Domain adaptation data governance standard recommendation system based on multi-modal large model and nlu-hgsc

Abstract

This invention discloses a domain-adaptive data governance standard recommendation system based on a multimodal large model and NLU-HGSC. The invention includes a standard management module, a directory preprocessing module, an NLU-HGSC module, and a data storage module. The standard management module is used for collecting and parsing standards. The directory preprocessing module is used for collecting and preprocessing metadata to be matched. The NLU-HGSC module is used to calculate the similarity between heterogeneous graph structures generated by the standard management module and the directory preprocessing module, and recommend corresponding data standards for the directory data in the data storage module as the basis for data governance. The NLU-HGSC heterogeneous graph similarity calculation technology in this invention requires no training; even if standards are added or modified, the recommendation results will adaptively adjust without requiring model retraining as the standard data changes.

Description

Technical Field

[0001] This invention belongs to the field of computer information technology and relates to a domain-adaptive data governance standard recommendation system based on multimodal large models and NLU-HGSC. Background Technology

[0002] In data governance, the establishment of data governance standards is crucial, providing an important basis for subsequent data quality and consistency. With the explosive growth of data scenarios and the diversification of fields, the traditional method of manually formulating data governance standards has gradually revealed some limitations, such as inefficiency, difficulty in adapting to the data needs of different fields, lag in standard formulation and updates, and low accuracy of recommendation results.

[0003] In practical data governance operations, the following challenges are faced: First, manually developing data governance standards is inefficient. Second, data from different systems involves numerous domains, and even a single system may encompass multiple domains. This domain diversity leads to a variety of data formats and types, making the development of data governance standards complex and time-consuming. Traditional methods struggle to efficiently output domain-appropriate standards. Third, data standards in various domains are constantly being updated and changed; manually developing data governance standards is inefficient and fails to meet the timeliness requirements of the standards.

[0004] Patent No. CN 118012890 A, entitled "A Matching Method for Data Fields and Data Standards and a Readable Storage Medium", describes a method for matching data to be processed and standard data based on a language model. In steps S202 to S205, the method for matching based on data fields and standards is described. Although the lineage graph relationship of the data to be matched (i.e., the source information of the data) is considered, which can supplement the characteristics of the data to be matched, the more critical data standards and the graph structure information contained in the data to be explored itself are ignored.

[0005] Patent CN 115794899 A, entitled "Data Standard Matching Method, Apparatus, Device, Medium and Product", proposes in S110 to obtain multiple features (name, data format and remarks) of a certain field data item for matching calculation. Compared with the previous patent, it introduces more features for standard matching and introduces a step of matching the corresponding standard by data item synonyms. However, compared with the heterogeneous graph proposed in this patent, which is constructed from multiple entities such as domain, standard name, table data item, field data item and field data item features, patent CN 115794899 A only considers the local information in this heterogeneous graph, namely the matching of field data item and field data item features, without fully considering table name and domain information, thus affecting the accuracy of matching.

[0006] Patent No. CN 114840576 A, "Data Standard Matching Method and Device", proposes a method based on field and table name decomposition, and determines the degree of matching with the standard based on the number of associations between the decomposition results and the standard library. Although it considers the matching degree of table name and field name information at the same level, it loses the hierarchical relationship by matching the table name and field name at the same level, which ultimately affects the accuracy of matching.

[0007] Patent No. CN 115374108 B, entitled "A Data Standard Generation and Automatic Mapping Method Based on Knowledge Graph Technology," describes a data standard mapping method. First, it proposes a knowledge graph-based matching method. In S201, it describes a graph structure consisting of standard number, standard English name, standard name, standard classification, standard type, creator, status, and operation. This structure focuses more on the basic attribute information of the standard, omitting multi-level content information of the standard itself (such as fields and field types). Second, in S301, it describes a rule-based method for mapping data standards and metadata feature datasets, and generating training and validation sets based on the mapping results to train the mapping model. However, this method based on the trained mapping model has a significant limitation: it lacks adaptive capability for standard mapping. Whenever a new standard is introduced or the content of existing standards changes, the model needs to be retrained because it learns from past knowledge and cannot automatically adapt to new or modified content.

[0008] Patent CN 117951336 A, entitled "Enterprise-level Data Standard Processing Method, Apparatus, Equipment and Media," describes a matching method based on multiple steps and various rules. However, many of these steps involve manual intervention and lack adaptive capabilities. The specific composition of the data items to be processed and the data standards in step S202 is not publicly described, nor is the method for determining the business scope (domain) in step S301 clearly described.

[0009] In summary, the existing technical solutions have the following problems:

[0010] 1. Missing Matching Information: For example, CN 118012890 A ignores the graph structure information contained in the data standard and the data to be explored itself; CN 115794899 A mentions field data items and data item characteristics, but still ignores the graph structure information of the domain, standard, and table name; CN 114840576 A considers table name and field information, but matches them at the same level, losing the hierarchical relationship; CN 115374108 B only focuses on the basic information of the standard, without paying attention to the multi-level content information of the standard itself. In summary, the existing technical solutions have certain deficiencies in the description of the matching objects, ultimately affecting the accuracy of the matching.

[0011] 2. Lack of adaptive domain matching capability: As described in CN 115374108 B, the trained model cannot automatically adapt to new standards or domain types, resulting in the need to retrain the model after each change; Although CN117951336 A mentions determining the business scope (domain matching), it does not propose specific technical solutions.

[0012] 3. Lack of Standard Acquisition and Parsing Technology: High-accuracy standard recommendation systems rely on high-quality standard data. However, existing solutions generally lack descriptions of the standard data acquisition and parsing process. Describing the standard data during the matching process can reveal missing matching information. Incomplete data during parsing is the main reason why "missing matching information" ultimately affects "low standard matching accuracy." For example, CN 117951336 A mentions acquiring publicly available standard data within the defined business scope, but does not explain how to parse these standard documents. Similarly, CN 115794899 A and CN114840576 A also do not mention the standard data acquisition and parsing process. While CN 115374108 B proposes a method based on data frequency and importance analysis in S102, S101 does not explicitly explain how to parse metadata. Summary of the Invention

[0013] To address the shortcomings of the prior art, this invention provides a domain-adaptive data governance standard recommendation system based on multimodal large models and NLU-HGSC.

[0014] This invention includes a standard management module, a directory preprocessing module, an NLU-HGSC module, and a data storage module;

[0015] The standard management module is used for the collection and parsing of standards;

[0016] The directory preprocessing module is used to collect and preprocess the metadata to be matched;

[0017] The NLU-HGSC module is used to calculate the similarity between the heterogeneous graph structures generated by the standard management module and the directory preprocessing module, and to recommend corresponding data standards for the directory data in the data storage module as a basis for data governance.

[0018] Compared with the prior art, the beneficial effects of this invention are:

[0019] 1. This invention uses multimodal large model technology combined with layout analysis technology to parse standard information from standard files. Compared with existing technical solutions, the standard data parsed by this invention has richer node information and hierarchical information, providing data support for subsequent high-precision matching calculations.

[0020] 2. This invention employs a few-shot learning technique based on multimodal large model technology to achieve adaptive domain classification. Even if a new domain classification is added, the model does not need to be retrained; the domain classification can be achieved simply by adjusting the prompt.

[0021] 3. This invention employs NLU-HGSC, a heterogeneous graph similarity calculation technique based on natural language understanding, to recommend corresponding data standards for all tables and fields in the database to be governed, thereby improving matching accuracy. Because the NLU-HGSC heterogeneous graph similarity calculation technique requires no training, even if new standards are added or modified, the recommendation results will adaptively adjust, without needing to retrain the model as the standard data changes, thus possessing strong adaptive standard matching capabilities.

[0022] In summary, compared with traditional technologies, this invention can extract richer standard information from standard files of various formats, has higher matching accuracy, and has domain-adaptive capabilities. Attached Figure Description

[0023] Figure 1 This is a system structure diagram of an embodiment of this application;

[0024] Figure 2 This is a flowchart illustrating the implementation process of the standard management module in an embodiment of this application.

[0025] Figure 3 This is a schematic diagram illustrating the similarity between levels in an embodiment of this application. Detailed Implementation

[0026] The present invention will be further described below with reference to the accompanying drawings and examples.

[0027] like Figure 1As shown in the embodiments of this application, the domain-adaptive data governance standard recommendation system based on multimodal large models and NLU-HGSC mainly consists of four modules: a standard management module, a catalog preprocessing module, an NLU-HGSC module, and a data storage module. This system aims to recommend corresponding data standards for all tables and fields in the database to be governed using NLU-HGSC (heterogeneous graph similarity computation based on natural language understanding). It provides authoritative standard information such as national standards, local standards, industry standards, enterprise standards, and group standards as important bases for subsequent data governance work. Through this technology, standard documents can be parsed more accurately, and the domain scope of the recommended data can be adaptively suggested, recommending suitable data standards for various data scenarios.

[0028] In the embodiments of this application, the standard management module is responsible for the collection and parsing of standards. First, through the standard collection submodule, relevant standard documents are collected from multiple sources, and the publication, implementation, and update status of standards are monitored. Then, layout analysis technology is used to parse the standard documents into elements. Among them, the image-formatted tables are further parsed into text format using a multimodal large model, providing a data foundation for the similarity graph calculation module in the subsequent NLU-HGSC module.

[0029] Furthermore, such as Figure 2 As shown, the standard management module includes two sub-modules: standard collection and standard parsing. First, through the standard information collection function, the system automatically collects standard documents and monitors their implementation and update status. If a new standard is released, its document is automatically downloaded, and its recommendation is determined based on its implementation status. If a standard document is updated, the latest document is retrieved and a new standard library is added. These standard documents are available in formats including but not limited to PDF, Word, JPG, JPEG, PNG, BMP, HTML, XLSX, and JSON, covering various types such as national standards, local standards, and industry standards.

[0030] Secondly, the collected standard documents are input into the layout analysis unit, which uses object detection technology to parse the content of the standard documents. The parsed content includes, but is not limited to, key-value pairs such as the standard's Chinese name, English name, publication date, implementation date, issuing department, standard number, and standard type, as well as table names, charts, and images. Tables in image format are input into a multimodal large model for table parsing, outputting text information such as the table's field names, field types, and field lengths. By employing layout analysis technology combined with table parsing and other multimodal large model technologies, rich node information and hierarchical relationships are extracted from standard documents of various formats, providing a solid data foundation for subsequent standard matching.

[0031] Finally, the parsing results are stored in the data storage module and input into the NLU-HGSC module to add domain classifications and generate graph structures.

[0032] like Figure 1 As shown, in the embodiments of this application, the directory preprocessing module includes a metadata collection and data cleaning module for the database to be matched. The main function of this module is to preprocess the data directory to be matched, including the collection and cleaning of metadata. Finally, these data will be used in the heterogeneous graph generation module of the NLU-HGSC module to generate a graph structure suitable for subsequent similarity calculation with standard heterogeneous graphs.

[0033] Furthermore, firstly, the metadata collection module acquires metadata information from the database, including table names, field names, and field types. This data is then cleaned, including removing duplicate fields and invalid characters from field names. Subsequently, this data serves as input to the NLU-HGSC module, recommending appropriate standard data for each table in the catalog as the basis for governance.

[0034] like Figure 1 As shown in the embodiments of this application, the NLU-HGSC module is the core part of the system, used to recommend corresponding data standards for the directory data in the database as the basis for data governance. First, the domain adaptation module performs domain classification on the standard data and directory metadata from the standard management module and the directory preprocessing module. Then, the heterogeneous graph generation module transforms the two inputs into two heterogeneous graph structures, adding the domain classification results as nodes to their respective graphs. Subsequently, the semantic understanding module is responsible for converting each natural language node in the above two graph structures into a computable mathematical form. Finally, the graph computation module is responsible for calculating the similarity of the semantically encoded graph data and outputting the standard data corresponding to the directory.

[0035] The data storage module described in the embodiments of this application utilizes various database technologies, including the relational database MySQL and the graph database Neo4j. The combined use of these database technologies ensures that the system can efficiently store and retrieve different types of data.

[0036] In some embodiments, the domain adaptation module is responsible for domain classification of table names in the standard management module and the directory preprocessing module. Specifically, this module implements this function through its internal domain adaptation submodule. This submodule is based on multimodal large model Few-Shot Learning technology, enabling flexible domain classification without model retraining. Its core advantage lies in its ability to quickly adapt to and learn the classification capabilities of new domains by adjusting the parameters of the input prompt, thereby ensuring high accuracy and coverage of the classification results. This adaptability not only significantly improves the system's scalability but also ensures that when faced with new domains and new standards, the system can respond quickly and achieve efficient and accurate classification processing without retraining.

[0037] In some embodiments, the heterogeneous graph generation module utilizes the functionality of the Neo4j graph database to transform the output data from the standards management module, catalog preprocessing module, and domain adaptation module into graph data. Specifically, the standards management module outputs all parsed standard information, including standard domain classifications, table names, field names, their types, and lengths; the catalog preprocessing module outputs metadata of the database tables in the database to be governed, including table names, Chinese field names, and field types. The heterogeneous graph generation module uses Neo4j's graph processing engine to map the entities, relationships, and attributes generated by each module into a heterogeneous graph structure, and fully leverages Neo4j's native graph algorithms and query language (Cypher) to achieve complex associations and calculations between nodes and edges. Simultaneously, during data import, the system supports dynamic adjustment of the graph data model to ensure seamless integration and interoperability of data from different sources and types within the same graph, thus providing a solid data foundation for subsequent data standard recommendations.

[0038] In some embodiments, the semantic understanding submodule provides semantic understanding for similarity graph computation within the NLU-HGSC module. This submodule is primarily responsible for embedding all nodes in the graph. Specifically, relying on a deep learning language model, this submodule transforms natural language elements in the graph into highly abstract and computable mathematical representations, thus laying the foundation for subsequent similarity graph computation. Through this process, the semantic understanding submodule can effectively capture and characterize the potential semantic relationships and complex associations between nodes, enabling the system to maintain high-precision computational capabilities when processing diverse and complex heterogeneous data. This provides solid technical support for the accuracy of automated recommendation based on data standards.

[0039] In some embodiments, during the heterogeneous graph generation module's heterogeneous graph similarity calculation process, the standard graph data and metadata graph data are divided into four levels based on different node types: domain layer, table name layer, field layer, and field type layer, and the similarity between each level is calculated one by one.

[0040] like Figure 3 As shown, taking field-level similarity calculation as an example, firstly, each field node in the metadata graph data is compared one by one with all field nodes in the standard graph data, and the similarity between fields is calculated through semantic similarity. Next, these similarity results are grouped according to the table names, and the standard field matching result with the highest similarity in each group is selected as the final similarity value of the corresponding field node in the current standard. Subsequently, each standard table name is grouped, and the average of the similarities of all fields within the group is calculated, which is used as the overall similarity value of that standard table at the field level. The similarity calculation method for the field type layer is consistent with that for the field layer. Finally, the semantic similarity of the domain layer and the table name layer is calculated. This hierarchical approach to similarity calculation for heterogeneous graph data fully considers the impact of information at different levels on the similarity results, ensuring both accuracy and reasonableness.

[0041] Furthermore, to ensure that the overall similarity calculation process fully reflects the importance of each level, different weights are assigned to each level based on expert experience. This weighting strategy aims to improve the reliability of the standard recommendation. Specifically, after completing the similarity calculation for each level, the graph calculation module integrates the similarity values ​​of each level through a weighted summation to generate a comprehensive similarity score. This score is used to quantitatively assess the similarity between the heterogeneous graph structure of each table in the database to be governed and the heterogeneous graph structure of the standard data, ensuring accurate identification of the most matching standard data. Subsequently, the standard with the highest similarity to the table to be matched is obtained by grouping according to the database table name and taking the maximum value within each group, which serves as the final recommendation result. Simultaneously, the domain corresponding to the most matching standard result is the domain to which the table to be matched belongs. This calculation method not only significantly improves the accuracy of the standard recommendation but also enhances the system's flexibility and adaptability. Especially when facing the implementation of new standards and updates to existing standards, the system can automatically adapt to these changes, thereby avoiding the tedious retraining process of the model and further improving the system's efficiency and intelligence.

[0042] Based on multi-node type and multi-level relationship information, this application uses a heterogeneous graph similarity calculation method based on natural language understanding technology. This method can accurately identify and match isomorphic relationships between different graph structures through semantic understanding. Even when new standards are implemented, existing standards are updated, or new domain classifications are added, the standard recommendation results and domain classification results can be quickly adjusted, avoiding the drawback of needing to retrain the model in traditional methods.

[0043] In summary, this application relates to automated standard parsing technology and NLU-HGSC standard matching technology. The automated standard parsing technology employs a multimodal large model combined with layout analysis technology to extract standard information and uses a heterogeneous graph similarity calculation method based on semantic understanding to recommend standard information with high accuracy. First, layout analysis technology extracts various elements from images, such as images, tables, and table names. Then, multimodal large model technology is used to parse image-formatted tables into numerical tables, providing a high-quality data foundation for subsequent matching calculations. Heterogeneous graph similarity calculation is a graph theory-based method that can identify and match isomorphic relationships between different graph structures. Semantic understanding technology is introduced on top of heterogeneous graph similarity calculation, using deep learning algorithms from natural language processing to understand the meaning of standard names, standard table names, and standard field names. This process further improves the accuracy of heterogeneous graph similarity calculation, ultimately enhancing the precision of governance standard recommendations.

[0044] Please note that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. The above embodiments only illustrate several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be pointed out that for those skilled in the art, several modifications and improvements can be made without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A domain-adaptive data governance standard recommendation system based on multimodal large models and NLU-HGSC, characterized by: It includes a standard management module, a catalog preprocessing module, an NLU-HGSC module, and a data storage module; The standard management module is used for the collection and parsing of standards; The directory preprocessing module is used to collect and preprocess the metadata to be matched; The NLU-HGSC module is used to calculate the similarity between the heterogeneous graph structures generated by the standard management module and the directory preprocessing module, and to recommend corresponding data standards for the directory data in the data storage module as a basis for data governance. The NLU-HGSC module includes a domain adaptation module, a heterogeneous graph generation module, a semantic understanding module, and a graph computation module; The domain adaptive module is used to classify table names in the standard management module and the directory preprocessing module by domain. The heterogeneous graph generation module is used to convert the output data of the standard management module, the catalog preprocessing module, and the domain adaptation module into graph data. The semantic understanding module is used to transform each natural language node in the graph data into an abstract and computable mathematical form; The graph calculation module is used to perform similarity calculation on the graph data output by the semantic understanding module and output the standard data corresponding to the directory. The domain adaptation module employs a few-shot learning technique based on a multimodal large model, enabling it to quickly adapt to and learn the classification capabilities of new domains, thereby ensuring the accuracy and coverage of the classification results. In the similarity calculation process, the standard graph data and the metadata graph data are divided into four levels according to different node types: domain layer, table name layer, field layer and field type layer, and the similarity between each level is calculated one by one.

2. The domain-adaptive data governance standard recommendation system based on multimodal large models and NLU-HGSC as described in claim 1, characterized in that: The standard management module includes a standard collection module and a standard parsing module; The standard collection module is used to collect relevant standard documents from multiple sources and monitor the implementation and update status of standards; The standard parsing module is used to parse the elements of the standard file.

3. The domain-adaptive data governance standard recommendation system based on multimodal large models and NLU-HGSC according to claim 2, characterized in that: The standard document is parsed using layout analysis technology, and the table content in image format is parsed into text format using a multimodal large model.

4. The domain-adaptive data governance standard recommendation system based on multimodal large models and NLU-HGSC according to any one of claims 1 to 3, characterized in that: The directory preprocessing module includes a metadata acquisition module and a metadata cleaning module; The metadata acquisition module is used to acquire metadata information from the data storage module, including table name, field name, and field type; The metadata cleaning module is used to clean the acquired metadata information, including removing duplicate fields and invalid characters in field names.

5. The domain-adaptive data governance standard recommendation system based on multimodal large models and NLU-HGSC according to claim 1, characterized in that: After calculating the similarity at each level, the similarity values ​​at each level are combined by weighted summation to generate a comprehensive similarity score.

6. The domain-adaptive data governance standard recommendation system based on multimodal large models and NLU-HGSC according to claim 1, characterized in that: The data storage module includes the relational database MySQL and the graph database Neo4j.

7. The domain-adaptive data governance standard recommendation system based on multimodal large models and NLU-HGSC according to claim 6, characterized in that: The heterogeneous graph generation module uses the graph processing engine of the graph database Neo4j to map entities, relations, and attributes to the heterogeneous graph structure, and uses Neo4j's native graph algorithm and query language to realize the association and calculation of nodes and edges.

Images