Metadata missing attribute automatic filling method, system and device based on large model, processor and storage medium thereof

Abstract

The invention relates to a metadata missing attribute automatic filling method based on a large model, and the method comprises the following steps: (S1) a system collects original metadata from a target data source, and carries out the data preprocessing of the original metadata; (S2) analyzing the obtained metadata structure, identifying missing attribute items of the metadata structure, and generating a missing item list; (S3) calling a pre-trained large language model to perform semantic completion on the missing attribute items obtained in the missing item list through semantic understanding and association reasoning, and generating corresponding attribute candidate labels and confidence scores; (S4) through a preset rule and / or manual verification, performing completion verification and correction processing on the metadata attributes after semantic completion processing; and (S5) for the verified attribute labels, the system automatically fills and writes a metadata management library, and updates corresponding records to form a training sample closed loop. The invention also relates to a corresponding system, device, processor and storage medium.

Description

Technical Field

[0001] This invention relates to the field of large language models, and more particularly to the intersection of data governance and artificial intelligence technologies. Specifically, it refers to a method, system, device, processor, and computer-readable storage medium for automatically filling missing metadata attributes based on large models. Background Technology

[0002] With the rapid expansion of enterprise data scale and the increasing complexity of data architecture, metadata, as "data about data," has become a core element of data assetization in terms of management and governance. High-quality metadata is the foundation for achieving data discovery, data lineage, data quality management, and data security compliance.

[0003] In the metadata management lifecycle, "collection" is the first and crucial step. However, in practice, metadata collected from various heterogeneous data sources (such as databases, APIs, and files) often suffers from missing attributes. For example, the "business owner" field in a database table may be empty, the "security level" of a data model may not be labeled, and the "update frequency" information for an API interface may be missing. This incompleteness severely restricts the effectiveness of subsequent metadata applications, leading to inaccurate data catalogs, interrupted lineage analysis, and difficulties in implementing data governance strategies.

[0004] Currently, various solutions have emerged in existing technologies that attempt to improve the intelligence of metadata management, but none of them have effectively solved the problem of missing attributes during the data collection phase: End-to-end intelligent management solutions, as shown in existing technology CN202411800707.0, focus on using large models to intelligently upgrade the entire metadata management process, including intelligent querying and lineage analysis. However, such solutions are based on relatively complete metadata and do not design specific, fine-grained identification and filling mechanisms for missing attributes at the data collection source.

[0005] Tool orchestration and semantic matching schemes: As shown in the existing technology CN202510681116.4, its core technology is to use large models for dynamic tool orchestration, focusing on the semantic matching of interface functions, which is completely different from the completion of metadata attributes in terms of technical goals and implementation paths.

[0006] Data security classification and grading schemes, as shown in the existing technology CN202510540885.2, use large models to determine the data security level. The application scenarios are limited to the field of data security, and no general methodology is provided for repairing the integrity of metadata itself.

[0007] Multimodal vector retrieval schemes, as shown in the existing technology CN202510765415.6, focus on the vectorization representation and retrieval optimization of multimodal data, which differs significantly from the technical means of solving the attribute missing problem of structured metadata.

[0008] In summary, existing technologies generally suffer from a technological gap: the lack of a specialized solution capable of automatically and accurately identifying missing attributes during the metadata collection phase, and intelligently generating and filling in missing values ​​based on semantic understanding. Existing solutions are either comprehensive but lack depth, or their application scenarios do not align with the objectives.

[0009] Therefore, there is an urgent need in this field for an innovative technology that can directly address the pain points of metadata collection and achieve the transformation from "incomplete collection" to "intelligent completion". Summary of the Invention

[0010] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method, system, apparatus, processor and computer-readable storage medium for automatically filling missing metadata attributes based on a large model.

[0011] To achieve the above objectives, the present invention provides a method, system, apparatus, processor, and computer-readable storage medium for automatically filling missing metadata attributes based on a large model, as follows: The main feature of this method for automatically filling missing metadata attributes based on a large model is that the method includes the following steps: (S1) The system collects raw metadata from the target data source and performs data preprocessing on it; (S2) Parse the obtained metadata structure, identify its missing attribute items, and generate a list of missing items; (S3) Call the pre-trained large language model to perform semantic completion on the missing attribute items obtained from the missing item list through semantic understanding and association reasoning, and generate corresponding attribute candidate labels and confidence scores; (S4) Perform completion verification and correction processing on the metadata attributes after semantic completion processing through preset rules and / or manual verification; (S5) For the attribute annotations that pass the verification, the system automatically fills them into the metadata management library and updates the corresponding records to form a training sample closed loop.

[0012] Preferably, step (S1) includes: (S1.1) The system collects raw metadata from the target data source, including table structure, field names, data types, constraint information, comments, and source system; (S1.2) The collected metadata is preprocessed by formatting and unifying, cleaning special symbols and detecting missing items, in order to provide semantic context for the large language model.

[0013] Preferably, step (S2) specifically includes: The preprocessed metadata is parsed to automatically identify missing attribute items, and the fill state vector for each attribute item is calculated as follows: ; in, For the first Individual metadata attributes, For missing detection functions, when =0 indicates that the attribute is missing. =1 indicates that the attribute is complete, and a list of missing items for the currently obtained metadata is generated. The rules for automatically identifying missing attribute items include: The value is empty, the data type is undefined, the business description is missing, or the semantic information is inconsistent with similar metadata.

[0014] Preferably, step (S3) includes: (S3.1) Input the obtained list of missing items into the pre-trained large language model for semantic completion processing; the list of missing items includes the context information of each current metadata and the known attribute sample information in the historical metadata knowledge base; (S3.2) The large language model described above outputs a set of candidate annotations through contextual understanding: ; in, For the first A set of candidate annotations for individual metadata attributes. It consists of n candidate labels; (S3.3) Calculate the semantic confidence score for each candidate: ; in, The semantic relevance score generated for the large language model; For the first The confidence probability of each candidate label.

[0015] Preferably, step (S4) includes: (S4.1) Candidate label verification: Automatically filter out candidates that do not meet the constraints according to the preset rule template, and / or manually confirm candidates with confidence scores below the threshold. (S4.2) After candidate validation is completed, the final corrected score is calculated as follows. : ; in, For model confidence, For rule matching degree, These are adaptive weighting coefficients; when hour, The system confirms the validity of the annotation results by setting a score threshold.

[0016] Preferably, step (S5) specifically includes: For attribute annotations that pass verification, they are automatically written into the metadata management library and the corresponding records are updated. At the same time, the prediction results of the large language model are compared with the actual verification results to form new training samples, which are then fed back to the large language model for fine-tuning.

[0017] The system for automatically filling missing metadata attributes based on a large model to implement the above-described method is characterized in that the system comprises: The metadata acquisition module is used to collect raw metadata from the target data source and output a standardized metadata structure. The missing item detection module is connected to the metadata acquisition module and is used to perform missing attribute item detection processing on the acquired metadata and output the corresponding missing item list. The semantic inference module, connected to the missing item detection module, is used to perform semantic understanding and reasoning on the missing attributes in the missing item list by calling a large language model, and generate attribute candidate labels. The verification and correction module, connected to the semantic inference module, is used to verify and correct candidate annotations based on rule templates and / or human feedback. An autofill module, connected to the verification and correction module, is used to automatically fill in the attribute annotations that have passed final verification by writing them into the metadata management system; and The knowledge base and feedback learning module, connected to the automatic filling module, is used to maintain historical labeled samples and large language model fine-tuning samples, and to realize the self-learning optimization of the large language model.

[0018] The main feature of this device for automatically filling missing metadata attributes based on a large model is that the device includes: A processor is configured to execute computer-executable instructions; The memory stores one or more computer-executable instructions, which, when executed by the processor, implement the steps of the method for automatically filling missing metadata attributes based on a large model as described above.

[0019] The main feature of this processor for automatically filling missing metadata attributes based on a large model is that the processor is configured to execute computer-executable instructions, and when the computer-executable instructions are executed by the processor, the steps of the method for automatically filling missing metadata attributes based on a large model described above are implemented.

[0020] The computer-readable storage medium is characterized in that it stores a computer program thereon, which can be executed by a processor to implement the steps of the method for automatically filling missing metadata attributes based on a large model as described above.

[0021] The present invention employs a method, system, apparatus, processor, and computer-readable storage medium for automatic attribute filling of metadata based on a large model, forming a closed-loop mechanism from missing attribute identification, semantic reasoning, candidate generation, result verification to attribute filling, effectively improving the completeness and usability of metadata collection. Simultaneously, based on automatically generated candidate annotations, the results are verified and corrected through a rule engine or manual review to ensure the accuracy and interpretability of the automatically filled results. This achieves the technical effect of automated and precise metadata annotation. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the method for automatically filling missing metadata attributes based on a large model according to the present invention.

[0023] Figure 2 This is an architecture diagram of the system for automatically filling missing metadata attributes based on a large model according to the present invention. Detailed Implementation

[0024] To more clearly describe the technical content of the present invention, the following description is provided in conjunction with specific embodiments.

[0025] Before describing the embodiments of the present invention in detail, it should be noted that, in the following, the terms “comprising,” “including,” or any other variations are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0026] Please see Figure 1 As shown, the overall processing flow of this method for automatically filling missing metadata attributes based on a large model is as follows: S1 Metadata Acquisition and Preprocessing: Obtain structured information from data sources and standardize the format and encoding.

[0027] S2 Missing Attribute Identification: Detects missing fields, descriptions, and types, and generates a list of missing items.

[0028] S3 large model semantic reasoning: Generate candidate labels and corresponding confidence scores through contextual understanding.

[0029] S4 Verification and Correction: Combine rules with manual verification to calculate the overall score.

[0030] S5 Autofill and Learning Feedback: Writes the results into the metadata system, while forming a closed loop of training samples to achieve continuous optimization.

[0031] This overall process enables automatic completion of missing metadata, ensures labeling accuracy, and continuously improves model capabilities.

[0032] The following will explain each processing step in more detail: Step S1: Metadata Acquisition and Preprocessing The system collects metadata from the target data source (database, API, data file, etc.), including table structure, field names, data types, constraint information, comments, source system, etc.

[0033] After data collection is complete, the raw metadata is preprocessed, including: Uniform formatting; Cleaning of special symbols; Missing item detection (such as field descriptions being empty, missing business meanings, undefined data types, etc.).

[0034] Technical Feature Interaction: This step directly interacts with the subsequent large-scale model inference process, and the preprocessing results provide semantic context for the input of the large-scale model.

[0035] Technical effects: Ensure that the metadata input to the large model is standardized and structured to improve the accuracy of subsequent inference.

[0036] Step S2: Identification of missing attributes By parsing the collected metadata structure, missing attribute items are automatically identified. The implementation method is as follows: 1. Construct a metadata attribute model Each piece of metadata is represented as a fixed set of attributes (such as field name, data type, business meaning, constraints, sensitivity level, responsible person, etc.).

[0037] Each attribute has fields to determine whether it is missing, such as: whether the value is empty, whether the type is valid, and whether the description conforms to the specification.

[0038] 2. Missing Detection Function ; For each attribute of metadata Perform missing detection: The value is empty (such as an empty string or null). Invalid values ​​(such as unknown data types or non-standard constraints). Lack of semantic description (e.g., fields have no comments, interfaces have no purpose description). The semantics are inconsistent with similar metadata (e.g., the user table does not have common attributes such as "creation time" and "modification time"). Output based on the test results: = 0 → Attribute missing = 1 → Complete attributes 3. Generate a list of missing items All The collection of attributes with a value of 0 forms a list of missing items, which serves as input for subsequent large model completion.

[0039] Here, "metadata attributes" refers to fixed semantic fields used to describe a data object, for example: Structure attributes: field name, data type, length, primary key / foreign key, and whether it is required; Business attributes: business meaning, business category, system to which it belongs, and person in charge of the business; Governance attributes: data quality level, security level, update frequency, and lifecycle; Descriptive attributes: comments, descriptions, example values, usage descriptions.

[0040] These attributes constitute the semantic features of the metadata object, which is the basis for this technical solution to determine missing items and complete automatic completion.

[0041] The system calculates the "filled state vector" for each attribute item accordingly: ; The state vector is used to quantify the completeness of metadata attributes and is a key intermediate result for generating a list of missing items. Its calculation result directly determines the scope of completion and the input of the completion process.

[0042] Technical effects: It enables automatic identification of missing items, eliminating the need for manual annotation and confirmation, and improving the efficiency of metadata quality inspection.

[0043] Step S3: Semantic Understanding and Relational Reasoning Based on Large Models For detected missing attribute items, the system calls a pre-trained Large Language Model (LLM) to perform semantic completion, including: The context of using structured metadata; Combined with historical metadata knowledge base; Perform semantic reasoning and completion for missing attributes; Output candidate labels + confidence distribution; Further corrections will be made based on rules and manual intervention.

[0044] The input to this large language model includes: The context information of the current metadata (such as table name, field name, business module, etc.); Samples of known attributes in the historical metadata knowledge base.

[0045] Model outputs a set of candidate annotations: ; The candidate annotation set The input used for subsequent confidence screening, rule validation, and final result selection essentially provides a space of options and enhances the interpretability and scalability of the system.

[0046] And calculate semantic confidence for each candidate: ; in: : A semantic relevance score generated by the large model based on semantic understanding; : the first in the candidate annotation set The confidence probability of each candidate label. Corresponding to a semantic confidence level This is used to reflect the rationality or credibility of the candidate in the current context.

[0047] Interaction relationship: This stage interacts with the metadata knowledge base for semantic comparison and outputs candidate annotation results to the verification and correction module.

[0048] Technical effects: Intelligent prediction of missing attributes can be achieved by leveraging the semantic understanding and contextual reasoning capabilities of large models.

[0049] Step S4: Candidate Label Verification and Correction The system performs verification based on two types of mechanisms: 1. Rule Validation Module: Automatically filters out candidates that do not meet the constraints based on preset rule templates (such as field name and description consistency rules, data type constraint rules, etc.); 2. Manual verification interface (optional): Experts review the model output results and manually confirm items with confidence levels below the threshold.

[0050] After verification, the system calculates the final corrected score: ; in: Model confidence score; During the validation phase, the system will select the confidence score corresponding to the final candidate annotations (annotations that will be written to the metadata management library). And record it as .Right now The selected candidate labels are in the candidate set The semantic confidence level reflects the model's confidence in the annotation. In the final corrected score calculation, Matching degree with rules This combination is used to quantify the credibility of annotations, thereby supporting decisions based on either automatic filling or manual review.

[0051] Rule matching degree: This feature represents the degree of matching between candidate annotations and the system's preset rule templates, and is used to quantify the compliance of candidate annotations under rule constraints. The method for determining this feature includes the following steps: 1. Rule Definition The system predefines several rule templates, for example: Semantic consistency rules between field names and field descriptions; Data type constraint rules; Business logic or dependency constraint rules.

[0052] 2. Candidate Matching For each candidate label The system checks whether it meets the above rules: If the rule is met, it is counted as 1; If the rule is not met, it is counted as 0; A weighted average can be calculated based on the importance of the rules to generate a rule matching degree with a continuous value between 0 and 1.

[0053] 3. Numerical output The final generated The rule matching score for candidate annotations is in the range of [0,1]. The larger the value, the more the candidate matches the rule constraints.

[0054] Technical effect: Ri can be used to verify the prediction results of large language models by rules, ensuring that the final annotation results conform to both semantic reasoning and business rules, thereby improving the reliability and interpretability of the completion results.

[0055] : System adaptive weighting coefficient (between 0 and 1).

[0056] when When the threshold is reached, the system confirms that the annotation results are valid.

[0057] Technical effects: By combining automated rules with human feedback, an "explainable completion verification mechanism" is formed to ensure the reliability of automatic completion.

[0058] Step S5: Autofill and Feedback Learning For attribute annotations that pass verification, the system automatically writes them to the metadata management database and updates the corresponding records.

[0059] Simultaneously, the model's predictions are compared with actual validation results to form new training samples, which are then fed back to the larger model for fine-tuning. Specifically: Model prediction results: refer to the candidate labels and their corresponding confidence scores output by the large model in step S3. (i.e., the original predictions generated by the model), reflecting the model's initial inferences about the missing attributes.

[0060] Actual verification results refer to the validity of candidate annotations confirmed by the verification and correction module, specifically including: Rule matching degree ; Final confirmation score after weighting the large model and rules ; Whether candidate annotations pass the threshold It is deemed valid.

[0061] The process of generating training samples and fine-tuning includes the following steps: 1. Compare and generate tags For each candidate label The predictions of large models Verification and correction results Compare the pass / fail status; Construct training samples: The input is the original metadata context (field name, table name, business module, etc.), and the label is the validation result (valid / invalid or specific annotation value).

[0062] 2. Create a fine-tuning dataset Add the newly generated training samples to the historical annotation knowledge base; Batch production is carried out as needed to form continuously updated datasets for incremental model training.

[0063] 3. Model fine-tuning Use feedback samples to perform small-scale gradient updates on the large language model (using methods such as Fine-tuning / LoRA / Prompt-tuning); By optimizing model parameters, the model can be made more accurate and more in line with business rules in future missing attribute completion tasks.

[0064] Technical Feature Interaction: The model prediction results are processed through a verification and correction module, followed by a knowledge base feedback and learning module, forming a dynamic optimization closed loop.

[0065] Technical effects: By comparing predictions with validation data to generate fine-tuning data, the model can learn itself, and the system's automatic completion accuracy can be gradually improved as the number of uses increases, forming an interpretable optimization loop.

[0066] like Figure 2 The diagram shown is a system architecture diagram of the present invention, which mainly includes the following modules: Metadata collection module (101) Function: Collect raw metadata from multi-source systems; Output: Standardized metadata structure.

[0067] Missing item detection module (102) Function: Detect missing attribute items in collected metadata; Output: List of missing items.

[0068] Semantic reasoning module (103) Function: Calls a large language model to perform semantic understanding and reasoning on missing attributes to generate candidate labels.

[0069] Verification and correction module (104) Function: Verify and correct candidate annotations based on rule templates and human feedback.

[0070] Autofill module (105) Function: Writes the finally verified annotations into the metadata management system.

[0071] Knowledge base and feedback learning module (106) Function: Maintain historical labeled samples and model fine-tuning samples to achieve self-learning optimization of the model.

[0072] Module connection relationship: Metadata Acquisition Module → Missing Detection Module → Semantic Inference Module → Verification and Correction Module → Autofill Module → Knowledge Base and Feedback Learning Module The modules interact through a data bus interface to achieve a closed-loop information flow of metadata throughout the entire process of collection, inference, verification, and updating.

[0073] In a specific embodiment of the present invention, taking the database table metadata collected by a data platform as an example, the field descriptions and business meanings in the collection results are missing.

[0074] Implementation process: 1. The system collected field information from the USER_ACCOUNT table, but the description of the acct_type field was missing; 2. The missing data detection module detected that the acct_type attribute description is empty; 3. The system inputs the table structure context and historical metadata into the large model; 4. Model inference outputs candidate descriptions: Account Type (Individual / Business) Account Types (Savings / Credit) Confidence level: =0.82, =0.54; 1. The rule verification module detected that the field values ​​were concentrated in "P" and "C", which conforms to the personal / enterprise rules; 2. Calculate the final score: = 0.9 × 0.82 + 0.1 × 1.0 = 0.838; 3. The system will select candidates The final annotations are written into the database. Simultaneously, the validation process is recorded as new samples to feed back into the large model.

[0075] In this process, for missing attributes, the system first generates a set of candidate annotations. ,For example The large model calculates a relevance score for each candidate label. This indicates the degree of matching between the candidate label and the metadata context. To obtain comparable probability values, the system normalizes the scores of all candidate labels, converting them into confidence scores. The normalized confidence score reflects the model's relative confidence in each candidate label. (Example:) and For example, after processing, the system obtains the confidence level. =0.82、 =0.54, indicating that the model believes It is more in line with the context, so it is given priority as the final annotation for verification and filling.

[0076] In the final score calculation, the system uses adaptive weighting coefficients. Adjusting model confidence Rule matching degree The proportion of contribution. Specifically: Adaptive weight coefficients The system can dynamically adjust based on historical verification results and metadata types, or it can be preset by the user. The value range is 0~1. The higher the confidence level, the greater the impact of the model's prediction confidence on the final score. (Example) =0.9 indicates that the model output accounts for the majority of the weight.

[0077] Rule matching degree The rule validation module determines whether candidate tags meet preset rules, such as consistency between field values ​​and descriptions, and data type constraints. If a candidate tag fully conforms to the rules, then... Set the value to 1; if some parts do not match, a value between 0 and 1 can be used to reflect the degree of rule matching.

[0078] Ultimately, the system will and according to Weighted composite final score This is used to determine whether candidate labels pass validation, ensuring the reliability of automatic filling. Test results (comparison table): ; Compared with existing technologies, this technical solution has the following technical advantages: Significantly improves the completeness, accuracy, and consistency of metadata; Continuously optimize model performance through feedback mechanisms; To form an automated and intelligent closed loop for metadata quality management.

[0079] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.

[0080] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution device.

[0081] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0082] The storage media mentioned above can be read-only memory, disk, or optical disk, etc.

[0083] In the description of this specification, references to terms such as "an embodiment," "some embodiments," "example," "specific example," or "embodiment," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0084] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

[0085] The present invention employs a method, system, apparatus, processor, and computer-readable storage medium for automatic attribute filling of metadata based on a large model, forming a closed-loop mechanism from missing attribute identification, semantic reasoning, candidate generation, result verification to attribute filling, effectively improving the completeness and usability of metadata collection. Simultaneously, based on automatically generated candidate annotations, the results are verified and corrected through a rule engine or manual review to ensure the accuracy and interpretability of the automatically filled results. This achieves the technical effect of automated and precise metadata annotation.

[0086] In this specification, the invention has been described with reference to specific embodiments thereof. However, it will be apparent that various modifications and variations can be made without departing from the spirit and scope of the invention. Therefore, the specification and drawings should be considered illustrative rather than restrictive.

Claims

1. A method for automatically filling missing metadata attributes based on a large model, characterized in that, The method includes the following steps: (S1) The system collects raw metadata from the target data source and performs data preprocessing on it; (S2) Parse the obtained metadata structure, identify its missing attribute items, and generate a list of missing items; (S3) Call the pre-trained large language model to perform semantic completion on the missing attribute items obtained from the missing item list through semantic understanding and association reasoning, and generate corresponding attribute candidate labels and confidence scores; (S4) Perform completion verification and correction processing on the metadata attributes after semantic completion processing through preset rules and / or manual verification; (S5) For the attribute annotations that pass the verification, the system automatically fills them into the metadata management library and updates the corresponding records to form a training sample closed loop.

2. The method for automatically filling missing metadata attributes based on a large model according to claim 1, characterized in that, The step (S1) includes: (S1.1) The system collects raw metadata from the target data source, including table structure, field names, data types, constraint information, comments, and source system; (S1.2) The collected metadata is preprocessed by formatting and unifying, cleaning special symbols and detecting missing items, in order to provide semantic context for the large language model.

3. The method for automatically filling missing metadata attributes based on a large model according to claim 1, characterized in that, The specific steps (S2) are as follows: The preprocessed metadata is parsed to automatically identify missing attribute items, and the fill state vector for each attribute item is calculated as follows: ； in, For the first Individual metadata attributes, For missing detection functions, when =0 indicates that the attribute is missing. =1 indicates that the attribute is complete, and a list of missing items for the currently obtained metadata is generated. The rules for automatically identifying missing attribute items include: The value is empty, the data type is undefined, the business description is missing, or the semantic information is inconsistent with similar metadata.

4. The method for automatically filling missing metadata attributes based on a large model according to claim 3, characterized in that, The step (S3) includes: (S3.1) Input the obtained list of missing items into the pre-trained large language model for semantic completion processing; the list of missing items includes the context information of each current metadata and the known attribute sample information in the historical metadata knowledge base; (S3.2) The large language model described above outputs a set of candidate annotations through contextual understanding: ； in, For the first A set of candidate annotations for individual metadata attributes. It consists of n candidate labels; (S3.3) Calculate the semantic confidence score for each candidate: ； in, The semantic relevance score generated for the large language model; For the set of The confidence probability of each candidate label.

5. The method for automatically filling missing metadata attributes based on a large model according to claim 4, characterized in that, The step (S4) includes: (S4.1) Candidate label verification: Automatically filter out candidates that do not meet the constraints according to the preset rule template, and / or manually confirm candidates with confidence scores below the threshold. (S4.2) After candidate validation is completed, the final corrected score is calculated as follows. : ； in, For model confidence, For rule matching degree, These are adaptive weighting coefficients; when hour, The system confirms the validity of the annotation results by setting a score threshold.

6. The method for automatically filling missing metadata attributes based on a large model according to claim 5, characterized in that, The specific steps (S5) are as follows: For attribute annotations that pass verification, they are automatically written into the metadata management library and the corresponding records are updated. At the same time, the prediction results of the large language model are compared with the actual verification results to form new training samples, which are then fed back to the large language model for fine-tuning.

7. A system for automatically filling missing metadata attributes based on a large model to implement the method described above, characterized in that, The system includes: The metadata acquisition module is used to collect raw metadata from the target data source and output a standardized metadata structure. The missing item detection module is connected to the metadata acquisition module and is used to perform missing attribute item detection processing on the acquired metadata and output the corresponding missing item list. The semantic inference module, connected to the missing item detection module, is used to perform semantic understanding and reasoning on the missing attributes in the missing item list by calling a large language model, and generate attribute candidate labels. The verification and correction module, connected to the semantic inference module, is used to verify and correct candidate annotations based on rule templates and / or human feedback. An autofill module, connected to the verification and correction module, is used to automatically fill in the attribute annotations that have passed final verification by writing them into the metadata management system; and The knowledge base and feedback learning module, connected to the automatic filling module, is used to maintain historical labeled samples and large language model fine-tuning samples, and to realize the self-learning optimization of the large language model.

8. A device for automatically filling missing metadata attributes based on a large model, characterized in that, The device includes: A processor is configured to execute computer-executable instructions; A memory that stores one or more computer-executable instructions, which, when executed by the processor, implement the steps of the method for automatically filling missing metadata attributes based on a large model, as described in any one of claims 1 to 6.

9. A processor for automatically filling missing metadata attributes based on a large model, characterized in that, The processor is configured to execute computer-executable instructions, which, when executed by the processor, implement the steps of the method for automatically filling missing metadata attributes based on a large model, as described in any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that, It stores a computer program that can be executed by a processor to implement the steps of the method for automatically filling missing metadata attributes based on a large model, as described in any one of claims 1 to 6.

Images