials, in step 2, receives the multi-source heterogeneous assessment input materials and outputs a structured assessment input model. The specific steps are as follows:

[0074] 2.1 Constructing a multi-source assessment material set The specific formula is as follows:

[0075] ;

[0076] in, A subset of structured materials; It is a subset of semi-structured materials; It is a subset of unstructured materials; A subset of other auxiliary materials;

[0077] 2.2 For different types of materials in the multi-source assessment material set, a differentiated intelligent analysis strategy is adopted to automatically extract core attribute information;

[0078] 2.3. Fragmented information extracted from different materials will be correlated, integrated, and aligned to eliminate information conflicts, fill in missing information, and generate a unified structured evaluation input model. The specific formula is as follows:

[0079] ;

[0080] in, This constitutes a list of data assets, including the name, identifier, and source information of all data assets to be evaluated; For data structure details, including field definitions, types, and constraint information for each data asset; This includes business background information, such as business scenario descriptions, business rules, and data flow relationships; This is a list of compliance requirements, including all applicable compliance clauses, mandatory requirements, and qualification standards.

[0081] The aforementioned data quality assessment method based on multi-source heterogeneous materials, in step 2.2, employs differentiated intelligent parsing strategies to automatically extract core attribute information for different types of materials in the multi-source assessment material set, as detailed below:

[0082] 2.2.1 For structured materials such as data dictionaries, database metadata, and API interface documents, directly extract the core attribute information of the data assets for structured material parsing;

[0083] Furthermore, the core attribute information includes table name, field name, field type, field description, primary key identifier, foreign key association, index information, and data length constraints;

[0084] 2.2.2. Perform semi-structured material parsing on semi-structured materials in Word, PDF, and Markdown document formats;

[0085] Specifically, it employs a large language model for text-level semantic understanding, automatically extracting information such as business rules, data flow relationships, data quality requirements, mandatory compliance clauses, and business constraints.

[0086] 2.2.3. Perform unstructured material analysis on system architecture diagrams, business process diagrams, and scanned unstructured materials;

[0087] Specifically, optical character recognition technology is used in conjunction with a multimodal large model to analyze image content and automatically extract information on system components, data flow, storage location, processing nodes, and interface call relationships.

[0088] 2.2.4. Analyze system logs, data samples, compliance documentation, and other supporting materials.

[0089] Specifically, this includes: automatically extracting statistical features, anomaly records, and compliance qualification information from data;

[0090] In step 2.3, the fused structured evaluation input model needs to undergo multi-dimensional quality verification to check the completeness, consistency, and accuracy of its information before a complete, quality-verified structured evaluation input model can be output. The specific multi-dimensional quality verification process is as follows:

[0091] Integrity verification: Checks for missing information in core fields, and marks and prompts for any missing information;

[0092] Consistency check: Check whether there are conflicts in the same information extracted from different materials, mark the conflicting information and provide correction suggestions;

[0093] Accuracy verification: Check whether the extracted information conforms to logical norms and industry common sense, and mark obviously erroneous information.

[0094] The aforementioned data quality assessment method driven by multi-source heterogeneous materials, in step 3, automatically maps data fields to corresponding standardized indicator nodes based on the multi-dimensional association between the structured assessment input model and the assessment knowledge graph. The specific steps are as follows:

[0095] 3.1 For each data field, calculate its matching degree with all standardized indicators in the evaluation knowledge graph across four dimensions. The specific formula is as follows:

[0096] ;

[0097] in, For the first The data field and the first A set of multi-dimensional matching degrees for standardized indicators; For semantic similarity, calculate the textual semantic similarity between the data field description and the standardized metric definition; For type similarity, standardized metrics that support the type of data field are matched based on the data field type. For business similarity, based on business background information Match standardized metrics applicable to the corresponding business scenarios; For compliance similarity, based on the compliance requirements list Match standardized metrics that meet mandatory compliance requirements;

[0098] 3.2 Assign preset weights to the matching degree of the four dimensions, and calculate the final matching degree score of each data field with each standardized indicator using a weighted average. The specific formula is as follows:

[0099] ;

[0100] in, For the first The data field and the first The final matching score of each standardized indicator; For semantic weight coefficients, Weight coefficients for types, Weighting coefficients for business operations For the compliance dimension, the weighting coefficients are set, and they satisfy the following conditions: ;

[0101] Furthermore, the weighting coefficients can be dynamically adjusted based on business scenarios and compliance requirements;

[0102] 3.3. Score all standardized metrics for each data field based on the final matching degree. Sort the data from high to low, select the top N indicators with the highest matching degree as the recommendation results, and generate an initial field-indicator mapping table;

[0103] The value of N can be flexibly set according to the evaluation accuracy requirements and the workload of expert review;

[0104] 3.4 When multiple standardized metrics are mapped to the same data field, the optimal matching metric is automatically selected based on metric priority, business scenario importance, and compliance requirements. The specific formula is as follows:

[0105] ;

[0106] in, For the first The optimal matching metric for each data field; For the first The overall priority of each standardized indicator is calculated by weighting the inherent priority of the indicator, the priority of the business scenario, and the priority of compliance.

[0107] Furthermore, for mandatory compliance indicators that have a veto requirement, their priority is automatically set to the highest.

[0108] 3.5 Generate detailed matching reasons for each initial field-indicator mapping result, and output the target field-indicator mapping result set;

[0109] The matching reasons include: the matching score of each dimension, the matching basis and the reasons for priority determination, which are presented in a structured form for expert review and reference, thereby ensuring the interpretability of the mapping results and supporting expert review.

[0110] The target field-indicator mapping result set includes the correspondence between data fields and standardized indicators, matching scores, matching reasons, and priority determination results.

[0111] The aforementioned data quality assessment method based on multi-source heterogeneous materials automatically generates an initial executable assessment rule set adapted to multiple execution engines in step 4, according to the assessment rule templates corresponding to the standardized index nodes obtained by mapping. The specific steps are as follows:

[0112] 4.1 Based on the optimal matching index The evaluation rule templates associated with this indicator are obtained from the evaluation knowledge graph. The specific formula is as follows:

[0113] ;

[0114] in, For the first A set of evaluation rule templates corresponding to each data field. For the first An evaluation rule template, and The calculation formula corresponds to; For the first A globally unique identifier for each evaluation rule template. For the first The set of rule template identifiers corresponding to the best matching index for each data field, and the calculation The definitions in the formula are consistent.

[0115] 4.2 Using a structured evaluation input model, automatically infer and populate all parameters in the evaluation rule template to generate parameterized rule instances. The specific formula is as follows:

[0116] ;

[0117] in, For the first The first data field corresponding to the A set of parameters for each evaluation rule template; This is a set of data location parameters, including table name, field name, and database connection information; This is a set of business logic parameters, including qualification thresholds, judgment conditions, and filtering rules. This is a set of execution parameters, including the execution engine identifier, timeout, and number of retries.

[0118] 4.3. Based on the client's specific compliance requirements and business needs, automatically adjust the rules' parameters, logic, and priorities to generate customized rule parameters adapted to the client's scenario. The specific formula is as follows:

[0119] ;

[0120] in, For the adjusted set of customized parameters, Adjust the function for the rules;

[0121] For business background information, For the compliance requirements list, for clauses with mandatory compliance requirements, the corresponding rules are automatically set to the highest priority;

[0122] 4.4. Based on the execution engine types supported by the evaluation rule template, convert the rule instance with customized rule parameters into the native format supported by the target execution engine. The specific formula is as follows:

[0123] ;

[0124] in, For the first The first data field, the first The first evaluation rule template, the first The native executable rules corresponding to each execution engine; This is a format conversion function; For the first Each execution engine, and the execution engine node correspond;

[0125] 4.5 Generate complete metadata information for each executable rule for subsequent rule management, traceability, and auditing. Automatically generate a metadata set using the following formula:

[0126] ;

[0127] in, For the first The first data field, the first The first evaluation rule template, the first Metadata collection corresponding to each execution engine This serves as a globally unique identifier for the rule. The standardized indicator identifier corresponding to the rule. For the rule generation time, This is the identifier of the execution engine corresponding to the rule;

[0128] 4.6 Integrate all generated executable rules and their metadata to obtain a complete initial executable evaluation rule set, as shown in the following formula:

[0129] ;

[0130] in, This is the initial set of executable evaluation rules; This represents the total number of data fields to be evaluated. For the first A set of evaluation rule templates corresponding to each data field.

[0131] The aforementioned data quality assessment method based on multi-source heterogeneous materials, in step 5, performs multi-dimensional cross-validation and intelligent correction on the initial executable assessment rule set to obtain the target executable assessment rule set. The specific steps are as follows:

[0132] 5.1 Perform multi-dimensional automatic validation on each executable evaluation rule to generate a rule validation result set. The specific formula is as follows:

[0133] ;

[0134] in, For the first The data field, the first The first evaluation rule template, the first The set of verification results for each rule corresponding to an execution engine; To verify the grammatical validity, check whether the rules conform to the grammatical specifications of the target execution engine; To verify the semantic correctness of the results, check whether the rule logic is consistent with the definition of the corresponding standardized indicator. To verify the existence of data, check whether the tables, fields, and database resources referenced in the rules actually exist and are accessible. To verify the logical consistency of the result, check whether there are logical conflicts, duplicate calculations, or contradictory results between this rule and other rules; To verify the compliance results, check whether the rules meet the compliance clauses and mandatory requirements extracted in step 2;

[0135] 5.2 Intelligent automatic error correction: For rules that fail validation, the system automatically identifies the cause of the error and generates a correction plan. It automatically corrects common errors using the following formula:

[0136] ;

[0137] in, The revised executable rules; This is a smart correction function; The evaluation knowledge graph in step 1 is used to provide semantic and logical basis for error correction;

[0138] 5.3 Deploy the revised rules to an isolated sandbox environment for testing to verify the stability of the rules and the reasonableness of the results. The specific formula is as follows:

[0139] ;

[0140] in, This is a function executed in the sandbox that returns two results: This indicates the sandbox execution status, with values ​​of "passed" or "failed". The results are for sandbox execution; the sandbox environment provides a simulated operating environment consistent with the production environment and will not have any impact on real data;

[0141] 5.4 For rules that fail verification after multiple automatic corrections or fail to execute in the sandbox, an exception downgrade process will be implemented. The specific formula is as follows:

[0142] ;

[0143] in, The rules for the final output; For the alternative rules recommended by the system that have been verified in similar historical scenarios, rules that cannot be automatically corrected and have no effective alternative rules are automatically marked as "pending manual review" and a detailed error report is generated for experts to handle.

[0144] 5.5 Generate a set of executable evaluation rules and a full-process verification report. Integrate all verified rules to generate a complete verification report. The specific formula is as follows:

[0145] ;

[0146] The full-process verification report includes the verification process, error logs, corrections, and execution results.

[0147] This is the final, fully validated set of target executable evaluation rules; This is a full-process verification report; The rule correction log records the cause of the error, the correction details, and the correction time.

[0148] The aforementioned data quality assessment method based on multi-source heterogeneous materials, in step 6, assigns assessment weights to each indicator based on a unified basic weight pre-fixed by experts, combined with business scenarios and compliance requirements. It then optimizes and executes the target executable assessment rule set in parallel through rule dependency optimization to generate a standardized data quality assessment report. The specific steps are as follows:

[0149] 6.1 Experts pre-fix the basic weight configuration to construct a globally unified benchmark weight set, as shown in the following formula:

[0150] ;

[0151] in, For the first The basic weights of each standardized indicator This is a function for assigning values ​​to both qualitative and quantitative data.

[0152] 6.2 Adaptive Adjustment of Business and Compliance Weights: Based on the business scenario and compliance requirements of the current assessment task, the basic weights are adjusted to obtain the adjusted indicator weights. The specific formula is as follows:

[0153] ;

[0154] in, This is a business adjustment coefficient that assigns weights to core business metrics; the default value is 1. To adjust the compliance coefficient, the mandatory compliance indicators are weighted more heavily, with a default value of 1.

[0155] 6.3 Perform global weight normalization on the corrected indicator weights to generate the final standardized weights that can be used for comprehensive scoring. The specific formula is as follows:

[0156] ;

[0157] in, For the first The final standardized weight of each standardized indicator, with the sum of all indicator weights being 1; This represents the total number of standardized indicators.

[0158] 6.4 Perform dependency resolution on the output executable rule set, construct a rule execution dependency graph, merge repetitive computation tasks, and determine the optimal execution topology order. The specific formula is as follows:

[0159] ;

[0160] in, For rule execution dependency graph; It is a set of rule nodes, each node corresponding to an executable evaluation rule; This is a set of dependencies, where each edge represents a data dependency or computational dependency between two rules;

[0161] 6.5. Distribute the optimized evaluation task to the distributed execution cluster, execute all verified evaluation rules in batches, and generate single-rule evaluation scores. The specific formula is as follows:

[0162] ;

[0163] in, For the first The data field, the first The first evaluation rule template, the first The evaluation score for each rule corresponding to each execution engine; For final executable rules; The dataset to be evaluated;

[0164] 6.6. Based on the single-rule evaluation score, calculate the sub-indicator scores and the overall data quality score. The specific formula is as follows:

[0165] ;

[0166] ;

[0167] in, For the first The sub-evaluation scores of each standardized indicator; This is an aggregation function for single-indicator, multi-rule results; This represents the overall data quality score for the target dataset.

[0168] Step 6.7: Automatically locate the location, type, and severity of data quality issues, generate rectification suggestions, and output a standardized assessment report that meets the requirements of third-party data quality certification.

[0169] A system based on a data quality assessment method driven by multi-source heterogeneous materials includes an assessment knowledge graph construction module for constructing an assessment knowledge graph with a standardized indicator system as its core.

[0170] The multi-source heterogeneous material access module is used to receive multi-source heterogeneous evaluation input materials and supports the unified access of structured materials, semi-structured materials, unstructured materials and other auxiliary materials.

[0171] The intelligent material analysis and cross-document information fusion module is used to evaluate different types of materials in a multi-source material set and automatically extract core attribute information using differentiated intelligent analysis strategies.

[0172] The multi-dimensional standardized indicator automatic mapping module is used to automatically complete the mapping and matching of data fields and standardized indicator nodes and resolve conflicts.

[0173] The metric-driven executable evaluation rule automatic generation module is used to generate an initial executable evaluation rule set that is compatible with multiple execution engines;

[0174] The rule cross-validation and intelligent correction module is used to perform multi-dimensional cross-validation on the generated initial executable evaluation rule set and output the target executable evaluation rule set.

[0175] The weight adaptive allocation and dependency-aware parallel evaluation module is used to execute the target executable evaluation rule set;

[0176] The standardized data quality assessment report generation module is used to generate standardized data quality assessment reports that conform to third-party data quality certification specifications.

[0177] The data storage and version management module is used to store the evaluation knowledge graph, structured evaluation input, target executable evaluation rule set, evaluation results, and full-process operation logs, and supports version traceability and query.

[0178] Compared with the prior art, the beneficial effects of the present invention are:

[0179] This invention constructs an evaluation knowledge graph with a standardized indicator system at its core, then receives multi-source heterogeneous evaluation input materials and performs intelligent parsing and cross-document information fusion, eliminating information conflicts and filling information gaps. It can generate a unified, machine-understandable structured evaluation input model, which is highly efficient and less prone to information omissions and comprehension biases, thus ensuring the accuracy of indicator mapping and rule generation.

[0180] By automatically mapping data fields to corresponding standardized indicator nodes, an initial executable set of evaluation rules is generated, which enables the mapping between standardized indicators and customer data fields to be completed automatically. The evaluation rules do not need to be manually written by domain experts, thereby improving evaluation efficiency and accuracy. They can be uniformly maintained and reused to meet the needs of large-scale compliance evaluation.

[0181] By performing multi-dimensional cross-validation and intelligent correction on the initial executable evaluation rule set, the evaluation rules have a systematic closed-loop verification mechanism, which optimizes evaluation efficiency, ensures the accuracy of evaluation results, and has good traceability.

[0182] Finally, by using a unified basic weight pre-fixed by experts, and combining business and compliance requirements to allocate assessment weights, the assessment is executed in parallel through rule dependency optimization and a standardized data quality assessment report is generated. This ensures that rule dependencies are fully considered during the assessment execution, avoids a large amount of redundant calculations, and makes the assessment results of different batches comparable.

[0183] In summary, this invention automates the entire process from inputting multi-source assessment materials to outputting a standardized data quality assessment report, shortening the overall assessment cycle and improving the efficiency and accuracy of data quality assessment. Attached Figure Description

[0184] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0185] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0186] This invention provides the following: Figure 1 The data quality assessment method based on multi-source heterogeneous materials, as shown, includes the following steps:

[0187] Step 1: Construct an evaluation knowledge graph with a standardized indicator system at its core. The specific steps are as follows:

[0188] 1.1 Constructing a node set The specific formula is as follows:

[0189] ;

[0190] in, A standardized set of indicator nodes; To evaluate the set of rule template nodes; A collection of data type nodes; A collection of business scenario nodes; For the set of execution engine nodes;

[0191] The standardized indicator node set, evaluation rule template node set, data type node set, business scenario node set, and execution engine node set are constructed in the following ways:

[0192] 1.1.1 Constructing a standardized set of indicator nodes Specifically, the calculation formula for each standardized indicator node is as follows:

[0193] ;

[0194] in, For the first One standardized indicator node; For the first A globally unique identifier for each indicator; For the first The standard name of each indicator; For the first Standardized definition of each indicator; For the first Description of the calculation method for each indicator; For the first The pass / fail threshold for each indicator; For the first The compliance documents and clause numbers corresponding to each indicator; For the first The priority of each indicator ranges from 1 to 5, with higher values ​​indicating higher priority. For the first The set of data type identifiers applicable to each indicator; For the first A set of business scenario identifiers applicable to each indicator; For the first A set of evaluation rule template identifiers associated with each indicator;

[0195] 1.1.2 Constructing a set of evaluation rule template nodes Specifically, the calculation formula for each evaluation rule template node is as follows:

[0196] ;

[0197] in, For the first One evaluation rule template node; For the first A globally unique identifier for each template; For the first The name of the template; For the first The rule content of a template with parameter placeholders; For the first Each template contains a list of parameter definitions, and each parameter includes a parameter name, parameter type, and valid parameter range. For the first The syntax rules corresponding to each template; For the first Each template supports a set of execution engine identifiers; For the first Automatic validation rules corresponding to each template;

[0198] 1.1.3 Constructing a collection of data type nodes Specifically, the calculation formula for each data type node is as follows;

[0199] ;

[0200] in, For the first Each data type node; For the first A globally unique identifier for each data type; For the first The name of each data type; For the first A format specification description for each data type; For the first Description of common quality issues for each data type; For the first A set of evaluation method identifiers corresponding to each data type;

[0201] 1.1.4 Constructing a subset of business scenario nodes Specifically, the calculation formula for each business scenario node is as follows:

[0202] ;

[0203] in, For the first Each business scenario node; For the first A globally unique identifier for each business scenario; For the first The name of each business scenario; For the first A detailed description of each business scenario; For the first A set of standardized indicators that are of core concern in each scenario; For the first Description of mandatory compliance requirements in various scenarios;

[0204] 1.1.5 Constructing the Execution Engine Node Set Specifically, the calculation formula for each execution engine node is as follows:

[0205] ;

[0206] in, For the first One execution engine node; For the first A globally unique identifier for each execution engine; For the first The name of the execution engine; For the first The set of rule types supported by each engine; For the first Performance parameters of each engine, such as execution time of a single rule and maximum number of concurrent connections supported; For the first The default configuration parameters for each engine;

[0207] 1.2 Construct a set of relationships to describe the semantic, logical, and dependency relationships between nodes. The specific formula is as follows:

[0208] ;

[0209] in, This is a subset of the relationships between indicators and applicable data types. This refers to a subset of the relationships between metrics and applicable business scenarios. This is a subset of the relationships between indicators and their corresponding rule templates; This refers to a subset of the relationship between rule templates and supporting execution engines. This is a subset of the relationships between business scenarios and core metrics of interest.

[0210] The following methods are used to construct the subsets of relationships between metrics and applicable data types, metrics and applicable business scenarios, metrics and corresponding rule templates, rule templates and supporting execution engines, and business scenarios and core metrics of interest:

[0211] 1.2.1 Construct a subset of relationships between indicators and applicable data types. Each relationship represents a standardized indicator that is applicable to one or more data types. The specific formula is as follows;

[0212] ;

[0213] in, To start from standardized indicator nodes Pointer to data type node A directed edge, representing the first... The first indicator applies to the first Data types;

[0214] 1.2.2 Construct a subset of relationships between metrics and applicable business scenarios. Each relationship represents a standardized metric that is applicable to one or more business scenarios. The specific formula is as follows:

[0215] ;

[0216] in, To start from standardized indicator nodes Pointing to business scenario nodes A directed edge, representing the first... The first indicator applies to the first One business scenario;

[0217] 1.2.3 Constructing a subset of relationships between indicators and corresponding rule templates Each relation represents a standardized indicator corresponding to one or more evaluation rule templates, as shown in the following formula:

[0218] ;

[0219] in, To start from standardized indicator nodes Point to the evaluation rule template node A directed edge, representing the first... The first indicator applies to the first A rule template generates evaluation rules. For the first A set of rule template identifiers corresponding to each indicator;

[0220] 1.2.4. Relationship between rule templates and supporting execution engines (subset) Each relation represents an evaluation rule template that can be run on one or more execution engines, as shown in the following formula:

[0221] ;

[0222] in, To evaluate the rule template node Point to the execution engine node A directed edge, representing the first... The rule template supports the first one. It runs on an execution engine;

[0223] 1.2.5 Constructing a subset of relationships between business scenarios and core metrics of interest Each relation represents a core metric of interest in a business scenario, and the specific formula is as follows:

[0224] ;

[0225] in, To start from business scenario nodes Pointing to standardized indicator nodes A directed edge, representing the first... The first indicator is the Key metrics to focus on for each business scenario;

[0226] 1.3 Generate text descriptions for all nodes. Dense vector representation, and constructing vector indexes based on the approximate nearest neighbor algorithm. The specific steps are as follows:

[0227] 1.3.1. Preprocess the core text description of each node, including word segmentation, stop word removal, and normalization, to generate a standardized text representation;

[0228] 1.3.2. A pre-trained language model is used to convert the standardized text representation of each node into... The 3D dense vector representation is given by the following formula:

[0229] ;

[0230] in, For the first Each node The dimension is represented as a dense vector. For the first Standardized text representation of each node, For pre-trained language model encoders, , for 3D real vector space;

[0231] The pre-trained language model encoder is a semantic encoding model pre-trained on a large-scale general text corpus. It can map semantically similar texts to positions with similar distances in a high-dimensional vector space, and has cross-domain semantic generalization ability. It can achieve unified semantic representation of index definitions and rule description texts with different expressions without fine-tuning for specific data quality assessment scenarios, and provide a unified vector benchmark for subsequent cross-document semantic matching.

[0232] 1.3.3 Constructing a vector index using an approximate nearest neighbor algorithm , all nodes The dense vector representation is indexed to support millisecond-level semantic similarity retrieval, as shown in the following formula;

[0233] ;

[0234] in, Functions for constructing approximate nearest neighbor indices. To evaluate the total number of nodes in a knowledge graph, this vector index reduces the time complexity of semantic retrieval from linear to approximately logarithmic by partitioning the high-dimensional vector space into blocks and performing hierarchical clustering. It can achieve millisecond-level recall even with a large number of nodes, ensuring the response efficiency of subsequent automatic mapping of metrics.

[0235] 1.4. Provides CRUD (Create, Read, Update, Delete) interfaces for evaluating knowledge graphs, supporting version tracking and incremental updates based on node sets. Relation sets Vector index Obtain the evaluation knowledge graph The specific steps are as follows:

[0236] 1.4.1 Generate a unique version number for each evaluation of knowledge graph changes. Record the changes, the time of change, and the person making the change, for use in evaluating the version traceability of the knowledge graph;

[0237] 1.4.2 Supports incremental updates of the knowledge graph for evaluation. When a knowledge node or relationship is added or modified, the corresponding vector index is automatically updated. ;

[0238] 1.4.3. Provides a batch import / export interface for evaluation knowledge graphs, supporting batch import of evaluation knowledge graphs from external files, and also supporting the export of existing evaluation knowledge graphs to a standardized format, thus obtaining a complete evaluation knowledge graph capable of machine reasoning. The specific formula is as follows:

[0239] .

[0240] In step 1, an evaluation knowledge graph with a standardized indicator system at its core is constructed. This is a directed attribute graph containing a set of nodes, a set of relationships, and a vector index. It is used to uniformly store and manage all evaluation-related knowledge, thereby realizing the structured accumulation and machine-reasonable nature of the evaluation knowledge graph.

[0241] Step 2: Receive multi-source heterogeneous evaluation input materials and output a structured evaluation input model. The specific steps are as follows:

[0242] 2.1 Constructing a multi-source assessment material set The specific formula is as follows:

[0243] ;

[0244] in, A subset of structured materials; It is a subset of semi-structured materials; It is a subset of unstructured materials; A subset of other auxiliary materials;

[0245] In step 2.1, the above-mentioned material types are not limited to any input material related to data quality assessment. Any input material can be accessed and analyzed by this invention.

[0246] 2.2 For different types of materials in the multi-source assessment material set, differentiated intelligent analysis strategies are used to automatically extract core attribute information, as follows:

[0247] 2.2.1 For structured materials such as data dictionaries, database metadata, and API interface documents, directly extract the core attribute information of the data assets for structured material parsing;

[0248] Furthermore, the core attribute information includes table name, field name, field type, field description, primary key identifier, foreign key association, index information, and data length constraints;

[0249] 2.2.2. Perform semi-structured material parsing on semi-structured materials in Word, PDF, and Markdown document formats;

[0250] Specifically, it employs a large language model for text-level semantic understanding, automatically extracting information such as business rules, data flow relationships, data quality requirements, mandatory compliance clauses, and business constraints.

[0251] 2.2.3. Perform unstructured material analysis on system architecture diagrams, business process diagrams, and scanned unstructured materials;

[0252] Specifically, optical character recognition technology is used in conjunction with a multimodal large model to analyze image content and automatically extract information on system components, data flow, storage location, processing nodes, and interface call relationships.

[0253] Among them, optical character recognition technology is used to accurately extract printed and handwritten text and annotation information in images; multimodal large model is used to identify graphic elements, topological connections and business flow logic in images. The combination of the two can automatically complete the extraction of structured information from unstructured image materials without the need for manual image reading and sorting. While improving the parsing efficiency, it can effectively avoid information omissions and misunderstandings caused by manual image reading, and ensure the accuracy of information for subsequent indicator mapping and rule generation.

[0254] 2.2.4. Analyze system logs, data samples, compliance documentation, and other supporting materials.

[0255] Specifically, this includes: automatically extracting statistical features, anomaly records, and compliance qualification information from data;

[0256] In step 2.2, the above material classification method and corresponding analysis strategy are only one embodiment of the present invention. When applying the present invention, it is not limited to the above material classification method and corresponding analysis strategy. Any technical means that can extract the core information required for data quality assessment from the input materials can be applied to the present invention.

[0257] 2.3. Fragmented information extracted from different materials will be correlated, integrated, and aligned to eliminate information conflicts, fill in missing information, and generate a unified structured evaluation input model. The specific formula is as follows:

[0258] ;

[0259] in, This constitutes a list of data assets, including the name, identifier, and source information of all data assets to be evaluated; For data structure details, including field definitions, types, and constraint information for each data asset; This includes business background information, such as business scenario descriptions, business rules, and data flow relationships. This is a list of compliance requirements, including all applicable compliance clauses, mandatory requirements, and qualification standards.

[0260] In step 2.3, the fused structured evaluation input model needs to undergo multi-dimensional quality verification to check the completeness, consistency, and accuracy of its information. Only then can a complete, quality-verified structured evaluation input model be output, which will serve as a unified input for the automatic mapping of standardized indicators in step 3. The specific multi-dimensional quality verification process is as follows:

[0261] Integrity verification: Checks for missing information in core fields, and marks and prompts for any missing information;

[0262] Consistency check: Check whether there are conflicts in the same information extracted from different materials, mark the conflicting information and provide correction suggestions;

[0263] Accuracy verification: Check whether the extracted information conforms to logical norms and industry common sense, and mark obviously erroneous information;

[0264] Furthermore, in step 2.3, cross-document multi-source information fusion is achieved through entity alignment, relation completion, and conflict resolution, which can map fragmented information from different sources into a unified semantic space;

[0265] In step 2, a unified set of multi-source evaluation materials is established. Differentiated intelligent parsing strategies are adopted for different types of materials to automatically extract structured information. Cross-document information fusion technology is used to eliminate information conflicts and fill in missing information, ultimately generating a unified, machine-understandable structured evaluation input model. This avoids the situation where the evaluation input materials provided by customers are diverse in format, scattered in source, and fragmented in information. Domain experts need to spend a lot of time reading, sorting, and extracting information, which is inefficient and prone to information omissions and misunderstandings. This ensures the accuracy of subsequent indicator mapping and rule generation.

[0266] Step 3: Based on the multi-dimensional association between the structured evaluation input model and the evaluation knowledge graph, automatically map data fields to corresponding standardized indicator nodes. The specific steps are as follows:

[0267] 3.1 For each data field, calculate its matching degree with all standardized indicators in the evaluation knowledge graph across four dimensions. The specific formula is as follows:

[0268] ;

[0269] in, For the first The data field and the first A set of multi-dimensional matching degrees for standardized indicators; For semantic similarity, calculate the textual semantic similarity between the data field description and the standardized metric definition; For type similarity, standardized metrics that support the type of data field are matched based on the data field type. For business similarity, based on business background information Match standardized metrics applicable to the corresponding business scenarios; For compliance similarity, based on the compliance requirements list Match standardized metrics that meet mandatory compliance requirements;

[0270] 3.2 Assign preset weights to the matching degree of the four dimensions, and calculate the final matching degree score of each data field with each standardized indicator using a weighted average. The specific formula is as follows:

[0271] ;

[0272] in, For the first The data field and the first The final matching score of each standardized indicator; For semantic weight coefficients, Weight coefficients for types, Weighting coefficients for business operations For the compliance dimension, the weighting coefficients are set, and they satisfy the following conditions: ;

[0273] Furthermore, the weighting coefficients can be dynamically adjusted based on business scenarios and compliance requirements;

[0274] 3.3. Score all standardized metrics for each data field based on the final matching degree. Sort the data from high to low, select the top N indicators with the highest matching degree as the recommendation results, and generate an initial field-indicator mapping table;

[0275] The value of N can be flexibly set according to the evaluation accuracy requirements and the workload of expert review;

[0276] 3.4 When multiple standardized metrics are mapped to the same data field, the optimal matching metric is automatically selected based on metric priority, business scenario importance, and compliance requirements. The specific formula is as follows:

[0277] ;

[0278] in, For the first The optimal matching metric for each data field; For the first The overall priority of each standardized indicator is calculated by weighting the inherent priority of the indicator, the priority of the business scenario, and the priority of compliance.

[0279] Furthermore, for mandatory compliance indicators that have a veto requirement, their priority is automatically set to the highest.

[0280] 3.5 Generate detailed matching reasons for each initial field-indicator mapping result, and output the target field-indicator mapping result set;

[0281] The matching reasons include: the matching score of each dimension, the matching basis and the reasons for priority determination, which are presented in a structured form for expert review and reference, thereby ensuring the interpretability of the mapping results and supporting expert review.

[0282] The target field-indicator mapping result set contains the correspondence between data fields and standardized indicators, matching scores, matching reasons, and priority determination results, which serve as the input basis for the automatic generation of executable evaluation rules in step 4.

[0283] In step 3, based on the evaluation of the knowledge graph, the matching degree between data fields and standardized indicators is calculated from four dimensions: semantics, type, business, and compliance. The final matching degree score is obtained through weighted fusion, the optimal matching indicator is automatically recommended, mapping conflicts are resolved, and an explainable matching reason is generated. This realizes the automation and standardization of indicator mapping, so that the mapping between standardized indicators and customer data fields can be completed automatically, with high efficiency and stable accuracy. At the same time, it avoids the situation where different experts have different understandings of indicator definitions, which leads to poor consistency of evaluation results. This meets the legal requirements of repeatability and traceability of quality assessment results.

[0284] Step 4: Based on the evaluation rule templates corresponding to the standardized indicator nodes obtained from the mapping, automatically generate an initial executable evaluation rule set that is compatible with multiple execution engines. The specific steps are as follows:

[0285] 4.1 Based on the optimal matching index The evaluation rule templates associated with this indicator are obtained from the evaluation knowledge graph. The specific formula is as follows:

[0286] ;

[0287] in, For the first A set of evaluation rule templates corresponding to each data field. For the first An evaluation rule template, and The calculation formula corresponds to; For the first A globally unique identifier for each evaluation rule template. For the first The set of rule template identifiers corresponding to the best matching index for each data field, and the calculation The definitions in the formula are consistent.

[0288] 4.2 Using a structured evaluation input model, automatically infer and populate all parameters in the evaluation rule template to generate parameterized rule instances. The specific formula is as follows:

[0289] ;

[0290] in, For the first The first data field corresponding to the A set of parameters for each evaluation rule template; This is a set of data location parameters, including table name, field name, and database connection information; This is a set of business logic parameters, including qualification thresholds, judgment conditions, and filtering rules. This is a set of execution parameters, including the execution engine identifier, timeout, and number of retries.

[0291] 4.3. Based on the client's specific compliance requirements and business needs, automatically adjust the rules' parameters, logic, and priorities to generate customized rule parameters adapted to the client's scenario. The specific formula is as follows:

[0292] ;

[0293] in, For the adjusted set of customized parameters, Adjust the function for the rules;

[0294] For business background information, For the compliance requirements list, for clauses with mandatory compliance requirements, the corresponding rules are automatically set to the highest priority;

[0295] 4.4. Based on the execution engine types supported by the evaluation rule template, convert the rule instance with customized rule parameters into the native format supported by the target execution engine. The specific formula is as follows:

[0296] ;

[0297] in, For the first The first data field, the first The first evaluation rule template, the first The native executable rules corresponding to each execution engine; This is a format conversion function; For the first Each execution engine, and the execution engine node correspond;

[0298] The execution engines supported by this invention include: Dingo, Great Expectations, Spark SQL, and Python;

[0299] 4.5 Generate complete metadata information for each executable rule for subsequent rule management, traceability, and auditing. Automatically generate a metadata set using the following formula:

[0300] ;

[0301] in, For the first The first data field, the first The first evaluation rule template, the first Metadata collection corresponding to each execution engine This serves as a globally unique identifier for the rule. The standardized indicator identifier corresponding to the rule. For the rule generation time, This is the identifier of the execution engine corresponding to the rule;

[0302] 4.6 Integrate all generated executable rules and their metadata to obtain a complete initial executable evaluation rule set, as shown in the following formula:

[0303] ;

[0304] in, This is the initial set of executable evaluation rules; This represents the total number of data fields to be evaluated. For the first A set of evaluation rule templates corresponding to each data field;

[0305] In step 4, based on the step field-metric mapping results, the corresponding assessment rule template is automatically matched from the assessment knowledge graph. The structured information obtained from the parsing is used to automatically fill in the template parameters. Custom adjustments are made according to customer compliance requirements and business needs. The rules are automatically converted into native rule formats supported by multiple execution engines and complete rule metadata is generated. This realizes the automated generation and standardized management of assessment rules, so that assessment rules do not need to be manually written by domain experts, improving the efficiency and accuracy of assessment. Moreover, the rule style and logic are consistent, which can be uniformly maintained and reused, and can meet the needs of large-scale compliance assessment.

[0306] Step 5: Perform multi-dimensional cross-validation and intelligent correction on the initial executable evaluation rule set to obtain the target executable evaluation rule set. The specific steps are as follows:

[0307] 5.1 Perform multi-dimensional automatic validation on each executable evaluation rule to generate a rule validation result set. The specific formula is as follows:

[0308] ;

[0309] in, For the first The first data field, the first The first evaluation rule template, the first The set of verification results for each rule corresponding to an execution engine; To verify the grammatical validity, check whether the rules conform to the grammatical specifications of the target execution engine; To verify the semantic correctness of the results, check whether the rule logic is consistent with the definition of the corresponding standardized indicator. To verify the existence of data, check whether the tables, fields, and database resources referenced in the rules actually exist and are accessible. To verify the logical consistency of the result, check whether there are logical conflicts, duplicate calculations, or contradictory results between this rule and other rules; To verify the compliance results, check whether the rules meet the compliance clauses and mandatory requirements extracted in step 2;

[0310] 5.2 Intelligent automatic error correction: For rules that fail validation, the system automatically identifies the cause of the error and generates a correction plan. It automatically corrects common errors using the following formula:

[0311] ;

[0312] in, The revised executable rules; This is a smart correction function; The evaluation knowledge graph in step 1 is used to provide semantic and logical basis for error correction, and the errors that can be automatically corrected include syntax errors, parameter value errors, data reference path errors, and simple logic errors.

[0313] 5.3 Deploy the revised rules to an isolated sandbox environment for testing to verify the stability of the rules and the reasonableness of the results. The specific formula is as follows:

[0314] ;

[0315] in, This is a function executed in the sandbox that returns two results: This indicates the sandbox execution status, with values ​​of "passed" or "failed". The results are for sandbox execution; the sandbox environment provides a simulated operating environment consistent with the production environment and will not have any impact on real data;

[0316] 5.4 For rules that fail verification after multiple automatic corrections or fail to execute in the sandbox, an exception downgrade process will be implemented. The specific formula is as follows:

[0317] ;

[0318] in, The rules for the final output; For the alternative rules recommended by the system that have been verified in similar historical scenarios, rules that cannot be automatically corrected and have no effective alternative rules are automatically marked as "pending manual review" and a detailed error report is generated for experts to handle.

[0319] 5.5 Generate a set of executable evaluation rules and a full-process verification report. Integrate all verified rules to generate a complete verification report. The specific formula is as follows:

[0320] ;

[0321] The full-process verification report includes the verification process, error logs, corrections, and execution results.

[0322] This is the final, fully validated set of target executable evaluation rules; This is a full-process verification report; The rule correction log records the cause of the error, the correction details, and the correction time.

[0323] In step 5, based on the evaluation knowledge graph and the initial executable evaluation rule set, rule errors are automatically located and correction suggestions are generated. The effectiveness of the rules is verified by pre-execution through sandbox isolation. For rules that cannot be automatically corrected, anomaly degradation processing is performed. Finally, the target executable evaluation rule set and the full-process verification record are output, which enable the evaluation rules to have a systematic closed-loop verification mechanism. There are no problems such as syntax errors, logical conflicts, invalid data references, or lack of compliance. This further optimizes the evaluation efficiency, ensures the accuracy of the evaluation results, and meets the legal requirements for the accuracy and traceability of the results in quality evaluation.

[0324] Meanwhile, the aforementioned verification dimension division, intelligent correction method, optional sandbox execution mechanism, and anomaly degradation strategy are merely preferred embodiments of the present invention. The core protection scope of the present invention covers all rule verification schemes that include five-fold cross-validation and intelligent correction. Regardless of whether a sandbox pre-execution step is included, any technical solution that can achieve automatic verification and correction of evaluation rules is an equivalent substitution of the present invention and falls within the protection scope of the present invention.

[0325] Step 6: Based on the unified basic weights pre-fixed by experts, and in combination with business scenarios and compliance requirements, assign evaluation weights to each indicator. Optimize the parallel execution of the target executable evaluation rule set through rule dependencies to generate a standardized data quality evaluation report. The specific steps are as follows:

[0326] 6.1 Experts pre-fix the basic weight configuration to construct a globally unified benchmark weight set, as shown in the following formula:

[0327] ;

[0328] in, For the first The basic weights of each standardized indicator As a qualitative and quantitative assignment function, the basic weights of all indicators are fixed in advance and not changed arbitrarily, so as to ensure the consistency of the benchmark and the comparability of the results of the evaluation task;

[0329] 6.2 Adaptive Adjustment of Business and Compliance Weights: Based on the business scenario and compliance requirements of the current assessment task, the basic weights are adjusted to obtain the adjusted indicator weights. The specific formula is as follows:

[0330] ;

[0331] in, This is a business adjustment coefficient that assigns weights to core business metrics; the default value is 1. To adjust the compliance coefficients, the mandatory compliance indicators are weighted more heavily, with a default value of 1. This step only adjusts the proportion of the basic weights; the basic weights themselves remain unchanged and always serve as a unified benchmark for all assessment tasks.

[0332] 6.3 Perform global weight normalization on the corrected indicator weights to generate the final standardized weights that can be used for comprehensive scoring. The specific formula is as follows:

[0333] ;

[0334] in, For the first The final standardized weight of each standardized indicator, with the sum of all indicator weights being 1; This represents the total number of standardized indicators.

[0335] 6.4 Perform dependency resolution on the output executable rule set, construct a rule execution dependency graph, merge repetitive computation tasks, and determine the optimal execution topology order. The specific formula is as follows:

[0336] ;

[0337] in, For rule execution dependency graph; It is a set of rule nodes, each node corresponding to an executable evaluation rule; This is a set of dependencies, where each edge represents a data dependency or computational dependency between two rules;

[0338] 6.5. Distribute the optimized evaluation task to the distributed execution cluster, execute all verified evaluation rules in batches, and generate single-rule evaluation scores. The specific formula is as follows:

[0339] ;

[0340] in, For the first The first data field, the first The first evaluation rule template, the first The evaluation score for each rule corresponding to each execution engine; For final executable rules; The dataset to be evaluated;

[0341] 6.6. Based on the single-rule evaluation score, calculate the sub-indicator scores and the overall data quality score. The specific formula is as follows:

[0342] ;

[0343] ;

[0344] in, For the first The sub-evaluation scores of each standardized indicator; This is an aggregation function for single-indicator, multi-rule results; This represents the overall data quality score for the target dataset.

[0345] Step 6.7: Automatically locate the location, type, and severity of data quality issues, generate rectification suggestions, and output a standardized assessment report that meets the requirements of third-party data quality certification;

[0346] In step 6, a three-tiered standardized weighting system is constructed, consisting of fixed base weights, business scenario corrections, and mandatory compliance corrections. This ensures that the evaluation results are consistent and comparable. The execution order is optimized through rule dependency analysis, and the evaluation rules are executed in a distributed and parallel manner. The multi-dimensional evaluation results are weighted and aggregated, and problems are automatically located and standardized data quality reports are generated. This allows the evaluation to fully consider rule dependencies, avoid a large amount of repetitive calculations, improve evaluation efficiency, ensure consistency of indicator weights, and make the evaluation results of different batches comparable. At the same time, the data quality report format is unified, saving time and effort in writing and providing strong traceability.

[0347] Those skilled in the art should understand that the above-mentioned weight adjustment method, execution optimization method and report output format are only preferred embodiments of the present invention. The present invention is not limited to the above specific implementation methods. Any technical solution that is based on a pre-fixed unified benchmark weight, combines business scenarios and compliance requirements to adapt the weight, and achieves efficient evaluation through rule dependency optimization is an equivalent substitution of the present invention and falls within the protection scope of the present invention.

[0348] A system based on a data quality assessment method driven by multi-source heterogeneous materials includes:

[0349] The evaluation knowledge graph construction module is used to build an evaluation knowledge graph with a standardized indicator system as its core, so as to realize the structured accumulation, semantic retrieval and machine reasoning of evaluation knowledge;

[0350] The multi-source heterogeneous material access module is used to receive multi-source heterogeneous evaluation input materials and supports the unified access of structured materials, semi-structured materials, unstructured materials and other auxiliary materials.

[0351] The intelligent material analysis and cross-document information fusion module is used to evaluate different types of materials in a multi-source material set. It automatically extracts core attribute information using differentiated intelligent analysis strategies, performs cross-document entity alignment, conflict resolution and information completion, and generates a unified structured evaluation input model.

[0352] The multi-dimensional standardized indicator automatic mapping module is used to automatically complete the mapping and matching of data fields and standardized indicator nodes and resolve conflicts based on the multi-dimensional association matching between the structured evaluation input model and the evaluation knowledge graph.

[0353] The indicator-driven executable evaluation rule automatic generation module is used to automatically fill in parameters based on the evaluation rule template corresponding to the standardized indicator nodes obtained by mapping, combined with structured evaluation input information, to generate an initial executable evaluation rule set that is compatible with multiple execution engines.

[0354] The rule cross-validation and intelligent correction module is used to perform multi-dimensional cross-validation on the generated initial executable evaluation rule set, automatically locate and correct rule errors, and output the target executable evaluation rule set.

[0355] The weight adaptive allocation and dependency-aware parallel evaluation module is used to allocate evaluation weights to each indicator based on the unified basic weights pre-fixed by experts, combined with business scenarios and compliance requirements, to parse rule dependencies and optimize the execution order, and to execute the target executable evaluation rule set in parallel.

[0356] The standardized data quality assessment report generation module is used to generate standardized data quality assessment reports that conform to third-party data quality certification specifications.

[0357] The data storage and version management module is used to store the evaluation knowledge graph, structured evaluation input, target executable evaluation rule set, evaluation results, and full-process operation logs, and supports version traceability and query.

[0358] Verification experiment:

[0359] This experiment focuses on the quality evaluation of business datasets. The evaluation dimensions include completeness, accuracy, consistency, timeliness, uniqueness, effectiveness, as well as sub-indicators suitable for evaluating high-quality datasets, such as content cleanliness, annotation standardization, security standardization, structural completeness, content diversity, and content authenticity, to complete the dataset quality assessment.

[0360] This embodiment is an exemplary test used to illustrate the implementation process and technical effects of the technical solution of the present invention. The performance data therein are test results under specific test environments and test samples, and do not represent the performance of the present invention in all scenarios, nor do they constitute a limitation on the scope of protection of the present invention.

[0361] I. Experimental Environment and Test Samples:

[0362] 1. Hardware and software environment:

[0363] The hardware deployment consists of two Intel Xeon 6330 servers, each with 32 cores, 64 threads, and 128GB of memory, and equipped with NVIDIA A10 graphics cards for semantic vector computation and multimodal content parsing; a three-node Spark distributed computing cluster is built, with nodes connected via 10 Gigabit Ethernet.

[0364] The software deploys a semantic parsing module, a vector retrieval module, an evaluation knowledge graph module, and a dual-kernel rule execution engine; the indicator scoring system unifies and integrates general data foundation quality dimensions with high-quality dataset-specific evaluation dimensions, supporting simultaneous quality verification of multiple types of samples such as structured data, text, and images.

[0365] 2. Test Samples:

[0366] A comprehensive dataset containing structured business records, text annotation samples, and image annotation samples was selected as the test object, along with various heterogeneous evaluation materials: a structured data dictionary, business rule documents, annotation specification documents, and system architecture diagrams. The evaluation covers a complete quality indicator system, simultaneously verifying specific indicators such as content cleanliness, annotation standardization, content authenticity, security compliance, and sample diversity in addition to the traditional six basic dimensions.

[0367] II. Specific Implementation Process:

[0368] Step 1: Evaluate the pre-constructed knowledge graph (Step 1):

[0369] A unified and standardized evaluation knowledge graph is constructed. The graph's indicator nodes include traditional indicators such as completeness, accuracy, consistency, timeliness, uniqueness, and effectiveness, as well as high-quality dataset evaluation indicators such as content cleanliness, annotation standardization, security standardization, content authenticity, content diversity, and structural completeness. The graph is equipped with standardized verification rule templates, data type nodes, business scenario nodes, and execution engine nodes corresponding to various indicators, and defines the relationships such as indicator affiliation, field matching, and scenario adaptation.

[0370] A 768-dimensional dense semantic vector is generated for all indicator descriptions, judgment constraints, and business rule texts within the graph. An index is built based on the vector retrieval engine, and the graph versioning initialization is completed, forming an integrated evaluation knowledge base covering general data and high-quality labeled datasets.

[0371] Step 2: Multi-source heterogeneous material access and intelligent analysis fusion (Steps 2-3):

[0372] All heterogeneous evaluation data are integrated, and differentiated parsing strategies are adopted according to data type: For structured data, complete metadata such as field names, data types, value range constraints, primary keys, foreign keys, and metadata descriptions are extracted; for semi-structured rules and annotation specification documents, semantic extraction is performed by chapter through a large model to extract various indicator judgment standards, cleanliness judgment thresholds, and annotation compliance constraints; for unstructured data such as drawings, data flow and sample storage structure information are parsed using OCR combined with a multimodal model.

[0373] By unifying field definitions and indicator judgment criteria through cross-document entity alignment and conflict resolution algorithms, it automatically corrects field definition conflicts, fills in missing evaluation constraints, and outputs a unified and standardized structured evaluation input model.

[0374] Step 3: Automatic mapping of standardized indicators (Step 4):

[0375] For all fields and samples in the dataset, the system synchronously matches basic dimensions within the complete indicator system with extended dimensions from high-quality datasets. The matching process integrates four scores: semantic similarity, data type matching, business scenario adaptability, and indicator compliance. These scores are weighted at a ratio of 0.4:0.3:0.15:0.15 to obtain the total matching score. The standardized indicator with the highest matching score is selected as the optimal mapping result, and multi-indicator matching conflicts are automatically resolved. In this embodiment, the automatic mapping accuracy between fields, samples, and standard indicators reaches approximately 93.2%.

[0376] Step 4: Automatic generation of validation rules and multi-level cross-validation (Step 5):

[0377] Based on the mapping results of fields, samples and standardized indicators, the system matches the corresponding rule templates from the evaluation knowledge graph, automatically fills in parameters such as data tables, sample paths, field names, thresholds for various indicators, and cleanliness judgment standards, and generates executable verification rules adapted to dual execution engines in batches.

[0378] A five-layer progressive cross-validation mechanism is implemented: syntax validation corrects format and function call errors; data existence validation corrects misaligned table and field references; logical consistency validation merges duplicate and redundant rules; indicator dimension classification validation corrects indicator mismatch issues; and rule compliance validation corrects judgment logic that does not conform to general data quality standards. After multi-layer validation is completed, it can usually be directly deployed and executed without manual modification, with an automatic pass rate of approximately 98.4% for the first-time validation.

[0379] Step 5: Hierarchical weight allocation and dependency-aware parallel evaluation (Step 6):

[0380] Based on industry-standard basic weights, a hierarchical weight allocation logic is set up: the basic layer configures fixed basic weights for each quality dimension; the scenario adjustment layer is designed for high-quality dataset evaluation scenarios, and improves the weights of extended indicators such as content cleanliness, annotation standardization, and content authenticity. All weights are processed through global normalization to obtain the final score weight for each single indicator.

[0381] The system analyzes the data reading and computation dependencies of all validation rules, constructs a directed graph of rule execution dependencies, identifies and merges nodes with repeated readings and computations, and generates an acyclic optimal execution topology. The optimized evaluation task is then distributed to a distributed cluster for parallel computation, enabling batch processing of massive amounts of structured data and multimodal labeled samples across all dimensions for quality validation, outputting independent evaluation scores for single rules and single quality dimensions.

[0382] Hierarchical results aggregation and standardized evaluation report generation:

[0383] The evaluation results are aggregated and analyzed in a hierarchical manner according to a complete indicator system: the bottom layer statistics include the number of violations and the percentage of violations for each rule;

[0384] The middle layer summarizes the weighted scores of all basic dimensions and extended dimensions of high-quality datasets;

[0385] The overall quality score of the top-level output dataset;

[0386] The system automatically identifies multiple types of data quality issues, classifies them into risk levels as severe, moderate, and minor, and matches corresponding standardized rectification suggestions. Finally, it outputs a unified standardized assessment report, which includes an assessment overview, scores for each dimension, a detailed list of issues, graded rectification suggestions, and overall quality judgment. It can cover the quality assessment needs of two types of objects: general business data and high-quality labeled datasets.

[0387] III. Comparison and Verification of Implementation Results:

[0388] 1. Comparison benchmarks and evaluation indicators:

[0389] Three mainstream evaluation schemes were selected from the same dataset as comparison benchmarks, and the core evaluation indicators are defined as follows:

[0390] Indicator mapping accuracy: the percentage of overlap between the automatic field, sample and standard indicator mapping results and the expert manual annotation standards;

[0391] Rule effectiveness rate: The percentage of verification rules that can be directly issued and executed without manual modification to the total number of generated rules;

[0392] Consistency of assessment results: The degree of agreement between the automated assessment conclusions and the conclusions of a joint human review by two senior industry experts;

[0393] Single task processing time: The total time spent on the entire process from heterogeneous material access to the complete generation of a standardized evaluation report.

[0394] 2. Comparison of evaluation methods:

[0395] The entire process of expert manual evaluation involves senior data quality experts manually reading all heterogeneous specification documents, manually matching field quality indicators one by one, manually writing verification rules, manually verifying all sample data offline, and manually summarizing and generating evaluation reports. The entire process is conducted without the assistance of automated tools and serves as the gold standard for experiments.

[0396] Commercial semi-automated data quality tools: These tools only have built-in fixed structured field validation templates and only support direct import of data tables. They do not have the ability to automatically read business specifications, automatically map indicators, or parse unstructured drawings / specification documents. Rules need to be manually written and entered one by one.

[0397] For a single evaluation of a general large model: there is no pre-built solidified quality indicator knowledge base, and it is necessary to upload heterogeneous materials one by one and call the large model prompt words in a loop for each field to complete indicator matching and rule generation; for multimodal labeled samples, there is no dedicated fixed verification logic for cleanliness verification, and prompt words need to be repeatedly constructed and model output deviations need to be manually reviewed for each type of verification requirement.

[0398] The integrated evaluation scheme of this invention: relying on a pre-built unified evaluation knowledge graph, it completes the entire chain of automated processing in one go, including multi-source heterogeneous material analysis, automatic field-indicator mapping, batch rule generation, distributed parallel data verification, and automatic report output.

[0399] Table 1 Comparison of the effects of different evaluation methods:

[0400]

[0401] The above data represents typical results from this test sample. Actual performance may vary depending on the dataset size, material complexity, and hardware configuration.

[0402] 3. Comparison Results and Analysis of Innovative Advantages:

[0403] As shown in Table 1, the present invention significantly outperforms general large models and commercial semi-automated tools in the three core indicators of indicator mapping accuracy, rule effectiveness, and evaluation result consistency. The evaluation effect is close to the standard of expert manual review, and the processing time of the whole process is reduced by more than 85% compared with manual evaluation.

[0404] Existing conventional data quality assessment schemes only cover the traditional six basic dimensions, lacking key evaluation dimensions such as content cleanliness, annotation standardization, and sample authenticity adapted to high-quality datasets, and thus cannot achieve integrated, multi-dimensional assessment; general-purpose large models lack a solidified standardized indicator knowledge base, indicator matching is highly random, and the verification of extended indicators requires a lot of manual intervention; commercial tools are only adapted to simple structured data tables and lack specialized automatic verification capabilities such as multimodal sample cleanliness.

[0405] This invention constructs an integrated unified indicator knowledge graph, which integrates traditional basic quality dimensions with the extended evaluation dimensions required for high-quality datasets into a complete system. Relying on intelligent analysis of heterogeneous materials, automatic matching of four-dimensional indicators, multi-layer rule cross-validation, and distributed parallel evaluation, it achieves end-to-end automated processing. A single architecture supports full-dimensional quality evaluation of both ordinary business data and high-quality labeled datasets, overcoming the shortcomings of existing technologies such as incomplete indicator coverage, low automation, inconsistent evaluation standards, and limited applicable scenarios from the bottom up.

[0406] In summary, this invention constructs a unified evaluation knowledge graph that integrates traditional basic quality dimensions with extended evaluation dimensions of high-quality datasets. It receives multi-source heterogeneous evaluation materials and performs intelligent parsing and cross-document information fusion based on their types. Based on multi-dimensional matching scores, it automatically maps data fields and samples to a full set of standardized indicators. It batch-generates executable verification rules and completes intelligent rule correction through multi-layer cross-validation. It achieves differentiated weighted scoring for each dimension through a hierarchical weight allocation mechanism, and completes distributed parallel batch evaluation by combining rule-dependent topology optimization. It then aggregates evaluation results hierarchically and outputs a unified standardized quality evaluation report. Thus, through an integrated indicator knowledge base, it enables general data and high-quality datasets to share a single evaluation framework, fully covering extended evaluation dimensions such as content cleanliness and annotation standards. Relying on solidified standardized judgment logic, it significantly improves the stability and accuracy of indicator matching and rule generation, shortens the entire evaluation cycle, and allows evaluation conclusions to be directly used for data governance rectification and high-quality dataset quality verification, resulting in a wider range of applications and more comprehensive indicator coverage.

[0407] This invention can be widely applied to various scenarios such as large model pre-training and fine-tuning corpus data quality assessment, high-quality dataset construction and quality verification, generative AI output content quality assessment, data asset table entry quality evaluation, data transaction target quality assessment, financial risk control multimodal data quality audit, medical imaging and electronic medical record quality detection, e-commerce content data quality screening, government information disclosure multimodal content quality verification, online education resource quality assessment, autonomous driving perception data quality detection, and industrial defect detection data quality assessment.

[0408] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used above are only some embodiments described in this invention. Obviously, those skilled in the art can obtain other drawings based on these drawings.

[0409] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A data quality assessment method based on multi-source heterogeneous materials, characterized in that: Includes the following steps: Step 1: Construct an evaluation knowledge graph with a standardized indicator system at its core; Step 2: Receive multi-source heterogeneous evaluation input materials and output a structured evaluation input model; Step 3: Based on the multi-dimensional association between the structured evaluation input model and the evaluation knowledge graph, automatically map data fields to the corresponding standardized indicator nodes; Step 4: Based on the evaluation rule templates corresponding to the standardized indicator nodes obtained from the mapping, automatically generate an initial executable evaluation rule set that is compatible with multiple execution engines; Step 5: Perform multi-dimensional cross-validation and intelligent correction on the initial executable evaluation rule set to obtain the target executable evaluation rule set; Step 6: Based on the unified basic weights pre-fixed by experts, assign evaluation weights to each indicator in combination with business scenarios and compliance requirements, optimize the parallel execution of the target executable evaluation rule set through rule dependency optimization, and generate a standardized data quality evaluation report.

2. The data quality assessment method based on multi-source heterogeneous materials as described in claim 1, characterized in that: In step 1, an evaluation knowledge graph centered on a standardized indicator system is constructed. The specific steps are as follows: 1.1 Constructing a node set ; 1.2 Construct a relation set to describe the semantic, logical, and dependency relationships between nodes. ; 1.3 Generate text descriptions for all nodes. Dense vector representation, and constructing vector indexes based on the approximate nearest neighbor algorithm. ; 1.

4. Provides CRUD (Create, Read, Update, Delete) interfaces for evaluating knowledge graphs, supporting version tracking and incremental updates based on node sets. Relation sets Vector index Obtain the evaluation knowledge graph .

3. The data quality assessment method based on multi-source heterogeneous materials as described in claim 2, characterized in that: In step 1.1, a node set is constructed. The specific formula is as follows: ； in, A standardized set of indicator nodes; To evaluate the set of rule template nodes; A collection of data type nodes; A collection of business scenario nodes; For the set of execution engine nodes; The standardized indicator node set, evaluation rule template node set, data type node set, business scenario node set, and execution engine node set are constructed as follows: 1.1.1 Constructing a standardized set of indicator nodes Specifically, the calculation formula for each standardized indicator node is as follows: ； in, For the first One standardized indicator node; For the first A globally unique identifier for each indicator; For the first The standard name of each indicator; For the first Standardized definition of each indicator; For the first Description of the calculation method for each indicator; For the first The pass / fail threshold for each indicator; For the first The compliance documents and clause numbers corresponding to each indicator; For the first Priority of each indicator; For the first The set of data type identifiers applicable to each indicator; For the first A set of business scenario identifiers applicable to each indicator; For the first A set of evaluation rule template identifiers associated with each indicator; 1.1.2 Constructing a set of evaluation rule template nodes Specifically, the calculation formula for each evaluation rule template node is as follows: ； in, For the first One evaluation rule template node; For the first A globally unique identifier for each template; For the first The name of the template; For the first The rule content of a template with parameter placeholders; For the first Each template contains a list of parameter definitions, and each parameter includes a parameter name, parameter type, and valid parameter range. For the first The syntax rules corresponding to each template; For the first The set of execution engine identifiers supported by each template; For the first Automatic validation rules corresponding to each template; 1.1.3 Constructing a collection of data type nodes Specifically, the calculation formula for each data type node is as follows; ； in, For the first Each data type node; For the first A globally unique identifier for each data type; For the first The name of each data type; For the first A format specification description for each data type; For the first Description of common quality issues for each data type; For the first A set of evaluation method identifiers corresponding to each data type; 1.1.4 Constructing a subset of business scenario nodes Specifically, the calculation formula for each business scenario node is as follows: ； in, For the first Each business scenario node; For the first A globally unique identifier for each business scenario; For the first The name of each business scenario; For the first A detailed description of each business scenario; For the first A set of standardized indicators that are of core concern in each scenario; For the first Description of mandatory compliance requirements in various scenarios; 1.1.5 Constructing the Execution Engine Node Set Specifically, the calculation formula for each execution engine node is as follows: ； in, For the first One execution engine node; For the first A globally unique identifier for each execution engine; For the first The name of the execution engine; For the first The set of rule types supported by each engine; For the first Performance parameters of each engine; For the first The default configuration parameters for each engine; In step 1.2, a set of relationships is constructed to describe the semantic associations, logical associations, and dependencies between the nodes. The specific formula is as follows: ； in, This is a subset of the relationships between indicators and applicable data types. This refers to a subset of the relationships between metrics and applicable business scenarios. This is a subset of the relationships between indicators and their corresponding rule templates; This refers to a subset of the relationship between rule templates and supporting execution engines. This is a subset of the relationships between business scenarios and core metrics of interest. The following methods are used to construct the subsets of relationships between metrics and applicable data types, metrics and applicable business scenarios, metrics and corresponding rule templates, rule templates and supporting execution engines, and business scenarios and core metrics of interest: 1.2.1 Construct a subset of relationships between indicators and applicable data types. Each relationship represents a standardized indicator that is applicable to one or more data types. The specific formula is as follows; ； in, To start from standardized indicator nodes Pointer to data type node A directed edge, representing the first... The first indicator applies to the first Data types; 1.2.2 Construct a subset of relationships between metrics and applicable business scenarios. Each relationship represents a standardized metric that is applicable to one or more business scenarios. The specific formula is as follows: ； in, To start from standardized indicator nodes Pointing to business scenario nodes A directed edge, representing the first... The first indicator applies to the first One business scenario; 1.2.3 Constructing a subset of relationships between indicators and corresponding rule templates Each relation represents a standardized indicator corresponding to one or more evaluation rule templates, as shown in the following formula: ； in, To start from standardized indicator nodes Point to the evaluation rule template node A directed edge, representing the first... The first indicator applies to the first A rule template generates evaluation rules. For the first A set of rule template identifiers corresponding to each indicator; 1.2.

4. Relationship between rule templates and supporting execution engines (subset) Each relation represents an evaluation rule template that can be run on one or more execution engines, as shown in the following formula: ； in, To evaluate the rule template node Point to the execution engine node A directed edge, representing the first... The rule template supports the first one. It runs on an execution engine; 1.2.5 Constructing a subset of relationships between business scenarios and core metrics of interest Each relation represents a core metric of interest in a business scenario, and the specific formula is as follows: ； in, To start from business scenario nodes Pointing to standardized indicator nodes A directed edge, representing the first... The first indicator is the Key metrics to focus on for each business scenario; In step 1.3, text descriptions are generated for all nodes. Dense vector representation, and constructing vector indexes based on the approximate nearest neighbor algorithm. The specific steps are as follows: 1.3.

1. Preprocess the core text description of each node, including word segmentation, stop word removal, and normalization, to generate a standardized text representation; 1.3.

2. A pre-trained language model is used to convert the standardized text representation of each node into... The 3D dense vector representation is given by the following formula: ； in, For the first Each node The dimension is represented as a dense vector. For the first Standardized text representation of each node, For pre-trained language model encoders, , for 3D real vector space; 1.3.3 Constructing a vector index using an approximate nearest neighbor algorithm , all nodes The dense vector representation is indexed to support millisecond-level semantic similarity retrieval, as shown in the following formula; ； in, Functions for constructing approximate nearest neighbor indices. To assess the total number of nodes in a knowledge graph; In step 1.4, CRUD (Create, Read, Update, Delete) interfaces for evaluating the knowledge graph are provided, supporting version tracking and incremental updates based on node sets. Relation sets Vector index Obtain the evaluation knowledge graph The specific steps are as follows: 1.4.1 Generate a unique version number for each evaluation of knowledge graph changes. Record the changes, the time of change, and the person making the change, for use in evaluating the version traceability of the knowledge graph; 1.4.2 Supports incremental updates of the knowledge graph for evaluation. When a knowledge node or relationship is added or modified, the corresponding vector index is automatically updated. ; 1.4.

3. Provides a batch import / export interface for evaluation knowledge graphs, supporting batch import of evaluation knowledge graphs from external files, and also supporting export of evaluation knowledge graphs to a standardized format, thus obtaining a complete evaluation knowledge graph capable of machine reasoning. The specific formula is as follows: 。 4. The data quality assessment method based on multi-source heterogeneous materials as described in claim 1, characterized in that: In step 2, the multi-source heterogeneous evaluation input material is received, and the structured evaluation input model is output. The specific steps are as follows: 2.1 Constructing a multi-source assessment material set The specific formula is as follows: ； in, A subset of structured materials; It is a subset of semi-structured materials; It is a subset of unstructured materials; A subset of other auxiliary materials; 2.2 For different types of materials in the multi-source assessment material set, a differentiated intelligent analysis strategy is adopted to automatically extract core attribute information; 2.

3. Fragmented information extracted from different materials will be correlated, integrated, and aligned to eliminate information conflicts, fill in missing information, and generate a unified structured evaluation input model. The specific formula is as follows: ； in, This constitutes a list of data assets, including the name, identifier, and source information of all data assets to be evaluated; For data structure details, including field definitions, types, and constraint information for each data asset; This includes business background information, such as business scenario descriptions, business rules, and data flow relationships; This is a list of compliance requirements, including all applicable compliance clauses, mandatory requirements, and qualification standards.

5. The data quality assessment method based on multi-source heterogeneous materials as described in claim 4, characterized in that: In step 2.2, for different types of materials in the multi-source evaluation material set, a differentiated intelligent analysis strategy is used to automatically extract core attribute information, as follows: 2.2.1 For structured materials such as data dictionaries, database metadata, and API interface documents, directly extract the core attribute information of the data assets for structured material parsing; Furthermore, the core attribute information includes table name, field name, field type, field description, primary key identifier, foreign key association, index information, and data length constraints; 2.2.

2. Perform semi-structured material parsing on semi-structured materials in Word, PDF, and Markdown document formats; Specifically, it employs a large language model for text-level semantic understanding, automatically extracting information such as business rules, data flow relationships, data quality requirements, mandatory compliance clauses, and business constraints. 2.2.

3. Perform unstructured material analysis on system architecture diagrams, business process diagrams, and scanned unstructured materials; Specifically, optical character recognition technology is used in conjunction with a multimodal large model to analyze image content and automatically extract information on system components, data flow, storage location, processing nodes, and interface call relationships. 2.2.

4. Analyze system logs, data samples, compliance documentation, and other supporting materials. Specifically, this includes: automatically extracting statistical features, anomaly records, and compliance qualification information from data; In step 2.3, the fused structured evaluation input model needs to undergo multi-dimensional quality verification to check the completeness, consistency, and accuracy of its information before a complete, quality-verified structured evaluation input model can be output. The specific multi-dimensional quality verification process is as follows: Integrity verification: Checks for missing information in core fields, and marks and prompts for any missing information; Consistency check: Check whether there are conflicts in the same information extracted from different materials, mark the conflicting information and provide correction suggestions; Accuracy verification: Check whether the extracted information conforms to logical norms and industry common sense, and mark obviously erroneous information.

6. The data quality assessment method based on multi-source heterogeneous materials as described in claim 1, characterized in that: In step 3, based on the multi-dimensional association between the structured evaluation input model and the evaluation knowledge graph, data fields are automatically mapped to corresponding standardized indicator nodes. The specific steps are as follows: 3.1 For each data field, calculate its matching degree with all standardized indicators in the evaluation knowledge graph across four dimensions. The specific formula is as follows: ； in, For the first The data field and the first A set of multi-dimensional matching degrees for standardized indicators; For semantic similarity, calculate the textual semantic similarity between the data field description and the standardized metric definition; For type similarity, standardized metrics that support the type of data field are matched based on the data field type. For business similarity, based on business background information Match standardized metrics applicable to the corresponding business scenarios; For compliance similarity, based on the compliance requirements list Match standardized metrics that meet mandatory compliance requirements; 3.2 Assign preset weights to the matching degree of the four dimensions, and calculate the final matching degree score of each data field with each standardized indicator using a weighted average. The specific formula is as follows: ； in, For the first The data field and the first The final matching score of each standardized indicator; For semantic weight coefficients, Weight coefficients for types, Weighting coefficients for business operations For the compliance dimension, the weighting coefficients are set, and they satisfy the following conditions: ; Furthermore, the weighting coefficients can be dynamically adjusted based on business scenarios and compliance requirements; 3.

3. Score all standardized metrics for each data field based on the final matching degree. Sort the data from high to low, select the top N indicators with the highest matching degree as the recommendation results, and generate an initial field-indicator mapping table; The value of N is set according to the evaluation accuracy requirements and the workload of expert review; 3.4 When multiple standardized metrics are mapped to the same data field, the optimal matching metric is automatically selected based on metric priority, business scenario importance, and compliance requirements. The specific formula is as follows: ； in, For the first The optimal matching metric for each data field; For the first The overall priority of each standardized indicator is calculated by weighting the inherent priority of the indicator, the priority of the business scenario, and the priority of compliance. Furthermore, for mandatory compliance indicators that have a veto requirement, their priority is automatically set to the highest. 3.5 Generate detailed matching reasons for each initial field-indicator mapping result, and output the target field-indicator mapping result set; The matching reasons include: the matching score of each dimension, the matching basis and the reasons for priority determination, which are presented in a structured form for expert review and reference, thereby ensuring the interpretability of the mapping results and supporting expert review. The target field-indicator mapping result set includes the correspondence between data fields and standardized indicators, matching scores, matching reasons, and priority determination results.

7. The data quality assessment method based on multi-source heterogeneous materials as described in claim 1, characterized in that: In step 4, based on the evaluation rule templates corresponding to the standardized indicator nodes obtained from the mapping, an initial executable evaluation rule set adapted to multiple execution engines is automatically generated. The specific steps are as follows: 4.1 Based on the optimal matching index The evaluation rule templates associated with this indicator are obtained from the evaluation knowledge graph. The specific formula is as follows: ； in, For the first A set of evaluation rule templates corresponding to each data field. For the first An evaluation rule template, and The calculation formula corresponds to; For the first A globally unique identifier for each evaluation rule template. For the first The set of rule template identifiers corresponding to the best matching index for each data field, and the calculation The definitions in the formula are consistent; 4.2 Using a structured evaluation input model, automatically infer and populate all parameters in the evaluation rule template to generate parameterized rule instances. The specific formula is as follows: ； in, For the first The first data field corresponding to the A set of parameters for each evaluation rule template; This is a set of data location parameters, including table name, field name, and database connection information; This is a set of business logic parameters, including qualification thresholds, judgment conditions, and filtering rules. This is a set of execution parameters, including the execution engine identifier, timeout, and number of retries. 4.

3. Based on the client's specific compliance requirements and business needs, automatically adjust the rules' parameters, logic, and priorities to generate customized rule parameters adapted to the client's scenario. The specific formula is as follows: ； in, For the adjusted set of customized parameters, Adjust the function for the rules; For business background information, For the compliance requirements list, for clauses with mandatory compliance requirements, the corresponding rules are automatically set to the highest priority; 4.

4. Based on the execution engine types supported by the evaluation rule template, convert the rule instance with customized rule parameters into the native format supported by the target execution engine. The specific formula is as follows: ； in, For the first The data field, the first The first evaluation rule template, the first The native executable rules corresponding to each execution engine; This is a format conversion function; For the first Each execution engine, and the execution engine node correspond; 4.5 Generate complete metadata information for each executable rule for subsequent rule management, traceability, and auditing. Automatically generate a metadata set using the following formula: ； in, For the first The data field, the first The first evaluation rule template, the first Metadata collection corresponding to each execution engine This serves as a globally unique identifier for the rule. The standardized indicator identifier corresponding to the rule. For the rule generation time, This is the identifier of the execution engine corresponding to the rule; 4.6 Integrate all generated executable rules and their metadata to obtain a complete initial executable evaluation rule set, as shown in the following formula: ； in, This is the initial set of executable evaluation rules; This represents the total number of data fields to be evaluated. For the first A set of evaluation rule templates corresponding to each data field.

8. The data quality assessment method based on multi-source heterogeneous materials as described in claim 1, characterized in that: In step 5, the initial executable evaluation rule set is subjected to multi-dimensional cross-validation and intelligent correction to obtain the target executable evaluation rule set. The specific steps are as follows: 5.1 Perform multi-dimensional automatic validation on each executable evaluation rule to generate a rule validation result set. The specific formula is as follows: ； in, For the first The data field, the first The first evaluation rule template, the first The set of verification results for each rule corresponding to an execution engine; To verify the grammatical validity, check whether the rules conform to the grammatical specifications of the target execution engine; To verify the semantic correctness of the results, check whether the rule logic is consistent with the definition of the corresponding standardized indicator. To verify the existence of data, check whether the tables, fields, and database resources referenced in the rules actually exist and are accessible. To verify the logical consistency of the result, check whether there are logical conflicts, duplicate calculations, or contradictory results between this rule and other rules; To verify the compliance results, check whether the rules meet the compliance clauses and mandatory requirements extracted in step 2; 5.2 Intelligent automatic error correction: For rules that fail validation, the system automatically identifies the cause of the error and generates a correction plan. It automatically corrects common errors using the following formula: ； in, The revised executable rules; This is a smart correction function; The evaluation knowledge graph in step 1 is used to provide semantic and logical basis for error correction; 5.3 Deploy the revised rules to an isolated sandbox environment for testing to verify the stability of the rules and the reasonableness of the results. The specific formula is as follows: ； in, This is a function executed in the sandbox that returns two results: This indicates the sandbox execution status, with values ​​of "passed" or "failed". The results are for sandbox execution; the sandbox environment provides a simulated operating environment consistent with the production environment and will not have any impact on real data; 5.4 For rules that fail verification after multiple automatic corrections or fail to execute in the sandbox, an exception downgrade process will be implemented. The specific formula is as follows: ； in, The rules for the final output; For the alternative rules recommended by the system that have been verified in similar historical scenarios, rules that cannot be automatically corrected and have no effective alternative rules are automatically marked as "pending manual review" and a detailed error report is generated for experts to handle. 5.5 Generate a set of executable evaluation rules and a full-process verification report. Integrate all verified rules to generate a complete verification report. The specific formula is as follows: ； The full-process verification report includes the verification process, error logs, corrections, and execution results. This is the final, fully validated set of target executable evaluation rules; This is a full-process verification report; The rule correction log records the cause of the error, the correction details, and the correction time.

9. The data quality assessment method based on multi-source heterogeneous materials as described in claim 1, characterized in that: In step 6, based on the unified basic weights pre-fixed by experts, evaluation weights are assigned to each indicator in combination with business scenarios and compliance requirements. The executable evaluation rule set of the target is then optimized and executed in parallel through rule dependency optimization to generate a standardized data quality evaluation report. The specific steps are as follows: 6.1 Experts pre-fix the basic weight configuration to construct a globally unified benchmark weight set, as shown in the following formula: ； in, For the first The basic weights of each standardized indicator This is a function for assigning values ​​to both qualitative and quantitative data. 6.2 Adaptive Adjustment of Business and Compliance Weights: Based on the business scenario and compliance requirements of the current assessment task, the basic weights are adjusted to obtain the adjusted indicator weights. The specific formula is as follows: ； in, This is a business adjustment coefficient that assigns weights to core business metrics; the default value is 1. To adjust the compliance coefficient, the mandatory compliance indicators are weighted more heavily, with a default value of 1. 6.3 Perform global weight normalization on the corrected indicator weights to generate the final standardized weights that can be used for comprehensive scoring. The specific formula is as follows: ； in, For the first The final standardized weight of each standardized indicator, with the sum of all indicator weights being 1; This represents the total number of standardized indicators. 6.4 Perform dependency resolution on the output executable rule set, construct a rule execution dependency graph, merge repetitive computation tasks, and determine the optimal execution topology order. The specific formula is as follows: ； in, For rule execution dependency graph; It is a set of rule nodes, each node corresponding to an executable evaluation rule; This is a set of dependencies, where each edge represents a data dependency or computational dependency between two rules; 6.

5. Distribute the optimized evaluation task to the distributed execution cluster, execute all verified evaluation rules in batches, and generate single-rule evaluation scores. The specific formula is as follows: ； in, For the first The data field, the first The first evaluation rule template, the first The evaluation score for each rule corresponding to each execution engine; For final executable rules; The dataset to be evaluated; 6.

6. Based on the single-rule evaluation score, calculate the sub-indicator scores and the overall data quality score. The specific formula is as follows: ； ； in, For the first The sub-evaluation scores of each standardized indicator; This is an aggregation function for single-indicator, multi-rule results; This represents the overall data quality score for the target dataset. Step 6.7: Automatically locate the location, type, and severity of data quality issues, generate rectification suggestions, and output a standardized assessment report that meets the requirements of third-party data quality certification.

10. A system based on a data quality assessment method driven by multi-source heterogeneous materials as described in any one of claims 1-9, characterized in that, include: The evaluation knowledge graph construction module is used to build an evaluation knowledge graph with a standardized indicator system as its core. The multi-source heterogeneous material access module is used to receive multi-source heterogeneous evaluation input materials and supports the unified access of structured materials, semi-structured materials, unstructured materials and other auxiliary materials. The intelligent material analysis and cross-document information fusion module is used to evaluate different types of materials in a multi-source material set and automatically extract core attribute information using differentiated intelligent analysis strategies. The multi-dimensional standardized indicator automatic mapping module is used to automatically complete the mapping and matching of data fields and standardized indicator nodes and resolve conflicts. The metric-driven executable evaluation rule automatic generation module is used to generate an initial executable evaluation rule set that is compatible with multiple execution engines; The rule cross-validation and intelligent correction module is used to perform multi-dimensional cross-validation on the generated initial executable evaluation rule set and output the target executable evaluation rule set. The weight adaptive allocation and dependency-aware parallel evaluation module is used to execute the target executable evaluation rule set; The standardized data quality assessment report generation module is used to generate standardized data quality assessment reports that conform to third-party data quality certification specifications. The data storage and version management module is used to store the evaluation knowledge graph, structured evaluation input, target executable evaluation rule set, evaluation results, and full-process operation logs, and supports version traceability and query.

Images