A field recommendation based label generation system

Abstract

The application discloses a label generation system based on field recommendation, a data layer is used for establishing and maintaining a physical connection with various database systems and providing a unified connection interface; a collection layer is used for collecting database metadata, data samples and historical label records from a target data system through the data layer and pushing to a feature layer; the feature layer is used for vectorizing the data pushed by the collection layer; a recommendation layer is used for generating candidate labels based on the feature vectors generated by the feature layer; and a display layer is used for presenting the candidate labels generated by the recommendation layer to a user. The application can solve the problems that field information utilization stays in the level of security classification and standard management, lacks deep mining of field business semantics, lacks quantitative representation of semantic association between fields and labels, depends on manual rule configuration, has low recommendation accuracy and recall rate, lacks feedback loop, fails to fully utilize historical label annotation experience, has high coupling degree, lacks self-adaptability and scalability and the like.

Description

Technical Field

[0001] This invention relates to the field of data governance technology, and more specifically to a tag generation system based on field recommendation. Background Technology

[0002] A field is the smallest unit of data in a database table used to store a specific type of data, possessing metadata attributes such as field name, data type, length, default value, and comments. Data tags, on the other hand, are standardized descriptions of a specific characteristic or business semantic of a data asset, used for data classification, retrieval, governance, and analysis, and are the core identifiers in a data governance system. Current data tag generation schemes mainly have the following drawbacks: (1) The use of field information is limited to security classification, standard governance or access control. There is a lack of in-depth mining of field business semantics and data characteristics. It is impossible to automatically recommend appropriate data tags based on the field's own multi-dimensional attributes (such as name semantics, data type, value distribution, association relationship, business context).

[0003] (2) Existing tag generation technologies mostly rely on manual rule configuration or domain expert experience. When the database table structure is complex and the number of fields is large, the efficiency of manual tag annotation is extremely low, and problems such as inconsistent tag definitions, incomplete coverage, and semantic understanding deviation are likely to occur.

[0004] (3) The lack of a quantitative representation method for the semantic relationship between fields and tags makes it difficult to accurately match by calculating the similarity between field features and tag semantics, resulting in low accuracy and recall of tag recommendation.

[0005] (4) Most of them adopt one-time tag generation or static rule matching schemes, lacking a user feedback closed loop mechanism. They cannot dynamically optimize the recommendation model based on users' adoption, rejection and correction of recommended tags, resulting in the recommendation effect not being able to continuously improve with the use process.

[0006] (5) Failed to make full use of the labeling experience of similar fields in the same business domain or the same data table, and was unable to transfer the label knowledge of existing fields to new fields through collaborative recommendation, resulting in the duplication of label resources and waste of knowledge.

[0007] (6) The existing tag generation scheme is highly coupled with the specific business system. When the database structure changes or the business requirements are adjusted, the rules need to be reconfigured or the code needs to be modified. It lacks adaptability and scalability. Summary of the Invention

[0008] In view of the shortcomings of existing technologies, the present invention aims to provide a tag generation system based on field recommendation.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: A tag generation system based on field recommendation comprises a data layer, a collection layer, a feature layer, a recommendation layer, a presentation layer, and a feedback layer. The data layer is used to establish and maintain physical connections with various database systems, providing a unified connection interface. The collection layer is used to collect database metadata, data samples, and historical tag records from the target data system through the data layer and push them to the feature layer. The feature layer is used to vectorize the data pushed by the collection layer. The recommendation layer is used to generate candidate tags based on the feature vectors generated by the feature layer. The presentation layer is used to present the candidate tags generated by the recommendation layer to the user.

[0010] Furthermore, the acquisition layer includes a field metadata acquisition module, which includes a database connector, a sampler, a relation parser, and a document parser; The database connector is used to execute the metadata acquisition protocol through the database connection established in the data layer, read the system directory table of the target database system, and obtain the basic metadata of the fields, including field name, data type, length, precision, whether it is nullable, default value, and comments. The sampler is used to obtain actual value samples of a field by executing SQL sampling queries, which are then used for subsequent value distribution analysis and semantic inference. The relation resolver is used to obtain the relationships between fields by parsing the foreign key constraint definitions and ER diagrams of the target database system; The document parser is used to collect data dictionary documents or interface definition documents of the target database system and extract the business context information of the fields, including the business module to which the field belongs, the description of the business meaning, the data entry rules, and the business validation rules. The outputs of the database connector, sampler, relation parser, and document parser converge into a unified structured field metadata record, stored in JSON format, containing four field groups: basic metadata of the field, value samples, relationships, and business context.

[0011] Furthermore, the feature layer includes a field feature parsing module, which is used to perform deep parsing on the collected structured field metadata records, extract multi-dimensional features, and construct field semantic feature vectors. The extracted features specifically include semantic features, statistical features, relational features, and business context features. (1) Semantic feature extraction: Natural language processing is performed on field names and field annotations. First, named entity recognition and dependency parsing are used to extract core entity words and modifiers from field names. Then, a pre-trained domain word vector model is used to convert the core entity words and modifiers extracted from field names into fixed-dimensional word vectors. After that, average pooling is used to average all word vectors to obtain fixed-dimensional field semantic feature vectors. Finally, the final semantic feature vector f_semantic is obtained by weighting the semantic vector by word frequency-inverse document frequency. (2) Statistical feature extraction: Statistical feature analysis is performed on the field value samples, and all statistical features of the field are uniformly encoded into a fixed-dimensional statistical feature vector f_statistic; (3) Relationship feature extraction: Construct a field relationship graph based on the field association relationship, extract the relationship features of each field, including degree, betweenness centrality, and clustering coefficient; at the same time, extract the position features of the field within the same table; all relationship features obtained from the field extraction are represented by scalar values ​​and uniformly encoded into a fixed-dimensional relationship feature vector f_relation; (4) Business context feature extraction: Extract keywords and perform latent LDA topic modeling on the business meaning descriptions in the data dictionary to generate the business topic distribution vector of the field; parse the regular expressions and value range constraints in the business validation rules and convert them into structured rule feature vectors; extract the role features of the field in the data entry rules; map the business topic distribution vector, structured rule feature vector and role feature vector of the field into a fixed-dimensional business context feature vector f_business through concatenation and filling operations; The obtained semantic feature vector, statistical feature vector, relational feature vector, and business context feature vector are concatenated and normalized to obtain a unified field semantic feature vector F=[f_semantic, f_statistic, f_relation, f_business]. Each feature vector is mapped to the [0,1] interval through Min-Max normalization, and the dimension of the field semantic feature vector F is the sum of the fixed dimensions of each feature vector.

[0012] Furthermore, the recommendation layer includes a tag knowledge graph module, a multi-strategy recommendation engine module, and a collaborative recommendation module. The tag knowledge graph module is used to construct and maintain a semantic knowledge network of tags, providing a semantic foundation for matching fields and tags. The multi-strategy recommendation engine module is used to calculate the matching score between fields and tags based on the semantic feature vectors of fields and the tag knowledge graph, and then fuse them to generate the final recommendation result. The collaborative recommendation module utilizes the existing tag annotation records of fields and recommends tags for target fields based on the historical tag annotation experience of similar fields.

[0013] Furthermore, the tag knowledge graph module includes a tag ontology management submodule, a tag semantic vector generation submodule, a tag statistical feature template management submodule, a tag topic distribution generation submodule, and a graph storage and query submodule; The tag ontology management submodule is used to define the ontology structure of tags, including tag name, tag definition, tag category and tag relationship. Tag relationship includes tag hierarchy relationship, tag mutual exclusion relationship and tag association relationship; The label semantic vector generation submodule is used to generate a semantic embedding vector for each label using a pre-trained language model. The label vector and the field semantic feature vector are located in the same vector space, which facilitates the calculation of cosine similarity. The tag statistical feature template management submodule is used to automatically learn the statistical feature template for each tag based on historical annotation records: collect the statistical feature vectors of all fields of the tag that have been annotated, calculate the mean and standard deviation of each statistical feature, and form a statistical feature template; for new tags or tags with insufficient annotation records, the statistical feature template is preset manually by domain experts. The tag topic distribution generation submodule is used to generate business topic distribution vectors for tags based on the tag definition text using the LDA topic model, ensuring that the tag topic distribution and the field topic distribution are in the same vector space; The graph storage and query submodule is used to store tag knowledge graphs in a graph database or relational database graph extension mode. It supports path query, subgraph matching and semantic reasoning based on tag relationships; the output is tag semantic vector, tag hierarchy structure and tag association path.

[0014] Furthermore, the specific strategies employed by the multi-strategy recommendation engine module include semantic matching, statistical feature matching, and business relevance matching; Semantic matching: Calculate the cosine similarity between the semantic feature vector of the field and the semantic vector of the label to obtain the semantic matching score S_semantic; for polysemous words and abbreviations, a context-aware disambiguation mechanism is adopted, which combines the table name to which the field belongs and the context information in the comments to select the most appropriate semantic interpretation; Statistical feature matching: The statistical feature vector of the field is compared with the statistical feature template corresponding to the label. The Gaussian matching function is used to calculate the matching statistical feature matching score S_statistic: For each statistical feature item in the statistical feature template, the deviation between the actual value of the field and the mean of the statistical feature template is calculated. The matching score is 1 if the deviation is within the standard deviation of the statistical feature template, and 0 if the deviation exceeds 2 times the standard deviation. Linear interpolation is used if the deviation is in between. Business relevance matching: Matching based on the business context feature vector of the field and the business scenario constraint of the label: Calculate the KL divergence between the business topic distribution vector of the field and the business topic distribution vector of the label, and convert the KL divergence into a business relevance score in the range [0,1] using an exponential function: S_business = exp(-KL(P_field || P_label)), which is used as the business relevance matching score; By querying the historical frequency of how tags are applied to similar fields in the tag knowledge graph, as well as the popularity of the tag in the overall tag system, tags that are used frequently are given an additional frequency score S_frequency; Finally, a weighted summation method is used to merge the matching scores to obtain the comprehensive recommendation score of the field and the tag: S_total=w1×S_semantic+w2×S_statistic+w3×S_business+w4×S_frequency; w1, w2, w3, and w4 are weight parameters. For labels whose S_total exceeds the threshold T, they are included in the candidate recommendation list and output in descending order of score.

[0015] Furthermore, the weight parameters w1, w2, w3, and w4 are determined in the following way: During the initial deployment of the system, default weights are preset manually. After accumulating enough manually labeled samples, a grid search is performed in the weight space with the F1 score as the optimization objective. The step size is 0.05, and the search range is [0.05, 0.85]. The constraints are that the sum of the weights is 1 and each weight is not less than 0.05. The weight combination that maximizes the F1 score is selected as the current optimal weight. When the number of user feedback samples exceeds 500, the system switches to an online optimization mode based on gradient descent.

[0016] Furthermore, the specific execution process of the collaborative recommendation module is as follows: A1. Field Similarity Calculation: Based on the semantic feature vector of the fields, calculate the comprehensive similarity Sim(Fi, Fj) between each labeled field in the labeled field set and the target field of the tag to be recommended; the similarity calculation formula is: Sim(Fi, Fj)=0.5×cos(f_semantic_i, f_semantic_j)+0.3×(1-|f_statistic_i-f_statistic_j|)+0.1×(1-|f_relation_i-f_relation_j|)+0.1×(1-|f_business_i-f_business_j|), where cos represents cosine similarity, and |·| represents the L1 norm normalized distance of the vectors; A2. Neighbor retrieval: For the target field of the tag to be recommended, retrieve the top-K neighbor fields with the highest similarity in the set of labeled fields to form a neighbor set N(F_target). Only fields with similarity exceeding the threshold θ are considered as valid neighbors. A3. Weighted summarization of labels: The labeled labels of the neighbor fields in the neighbor set N(F_target) are weighted and summed according to their frequency of occurrence to generate a collaborative recommendation score S_coll=Σ(Li∈N) Sim(F_target, Fi)×I(Li, L) / Σ(Li∈N)Sim(F_target, Fi), where I(Li, L) is the indicator function and the denominator is the normalization factor to ensure that the value range of the collaborative recommendation score is within the interval [0,1]. A4. Fusion Recommendation: The collaborative recommendation score is further fused with the comprehensive score of the multi-strategy recommendation engine. The final recommendation score is S_final=α×S_total+(1-α)×S_coll, where α is the fusion coefficient.

[0017] Furthermore, the presentation layer includes a tag recommendation display module, which is used to present the recommendation results to the user in a visual way and supports user interaction. (1) Recommendation list display: Displays the field list of the current data table. After selecting a field, the top-N recommended tags for that field are displayed in the middle recommendation tag display area. Each recommendation tag displays the tag name, recommendation score, recommendation reason, tag category and related tags. (2) Tag details display: The tag details area displays the complete definition, applicable scenarios, associated tags, historical usage count and typical application cases of the currently selected tag; when the user clicks on a recommended tag, detailed information is displayed in real time in the right area; (3) Batch recommendation display: Supports batch display of recommendation tags for all fields under the data table by data table dimension, presented in tabular form, and supports filtering and sorting by field type, recommendation score, tag category and other dimensions; (4) User interaction operations: Users can perform three operations on each recommended tag: “accept”, “reject”, and “correct”. Each recommended tag card contains an accept / reject / correct button. The user’s operation behavior is recorded in the feedback database for subsequent model optimization.

[0018] Furthermore, the system also includes a feedback layer, which is used to dynamically adjust model parameters based on user operations to form a complete closed-loop optimization process; the feedback layer includes a user feedback optimization module, which is used to record all user interactions with recommended tags and dynamically adjust recommendation strategy parameters and weights based on feedback data; the user feedback optimization module includes a feedback data collection submodule, a statistical analysis submodule, a parameter optimization submodule, and a tag knowledge graph update submodule; The feedback data collection submodule records user operation data, including user interaction field ID, recommendation tag ID, user operation type, operation timestamp, user ID, and corrected tag content. Among them, adoption operations are recorded as positive feedback, rejection operations are recorded as negative feedback, and correction operations are recorded as partially positive feedback. The feedback data collection submodule aggregates user operation data into the feedback database. The statistical analysis submodule is used to periodically calculate the accuracy, recall, F1 score, and user adoption rate of each recommendation strategy; using the tags adopted by users as the ground truth and the tags recommended by the system as the predicted values; and comparing the recommendation effects under different weight parameter configurations through A / B testing: users are randomly divided into a control group and an experimental group, and the difference in adoption rate between the two groups within a week is compared. If the adoption rate of the experimental group increases by more than 5% and is statistically significant, the candidate weight is switched to the current weight. The parameter optimization submodule automatically increases the weight w1 when the recommended label adoption of the semantic matching strategy is consistently higher than that of other strategies, and decreases it otherwise. The weight adjustment adopts a constrained gradient descent method, with the optimization objective being to maximize the F1 value. The loss function is L=1-F1+λ×(Σwi-1)², where λ is the constraint penalty coefficient, ensuring that the sum of the weights approaches 1. At the same time, the projection operation ensures that each weight wi is not lower than the minimum threshold of 0.05. The learning rate is initially 0.01, and decays to 0.9 times the original rate every 100 iterations. The tag knowledge graph update submodule is used to add semantic association edges between two tags in the tag knowledge graph and update the tag vector when a user frequently changes a tag to another tag; when a tag has not been adopted for a long time, its activity weight in the graph is reduced.

[0019] The beneficial effects of this invention are as follows: (1) By automatically collecting basic metadata of fields in the database system catalog table (field name, data type, length, precision, whether it is empty, default value, comment), executing SQL sampling queries to obtain actual value samples of fields, parsing database foreign key constraints to obtain field association relationships, and collecting data dictionary of business system to obtain field business context information, the full-scale automated collection of multidimensional metadata of fields is realized, which solves the problem that the use of field information in the existing technology is limited to the security classification and standard governance level and lacks in-depth mining of field business semantics.

[0020] (2) By performing natural language processing on field names and annotations to extract semantic features (named entity recognition, dependency parsing, word vector conversion, word frequency-inverse document frequency weighting), performing statistical analysis on field value samples to extract statistical features (mean, standard deviation, quantiles, regular pattern matching degree, entropy), extracting relational features (degree, betweenness centrality, clustering coefficient) based on field association graphs, and performing latent Dirichlet Allocation (LDA) topic modeling on data dictionary business descriptions to extract business context features, the automatic construction and normalization fusion of multidimensional feature vectors of fields are realized, solving the problem that existing technologies lack a method for quantitatively representing the semantic association between fields and labels.

[0021] (3) By constructing a tag knowledge graph to express the hierarchical relationship, semantic association and business constraints of tags, and designing a multi-strategy recommendation engine to integrate the weighted fusion score of four strategies: semantic matching (cosine similarity), statistical feature matching (template comparison), business relevance matching (Kullback-Leibler Divergence (KL divergence)) and historical usage frequency, a multi-dimensional matching recommendation with high accuracy and high recall between fields and tags is achieved, solving the problem that existing tag generation technology relies on manual rule configuration and has low recommendation accuracy and recall.

[0022] (4) By calculating the comprehensive similarity between fields based on the semantic feature vector of the fields, constructing a neighbor set of similar fields, weighting and summarizing the labeled tags of neighbor fields according to the frequency of occurrence to generate a collaborative recommendation score, and merging it with the multi-strategy recommendation score to generate the final recommendation score, the tag knowledge transfer and reuse of similar fields in the same business domain or the same data table can be realized, solving the problem that the existing technology has failed to make full use of historical tag labeling experience and caused the duplication of tag resources.

[0023] (5) By using standardized field metadata collection protocols such as Java Database Connectivity (JDBC) / Open Database Connectivity (ODBC), unified field semantic feature vector representation, configurable tag knowledge graph, and parameterized multi-strategy recommendation engine, the tag generation system can be decoupled from specific business systems, enabling the system to adapt to changes in database structure, adjustments in business requirements, and evolution of the tag system, thus solving the problems of high coupling, lack of adaptability, and scalability of existing tag generation solutions.

[0024] (6) By establishing a user feedback closed-loop mechanism, the user's adoption, rejection and correction behavior of recommended tags are recorded. The accuracy (Precision@N), recall (Recall@N), F1 score and user adoption rate of each recommendation strategy are calculated regularly. The weight of multi-strategy fusion is dynamically adjusted by using the constrained gradient descent method. The tag association relationship and activity weight are updated in the tag knowledge graph to achieve continuous self-optimization of the recommendation model and solve the problem that the existing technology lacks a feedback closed loop and the recommendation effect cannot be continuously improved with the use process. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the system structure of Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the field metadata collection module in Embodiment 1 of the present invention. Figure 3 This is a schematic diagram illustrating the implementation principle of the field feature parsing module in Embodiment 1 of the present invention. Figure 4 This is a schematic diagram illustrating the implementation principle of the tag knowledge graph module in Embodiment 1 of the present invention. Figure 5 This is a schematic diagram illustrating the implementation principle of the multi-strategy recommendation engine module in Embodiment 1 of the present invention. Figure 6 This is a schematic diagram illustrating the implementation principle of the collaborative recommendation module in Embodiment 1 of the present invention. Figure 7 This is a schematic diagram of the interactive interface generated by the tag recommendation and display module in Embodiment 1 of the present invention; Figure 8 This is a schematic diagram of the user feedback optimization module in Embodiment 1 of the present invention; Figure 9 This is a flowchart of the implementation method of Embodiment 2 of the present invention. Detailed Implementation

[0026] The present invention will be further described below with reference to the accompanying drawings. It should be noted that this embodiment is based on the present technical solution and provides detailed implementation methods and specific operation processes, but the protection scope of the present invention is not limited to this embodiment.

[0027] Example 1 This embodiment provides a tag generation system based on field recommendation. By deeply mining the multi-dimensional metadata features of database fields, a field-tag recommendation matrix is ​​constructed. Combined with a multi-strategy recommendation engine, collaborative recommendation, and user feedback closed-loop mechanism, intelligent, automated, and continuously optimized generation of data tags is achieved.

[0028] like Figure 1As shown, the field-based tag generation system is divided into a data layer, a collection layer, a feature layer, a recommendation layer, a display layer, and a feedback layer from top to bottom.

[0029] The data layer is used to establish and maintain physical connections with various database systems, providing a unified connection interface (JDBC / ODBC driver); the acquisition layer is used to collect database metadata, data samples, and historical tag records from the target data system through the data layer and push them to the feature layer; the feature layer is used to vectorize the data pushed by the acquisition layer; the recommendation layer is used to generate candidate tags based on the feature vectors generated by the feature layer; the presentation layer is used to present the candidate tags generated by the recommendation layer to the user; and the feedback layer is used to dynamically adjust model parameters according to user operations, forming a complete closed-loop optimization process.

[0030] The acquisition layer includes a field metadata acquisition module, such as... Figure 2 As shown, the field metadata acquisition module executes the metadata acquisition protocol through the established database connection (JDBC / ODBC driver) in the data layer to collect metadata information of table structure and fields from the target database system. The field metadata acquisition module includes a database connector, a sampler, a relation parser, and a document parser.

[0031] Database connectors are used to read the system catalog tables of a target database system and obtain basic metadata of fields, including field name, data type, length, precision, whether it is nullable, default value, and comments, via JDBC / ODBC connection protocols or native database APIs (such as MySQL's INFORMATION_SCHEMA, Oracle's ALL_TAB_COLUMNS, PostgreSQL's pg_catalog, etc.).

[0032] The sampler is used to obtain actual value samples of a field by executing SQL sampling queries (such as SELECT COLUMN_NAME FROM TABLE_NAMELIMIT N), which are then used for subsequent value distribution analysis and semantic inference. The sampling strategy is stratified random sampling. For tables with less than 100,000 data entries, full sampling is performed. For tables with more than 100,000 data entries, stratified sampling is performed according to data type, with at least 1,000 samples extracted from each type.

[0033] The relation resolver is used to obtain the relationships between fields by parsing the foreign key constraint (FOREIGN KEY) definition and ER relationship diagram of the target database system, including primary and foreign key relationships, relationships between fields within the same table, and cross-table relationship paths.

[0034] The document parser is used to collect data dictionary documents or interface definition documents of the target database system and extract the business context information of the fields, including the business module to which the field belongs, the description of the business meaning, the data entry rules, the business validation rules, etc.

[0035] The outputs of the database connector, sampler, relation parser, and document parser converge into a unified structured field metadata record, stored in JSON format, containing four field groups: basic metadata of the field, value samples, association relationships, and business context.

[0036] In this embodiment, the feature layer includes a field feature parsing module, such as... Figure 3 As shown, the field feature parsing module is used to perform deep parsing on the collected structured field metadata records, extract multi-dimensional features, and construct field semantic feature vectors. The extracted features specifically include semantic features, statistical features, relational features, and business context features: (1) Semantic feature extraction: Natural language processing is performed on the field names (e.g., "customer_name" is decomposed into the core entity word "customer" and the modifier word "name") and the field annotations. First, the core entity words and modifier words in the field names are extracted using Named Entity Recognition (NER) and Dependency Parsing. Then, a pre-trained domain word vector model (e.g., Word2Vec or BERT) is used to convert the core entity words and modifier words extracted from the field names into fixed-dimensional word vectors. Subsequently, average pooling is used to average all word vectors to obtain a fixed-dimensional field semantic feature vector. Finally, the semantic vector is weighted by term frequency-inverse document frequency (TF-IDF) to obtain the final semantic feature vector f_semantic.

[0037] (2) Statistical Feature Extraction: Perform statistical feature analysis on the field's value samples. For numeric fields, the statistical features to be analyzed include mean, standard deviation, quantiles, extreme values, distribution skewness, and kurtosis. For character fields, the statistical features to be analyzed include average length, maximum length, encoding type, and regular expression pattern matching degree. For enumerated fields, the statistical features to be analyzed include the number of value categories, frequency distribution of each value, and entropy value. For date fields, the time range, time granularity, and periodicity features are also analyzed. All statistical features of the field are uniformly encoded into a fixed-dimensional statistical feature vector f_statistic, where feature dimensions that are not applicable to the current field type are filled with a placeholder value of -1, and applicable feature dimensions are filled with the actual calculated values.

[0038] (3) Relationship feature extraction: Based on the field association relationships, a field relationship graph is constructed, and the relationship features of each field are extracted, including degree (number of associated fields), betweenness centrality (the bridging role in the association path), and clustering coefficient (the degree of closeness between the field and the associated field). At the same time, the positional features of the field within the same table are extracted (such as whether it is a primary key, whether it belongs to a composite index, and its ordinal position in the table). All relationship features obtained from the field extraction are represented by scalar values ​​and uniformly encoded into a fixed-dimensional relationship feature vector f_relation.

[0039] (4) Business context feature extraction: Keyword extraction and potential LDA topic modeling are performed on the business meaning descriptions in the data dictionary to generate the business topic distribution vector of the field; the regular expressions and value range constraints in the business validation rules are parsed and converted into structured rule feature vectors; the role features of the field in the data entry rules (such as entry field, calculated field, derived field, audit field) are extracted. The business topic distribution vector, structured rule feature vector and role feature vector of the field are mapped to a fixed-dimensional business context feature vector f_business through concatenation and filling operations.

[0040] The obtained semantic feature vector, statistical feature vector, relational feature vector, and business context feature vector are concatenated and normalized to obtain a unified field semantic feature vector F = [f_semantic, f_statistic, f_relation, f_business]. Each feature vector is mapped to the [0,1] interval through Min-Max normalization. The dimension of the field semantic feature vector F is the sum of the fixed dimensions of each feature vector (e.g., 300+50+20+20=390 dimensions).

[0041] like Figure 1 As shown, the recommendation layer includes a tag knowledge graph module, a multi-strategy recommendation engine module, and a collaborative recommendation module. The tag knowledge graph module is used to construct and maintain a semantic knowledge network of tags, providing a semantic foundation for matching fields with tags. For example... Figure 4 As shown, the tag knowledge graph module includes a tag ontology management submodule, a tag semantic vector generation submodule, a tag statistical feature template management submodule, a tag topic distribution generation submodule, and a graph storage and query submodule.

[0042] The tag ontology management submodule is used to define the tag ontology structure, including tag name, tag definition, tag category (such as data security tag, business attribute tag, quality level tag, source system tag, etc.), and tag relationships. Tag relationships include tag hierarchy (such as "Personal Information" being a superior tag to "User Basic Information"), tag mutual exclusion (such as "Highly Sensitive" and "Public" being mutually exclusive), and tag association (such as "ID Number" being related to "Real-Name Authentication"). The tag ontology management submodule takes as input manually entered tag definitions and externally imported standard tag systems.

[0043] The tag semantic vector generation submodule generates semantic embedding vectors for each tag using a pre-trained language model. The tag vectors and field semantic feature vectors reside in the same vector space, facilitating the calculation of cosine similarity. Simultaneously, it enhances the semantic coverage of tags through tag synonym expansion and near-synonym association. Specifically, it calculates the semantic similarity between tag names based on a pre-trained language model (such as BERT or Word2Vec). When the semantic similarity (cosine similarity) between two tag names exceeds the threshold θ_synonym (default 0.85), they are determined to be near-synonyms; when the similarity exceeds the threshold θ_equivalent (default 0.92), they are determined to be near-synonyms. The tag semantic vector generation submodule automatically establishes a set of synonyms and near-synonyms for each tag. Therefore, during semantic matching recommendations, while matching a field with tag A, it also automatically considers matching its synonyms and near-synonyms with tag A, thereby expanding the semantic coverage of the tags.

[0044] The tag statistical feature template management submodule is used to automatically learn the statistical feature template for each tag based on historical annotation records. Specifically, it collects the statistical feature vectors of all fields that have been annotated for the tag, calculates the mean and standard deviation of each statistical feature, and forms a statistical feature template (with the mean ± 2 times the standard deviation as the value range). For new tags or tags with insufficient annotation records (less than 5), domain experts manually preset the statistical feature template (e.g., the template for the "ID number" tag is: character length 18 or 15, conforms to a specific regular expression pattern, character type, and value uniqueness > 95%).

[0045] The tag topic distribution generation submodule is used to generate a business topic distribution vector for tags based on the tag definition text (i.e., the tag name, tag description, and business description text filled in by domain experts when entering tags, collectively referred to as tag definition text). This is achieved through the LDA topic model (using the same number of topics K and model parameters as the field feature parsing), ensuring that the tag topic distribution and the field topic distribution are in the same vector space, which facilitates the calculation of KL divergence.

[0046] The graph storage and query submodule stores tag knowledge graphs using a graph database (such as Neo4j) or a graph extension pattern for relational databases. It supports path queries, subgraph matching, and semantic reasoning based on tag relationships. For example, when a field matches the tag "phone number," the system can automatically associate it with the tags "contact information" and "personal information" through graph reasoning. The output includes tag semantic vectors, tag hierarchical structures, and tag association paths.

[0047] Furthermore, the multi-strategy recommendation engine module is used to calculate the matching score between fields and tags based on field semantic feature vectors and tag knowledge graphs, and then fuse them to generate the final recommendation result. Specific strategies include semantic matching, statistical feature matching, and business relevance matching. Semantic matching: The cosine similarity between the semantic feature vector of a field and the semantic vector of a label is calculated to obtain a semantic matching score S_semantic. For polysemous words and abbreviations, a context-aware disambiguation mechanism is used, combining the table name to which the field belongs and the contextual information in the comments to select the most appropriate semantic interpretation.

[0048] Statistical feature matching: The statistical feature vector of the field is compared with the statistical feature template corresponding to the label (for example, the statistical feature template of the "ID number" label is: character length 18 or 15, conforms to a specific regular expression pattern, character type, and high value uniqueness). The matching statistical feature matching score S_statistic is calculated using the Gaussian matching function: For each statistical feature item in the statistical feature template, the deviation between the actual value of the field and the mean of the statistical feature template is calculated. The matching score is 1 if the deviation is within the standard deviation of the statistical feature template, and 0 if the deviation exceeds 2 times the standard deviation. Linear interpolation is used between the two.

[0049] Business relevance matching: Matching is performed based on the business context feature vector of a field and the business scenario constraints of the label. For example, the business module to which the "order amount" field belongs is "order management," and the business scenario to which the label "transaction amount" belongs is "e-commerce transaction." When the KL divergence of the business theme distributions of the two is small, the business relevance S_business is high. Specifically, the KL divergence between the business theme distribution vector of the field and the business theme distribution vector of the label is calculated, and the KL divergence is converted into a business relevance score in the interval [0,1] using an exponential function: S_business = exp(-KL(P_field || P_label)), which is used as the business relevance matching score. For example, if the KL divergence between the business theme distribution of a certain field and the label "transaction amount" is 0.08, then S_business = exp(-0.08) ≈ 0.923.

[0050] In addition, in this embodiment, by querying the historical frequency of a tag being applied to similar fields in the tag knowledge graph, and the popularity of the tag in the overall tag system, tags that are used frequently are given an extra frequency score S_frequency.

[0051] Specifically, in this embodiment, a weighted summation method is used to fuse the matching scores to obtain the comprehensive recommendation score of the field and the tag, S_total=w1×S_semantic+w2×S_statistic+w3×S_business+w4×S_frequency.

[0052] The weight parameters w1, w2, w3, and w4 are determined in the following way: During initial system deployment, manually preset default weights (w1=0.4, w2=0.3, w3=0.2, w4=0.1) are used. Once sufficient manually labeled samples (at least 100 field-label pairs, with accepted records as positive samples and rejected records as negative samples) are accumulated, a grid search is performed in the weight space with the F1 score as the optimization objective. The step size is 0.05, and the search range is [0.05, 0.85]. The constraints are that the sum of the weights is 1 and each weight is not less than 0.05. The weight combination that maximizes the F1 score is selected as the current optimal weight. When the number of user feedback samples exceeds 500, the system switches to an online optimization mode based on gradient descent. For tags with scores exceeding a threshold T (default 0.65, determined based on the 25th percentile of the S_total distribution of accepted tags in historical labeled data, ensuring that at least 75% of the tags in the recommendation list have high credibility), they are included in the candidate recommendation list and output in descending order of score.

[0053] In this embodiment, as Figure 6 As shown, the collaborative recommendation module utilizes existing field labeling records and recommends labels for the target field based on historical labeling experience of similar fields. The specific implementation is as follows: A1. Field Similarity Calculation: Based on the semantic feature vectors of the fields, calculate the comprehensive similarity Sim(Fi, Fj) between each labeled field in the labeled field set and the target field of the tag to be recommended. The similarity calculation formula is: Sim(Fi, Fj) = 0.5×cos(f_semantic_i, f_semantic_j)+0.3×(1-|f_statistic_i-f_statistic_j|)+0.1×(1-|f_relation_i-f_relation_j|)+0.1×(1-|f_business_i-f_business_j|), where cos represents cosine similarity, and |·| represents the L1 norm normalized distance of the vectors. The semantic feature dimension has the highest weight (0.5), followed by the statistical feature dimension (0.3), while the relational feature and business context feature dimensions have lower weights (0.1 each).

[0054] A2. Neighbor Retrieval: For the target field of the tag to be recommended, retrieve the top-K neighbor fields with the highest similarity from the set of labeled fields. K is set to 5 by default, forming the neighbor set N(F_target). Only fields with similarity exceeding the threshold θ (default 0.60) are considered as valid neighbors. Fields with similarity below the threshold are not included in the neighbor set to ensure the quality of neighbors.

[0055] A3. Weighted Tag Summarization: The labeled tags of the neighbor fields in the neighbor set N(F_target) are weighted and summed according to their frequency of occurrence to generate a collaborative recommendation score S_coll=Σ(Li∈N) Sim(F_target, Fi)×I(Li, L) / Σ(Li∈N)Sim(F_target, Fi), where I(Li, L) is an indicator function (takes 1 when the neighbor Fi has been labeled with tag L, otherwise takes 0), and the denominator is a normalization factor (the sum of the similarities of all effective neighbors), ensuring that the collaborative recommendation score is within the range of [0,1].

[0056] A4. Fusion Recommendation: The collaborative recommendation score is further fused with the comprehensive score of the multi-strategy recommendation engine. The final recommendation score S_final = α × S_total + (1-α) × S_coll, where α is the fusion coefficient (default 0.7), which can be dynamically adjusted based on user feedback. When the number of effective neighbors in the labeled field library is insufficient (less than 3), the weight of collaborative recommendation is reduced, setting α = 0.9, making the system more reliant on the multi-strategy recommendation engine's score. When the number of effective neighbors is sufficient, the weight of collaborative recommendation is increased, setting α = 0.6, fully leveraging the transfer effect of historical labeling experience.

[0057] In this embodiment, as Figure 1As shown, the presentation layer includes a tag recommendation display module, which presents the recommendation results to the user in a visual way and supports user interaction. Figure 7 As shown, it specifically includes: (1) Recommendation List Display: The field list area on the left displays the field list of the current data table. After selecting a field, the middle recommendation tag display area displays the Top-N (5 by default) recommendation tags for that field. Each recommendation tag displays the tag name, recommendation score, recommendation reason (such as "semantic similarity 0.85", "statistical feature matching degree 0.72"), tag category, and associated tags. Recommendation tags are displayed intuitively by recommendation score using color depth or heat bars.

[0058] (2) Tag details display: The tag details area on the right displays the complete definition, applicable scenarios, associated tags, historical usage count, and typical application cases of the currently selected tag. When a user clicks on a recommended tag, detailed information is displayed in real time in the right area.

[0059] (3) Batch recommendation display: Supports batch display of recommended tags for all fields under the data table by data table dimension, presented in tabular form, and supports filtering and sorting by field type, recommendation score, tag category and other dimensions.

[0060] (4) User interaction: Users can perform three operations on each recommended tag: "Accept" (use the tag directly), "Reject" (remove from the recommendation list), and "Correct" (modify the recommended tag into a new tag). Each recommended tag card contains an accept / reject / correct button, and user actions are recorded in the feedback database for subsequent model optimization.

[0061] like Figure 1 As shown, the feedback layer includes a user feedback optimization module, which records all user interactions with recommended tags and dynamically adjusts recommendation strategy parameters and weights based on the feedback data. Figure 8 As shown, the user feedback optimization module includes a feedback data collection submodule, a statistical analysis submodule, a parameter optimization submodule, and a tag knowledge graph update submodule.

[0062] The feedback data collection submodule records user action data, including user interaction field ID, recommendation tag ID, user action type (accept / reject / correction), action timestamp, user ID, and corrected tag content (if it's a correction action). Acceptance actions are recorded as positive feedback, rejection actions as negative feedback, and correction actions as partially positive feedback (original tags are weakly positive samples, corrected tags are strongly positive samples). The feedback data collection submodule aggregates user action data into the feedback database.

[0063] The statistical analysis submodule is used to periodically (e.g., daily) calculate the precision (Precision@N), recall (Recall@N), F1 score, and user adoption rate of each recommendation strategy. The specific results are obtained through ablation experiments. Only a single strategy is enabled each time, such as enabling only the semantic matching strategy, and the weights of the other strategies are set to 0.

[0064] The user-adopted tags are used as the ground truth, and the system-recommended tags are used as the predicted values. A / B testing is conducted to compare the recommendation performance under different weight parameter configurations: users are randomly divided into a control group (using the current weights) and an experimental group (using candidate weights). The difference in adoption rates between the two groups over one week is compared. If the adoption rate of the experimental group increases by more than 5% and is statistically significant (p<0.05), the candidate weights are switched to the current weights.

[0065] The parameter optimization submodule automatically increases the weight w1 when the adoption rate of the recommended tags for the semantic matching strategy (determined separately through ablation experiments, i.e., the proportion of recommended tags adopted by users when only the semantic matching strategy is enabled) is consistently higher than that of other strategies; conversely, it decreases the weight w1. Weight adjustment employs constrained gradient descent, with the optimization objective being to maximize the F1 score. The loss function is L = 1 - F1 + λ × (Σwi - 1)², where λ is the constraint penalty coefficient (default 1000), ensuring the sum of weights approaches 1. Simultaneously, a projection operation ensures that each weight wi is not lower than the minimum threshold of 0.05. The initial learning rate is 0.01, decaying to 0.9 times the original rate every 100 iterations.

[0066] The tag knowledge graph update submodule is used to add semantic association edges between two tags in the tag knowledge graph and update the tag vector when a user frequently changes a tag to another tag. When a tag has not been adopted for a long time, its activity weight in the graph is reduced.

[0067] Example 2 This embodiment provides an application example of the system described in Embodiment 1.

[0068] A city-level population information management system contains a table named "person_info," which stores basic information on the city's permanent and transient populations. The table includes the following fields: id_card (ID card number), name, gender, nation, birth_date, hukou_addr (registered address), marital_status (marital status), and phone (contact number). The system contains approximately 12 million records. Field comments were entered by different developers at different times, resulting in inconsistent naming and incomplete comments. Figure 9The specific process is as follows: Step 1: The field metadata collection module connects to the Oracle database via JDBC and reads the field metadata of the person_info table from the ALL_TAB_COLUMNS table. The collected data includes: id_card: DATA_TYPE VARCHAR2, DATA_LENGTH 18, COMMENTS "ID card number"; hukou_addr: DATA_TYPE VARCHAR2, DATA_LENGTH 200, COMMENTS "registered address"; nation: DATA_TYPE VARCHAR2, DATA_LENGTH 20, COMMENTS "ethnicity, refer to GB / T 3304-1991"; phone: DATA_TYPE VARCHAR2, DATA_LENGTH 11, COMMENTS "contact number". The sampler performs stratified sampling on the phone field, extracting 10,000 samples. It was found that the value length is 11 digits, with the first three digits starting with "13", "15", "17", "18", or "19", conforming to the regular expression pattern of Chinese mobile phone numbers. The relation resolver found that the id_card field has a PRIMARY KEY constraint, and the phone field has a FOREIGN KEY association with the phone field in the address table.

[0069] Step 2: The field feature parsing module extracts multi-dimensional features from each field, which are then processed by four parallel processing channels and converged into a unified feature vector. Taking the id_card field as an example: In semantic feature extraction, the field name "id_card" is segmented into ["id", "card"] using camelCase. Combined with the annotation "ID card number", after Jieba segmentation and expansion using the public security domain dictionary, the semantic vector has a high projection value in the semantic direction of tags such as "ID card number", "citizen identity number", and "resident ID card". In statistical feature extraction, an analysis of 5,000 samples revealed that the id_card field has a fixed value length of 18 digits (with a small number of 15-digit old ID cards), a regular expression pattern matching degree of 0.999, a value uniqueness of 99.9% (with a very small number of duplicate entries), is a character type, and has an entropy value of 6.2. In relational feature extraction, the id_card field has a degree of 3 (associated with the phone, address, and family tables), a betweenness centrality of 0.15, and a clustering coefficient of 0.33. In business context feature extraction, the data dictionary describes "citizen identity number, used to uniquely identify permanent residents and migrant populations", and the probabilities of LDA topics distributed on the topics of "basic population information" and "identity authentication" are 0.45 and 0.30, respectively. The outputs of the four channels are concatenated and normalized by the feature fusion unit to generate a 390-dimensional semantic feature vector F_id_card.

[0070] Step 3: The multi-strategy recommendation engine module calculates the matching score between the id_card field and each tag through four parallel strategy calculation channels. In the semantic matching channel, the semantic similarity S_semantic with the "ID card number" tag is 0.96 (the field annotation "ID card number" completely matches the tag name), and the semantic similarity with the "citizen identity number" tag is 0.91; in the statistical feature matching channel, the statistical matching degree S_statistic with the "ID card number" tag is 0.98 (complete match of regular expression pattern, consistent length, and high uniqueness of value), and the statistical matching degree with the "document number" tag is 0.85; in the business relevance matching channel, the KL divergence between the "population basic information" topic distribution of the id_card field and the topic distribution of the "ID card number" tag is 0.08, and after conversion, S_business=0.92; in the historical frequency channel, the "ID card number" tag is used 156 times in the labeled field library, and S_frequency=0.85. The overall score S_total = 0.4 × 0.96 + 0.3 × 0.98 + 0.2 × 0.92 + 0.1 × 0.85 = 0.947.

[0071] Step 4: The collaborative recommendation module searches the labeled field library and finds that the "citizen_id" field in the population information systems of other branches of the bureau has a similarity of 0.93 with the "id_card" field (highly similar semantic features, only different field names). Furthermore, this field is labeled with both "ID card number" and "citizen identity number". Therefore, the "ID card number" label receives a collaborative recommendation score of S_coll = 0.93 × 1 = 0.93, and the "citizen identity number" label receives S_coll = 0.93 × 1 = 0.93. The final fusion recommendation score is S_final = 0.7 × 0.947 + 0.3 × 0.93 = 0.942.

[0072] Step 5: The tag recommendation and display module shows the data administrator the Top-5 recommended tags for the id_card field: 1. "ID Card Number" (0.942), 2. "Citizen Identity Number" (0.928), 3. "Personal Identifier" (0.72), 4. "ID Card Number" (0.68), 5. "Sensitive Data" (0.65). The data administrator accepts the tags "ID Card Number," "Citizen Identity Number," and "Sensitive Data," and rejects "Personal Identifier" and "ID Card Number."

[0073] Step 6: The user feedback optimization module records the above operations and detects that the adoption rates of recommended tags using both semantic matching and statistical feature matching strategies are high. It automatically adjusts the weights to w1=0.45, w2=0.35, w3=0.15, and w4=0.05. Simultaneously, the tag knowledge graph update module establishes a direct association edge (weight 0.85) between "ID number" and "sensitive data," automatically improving the recommendation ranking of the "sensitive data" tag when recommending similar fields to other systems.

[0074] For those skilled in the art, various corresponding changes and modifications can be made based on the above technical solutions and concepts, and all such changes and modifications should be included within the protection scope of the claims of this invention.

Claims

1. A tag generation system based on field recommendation, characterized in that, It includes a data layer, a collection layer, a feature layer, a recommendation layer, a presentation layer, and a feedback layer. The data layer is used to establish and maintain physical connections with various database systems and provides a unified connection interface. The collection layer is used to collect database metadata, data samples, and historical tag records from the target data system through the data layer and push them to the feature layer. The feature layer is used to vectorize the data pushed by the acquisition layer. The recommendation layer is used to generate candidate labels based on the feature vectors generated by the feature layer; the presentation layer is used to present the candidate labels generated by the recommendation layer to the user.

2. The method according to claim 1, characterized in that, The acquisition layer includes a field metadata acquisition module, which includes a database connector, a sampler, a relation parser, and a document parser. The database connector is used to execute the metadata acquisition protocol through the database connection established in the data layer, read the system directory table of the target database system, and obtain the basic metadata of the fields, including field name, data type, length, precision, whether it is nullable, default value, and comments. The sampler is used to obtain actual value samples of a field by executing SQL sampling queries, which are then used for subsequent value distribution analysis and semantic inference. The relation resolver is used to obtain the relationships between fields by parsing the foreign key constraint definitions and ER diagrams of the target database system; The document parser is used to collect data dictionary documents or interface definition documents of the target database system and extract the business context information of the fields, including the business module to which the field belongs, the description of the business meaning, the data entry rules, and the business validation rules. The outputs of the database connector, sampler, relation parser, and document parser converge into a unified structured field metadata record, stored in JSON format, containing four field groups: basic metadata of the field, value samples, relationships, and business context.

3. The system according to claim 2, characterized in that, The feature layer includes a field feature parsing module, which performs deep parsing on the collected structured field metadata records, extracts multi-dimensional features, and constructs field semantic feature vectors. The extracted features specifically include semantic features, statistical features, relational features, and business context features. (1) Semantic feature extraction: Natural language processing is performed on field names and field annotations. First, named entity recognition and dependency parsing are used to extract core entity words and modifiers from field names. Then, a pre-trained domain word vector model is used to convert the core entity words and modifiers extracted from field names into fixed-dimensional word vectors. After that, average pooling is used to average all word vectors to obtain fixed-dimensional field semantic feature vectors. Finally, the final semantic feature vector f_semantic is obtained by weighting the semantic vector by word frequency-inverse document frequency. (2) Statistical feature extraction: Statistical feature analysis is performed on the field value samples, and all statistical features of the field are uniformly encoded into a fixed-dimensional statistical feature vector f_statistic; (3) Relationship feature extraction: Construct a field relationship graph based on the field association relationship, extract the relationship features of each field, including degree, betweenness centrality, and clustering coefficient; at the same time, extract the position features of the field within the same table; all relationship features obtained from the field extraction are represented by scalar values ​​and uniformly encoded into a fixed-dimensional relationship feature vector f_relation; (4) Business context feature extraction: Extract keywords and perform latent LDA topic modeling on the business meaning descriptions in the data dictionary to generate the business topic distribution vector of the field; parse the regular expressions and value range constraints in the business validation rules and convert them into structured rule feature vectors; extract the role features of the field in the data entry rules; map the business topic distribution vector, structured rule feature vector and role feature vector of the field into a fixed-dimensional business context feature vector f_business through concatenation and filling operations; The obtained semantic feature vector, statistical feature vector, relational feature vector, and business context feature vector are concatenated and normalized to obtain a unified field semantic feature vector F=[f_semantic, f_statistic, f_relation, f_business]. Each feature vector is mapped to the [0,1] interval through Min-Max normalization, and the dimension of the field semantic feature vector F is the sum of the fixed dimensions of each feature vector.

4. The system according to claim 3, characterized in that, The recommendation layer includes a tag knowledge graph module, a multi-strategy recommendation engine module, and a collaborative recommendation module. The tag knowledge graph module is used to construct and maintain the semantic knowledge network of tags, providing a semantic basis for the matching of fields and tags. The multi-strategy recommendation engine module is used to calculate the matching score between fields and tags based on the semantic feature vector of the field and the tag knowledge graph, and fuse them to generate the final recommendation result. The collaborative recommendation module uses existing field labels to annotate records and recommends labels for the target field based on historical labeling experience of similar fields.

5. The system according to claim 4, characterized in that, The tag knowledge graph module includes a tag ontology management submodule, a tag semantic vector generation submodule, a tag statistical feature template management submodule, a tag topic distribution generation submodule, and a graph storage and query submodule. The tag ontology management submodule is used to define the ontology structure of tags, including tag name, tag definition, tag category and tag relationship. Tag relationship includes tag hierarchy relationship, tag mutual exclusion relationship and tag association relationship; The label semantic vector generation submodule is used to generate a semantic embedding vector for each label using a pre-trained language model. The label vector and the field semantic feature vector are located in the same vector space, which facilitates the calculation of cosine similarity. The tag statistical feature template management submodule is used to automatically learn the statistical feature template for each tag based on historical annotation records: collect the statistical feature vectors of all fields of the tag that have been annotated, calculate the mean and standard deviation of each statistical feature, and form a statistical feature template; for new tags or tags with insufficient annotation records, the statistical feature template is preset manually by domain experts. The tag topic distribution generation submodule is used to generate business topic distribution vectors for tags based on the tag definition text using the LDA topic model, ensuring that the tag topic distribution and the field topic distribution are in the same vector space; The graph storage and query submodule is used to store tag knowledge graphs in a graph database or relational database graph extension mode. It supports path query, subgraph matching and semantic reasoning based on tag relationships; the output is tag semantic vector, tag hierarchy structure and tag association path.

6. The system according to claim 4, characterized in that, The specific strategies employed by the multi-strategy recommendation engine module include semantic matching, statistical feature matching, and business relevance matching. Semantic matching: Calculate the cosine similarity between the semantic feature vector of the field and the semantic vector of the label to obtain the semantic matching score S_semantic; For polysemous words and abbreviations, a context-aware disambiguation mechanism is used, which combines the table name to which the field belongs and the context information in the comments to select the most appropriate semantic interpretation; Statistical feature matching: The statistical feature vector of the field is compared with the statistical feature template corresponding to the label. The Gaussian matching function is used to calculate the matching statistical feature matching score S_statistic: For each statistical feature item in the statistical feature template, the deviation between the actual value of the field and the mean of the statistical feature template is calculated. The matching score is 1 if the deviation is within the standard deviation of the statistical feature template, and 0 if the deviation exceeds 2 times the standard deviation. Linear interpolation is used if the deviation is in between. Business relevance matching: Matching based on the business context feature vector of the field and the business scenario constraint of the label: Calculate the KL divergence between the business topic distribution vector of the field and the business topic distribution vector of the label, and convert the KL divergence into a business relevance score in the range [0,1] using an exponential function: S_business = exp(-KL(P_field || P_label)), which is used as the business relevance matching score; By querying the historical frequency of how tags are applied to similar fields in the tag knowledge graph, as well as the popularity of the tag in the overall tag system, tags that are used frequently are given an additional frequency score S_frequency; Finally, a weighted summation method is used to merge the matching scores to obtain the comprehensive recommendation score of the field and the tag: S_total=w1×S_semantic+w2×S_statistic+w3×S_business+w4×S_frequency; w1, w2, w3, and w4 are weight parameters. For labels whose S_total exceeds the threshold T, they are included in the candidate recommendation list and output in descending order of score.

7. The system according to claim 6, characterized in that, The weight parameters w1, w2, w3, and w4 are determined in the following way: During the initial deployment of the system, default weights are preset manually. After accumulating enough manually labeled samples, a grid search is performed in the weight space with the F1 score as the optimization objective. The step size is 0.05, and the search range is [0.05, 0.85]. The constraints are that the sum of the weights is 1 and each weight is not less than 0.

05. The weight combination that maximizes the F1 score is selected as the current optimal weight. When the number of user feedback samples exceeds 500, the system switches to an online optimization mode based on gradient descent.

8. The system according to claim 4, characterized in that, The specific execution process of the collaborative recommendation module is as follows: A1. Field Similarity Calculation: Based on the semantic feature vector of the fields, calculate the comprehensive similarity Sim(Fi, Fj) between each labeled field in the labeled field set and the target field of the tag to be recommended; the similarity calculation formula is: Sim(Fi,Fj)=0.5×cos(f_semantic_i, f_semantic_j)+0.3×(1-|f_statistic_i-f_statistic_j|)+0.1×(1-|f_relation_i-f_relation_j|)+0.1×(1-|f_business_i-f_business_j|), where cos represents cosine similarity, and |·| represents the L1 norm normalized distance of the vectors; A2. Neighbor retrieval: For the target field of the tag to be recommended, retrieve the top-K neighbor fields with the highest similarity in the set of labeled fields to form a neighbor set N(F_target). Only fields with similarity exceeding the threshold θ are considered as valid neighbors. A3. Weighted summarization of labels: The labeled labels of the neighbor fields in the neighbor set N(F_target) are weighted and summed according to their frequency of occurrence to generate a collaborative recommendation score S_coll=Σ(Li∈N) Sim(F_target, Fi)×I(Li, L) / Σ(Li∈N) Sim(F_target, Fi), where I(Li, L) is the indicator function and the denominator is the normalization factor to ensure that the value range of the collaborative recommendation score is within the interval [0,1]. A4. Fusion Recommendation: The collaborative recommendation score is further fused with the comprehensive score of the multi-strategy recommendation engine. The final recommendation score S_final = α × S total + (1-α) × S_coll, where α is the fusion coefficient.

9. The system according to claim 8, characterized in that, The presentation layer includes a tag recommendation display module, which presents the recommendation results to users in a visual way and supports user interaction. (1) Recommendation list display: Displays the field list of the current data table. After selecting a field, the top-N recommended tags for that field are displayed in the middle recommendation tag display area. Each recommendation tag displays the tag name, recommendation score, recommendation reason, tag category and related tags. (2) Tag details display: The tag details area displays the complete definition, applicable scenarios, associated tags, historical usage count and typical application cases of the currently selected tag; when the user clicks on a recommended tag, detailed information is displayed in real time in the right area; (3) Batch recommendation display: Supports batch display of recommendation tags for all fields under the data table by data table dimension, presented in tabular form, and supports filtering and sorting by field type, recommendation score, tag category and other dimensions; (4) User interaction operations: Users can perform three operations on each recommended tag: "accept", "reject", and "correct". Each recommended tag card contains an accept / reject / correct button. The user's operation behavior is recorded in the feedback database for subsequent model optimization.

10. The system according to claim 9, characterized in that, It also includes a feedback layer, which is used to dynamically adjust model parameters based on user operations to form a complete closed-loop optimization process; the feedback layer includes a user feedback optimization module, which is used to record all user interactions with recommendation tags and dynamically adjust recommendation strategy parameters and weights based on feedback data. The user feedback optimization module includes a feedback data collection submodule, a statistical analysis submodule, a parameter optimization submodule, and a tag knowledge graph update submodule. The feedback data collection submodule records user operation data, including user interaction field ID, recommendation tag ID, user operation type, operation timestamp, user ID, and corrected tag content. Among them, adoption operations are recorded as positive feedback, rejection operations are recorded as negative feedback, and correction operations are recorded as partially positive feedback. The feedback data collection submodule aggregates user operation data into the feedback database. The statistical analysis submodule is used to periodically calculate the accuracy, recall, F1 score, and user adoption rate of each recommendation strategy; using the tags adopted by users as the ground truth and the tags recommended by the system as the predicted values; and comparing the recommendation effects under different weight parameter configurations through A / B testing: users are randomly divided into a control group and an experimental group, and the difference in adoption rate between the two groups within a week is compared. If the adoption rate of the experimental group increases by more than 5% and is statistically significant, the candidate weight is switched to the current weight. The parameter optimization submodule automatically increases the weight w1 when the recommended label adoption of the semantic matching strategy is consistently higher than that of other strategies, and decreases it otherwise. The weight adjustment adopts the constrained gradient descent method, and the optimization objective is to maximize the F1 value. The loss function is L=1-F1+λ×(Σwi-1). 2 , where λ is the constraint penalty coefficient, ensuring that the sum of the weights approaches 1; at the same time, the projection operation ensures that each weight wi is not lower than the minimum threshold of 0.05; the learning rate is initially 0.01, and decays to 0.9 times the original rate every 100 iterations; The tag knowledge graph update submodule is used to add semantic association edges between two tags in the tag knowledge graph and update the tag vector when a user frequently changes a tag to another tag; when a tag has not been adopted for a long time, its activity weight in the graph is reduced.

Images