Method and system for automatically constructing ontology and business knowledge base based on data catalog with AI

Abstract

The method for automatically constructing ontology and business knowledge base based on data directory by AI involves the technical fields of knowledge engineering and data governance. The construction process of the ontology in the method includes: collecting metadata and associated unstructured documents of each field from a data directory system, performing field grouping and semantic enhancement processing, and generating a semantic processing result; calling a field large model based on the result to identify the ontology, extract ontology concepts, identify attributes and relationships, generate axioms and rules, perform cross-domain semantic alignment, construct an ontology model and mapping relationship; performing bidirectional verification on the ontology model and mapping relationship, performing consistency verification using metadata changes, and verifying instance data using ontology axioms and rules to form an instance library and generate metadata quality feedback information. The application realizes the automatic construction of ontology, reduces the dependence on manual work, and improves the dynamic consistency and business adaptability of ontology and business data.

Description

Technical Field

[0001] This invention relates to a method and system for automatically constructing ontology and business knowledge bases based on data catalogs using AI, belonging to the fields of knowledge engineering, data governance and ontology engineering technology. Background Technology

[0002] Ontology is a formalized specification of business semantics, defining core concepts, attributes, relationships, axioms, and rules within a business domain. It serves as the core semantic foundation for enterprise data governance, knowledge assetization, and intelligent semantic applications. Data catalogs are the unified management carrier for all of an enterprise's data assets, aggregating technical, business, and management metadata from various business domains. They act as the core bridge connecting business data and business semantics. With the deepening of enterprise digital transformation, the demand for a unified enterprise-level ontology semantic system for cross-business domain data interoperability, knowledge reuse, and intelligent applications is becoming increasingly urgent. The automated and dynamic construction of ontology and business knowledge bases has become a core research direction in this field. Currently, industry solutions for ontology and business knowledge base construction mainly fall into two categories: traditional manual ontology construction and AI-automated ontology construction. Both have unresolved core technical defects, as detailed below: (I) Traditional ontology and knowledge base construction schemes, taking the prior patent with publication number CN120181206A and invention title "A Data Governance Knowledge Graph Construction Method Based on Metadata" as an example, this scheme constructs an ontology based on manually defined metadata mapping rules and knowledge graph semantic specifications, and has the following core defects: 1. High dependence on manual labor: Large enterprises often need months or even years to build cross-domain ontology, which is highly dependent on scarce domain experts and knowledge engineers, resulting in extremely high labor costs, making it virtually impossible for small and medium-sized enterprises to implement. 2. Severe disconnect between ontology and business data: Traditional ontology is mostly static and manually defined, which cannot respond in real time to changes in metadata in the data catalog and iterations of business data. Once the ontology is built, it enters a static and solidified state, which quickly becomes invalid as the business develops. The consistency with actual business data continues to decline, making it unable to adapt to the rapid iteration of enterprise business. The industry generally suffers from the persistent problem of "being useless after being built". 3. Cross-domain semantic barriers are difficult to break down: There are a large number of terms with the same name but different meanings and the same meaning but different names among different business departments and systems in an enterprise. Manual alignment is inefficient and has a high error rate. The ontologies and knowledge bases built independently by each department are fragmented and cannot form a unified semantic system at the enterprise level, resulting in the coexistence of data silos and knowledge silos. 4. Unable to achieve a two-way empowerment closed loop: It can only build ontology unidirectionally based on metadata, and cannot drive metadata quality optimization in reverse through the results of ontology construction. Data governance and knowledge construction are completely separated, and a virtuous cycle of two-way empowerment cannot be formed.

[0003] (II) Existing AI-automated ontology construction solutions, taking the prior patent with publication number CN119167085A and invention title "An Ontology Construction System Based on a Large Model" as an example, this solution achieves automatic generation of ontology concepts through a general large model, but has the following core shortcomings: 1. Lack of strong constraints from enterprise business context: Existing ontology construction solutions based on large models mostly rely on general large models to generate general domain ontology templates. They cannot deeply bind to real metadata and business documents in the enterprise data catalog, cannot adapt to the enterprise's own business scenarios and data asset status, and the generated ontology is seriously out of touch with the enterprise's actual business, with poor feasibility and serious AI illusion problems. 2. Lack of dynamic incremental update mechanism: It still adopts the static mode of "one-time construction", which cannot realize dynamic iterative updates of the ontology based on changes in metadata of the data directory. It fails to fundamentally solve the core industry pain point of static rigidity of the ontology and disconnection from business development. 3. Inability to achieve multimodal knowledge fusion: Existing solutions mostly focus on the ontology construction of structured data, and cannot automatically integrate unstructured business knowledge such as enterprise systems, process documents, and business experience based on the metadata association relationship of the data catalog, thus failing to form a complete enterprise-level business knowledge base covering both structured and unstructured data; 4. Lack of a full-process self-optimization closed loop: It is impossible to continuously iterate and optimize the ontology construction, knowledge extraction, and semantic reasoning capabilities of large models based on the quality assessment results of ontology and knowledge base, and the model performance continues to decline as business develops. Summary of the Invention

[0004] To at least partially overcome the problems existing in related technologies, this invention proposes a method and system for automatically constructing ontology and business knowledge base based on a data catalog using AI, aiming to solve the problems of low ontology construction efficiency, poor business adaptability, and insufficient linkage with the data catalog in the prior art.

[0005] First aspect This application proposes a method for automatically constructing ontology and business knowledge base based on a data catalog using AI. The ontology construction process includes: Step S1: Collect metadata from each domain in the data catalog system and obtain the associated unstructured business documents. Group the corresponding domains to form several domain groups. Perform semantic enhancement processing on each domain group based on the unstructured business documents to generate semantic processing results for each domain group. The semantic processing results include at least a structured metadata dataset, a metadata semantic vector library, and a mapping table between metadata and associated unstructured business documents. Step S2: Based on the semantic processing results, the domain big model is called to perform ontology recognition processing, and an ontology model and its supporting mapping relationship and axiom rule base are constructed based on the processing results. The ontology recognition processing includes extracting ontology concepts, identifying the attributes and relationships between ontology concepts, generating ontology axioms and standardized business rules, and performing cross-domain semantic alignment and fusion of ontology concepts from different domain groups. Step S3: Establish a bidirectional traceable mapping relationship between the ontology model and the metadata of the data directory, perform bidirectional verification processing, and form an ontology instance library based on the verification results and generate feedback information for optimizing metadata quality. The bidirectional verification processing includes performing consistency verification and incremental updates on the ontology model and mapping relationship based on newly added or changed metadata in the data directory, and performing compliance verification on ontology instance data extracted from business data sources based on the ontology axioms and rule base.

[0006] In one possible implementation, the specific steps for generating the semantic processing result in step S1 include: S11: Perform deduplication, completion, and format standardization on the collected metadata to form standardized metadata; S12: Classify and cluster the standardized metadata according to preset business domain rules to form several domain groups; S13: Based on the domain-wide model, standardized business semantic tags and semantic vectors are generated for the metadata entities in each domain group, forming a structured meta-dataset and a metadata semantic vector library; S14: Based on the standardized business semantic tags, extract the corresponding metadata business semantic supplementary information from the associated unstructured business documents, establish a mapping table between the metadata and the unstructured business documents, and provide business context constraints for subsequent ontology construction.

[0007] In one possible implementation, step S2 specifically includes: S21: Construct an AI agent, input the semantic processing result into the AI ​​agent, the AI ​​agent calls the domain big model, extracts the ontology concept of the business domain with the business semantics of the metadata as the core, and divides it into layers according to the three-level architecture of public ontology layer, domain ontology layer and scene ontology layer. S22: Identify the data attributes and object attributes of each ontology concept, and generate the domain, range, cardinality constraints of the data attributes, as well as the inheritance relationship, association relationship and semantic characteristics of the object attributes; S23: Based on the constraint rules and business rules of metadata, generate ontology logical axioms and output the rule base in SWRL form; S24: Perform semantic alignment and fusion of ontology concepts from different business domains. Based on semantic vector similarity calculation and domain-wide model semantic reasoning, distinguish between homonymous and synonymous concepts, merge synonymous concepts, generate ontology alignment mapping table, and optimize ontology concepts, attributes and axiom rules in reverse based on the alignment results. S25: Output a standardized ontology model, an ontology concept-metadata mapping table, an ontology alignment mapping table, and a rule base, which serve as the ontology model and its associated mapping relationships and axiomatic rule base.

[0008] In one possible implementation, the semantic alignment and fusion of ontology concepts from different business domains in step S24 further includes: S241: Generate semantic vectors of ontology concepts and metadata through the HST-Encoder semantic encoding model, calculate the initial similarity from three dimensions: text semantics, structural topology and instance distribution, and then fuse them according to preset weights to obtain the basic comprehensive similarity. S242: Obtain historically verified mapping information as prior knowledge, introduce a time decay mechanism to calculate the historical mapping enhancement factor, and combine it with the basic comprehensive similarity to calculate the comprehensive mapping confidence. S243: Generate a dynamic threshold based on the confidence statistics of historically verified mapping pairs. Based on the comparison between the comprehensive mapping confidence and the dynamic threshold, add the mapping pairs to the ontology concept-metadata mapping table, submit them for manual review, or discard them. The annotation results of the manual review serve as positive and negative samples for model fine-tuning. S244: When a new mapping pair is successfully verified, the semantic vector of the ontology concept is updated with weights, and the semantic encoding model is incrementally fine-tuned based on the positive and negative samples that have been verified. S245: Periodically clean up mapping pairs that have been retained for longer than a preset period and whose confidence level is lower than a preset threshold.

[0009] In one possible implementation, step S3 specifically includes: S31: Based on the ontology concept-metadata mapping table, establish a bidirectional traceable mapping relationship between the ontology model and the data catalog metadata. The bidirectional traceable mapping relationship includes: mapping between ontology classes and business tables, mapping between ontology data attributes and table fields, and mapping between ontology object attributes and table associations or data lineages. The bidirectional traceable mapping relationship is then solidified into the data catalog as the core basis for triggering incremental ontology updates when metadata changes occur. S32: Based on the bidirectional traceable mapping relationship, extract ontology instance data from the business data source associated with the data directory, generate standardized RDF triples, and form an ontology instance library, which serves as the core data foundation for the subsequent construction of the business knowledge base. S33: Perform bidirectional verification processing, specifically including: based on the newly added or changed metadata in the data catalog, perform dynamic consistency verification on the ontology model and mapping relationship through the bidirectional traceable mapping relationship; if the verification fails, trigger incremental update of the ontology. Based on the ontology axioms and rule base, an AI inference engine is used to automatically verify the ontology instance data. The automatic verification includes data consistency verification, semantic integrity verification, and identification of redundant or conflicting instances. Instances with abnormal verification are handled in a graded manner according to preset rules. S34: Feedback the verification results to the data catalog system, update the quality score and anomaly label of the corresponding metadata, and generate feedback information for optimizing metadata quality.

[0010] One possible implementation also includes step S4, which involves constructing a multimodal business knowledge base based on the constructed ontology model. Step S4 includes: S41: Using the ontology model as the semantic framework and the ontology instance library as the core data, construct the basic layer of the knowledge base. Based on the bidirectional traceable mapping relationship between the ontology model and the metadata of the data directory, connect the basic layer with the permission system and security level system of the data directory, and assign security tags, permission tags and business domain tags to knowledge entities. S42: Based on the domain big model and ontology semantic constraints, automatically extract business rules, business processes and business cases from unstructured business documents, semantically align the extracted knowledge entities with ontology concepts, and integrate them into the base layer to form a business knowledge layer. S43: Based on the ontology axioms and rule base, use AI reasoning engine and graph mining algorithm to perform compliance reasoning on explicit knowledge in the knowledge base, mine implicit knowledge corresponding to business scenarios, and update the verified implicit knowledge in reverse to the ontology axioms and rule base to form a reasoning knowledge layer. S44: Implement hierarchical classification and management of the basic layer, business knowledge layer and reasoning knowledge layer, construct interconnected semantic vector index, full-text index and ontology semantic index, and establish a synchronous update mechanism for the knowledge base, ontology model and data catalog, and output a hierarchical enterprise-level business knowledge base.

[0011] One possible implementation further includes an incremental iterative update step S5, which includes: S51: Real-time monitoring of preset metadata change events in the data catalog through an event-driven mechanism, as well as real-time acquisition of incremental data changes from business data sources, and filtering and classifying the monitored change events, filtering invalid changes and classifying update levels according to the scope of impact. S52: Automatically trigger the corresponding level of ontology incremental update based on the monitored change event. The incremental update includes adding or adjusting ontology concepts, attributes, relationships, axioms and cross-domain alignment mapping relationships, as well as synchronously updating the bidirectional mapping relationship between the ontology and the data directory. After the incremental update is completed, a consistency check is performed based on the bidirectional verification processing method in step S3. If the check fails, an automatic rollback is triggered. S53: Based on the incremental updates of the ontology and incremental business data, automatically synchronize and update the instance data, knowledge entities and corresponding indexes in the knowledge base, and extract knowledge from newly added unstructured documents for semantic alignment and fusion. S54: Perform canary releases of incremental updates, set verification rules and rollback mechanisms, and manage versions, supporting historical version backtracking and comparison.

[0012] One possible implementation also includes a step S6 of full-process closed-loop optimization, wherein step S6 includes: S61: Construct a multi-dimensional quality assessment system, triggering automated assessments according to preset cycles or incremental update events to generate quality assessment results. The assessment dimensions include metadata completeness, standardization, and accuracy; business adaptability, logical consistency, and cross-domain alignment accuracy of the ontology model; and knowledge accuracy, business coverage, and semantic consistency of the business knowledge base. S62: Feed the quality assessment results back to the data catalog system to drive the completion, optimization, and standardization of corresponding metadata, and simultaneously feed back ontology defects and knowledge conflict issues to the incremental update process of the ontology and knowledge base, achieving end-to-end reverse optimization. S63: Based on the quality assessment results, explicit user error correction feedback, and implicit business behavior feedback, automatically optimize the AI ​​agent prompts, HST-Encoder semantic encoding model parameters, and semantic alignment algorithm weights that are relied upon for ontology construction, knowledge extraction, semantic alignment, and reasoning.

[0013] Second aspect The system for automatically constructing ontology and business knowledge base based on a data catalog using AI proposed in this application is applicable to the ontology construction in the method for automatically constructing ontology and business knowledge base based on a data catalog in the first aspect mentioned above. The system includes: The metadata semanticization module is used to collect metadata from various domains in the data catalog system, obtain associated unstructured business documents, group the corresponding domains to form several domain groups, perform semantic enhancement processing on each domain group based on the unstructured business documents, and generate semantic processing results for each domain group. The semantic processing results include at least a structured metadata dataset, a metadata semantic vector library, and a mapping table between metadata and associated unstructured business documents. The ontology construction module communicates with the metadata semanticization module and is used to call the domain big model for ontology recognition processing based on the semanticization processing results, and to construct the ontology model and its supporting mapping relationship and axiom rule base based on the processing results. The ontology recognition processing includes extracting ontology concepts, identifying the attributes and relationships between ontology concepts, generating ontology axioms and standardized business rules, and performing cross-domain semantic alignment and fusion of ontology concepts from different domain groups. The bidirectional mapping and verification module communicates with the ontology construction module and the data catalog system. It is used to establish a bidirectional traceable mapping relationship between the ontology model and the metadata of the data catalog, perform bidirectional verification processing, and form an ontology instance library and generate feedback information for optimizing metadata quality based on the verification results. The bidirectional verification processing includes consistency verification and incremental updates of the ontology model and mapping relationship based on newly added or changed metadata in the data catalog, and compliance verification of ontology instance data extracted from business data sources based on the ontology axioms and rule base.

[0014] Compared with the prior art, the beneficial effects of the present invention are: 1. Achieved efficiency and cost optimization: This invention realizes full-process AI automation from metadata collection, semantic enhancement, ontology construction, bidirectional verification to knowledge base generation, shortening the traditional months-long ontology construction cycle to hours, significantly reducing manual input, lowering the professional threshold for ontology construction, and enabling small and medium-sized enterprises to implement it quickly.

[0015] 2. Solves the industry pain point of static and rigid ontology: This invention uses the data catalog as the core bridge to build a two-way mapping and real-time linkage mechanism between ontology and metadata and business data. Through event-driven incremental iterative updates, the ontology is transformed from a static model into a dynamic semantic model that adapts to business changes, effectively avoiding the problem of traditional solutions being "built and then discarded".

[0016] 3. Breaking through cross-domain semantic barriers and building an enterprise-level unified semantic system: This invention accurately distinguishes cross-domain concepts with "same name but different meaning" and "same meaning but different name" by performing cross-domain semantic alignment and fusion of ontology concepts from different domain groups, breaking down semantic silos between enterprise departments and forming a unified semantic standard for the entire enterprise; 4. Suppressing large model illusions through strong business constraints, significantly improving the business adaptability of the ontology: This invention uses the full metadata of the enterprise data catalog and business documents as strong constraints of the business context. By standardizing the output boundary of the large model through customized AI intelligent agents and combining cross-validation of ontology axiom rules, the illusion rate of the content generated by the large model is effectively reduced. The generated ontology is deeply adapted to the actual business scenarios of the enterprise, solving the core problem of poor implementability of existing AI ontology solutions. 5. Achieved multimodal knowledge fusion, forming a complete enterprise-level knowledge asset: This invention uses ontology as a unified semantic framework to achieve multimodal knowledge fusion of structured business data and unstructured business documents, and constructs an enterprise-level business knowledge base with a three-layer architecture of "basic layer - business knowledge layer - reasoning knowledge layer", covering the full amount of enterprise business knowledge assets, and can directly support full-scenario business applications such as intelligent semantic retrieval, intelligent question answering, and compliance auditing; 6. A two-way empowerment closed loop for data governance and knowledge construction has been constructed: This invention uses the quality assessment results of ontology construction and knowledge base to drive the completion, optimization and standardization of data catalog metadata, while realizing the continuous self-iteration of large model ontology construction capabilities, forming a virtuous cycle of "metadata quality improvement → ontology and knowledge base quality improvement → reverse-driven metadata optimization", which can simultaneously enhance the value of enterprise data assets and knowledge assets. Attached Figure Description

[0017] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the method flow in one embodiment of the present invention; Figure 3 This is a block diagram of the system structure of the present invention. Detailed Implementation

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] Based on the background technology, the technical problem to be solved by this invention is to address the six core industry pain points in existing ontology and business knowledge base construction schemes, namely, high dependence on manual work, static and rigid ontology that is disconnected from business data, difficulty in breaking down cross-domain semantic barriers, lack of business constraints in the construction of large model ontology leading to serious illusions, inability to achieve multimodal knowledge fusion, and lack of a closed loop of two-way empowerment throughout the entire process. The invention provides a method for constructing ontology and business knowledge bases automatically with AI based on a data catalog, so as to achieve the technical effects of AI automation of the entire ontology construction process, dynamic adaptation of ontology and business data, cross-domain semantic unification, multimodal knowledge fusion, and two-way empowerment of data governance and knowledge construction.

[0020] In one embodiment, the method for automatically constructing an ontology and business knowledge base based on a data catalog using AI proposed in this application involves first constructing the ontology, and then constructing a multimodal business knowledge base based on the constructed ontology. The method is executed based on a software information system. For example... Figure 1 As shown, in this embodiment, the ontology construction process includes: Step S1: Collect metadata from each domain in the data catalog system and obtain the associated unstructured business documents. Group the corresponding domains to form several domain groups, and perform semantic enhancement processing on each domain group based on the unstructured business documents to generate semantic processing results for each domain group. The semantic processing results include at least a structured metadata dataset, a metadata semantic vector library, and a mapping table between metadata and associated unstructured business documents.

[0021] In other words, step S1 is metadata semantic enhancement processing. Specifically, this step involves connecting to the enterprise data catalog system to collect metadata from the entire business domain and associated unstructured business documents. After standardizing and preprocessing the metadata, it clusters it into several domain groups based on the business domain. Based on the unstructured business documents, it supplements the metadata of each domain group with business context semantic information, generating semantic processing results for each domain group. The specific steps for generating semantic processing results in step S1 are S11-S14 executed sequentially: S11: Perform deduplication, completion, and format standardization on the collected metadata to form standardized metadata; In practice, the system connects to the enterprise's data catalog system via standard APIs or pre-built connectors to collect full metadata, including at least three types of core metadata: ① Technical metadata, including table structure, field names, data types, primary key / foreign key constraints, indexes, field lengths, enumeration value ranges, etc.; ② Business metadata, including business terminology, business rules, data standards, business tags, indicator definitions, business scenario descriptions, etc.; ③ Manage metadata, including data lineage, data ownership, security level, data lifecycle, etc.

[0022] Synchronously, through the association relationship of the data directory, it connects with the corresponding business data sources and unstructured business documents, including corporate rules and regulations, operation manuals, process documents, business cases, etc., to ensure data integrity.

[0023] After completing the metadata collection, standardization processing (preprocessing) is performed, which specifically includes: Deduplication: Using unique identifiers such as database table ID, field ID, and business term code in the data catalog as the core primary key, completely duplicate metadata is identified and removed; for metadata without a unique ID, multi-dimensional composite primary keys are constructed based on metadata type, business attribute, data type, and domain specification (including but not limited to "database table name + field name + business meaning + data type" and "business term code + business meaning + business domain") to identify and remove semantically duplicated metadata; Missing item completion: The system sequentially scans the required missing items in technical metadata, business metadata, and management metadata, including but not limited to field comments, data ownership, constraints, index types, and data lifecycles. It extracts valid information for the corresponding missing items from the associated business data sources and unstructured business documents through semantic matching and completes automatic completion. Missing items without matching information are marked as abnormal and pushed to the metadata administrator for confirmation. Format standardization: unify the naming conventions and data type formats of all metadata, and simultaneously unify the constraint rules such as non-null, unique, primary key, foreign key, and value range, to complete the format standardization of all metadata.

[0024] S12: Classify and cluster standardized metadata according to preset business domain rules to form several domain groups; In practice, S12 first divides the overall business of an enterprise into multiple independent business domains, including but not limited to the financial domain, human resources domain, sales domain, supply chain domain, and production domain. Based on the business attributes, system to which the metadata belongs, and data ownership, all standardized metadata is allocated to the corresponding business domain. After the business domain allocation is completed, within each business domain, the metadata is hierarchically subdivided according to the business processing flow to form a four-level structured hierarchy of "business domain - business system - metadata type - metadata item", thus completing the construction of domain groups.

[0025] Based on different scenario conditions, classification and clustering support two implementation modes: First, it uses rule-based classification based on the enterprise's pre-defined business domain rules, which is suitable for medium and large enterprises with clear business domain divisions. Second, it uses automatic clustering based on the DBSCAN density clustering algorithm, which is suitable for small and medium-sized enterprises with blurred business domain boundaries. The clustering features are the business semantic vector of metadata, the system to which it belongs, and the data lineage.

[0026] S13: Based on the domain-wide model, standardized business semantic tags and semantic vectors are generated for the metadata entities in each domain group, forming a structured meta-dataset and a metadata semantic vector library; In practical implementation, a hybrid semantic topology encoder, combining a domain-wide model and the HST-Encoder, is used. The domain-wide model is based on a domestically developed foundational model and fine-tuned using enterprise-annotated metadata corpora, industry-standard ontology libraries, and business rule documents. The core tasks of fine-tuning are metadata semantic understanding, business tag generation, and semantic feature extraction. The trained domain-wide model can accurately identify the business meaning of metadata and generate standardized semantic tags that conform to enterprise business specifications. The HST-Encoder encoder generates a fixed-dimensional semantic vector for each metadata entity. This semantic vector integrates three dimensions of metadata features: textual semantics, structural topology, and instance distribution, ultimately forming a standardized structured metadata dataset and a metadata semantic vector library.

[0027] S14: Based on standardized business semantic tags, extract the corresponding metadata business semantic supplementary information from the associated unstructured business documents, establish a mapping table between the metadata and the unstructured business documents, and provide business context constraints for subsequent ontology construction. In other words, S14 is to realize the construction of business document association mapping.

[0028] In practice, based on the standardized business semantic tags generated by S13, business semantic supplementary information of corresponding metadata is extracted from unstructured business documents through semantic retrieval. This includes business scenario descriptions, business rules, calculation logic, compliance requirements, etc. of the metadata. The metadata entities are then bound to the corresponding business semantic supplementary information and the associated unstructured business documents to generate a "metadata-business document association mapping table". The mapping table clearly records the business documents, associated fragments, and semantic supplementary information associated with each metadata entity, providing strong business context constraints for subsequent ontology construction and suppressing the illusion of large model-generated content.

[0029] Step S2: Based on the semantic processing results obtained in step S1, the domain big model is called to perform ontology recognition processing, and the ontology model and its supporting mapping relationship and axiom rule base are constructed based on the processing results. The ontology recognition processing here includes extracting ontology concepts, identifying the attributes and relationships between ontology concepts, generating ontology axioms and standardized business rules, and performing cross-domain semantic alignment and fusion of ontology concepts from different domain groups.

[0030] Step S2 specifically includes: S21: Construct an AI agent, input the semantic processing results into the AI ​​agent, the AI ​​agent calls the domain big model, extracts the ontology concept of the business domain with the business semantics of the metadata as the core, and divides it into layers according to a three-level architecture of public ontology layer, domain ontology layer and scenario ontology layer (such as following the DOLCE international meta-ontology specification).

[0031] First, a customized AI agent for ontology concept extraction is constructed. Its input consists of a semantically enhanced structured metadata dataset and a metadata semantic vector library, while its output is an enterprise-grade standardized ontology model conforming to the OWL international standard. Based on the implementation scenario, an appropriate technology stack is selected, including but not limited to Python, deep learning frameworks, and ontology modeling tools. The agent's internal architecture is then constructed, including a data input module, a model invocation module, a task execution module, and a result output module. This AI agent is used to uniformly manage the invocation process of large-scale domain models, configuring input constraints, output formats, and business context limitations to form a strong constraint mechanism on the large-scale domain models, reducing the AI ​​illusion during ontology generation.

[0032] Then, the AI ​​agent converts the semantically enhanced structured metadata dataset and metadata semantic vector library into the standard input format required by the domain-wide model. The domain-wide model, with metadata business semantics at its core, performs semantic parsing on each metadata entity, identifying and extracting ontology concepts within the corresponding business domain. Here, an ontology concept represents an abstract generalization of a class of business objects or business meanings, corresponding to a class in OWL, representing a set of business entities with common attributes and characteristics.

[0033] Next, the domain-wide model categorizes and organizes the extracted concepts based on business relationship logic and data lineage, and returns them to the AI ​​agent. The AI ​​agent loads a pre-built industry-standard ontology benchmark library, which contains recognized standard ontology concepts, attributes, and relationship rules for various business domains.

[0034] Furthermore, the AI ​​agent extracts standardized concepts from industry-standard ontology benchmark libraries and constructs a common ontology layer following the DOLCE meta-ontology specification, ensuring cross-domain compatibility and universality of concepts. Based on this, a domain ontology layer is constructed by combining core business concepts extracted from actual enterprise metadata using a domain-wide model, along with exclusive business elements from each business domain. Simultaneously, it allows for the addition of custom business concepts not covered by the benchmark library based on actual enterprise business scenarios, enabling deep adaptation of the ontology model to enterprise business. Then, the domain ontology is lightweighted and specialized based on specific business scenarios to construct a scenario ontology layer, ultimately forming a three-tiered ontology architecture of "common ontology layer—domain ontology layer—scenario ontology layer," reducing conceptual redundancy while ensuring semantic consistency.

[0035] Finally, the AI ​​agent formats and standardizes the ontology concepts that have completed the three-level architecture division, outputs an enterprise-level ontology model that conforms to the OWL standard, and records the mapping relationship between ontology concepts and corresponding metadata in the ontology concept metadata mapping table, providing a basis for subsequent ontology mapping, verification and iteration.

[0036] S22: Identify the data attributes and object attributes of each ontology concept, and generate the domain, range, cardinality constraints of the data attributes, as well as the inheritance relationship, association relationship and semantic characteristics of the object attributes.

[0037] Specifically, the system's ontology construction module identifies the data attributes and object attributes corresponding to each ontology concept based on the semantically enhanced metadata and the three-level ontology architecture. Among them, data attributes are used to describe the characteristics of the ontology concept itself, corresponding to the field characteristics in business metadata and technical metadata; object attributes are used to describe the business relationships between different ontology concepts, corresponding to business logic, data relationships, and data lineage.

[0038] For data attributes, information such as field name, data type, length limit, enumeration range, non-null / unique constraint, and business indicator definition are extracted from standardized metadata. Combined with the semantic parsing results of the domain model, the domain, value range, data format, and cardinality constraint of the corresponding data attributes are automatically generated to ensure that the attribute definition is consistent with the metadata rules.

[0039] For object attributes, a pre-fine-tuned domain model, combined with business semantics and data lineage, automatically identifies hierarchical inheritance relationships (e.g., individual customers and enterprise customers inherit from the customer class) and business association relationships (e.g., the purchase relationship between a customer and a product, and the inclusion relationship between an order and a product). Based on this, and incorporating business rules and constraints from the business metadata, the semantic characteristics of each object attribute are further defined, including transitivity, symmetry, mutual exclusion, reverse attribute association, and hierarchical constraints, forming a complete attribute relationship system that conforms to business logic.

[0040] S23: Based on metadata constraint rules and business rules, generate ontology logical axioms and output the rule base in SWRL form.

[0041] The system's ontology construction module extracts metadata constraint rules from the enterprise data catalog, business data sources, and data management platform. These constraint rules include mandatory restrictions on data-level format, value range, uniqueness, non-nullability, indexing, and data lineage. Simultaneously, it obtains business rules from business metadata, business documents, and business process definitions. These business rules include business operation logic, business process constraints, compliance requirements, indicator calculation logic, and business judgment conditions. Business metadata describes the meaning of data, business rules, business processes, and data usage from a business perspective and can be automatically obtained from the data catalog, business documents, and business system configuration.

[0042] Building upon this foundation, the ontology construction module takes metadata constraint rules and business rules as input. Through a pre-tuned domain-wide model, it performs deep semantic parsing and logical association analysis to extract business entities, business actions, attribute relationships, quantity constraints, time constraints, and conditional judgment elements from the business process. Furthermore, the domain-wide model, combined with the ontology hierarchy and attribute relationships, performs joint logical reasoning on the metadata constraint rules and business rules to generate implicit business constraints and ontology logical axioms. For example, based on the data constraint "product price is greater than zero" and the business rule "total order price = product price × quantity," the implicit business constraint "the total order price must be greater than zero" is automatically derived.

[0043] Finally, the ontology construction module standardizes and organizes the ontology axioms and business rules obtained through reasoning to form an ontology axiom and rule base, and outputs it in SWRL format conforming to W3C standards for subsequent instance verification, knowledge reasoning and quality control.

[0044] For example, the ontology axioms and rule base contain two main modules with a total of eight types of standardized constraints that can be executed by machines: Module 1: OWL Ontology Axioms and Rule Set (Core Constraints of the Logic Layer) Concept hierarchy and classification axioms, for example: Attribute feature constraint axioms include: ID card number is unique, age value range is [0, 120], and an order must contain at least one product (cardinality constraint).

[0045] For example, axioms for determining the necessary and sufficient conditions of a class: orders that satisfy both "paid" and "shipped" are automatically classified as orders awaiting receipt.

[0046] Cross-domain ontology alignment axioms include: Financial domain customers = Marketing domain users (equivalent mapping); E-commerce domain products ≠ Supply chain domain materials (semantic distinction).

[0047] Module 2: SWRL / SHACL Business Rule Set (Business Layer Implementation Constraints) Business access and compliance rules, for example: Age <18 → Cannot place prepaid orders.

[0048] Business processes and related dependency rules, for example: Paid orders → Trigger the shipping process.

[0049] Data quality pre-validation rules include, for example: invalid mobile phone number format → marked as abnormal data.

[0050] Implicit knowledge reasoning rules include: Customers who have purchased product A → Recommend product B.

[0051] All rules conform to international semantic web standards and can be directly executed by the AI ​​inference engine.

[0052] It should be noted that the process of identifying object attributes in S22 and the process of generating axioms and rules in S23 have an interdependent relationship. In actual execution, they can be processed in parallel or iteratively optimized. For the sake of clarity, this manual describes them step by step in logical order.

[0053] S24: Perform semantic alignment and fusion of ontology concepts from different business domains. Based on semantic vector similarity calculation and domain-wide model semantic reasoning, distinguish between homonymous and synonymous concepts, merge synonymous concepts, generate ontology alignment mapping table, and optimize ontology concepts, attributes and axiom rules in reverse based on the alignment results. Specifically, the ontology construction module collects the ontology concepts of each business domain output by S21, loads the formed ontology instance library, and performs automated semantic alignment and fusion on the ontology concepts of different business domains to solve the semantic conflict problem of "same name but different meaning" and "same meaning but different name" that is common in enterprises.

[0054] As a specific implementation method, the BM-SA bidirectional mapping self-enhancing semantic alignment algorithm can be introduced to strengthen and optimize the mapping relationship between ontology concepts and metadata. This algorithm includes: S241: Generate semantic vectors of ontology concepts and metadata through the HST-Encoder semantic encoding model. Calculate the initial similarity from three dimensions: textual semantics, structural topology, and instance distribution. Then, weighted and fused according to preset weights to obtain the basic comprehensive similarity. These three dimensions complement each other to ensure that the similarity calculation is consistent with the actual business scenarios of enterprises.

[0055] a. Textual semantic similarity S text (o) i m j Based on cosine similarity calculation and corrected by a domain thesaurus, textual semantic similarity is used to determine whether the textual meaning of ontology concepts and metadata is similar, avoiding misjudgments due to different expressions. The specific expression is as follows: Among them, o i To represent an ontology concept (such as "customer"), m j Representing metadata (such as "user"), , These represent the semantic vectors of the two respectively; This represents the synonym correction factor (default 0.15, value range [0, 0.2]). syn(o) i m j ) represents the ontology concept o i and metadata m j Similarity in the domain thesaurus (1 for synonyms, 0 for different synonyms). Based on a real-world business scenario, assuming the marketing domain ontology concept "user" and the customer service domain metadata "customer" have a cosine similarity of 0.85 in their semantic vectors, and since they are domain synonyms (syn=1), we can substitute them into the formula to calculate the text semantic similarity. =0.85x(1+0.15x1)=0.9775, indicating that the two texts are highly similar in semantics.

[0056] b. Calculate the structural topological similarity S based on the ontology concept and the hierarchical structure information of the data catalog. struct (o) i m j Structural topological similarity is used to determine whether the hierarchical association between ontology concepts and metadata is consistent. It combines the hierarchical structure of the data directory with the conceptual association of the ontology. The specific expression is as follows: in, It is an ontological concept. The first-order neighborhood concept set (e.g., the neighborhood of "customer" is "order, contact person, address"); N(m) j ) represents metadata m j The first-order neighborhood metadata set (e.g., the neighborhood of "user" is "order, mobile number, delivery address"); The formula calculates the ratio of the intersection to the union of the neighborhoods of the two entities, with a value range of [0, 1]. The higher the ratio, the more similar the structures.

[0057] In conjunction with actual business scenarios, the neighborhood set of "customers" ={orders, contacts, addresses}, the neighborhood set N(m) of the "user". j The set of elements is {order, phone number, shipping address}, the intersection of which is {order, address} (quantity 2), and the union of which is {order, contact person, address, phone number, shipping address} (quantity 5). Substituting these values ​​into the formula yields... =2 / 5=0.4, indicating that the two structures have a certain degree of similarity in topology.

[0058] c. Instance distribution similarity S inst (o) i m j This is used to determine whether the field sample distribution of instance data and metadata of an ontology concept is consistent. It is based on JS divergence calculation, and the formula is as follows: in, Representing ontology concepts Value distribution in the instance library (discrete values ​​are frequency distributions, continuous values ​​are kernel density estimates), such as the value distribution of the ontology concept "customer" in the instance library (e.g., the format of "customer number" is "KH + 8 digits", with a uniform frequency distribution); Represents metadata m j The sample value distribution of the corresponding field, such as the sample distribution of the corresponding field of the metadata "User" (e.g., the format of "User ID" is "YH+8 digits", and the frequency distribution is consistent with the customer number); express and The Jensen-Shannon divergence (JS divergence) of a probability distribution, with values ​​ranging from [value missing]. (The more similar the two distributions are, the smaller the JS divergence and the higher the instance similarity.) Based on a real-world business scenario, assuming the JS divergence between "Customer ID" and "User ID" is 0.1, substituting this into the formula yields... =1-0.1=0.9, indicating that the actual business data distribution of the two is highly consistent, further proving that the two are semantically similar.

[0059] Based on this, S242 is performed: historically verified mapping information is obtained as prior knowledge, a time decay mechanism is introduced to calculate the historical mapping enhancement factor, and the comprehensive mapping confidence is calculated in combination with the basic comprehensive similarity to ensure the accuracy and timeliness of the mapping results.

[0060] Step 1: Basic Comprehensive Similarity (Weighted Fusion Three-Dimensional Similarity) The similarity of text semantics, structural topology, and instance distribution is weighted and fused, with default weights of 0.5, 0.2, and 0.3 (summing up to 1, which can be adjusted according to business needs), as shown in the following formula: in, Both represent weighting coefficients. The sum of the three is 1. Users or relevant technical personnel can adjust the weighting coefficients according to the actual business scenario. The weights recommended by this invention are: ; If the recommended weights are used, and the three similarity values ​​for "customer" and "user" are substituted into the calculation above, then... =0.5x0.9775+0.2x0.4+0.3x0.9=0.48875+0.08+0.27=0.83875, indicating that the basic comprehensive similarity between the two is relatively high.

[0061] Step 2: Historical mapping enhancement factor (time decay mechanism) By incorporating historically verified mapping information, we attenuate outdated historical mappings to avoid interference from obsolete data. The formula is as follows: in, =0.003 (time decay coefficient, the weight decays to half of its original value after 3 months); Let be the confidence level of the k-th historical verification; Δt k This represents the time difference between the current time and the kth verification (in days).

[0062] Based on actual business scenarios, assume that the "customer" and "user" have two historical verification records: ① 60 days ago, with a confidence level of 0.9; ② 180 days ago, with a confidence level of 0.8.

[0063] Substituting into the formula: Numerator = 0.9xe^(-0.003x60) + 0.8xe^(-0.003x180) ≈ 0.9x0.835 + 0.8x0.582 ≈ 0.7515 + 0.4656 = 1.2171; Denominator = e^(-0.003x60) + e^(-0.003x180) ≈ 0.835 + 0.582 = 1.417; because ≈1.2171 / 1.417≈0.859.

[0064] Step 3: Final comprehensive mapping confidence level Combining the basic comprehensive similarity and historical mapping enhancement factor, the final comprehensive mapping confidence is calculated based on the following expression. : in, This represents the historical enhancement coefficient, used to control the degree of influence of historical mapping.

[0065] Such as historical enhancement coefficient Substituting 0.2 into the previous calculation result (S) base =0.83875, H=0.859), calculation process: =0.83875+0.2x0.859≈0.83875+0.1718≈1.00055 (the highest confidence level is 1, so we take 1), which means that the comprehensive mapping between "customer" and "user" has a very high confidence level and can be determined as synonyms.

[0066] Based on this, step S243 is performed: A dynamic threshold is generated based on the confidence statistics of historically validated mapping pairs. According to the comparison between the comprehensive mapping confidence and the dynamic threshold, the mapping pairs are added to the ontology concept-metadata mapping table, submitted for manual review, or discarded. The annotation results from the manual review serve as positive and negative samples for model fine-tuning. In other words, a dynamic threshold is generated based on the statistical results of historical mapping confidence, and the mapping pairs are graded according to the final comprehensive confidence (automatic entry, manual review, direct discard). The dynamic threshold formula is as follows: in, This represents the average confidence level (assumed to be 0.8) of historical verifications for the mapping pairs. T represents the standard deviation of the confidence level (assumed to be 0.1), therefore T high =0.9, T low =0.7; The corresponding classification rules and examples are as follows: 1. If ≥T high(As mentioned above, the confidence scores for "customer" and "user" are 1 ≥ 0.9), directly add them to the ontology concept - metadata mapping table; 2. If T low < <T high (e.g., confidence level = 0.85), submit for review by domain experts; 3. If ≤T low If the confidence level is 0.6, the mapping pair (e.g., "e-commerce domain product" and "supply chain domain material") is discarded directly.

[0067] Then proceed to S244: when the new mapping pair is successfully verified, the semantic vector of the ontology concept is updated with weights, and the semantic encoding model is incrementally fine-tuned based on the positive and negative samples that have been reviewed. Specifically, the system determines in real time whether the domain-wide model receives and verifies new mapping pairs. When a new mapping pair is successfully verified, the semantic vector of the ontology concept is automatically updated, and the HST-Encoder model is incrementally fine-tuned to achieve algorithm self-optimization. If the new mapping pair is not received, or is received but verification fails, no action is taken. In this process, the semantic vector is updated in the following way, which is a weighted average update, as shown in the following expression: in, This represents the learning rate (which controls the update magnitude), such as 0.05. Vold(o) i ) represents the original semantic vector of an ontology concept (such as "customer"); Vm j A semantic vector representing metadata such as "user".

[0068] In a real-world scenario, let's assume the original semantic vector for "customer" is [0.8, 0.2, 0.5], and the semantic vector for "user" is [0.85, 0.18, 0.52]. Substituting into the formula, we get Vnew=0.95x[0.8, 0.2, 0.5]+0.05x[0.85, 0.18, 0.52]=[0.8025, 0.199, 0.501]. After the update, the semantic vector of "customer" is closer to "user", and the recognition will be more accurate next time.

[0069] In this process, incremental model tuning employs triplet loss plus L2 regularization. The total loss function formula is as follows, used to avoid model overfitting and improve subsequent semantic alignment accuracy: in, Indicates the loss of the triplet. ; in, It is a semantic vector of a positive sample (such as a "user") that is successfully mapped to "customer"; It is the semantic vector of a negative sample (such as "material") that was not successfully mapped; margin=0.3 (interval parameter) =1e-4 (regularization coefficient).

[0070] By combining real-world scenarios and training with positive samples (users) and negative samples (materials), the model can more clearly distinguish between "synonymous" and "dissimilar" concepts, avoiding misjudging "customer" and "material" as synonyms.

[0071] S245: Periodically clean up mapping pairs that have been retained for longer than a preset period and whose confidence level is lower than a preset threshold, i.e., outdated mapping cleanup, to avoid mapping table expansion and affecting alignment accuracy.

[0072] For example, mapping entries that have been retained for more than one year and have a confidence level of less than 0.8 can be cleaned up on a monthly basis. For example, the mapping between "old system customer" and "new system user" (confidence level 0.75, retained for 15 months) can be deleted directly to ensure that the mapping table only retains highly reliable and timely mapping relationships.

[0073] Furthermore, the ontology concepts corresponding to each business domain are unified, that is, cross-domain concept unification.

[0074] After completing the above algorithm process, the ontology construction module unifies the ontology concepts of various business domains through a dual parallel path of "semantic similarity partitioning + domain-wide large-model semantic reasoning," focusing on resolving the issues of "homonyms and synonyms with different names." 1. Synonyms and synonyms (to be merged) For example, the marketing domain "user", customer service domain "customer", and financial domain "settlement object" are all determined to be synonymous after three-dimensional similarity calculation and large model inference, with a comprehensive confidence level of ≥0.9. The most frequently used "customer" in the business is selected as the main concept, and the three are merged to establish an equivalent mapping of "customer <-> user <-> settlement object".

[0075] 2. Differentiated names (handle accordingly) For example, the e-commerce domain "goods" (business meaning: selling items, with a comprehensive confidence level of only 0.5 compared to "materials") and the supply chain domain "goods" (business meaning: producing materials) have the same name, but their three-dimensional similarity is low and their comprehensive confidence level is <0.7. They are judged as different and their business domain affiliations are marked separately to avoid confusion.

[0076] Ultimately, the system assigns a unique identifier to each main concept, records the concept's origin, matching type (equivalence, inclusion), and tracing information, organizes these into an ontology alignment mapping table, and unifies ontology concept definitions based on this table, updates the ontology concept metadata mapping table, and achieves enterprise-wide semantic unification.

[0077] It should be noted that the attribute recognition in S24 and S22, and the axiom rule generation in S23 have cross dependencies. In actual execution, they can be optimized in parallel or iteratively. This manual describes them in logical order for clarity.

[0078] S25: Output a standardized ontology model, an ontology concept-metadata mapping table, an ontology alignment mapping table, and a rule base, which serve as the ontology model and its associated mapping relationships and axiomatic rule base.

[0079] After completing the aforementioned ontology concept extraction, concept attribute and relationship identification, ontology logical axioms and business rule generation, and cross-business domain ontology concept semantic alignment and fusion processing, the ontology construction module transforms the abstract business ontology definition into a standardized ontology model that can be parsed, reasoned, and executed by a computer. First, it clarifies the data attributes and object attributes corresponding to each ontology concept, and standardizes the data type, value range, uniqueness constraints, cardinality constraints, and other characteristics of each attribute. Then, it formally defines semantic relationships such as inheritance relationships, whole-part relationships, and business association relationships between different ontology concepts using the OWL ontology description language, and embeds the generated ontology logical axioms and business constraint rules.

[0080] Secondly, industry-standard ontology modeling tools, including but not limited to Protégé and TopBraid Composer, are utilized, employing international standard semantic web formal languages ​​such as OWL and RDF, to standardize and encode the defined ontology concepts, attributes, semantic relationships, logical axioms, and business rules, thus completing the structured construction of an enterprise-level ontology model. After the ontology model is built, logical consistency and business adaptability verification are performed, automatically identifying issues such as conceptual hierarchy conflicts, attribute definition domain value contradictions, axiom logical inconsistencies, and redundant cross-domain relationships, ensuring that the ontology model is logically rigorous, semantically conflict-free, and closely aligned with the enterprise's actual business rules.

[0081] After the consistency verification is passed, the ontology construction module uniformly organizes, standardizes, and formats all results, and outputs them in batches. The specific output content includes: 1) Output an enterprise-level standardized ontology model that conforms to the OWL international standard; 2) Solidify the correspondence between the ontology concepts and metadata generated in S21, forming an ontology concept-metadata mapping table; 3) Organize the results of cross-business domain concept synonym merging, homonymy differentiation and semantic matching in S24, and generate an ontology alignment mapping table; 4) Compile the ontology logic axioms and business constraint rules obtained from S23 reasoning, and output an ontology axiom and rule base that can be executed by machines in SWRL format. In practice, an example SWRL rule is shown below: illustrate: ?o, ?g, ?p, and ?n are all variables: ?o = Order ?g = product ?p = Unit price of the product ?n = Quantity purchased ∧ = and → = Derivation / Inference Greater than (x, y) = semantic relation: x is greater than y The semantics of this SWRL rule are: as long as an order contains a certain product, the product has a unit price, the order has a quantity, and the unit price is greater than 0, the total price of the order will be automatically inferred and calculated, and the total price will also necessarily be greater than 0.

[0082] Continue back Figure 1 Based on the ontology model and ontology concept-metadata mapping table constructed in step S2, the business data source associated with the data directory is obtained. Step S3 is then performed to establish a bidirectional traceable mapping relationship between the ontology model and the data directory metadata. Bidirectional verification processing is executed, and an ontology instance library is formed based on the verification results. Feedback information for optimizing metadata quality is generated. The bidirectional verification processing includes consistency verification and incremental updates of the ontology model and mapping relationship based on newly added or changed metadata in the data directory, and verification of instance data extracted from the business data source based on ontology axioms and rule bases. Step S3 specifically includes: S31: Based on the ontology concept-metadata mapping table, establish a two-way traceable mapping relationship between the ontology model and the data catalog metadata. The two-way traceable mapping relationship includes: the mapping between ontology classes and business tables, the mapping between ontology data attributes and table fields, and the mapping between ontology object attributes and table associations or data lineages. The two-way traceable mapping relationship is then solidified into the data catalog as the core basis for triggering incremental ontology updates when metadata changes occur.

[0083] For S31, the specific implementation process is as follows: First, the ontology construction module parses the classes, attributes, and their hierarchical relationships in the ontology model, establishing a one-to-one mapping between ontology classes and business tables, clarifying the corresponding business table in the data catalog for each ontology class (for example: the ontology class "Customer" maps to the "Customer Information Table" in the data catalog, and the ontology class "Order" maps to the "Order Details Table"). Second, it maps the data attributes of the ontology model to the fields of the corresponding business tables, ensuring that every data attribute in the ontology can find a matching field in the corresponding table in the data catalog (for example: the ontology data attribute "Customer Name" maps to the "Name" field in the "Customer Information Table"). The attribute "order amount" is mapped to the "amount" field in the "order details table"; finally, for the object attributes of the ontology model, the corresponding table relationships in the data directory are determined (implemented by foreign key association, for example: the ontology object attribute "order contains products" corresponds to the association between the "product ID" foreign key in the "order details table" and the "product ID" primary key in the "product information table" in the data directory), while tracing the source and flow of the data corresponding to the ontology concept, and clarifying the data lineage (for example: the "order amount" data comes from the "order details table", and its calculation depends on the "product unit price" in the "product information table" and the "purchase quantity" in the "order details table").

[0084] Through the above process, three standardized mapping relationships are ultimately formed: ontology class → business table, ontology data attribute → table field, and ontology object attribute → inter-table relationship (or data lineage). These three mapping relationships are then solidified into the data catalog, generating semantic mapping metadata to achieve bidirectional traceability between the ontology model and the data catalog, providing a foundation for subsequent instance extraction and verification.

[0085] S32: Based on the bidirectional traceable mapping relationship, extract ontology instance data from the business data source associated with the data directory, generate standardized RDF triples, and form an ontology instance library, which serves as the core data foundation for the subsequent construction of the business knowledge base.

[0086] For S32, the specific implementation process is as follows: The business data source associated with the data directory refers to the carrier storing the enterprise's actual business data, including but not limited to relational databases, file systems, etc., covering structured data (such as business tables in MySQL databases), semi-structured data (such as business files in JSON and XML formats), and unstructured data (such as business contract texts and customer feedback documents). The ontology construction module automatically reads the data records of the corresponding tables in each business data source according to the bidirectional mapping relationship, and converts each data record into an RDF triple conforming to the Semantic Web standard (example: extracting a data "Product ID: 2005, Product Name: Wireless Bluetooth Headset, Product Unit Price: 299 yuan" from the product information table, the corresponding RDF triple is "<product / 2005> <product name> <wireless Bluetooth headset>" and "<product / 2005> <product unit price> <299 yuan>"). The ontology instance library stores all ontology instances and their relationships in a graph structure, providing real and effective data support for subsequent automated verification and knowledge reasoning.

[0087] S33: Perform bidirectional verification processing, specifically including: based on the newly added or changed metadata in the data catalog, perform dynamic consistency verification on the ontology model and mapping relationship through the bidirectional traceable mapping relationship; if the verification fails, trigger incremental update of the ontology. Based on ontology axioms and rule base, an AI inference engine is used to automatically verify ontology instance data. The automatic verification includes data consistency verification, semantic integrity verification, and identification of redundant or conflicting instances. Instances with abnormal verification are handled in a graded manner according to preset rules. In this step, based on the ontology axioms and rule base generated in step S2, the ontology construction module calls the AI ​​inference engine to automatically verify the ontology instance data formed in step S32. Verification includes data consistency verification, semantic integrity verification, and identification of redundant or conflicting instances. The AI ​​inference engine is a deterministic inference mechanism based on symbolic logic and predefined rules, used only for executing the verification and inference of ontology axioms and SWRL rules, without involving semantic understanding or calling large domain models, to ensure the accuracy and efficiency of the verification results.

[0088] Semantic integrity verification is used to check whether an ontology instance is missing necessary semantic information (Example: Based on the actual e-commerce business operation needs, it can detect incomplete data such as "product instances for sale are missing product codes", "food products that require cold chain transportation are missing shelf life and storage temperature", and "home appliances are missing model specifications and after-sales warranty period". Such missing data will directly affect business processes such as product listing, warehousing, and after-sales service). The AI ​​inference engine applies ontology axioms and SWRL business rules to mark the three types of abnormal instances mentioned above, generating a detailed verification report that includes the abnormal type, abnormal location, violation rule, and rectification suggestions, providing a basis for subsequent metadata optimization.

[0089] S34: Feedback the verification results to the data catalog system, update the quality score and anomaly label of the corresponding metadata, and generate feedback information for optimizing metadata quality. In other words, this step feeds back the verification results, updates the metadata quality information, and triggers the rectification process.

[0090] For S34, the ontology construction module feeds back the verification results generated by S33 to the data directory, updates the quality score and anomaly tags of the corresponding metadata, and generates feedback information for optimizing metadata quality. For metadata with data quality problems, its quality score is automatically reduced, and a data quality rectification process is triggered, realizing two-way empowerment of ontology construction and data governance.

[0091] Warning Level: Data missing or minor anomalies that have no impact on business decisions (Example: The "marketing tag" attribute (such as "popular item" or "bestseller") of the product is missing, and the product's "remarks" are empty, which does not affect core business processes such as product listing, sales, and warehousing).

[0092] Secondly, the abnormal data corresponding to the abnormal tags is precisely matched with the metadata in the data catalog, and the abnormal information is pushed to the corresponding data owner or business manager. If the data owner returns the modified metadata within the specified period, the ontology construction module re-verifies the modified data until the data is qualified; if the data owner does not return the modified metadata, the system checks whether a metadata rule engine has been defined, and handles the situation according to the detection results: If a metadata rule engine has been defined: For severe-level anomalous metadata, a knowledge graph built based on historical metadata is loaded (this knowledge graph is constructed by extracting entities and relationships from various types of historical metadata, using RDF or graph database models, and pre-defined inference rules). Graph traversal algorithms are used to locate anomalous entities, construct local knowledge subgraphs, and the AI ​​inference engine matches the pre-defined rules to derive multiple correction schemes. All inference results are aggregated, and the scheme with the least conflict and highest feasibility is selected. After manual feasibility assessment, the corresponding metadata and data in the business data source are updated. For general-level and warning-level anomalous metadata, the rule engine directly performs automatic correction (such as data type conversion, missing value filling, format standardization, etc. For example, text characters in the "unit price" field of a product are automatically removed, and the missing "origin" field of a product is filled according to the historical data patterns associated with the product category, such as filling the corresponding production area for food products by default). After correction, the data is re-verified. If the verification passes, the data is updated; if the verification fails, the operation is rolled back and manual intervention is triggered.

[0093] If the metadata rules engine is not defined: continue to wait for the data owner to return the modified metadata until the rectification is completed or a timeout triggers manual supervision.

[0094] Through the above feedback and rectification mechanism, we can achieve two-way empowerment of data governance and ontology construction, which not only ensures the accuracy of the ontology instance library, but also continuously optimizes the quality of metadata, forming a closed loop of "ontology construction → instance verification → metadata optimization → ontology iteration".

[0095] To facilitate understanding of the technical solution of this application, the following description, in conjunction with embodiments, will introduce and explain the relevant content of constructing a multimodal business knowledge base based on the constructed ontology.

[0096] like Figure 2As shown, in one embodiment, after step S3, a step S4 is further included to construct a multimodal business knowledge base based on the constructed ontology model. In this step, the ontology model constructed in step S2 is used as the semantic framework, and the ontology instance library generated in step S3 is used as the core data to construct the basic layer of the business knowledge base. According to the pre-fine-tuned domain model and ontology semantic constraints, unstructured knowledge such as business rules, business processes, and business terms are automatically extracted from unstructured business documents (such as business manuals, contract texts, operating specifications, etc.). The extracted unstructured knowledge is semantically aligned with ontology concepts to ensure semantic unity of knowledge, and then it is integrated into the basic layer of the knowledge base. Through AI reasoning engine and graph mining algorithm, logical reasoning is performed on the explicit knowledge (such as explicit business rules and instance data) in the basic layer of the knowledge base to mine implicit knowledge (such as potential customer needs, product relationships, business process optimization points, etc.) to form a reasoning knowledge layer. The basic layer and the reasoning knowledge layer together constitute a complete multimodal business knowledge base, providing support for subsequent business decisions and knowledge reuse.

[0097] The specific steps of step S4 are as follows: S41: Using the ontology model as the semantic framework and the ontology instance library as the core data, the basic layer of the knowledge base is constructed. Based on the bidirectional traceable mapping relationship between the ontology model and the metadata of the data directory, the permission system and security level system of the basic layer and the data directory are connected to assign security tags, permission tags and business domain tags to knowledge entities.

[0098] Based on the OWL standard ontology model constructed in step S2 and the ontology instance library generated in step S3, this step is executed by the knowledge base construction module to build the foundation layer of the multimodal business knowledge base and complete deep integration with the data directory security and permission system. The specific implementation process is as follows: First, the business concepts, data attributes, object attributes, and semantic relationships between concepts defined in the ontology model are analyzed. Using the ontology model as a unified semantic framework, a three-layer mapping and matching process is performed on the core business data in the ontology instance library to ensure complete alignment between the knowledge base data and the ontology semantics. Here, core data refers to business data in the ontology instance library that is strongly related to the business concepts defined in the ontology model, covers business participants, business rules, execution processes, and other data that plays a decisive role in business operations (for example, in an e-commerce scenario, core data includes instance data corresponding to the ontology such as products, orders, categories, and suppliers). The three-layer mapping and matching specifically includes: Concept mapping: Accurately match the business entities in the core data with the concepts in the ontology model to determine the concept category to which each entity belongs (Example: match the instance of "XX brand 500ml laundry detergent" to the concept of "physical goods" in the ontology model, and match the instance of "XX brand flagship store" to the concept of "supplier"). Attribute mapping: Associating and binding the attributes of entities in the core data with the attributes of corresponding concepts in the ontology model, ensuring that numeric, enumerated, and text attributes are completely consistent with the data types and value range constraints defined in the ontology model (Example: binding the "unit price" attribute of a product instance with the "product unit price" data attribute of the "physical product" concept in the ontology model, and verifying that its value range conforms to the constraint rule of [0.01, 99999.99]). Relationship mapping: Identify the business relationships between entities in the core data and match and align them with the object attributes between concepts defined in the ontology model (Example: Match and align the instance relationship of "order contains goods" with the "contains" object attribute between the concepts of "order" and "goods" in the ontology model).

[0099] After the mapping and matching are completed, the knowledge base construction module determines the appropriate storage structure based on the enterprise's existing storage architecture and stores the core data that has been mapped in a structured manner: if a relational database is used for storage, the entity data is inserted into the corresponding business table, and the relationship between entities is linked through foreign keys; if a graph database is used for storage, the concepts in the ontology model are used as node types, business entities are used as node instances, entity attributes are used as node attributes, and the relationship between entities is used as edges to complete the graph structure storage.

[0100] After storage is completed, based on ontology axioms and rule base, the data stored in the base layer is checked for integrity and consistency. This checks for anomalies such as missing attributes, type mismatches, and relationship conflicts, ensuring that all data fully complies with the ontology model definition and business rule requirements.

[0101] Finally, the permission system and security level system of the knowledge base layer and the data directory are bidirectionally integrated, and each knowledge entity is assigned a corresponding standardized tag. The knowledge entity refers to a business object abstracted from core data based on the ontology model, possessing clear business semantics and independent meaning; the tag system specifically includes: Security Labels: Identify the sensitivity level and confidentiality requirements of knowledge entities, and are fully aligned with the security level of the corresponding metadata in the data catalog (Example: Core cost data of a product is marked as "highly sensitive", and publicly available sales information of a product is marked as "public"). Permission tags: Indicate the scope of users / user groups that can access and operate on this knowledge entity, and are fully synchronized with the permission control system of the data directory; Business domain label: Identifies the business domain to which the knowledge entity belongs, and corresponds completely to the business domains divided in S1 (Example: the "Home Appliance Products" entity is labeled "Home Appliance Business Domain", and the "Food Products" entity is labeled "Food Business Domain").

[0102] S42: Automated extraction and fusion of unstructured knowledge, which is based on the domain model and ontology semantic constraints, automatically extracts business rules, business processes and business cases from unstructured business documents, semantically aligns the extracted knowledge entities with ontology concepts, and integrates them into the base layer to form a business knowledge layer.

[0103] Specifically, this step is executed by the knowledge base construction module. It uses the full set of unstructured business documents collected in step S1 as the extraction object, and the ontology model constructed in step S2 as the unique semantic framework and extraction boundary constraints. Through a pre-fine-tuned domain-wide model, it completes the automated extraction, semantic alignment, and fusion of unstructured knowledge into the database. The specific implementation process is as follows: Unstructured document preprocessing: The document library of unstructured business documents in the enterprise is parsed, cleaned, and segmented into blocks. The document types include, but are not limited to, e-commerce business system specifications, product listing compliance manuals, category operation processes, after-sales processing rules, product quality control cases, supplier cooperation agreements, etc. At the same time, based on the business domain tags in S41, the documents are classified into business domains and correspond one-to-one with the domain ontology layer of the ontology model.

[0104] Knowledge extraction based on strong semantic constraints of ontology: An ontology-constrained knowledge extraction prompt template is constructed, injecting business concepts, data attributes, object attributes, and axioms defined in the ontology model as hard constraints into the prompt. This limits the domain-wide model to extracting only knowledge entities, attributes, and relationships within the defined scope of the ontology model, fundamentally suppressing the illusion of a large model. Through a pre-fine-tuned domain-wide model, three types of core unstructured knowledge are automatically extracted from the preprocessed document: Business rules: These include restrictive rules such as product compliance rules, category access rules, pricing rules, and after-sales compensation rules. Business processes: including step-by-step knowledge such as product listing review process, quality inspection process, warehousing and inbound / outbound process, and order fulfillment process; Business Cases: These include scenario-based knowledge such as case studies on best-selling product operations, customer complaint handling, and compliance risk management.

[0105] Standardized semantic alignment of extracted knowledge with ontology concepts: Reusing the BM-SA bidirectional mapping self-enhancing semantic alignment algorithm of S24, the extracted knowledge entities, attributes, and relations are semantically aligned with the corresponding concepts, data attributes, and object attributes in the ontology model, distinguishing between knowledge entities with the same name but different meanings and those with the same name but different names, ensuring that the extracted unstructured knowledge is completely consistent with the ontology semantic system; After alignment, the extracted knowledge entities are labeled with security tags, permission tags, and business domain tags consistent with the base layer.

[0106] Knowledge fusion and business knowledge layer construction: Unstructured knowledge that has been semantically aligned and tagged is integrated into the corresponding concept nodes of the knowledge base layer according to the semantic framework of the ontology model, realizing unified storage and semantic association between structured instance data and unstructured business knowledge; ultimately forming a business knowledge layer covering structured business data and unstructured business knowledge, providing full-dimensional business knowledge support for subsequent knowledge reasoning.

[0107] S43: Based on ontology axioms and rule base, use AI reasoning engine and graph mining algorithm to perform compliance reasoning on explicit knowledge in knowledge base, mine implicit knowledge corresponding to business scenarios, and update the verified implicit knowledge in reverse to ontology axioms and rule base to form reasoning knowledge layer.

[0108] Specifically, this step is executed by the knowledge base construction module. Using the business knowledge layer built in S42 as the data foundation and the ontology axioms and SWRL rule base generated in S23 as inference constraints, it combines a deterministic symbolic logic AI inference engine with graph mining algorithms to complete compliance inference of explicit knowledge and deep mining of implicit business knowledge, ultimately forming a standardized inference knowledge layer. The implicit knowledge mined in this step includes, but is not limited to, complementary relationships between products in e-commerce scenarios, potential combinations of best-selling products, product compliance risk points, abnormal inventory turnover patterns, and potential optimization points for category operations. The specific implementation process is as follows: I. Deterministic Logical Reasoning Based on Ontology Axioms and Rule Base The deterministic symbolic logic AI reasoning engine defined in S33 is invoked, and OWL ontology axioms and SWRL business rules from the ontology axioms and rule base are loaded to perform automated compliance reasoning and verification on explicit knowledge in the business knowledge layer, generating standardized reasoning conclusions: Compliance reasoning: Based on product compliance rules and category access rules, infer and identify product instances with compliance risks (Example: Based on the SWRL rule that "imported food must be labeled with importer information and food production license number", infer food products that lack the corresponding information and mark them as compliance risk instances). Category attribution reasoning: Based on the necessary and sufficient condition axiom of ontology concepts, product instances that are not explicitly classified are automatically categorized (Example: Based on the ontology axiom that "home appliances with cooling function and a volume ≥ 50L are classified as refrigerators", product instances that meet the conditions are automatically categorized under the "refrigerator" ontology concept). Business process reasoning: Based on the business rules of order fulfillment and after-sales processing, potential breakpoints and optimization nodes in the business process are inferred (example: based on the SWRL rule that "fresh produce orders must be shipped out within 24 hours of being placed," fresh produce orders that have not been shipped out within the time limit are inferred and marked as fulfillment risk nodes). All reasoning conclusions are synchronously updated to the attributes of the corresponding knowledge entities, and are also labeled with security tags, permission tags, and business domain tags consistent with the basic layer and business knowledge layer.

[0109] II. Implicit Association Knowledge Mining Based on Graph Mining Algorithms Using an ontology model as a semantic framework, structured instance data and unstructured business knowledge in the business knowledge layer are transformed into a standardized knowledge graph (with ontology concepts as node types, business entities as node instances, and ontology object attributes as edges). The Apriori association rule mining algorithm is then used to uncover hidden business association patterns in the knowledge graph. The specific implementation steps are as follows: Initialize algorithm parameters: Load the completed product business knowledge graph into a graph database (such as Neo4j), and preset the core thresholds of the algorithm: minimum support (default 0.05, used to filter frequent item combinations that meet the frequency criteria, which can be adjusted according to the business scale) and minimum confidence (default 0.7, used to evaluate the reliability of association rules, which can be adjusted according to business needs). Scanning the knowledge graph to generate frequent itemsets: The Apriori algorithm fully scans the product nodes, order nodes, category nodes, and related edges in the knowledge graph, counts the frequency of occurrence of various business combinations, and first filters out single-dimensional business items that meet the minimum support to obtain frequent 1-itemsets; based on the frequent 1-itemsets, candidate frequent 2-itemsets are generated through join and pruning operations, and the graph is scanned again to count the frequency, and frequent 2-itemsets that meet the minimum support are filtered out; the above iterative process is repeated until no new frequent itemsets can be generated; Generate high-value association rules: For each valid frequent itemset, generate all possible business association rules and calculate the confidence of each rule; select high-value association rules whose confidence meets the minimum confidence threshold as the tacit knowledge obtained (example: the high-confidence rule "85% of orders for infant formula will also include the purchase of infant diapers" is obtained, which is the tacit knowledge of complementary product association). Solidification of the reasoning knowledge layer: The above-mentioned deterministic logical reasoning conclusions and high-value implicit association rules are standardized and organized, classified and stored according to the semantic framework of the ontology model, and labeled and managed in sync, ultimately forming the reasoning knowledge layer, which together with the basic layer and business knowledge layer constitutes a complete three-layer architecture multimodal business knowledge base.

[0110] S44: Implement hierarchical and categorized management of the basic layer, business knowledge layer, and reasoning knowledge layer; construct interconnected semantic vector indexes, full-text indexes, and ontology semantic indexes; establish a synchronous update mechanism for the knowledge base, ontology model, and data catalog; and output a hierarchical enterprise-level business knowledge base.

[0111] In this sub-step, the knowledge base construction module executes the work, building upon the three-layer knowledge base construction results completed in S41-S43. Using the ontology model as a unified semantic framework and the data directory's permissions and security system as control benchmarks, it completes the standardized management, index construction, and final output of the knowledge base. The specific implementation process is as follows: I. Standardized Management of Three-Tier Knowledge Base (Hierarchical and Classified) Based on the business domain tags, security tags, and permission tags assigned to all knowledge entities by S41, a unified hierarchical classification and control system is implemented for the basic layer, business knowledge layer, and reasoning knowledge layer. Business domain classification: All knowledge is classified and collected according to e-commerce business domains (home appliance business domain, food business domain, apparel business domain, etc.), and corresponds one-to-one with the business domain division of S1 and the domain ontology layer of the ontology model to ensure clear business affiliation of knowledge. Security level classification: Knowledge is classified and controlled according to the sensitivity level of security labels (public, internal, sensitive, confidential), which is fully synchronized with the security level system of the data catalog. Highly sensitive knowledge is only open to authorized users. Knowledge type dimension layering: Strictly manage the knowledge in layers according to the hierarchical positioning of the basic layer, business knowledge layer, and reasoning knowledge layer, and clarify the update rules, verification rules and application boundaries of each layer of knowledge to ensure that the three layers of knowledge are semantically consistent, without redundancy and conflict.

[0112] II. Construction of a Multi-Dimensional Intelligent Indexing System To support efficient retrieval, semantic reasoning, and business applications of the knowledge base, three complementary intelligent indexes are constructed for the three-layer knowledge base. All indexes are bound to the tag system of knowledge entities, supporting accurate retrieval under access control. Semantic Vector Index: Reuses S24's HST-Encoder hybrid semantic encoder to generate standardized semantic vectors for all knowledge entities. Based on the vector database, it builds a semantic vector index to support core requirements such as product semantic matching, business rule semantic retrieval, and similar case recall in e-commerce scenarios. Full-text index: A full-text index is built for unstructured business knowledge (product compliance manuals, operation processes, business cases, etc.), supporting keyword search, fuzzy matching, and multi-condition combination search to meet the needs of business personnel for quick query of business rules and process documents; Ontology Semantic Index: Based on the concept hierarchy, attribute relationships, and semantic associations of the ontology model, an ontology semantic index is constructed. With ontology concepts as the core anchor points, it connects all corresponding knowledge entities, supports accurate knowledge retrieval based on business semantics, and cross-domain knowledge association queries, and completely solves the retrieval bias problem of "same name, different meaning, same meaning, different name".

[0113] III. Synchronous generation of standardized auxiliary products While completing knowledge base management and index construction, three types of standardized auxiliary artifacts are generated simultaneously to improve the knowledge base's control and application capabilities: Knowledge Index Repository: Integrates the above three types of intelligent indexes to provide the knowledge base with an efficient, accurate, and access-controlled retrieval mechanism, supporting rapid knowledge retrieval in scenarios such as business systems, intelligent question answering, and compliance auditing; Knowledge-Ontology Mapping Table: Records the one-to-one correspondence between all knowledge entities and concepts, attributes, and relationships in the ontology model. It works in conjunction with the ontology concept-metadata mapping table generated in S21 and the ontology alignment mapping table generated in S24 to help business personnel understand the semantic definition of knowledge and business boundaries. Knowledge-Data Catalog Association Table: Establishes a two-way association between knowledge entities and data catalog metadata, deeply integrates with the data catalog's permission system and security system, and realizes synchronous control of "metadata permissions → knowledge access permissions". It also supports reverse tracing of the source, lineage and quality information of corresponding business data based on knowledge.

[0114] IV. Definition and Boundaries of the Three-Tier Knowledge Base The three-tiered business knowledge base constructed in this invention, with the positioning, core content, and e-commerce scenario examples for each layer of knowledge, is as follows: Foundational knowledge: This is the semantic foundation of the entire knowledge base. It is built on the ontology model and ontology instance library and is a standardized definition and structured storage of the core basic elements of e-commerce business. The core content includes: the attribute and feature information of business entities (instance data such as products, orders, suppliers, and categories), and the standardized definition of business concepts (class, attribute, and relationship definitions in the ontology model), providing a unified semantic benchmark for upper-layer knowledge. Business knowledge layer: This is the application and extension of basic knowledge in actual e-commerce business scenarios. It is standardized knowledge that guides business execution. The core content includes: business rules such as product compliance access, pricing, and after-sales service; business process descriptions such as product listing, quality inspection, and fulfillment; and summaries of historical business experience such as best-selling product operation and customer complaint handling. It achieves the unified integration of structured business data and unstructured business knowledge. The reasoning knowledge layer consists of incremental implicit knowledge generated by AI reasoning engines and graph mining algorithms, based on the first two layers. It supplements and extends existing business knowledge. Its core content includes implicit business rules such as complementary product associations and category linkages, reasoning conclusions such as compliance risks and performance breakpoints, and the inherent knowledge relationships across business domains, providing data support for business decision-making, risk warning, and intelligent recommendation.

[0115] Ultimately, the knowledge base construction module will complete the hierarchical classification management and index construction of a three-layer knowledge base, and output it in a unified format along with three types of auxiliary products to form a complete three-layer architecture enterprise-level multimodal business knowledge base, providing standardized semantic support for subsequent incremental iterations and business applications.

[0116] In this embodiment, such as Figure 2 As shown, based on the completion of the above-mentioned hierarchical knowledge base construction, in order to enable the ontology and knowledge base to continuously adapt to business changes and achieve automated management throughout the entire lifecycle, some implementations also include an incremental iterative update step S5 (…). Figure 2 The dashed line between S4 and S5 indicates that S4 and S5 are not executed sequentially, but rather S5 is executed only when there is a relevant event. This step uses an event-driven mechanism to detect metadata change events in the data catalog and incremental data changes in the business data source in real time, automatically triggering incremental updates of the ontology model and synchronous iterations of the business knowledge base. The specific implementation process is as follows: S51: Real-time monitoring of preset metadata change events in the data catalog through an event-driven mechanism, as well as real-time acquisition of incremental data changes from business data sources, and filtering and classifying the monitored change events, filtering invalid changes and classifying update levels according to their impact scope.

[0117] For S51, an event-driven listening component is deployed in the incremental iteration module to capture change events of all types of metadata in the data catalog in real time, including but not limited to: adding / deleting business data tables, modifying field definitions / constraints, updating business terminology / data standards, adjusting business domain divisions, and changing business data source configurations, covering the full range of technical metadata, business metadata, and management metadata defined in S51. Simultaneously, incremental data changes in the business data source are obtained in real time through database log capture, CDC data change capture, and business system change notification mechanisms. All captured change events are encapsulated into standardized messages and pushed to a message queue for asynchronous and orderly processing in subsequent incremental update processes, avoiding system blocking caused by high-concurrency change events.

[0118] S52: Automatically trigger the corresponding level of incremental ontology update based on the detected change event. The incremental update includes adding or adjusting ontology concepts, attributes, relationships, axioms and cross-domain alignment mapping relationships, as well as synchronously updating the bidirectional mapping relationship between the ontology and the data directory. After the incremental update is completed, a consistency check is performed based on the bidirectional verification processing method in step S3. If the check fails, an automatic rollback is triggered.

[0119] For S52, the incremental iteration module automatically triggers the corresponding granularity of the ontology incremental update process based on the type and scope of the received change event. The incremental update content includes the addition, adjustment, and removal of ontology concepts, attributes, relationships, axioms, rules, and cross-domain alignment mapping relationships. The specific implementation rules are as follows: If a new business data table event is detected (for example, adding an "Imported Cold Chain Food Information Table" in an e-commerce scenario), the pre-fine-tuned domain model is invoked, and the semantic enhancement method of S1 is combined to analyze the table structure, field meaning and business semantics, identify and add corresponding ontology concepts, data attributes and object attributes, and update the ontology axioms and rule base synchronously. If a field definition / constraint modification event is detected (for example, in an e-commerce scenario, the value range of the "product unit price" field is adjusted from [0.01, 99999.99] to [0.01, 199999.99]), the value range constraint and cardinality constraint of the corresponding data attribute in the ontology will be automatically adjusted. If a business term / business domain adjustment event is detected, the name, definition, and business domain affiliation of the corresponding ontology concept will be modified synchronously to ensure that the ontology semantics and business terminology are completely consistent. If the change event affects the equivalence / distinction relationship of existing cross-domain concepts, then the BM-SA bidirectional mapping self-enhancing semantic alignment algorithm of S24 is reused to re-execute cross-domain semantic alignment and update the ontology alignment mapping table.

[0120] After all incremental modifications at the ontology level are completed, the incremental iteration module synchronously updates the bidirectional traceable mapping relationship between the ontology and the data catalog in S31, ensuring that the mapping relationship is completely consistent with the latest ontology model and metadata changes, and realizing real-time linkage between the ontology and the data catalog.

[0121] S53: Based on the incremental updates of the ontology and incremental business data, automatically synchronize and update the instance data, knowledge entities and corresponding indexes in the knowledge base, and extract knowledge from newly added unstructured documents for semantic alignment and fusion.

[0122] For S53, the incremental iteration module automatically updates the three-tier architecture business knowledge base based on the incremental ontology version generated by S52 and the incremental data from the business data source. Specific implementation details include: Base layer update: Based on changes in ontology concepts, attributes, and relationships, perform add, modify, and remove operations on RDF triples in the ontology instance library; based on bidirectional mapping relationships, extract new ontology instances from incremental business data, convert them into standardized RDF triples, and store them in the library; Business knowledge layer update: For newly added unstructured business documents (examples: "Compliance Manual for Imported Cold Chain Food" and "Pricing Rules for Major Promotional Products" added in e-commerce scenarios), reuse the ontology semantic strong constraint knowledge extraction method of S42, call the pre-fine-tuned domain big model to extract business rules, processes and cases, and after completing the semantic alignment with ontology concepts, integrate them into the business knowledge layer, and update the knowledge-ontology mapping table and knowledge-data directory association table simultaneously; Reasoning knowledge layer update: Based on the updated basic layer and business knowledge layer, reuse S43's AI reasoning engine and graph mining algorithm to re-execute logical reasoning and implicit knowledge mining, and update the content of the reasoning knowledge layer; Intelligent index synchronous update: Synchronously update the semantic vector index, full-text index, and ontology semantic index built by S44 to ensure the timeliness and accuracy of knowledge retrieval results.

[0123] S54: Perform canary releases of incremental updates, set verification rules and rollback mechanisms, and manage versions, supporting historical version backtracking and comparison.

[0124] For S54, the incremental iteration module performs full-process control over the incrementally updated ontology model and business knowledge base. The specific implementation process is as follows: Canary release and automated verification: The incrementally updated ontology and knowledge base are treated as new versions and deployed using a canary release strategy. First, the new version is deployed in a canary environment, reusing the ontology logical consistency verification rules of S25 and the instance data verification rules of S33 to perform automated verification; after verification, the canary release scope is gradually expanded according to business domains until full release. Automatic rollback for anomalies: If automated verification fails or business logic anomalies or semantic conflicts occur in the gray-scale environment, the rollback mechanism will be automatically triggered to restore to the previous stable version, so as to avoid affecting the operation of online business (Example: During e-commerce promotions, if the incremental update verification of the product category body fails, it will automatically roll back to the stable version to ensure that the core business of product listing and order fulfillment is not affected). Full lifecycle version management: Full lifecycle version management is performed on the ontology model, business knowledge base, mapping table, and rule base. It records the triggering event, change time, change content, operator, and verification results for each change. It supports one-click retrospective of historical versions and comparative analysis of version differences, ensuring the reliability, traceability, and auditability of enterprise-level knowledge assets. In some embodiments, to continuously assess the quality and optimize the ontology construction, knowledge base generation, and domain-wide model, a full-process closed-loop optimization step S6 is also included. Step S6 includes the following sub-steps: S61: Construct a multi-dimensional quality assessment system, trigger automated assessments according to preset cycles or incremental update events, and generate quality assessment results. The assessment dimensions include metadata completeness, standardization, and accuracy; business adaptability, logical consistency, and cross-domain alignment accuracy of the ontology model; and knowledge accuracy, business coverage, and semantic consistency of the business knowledge base.

[0125] For S61, this step is executed by the closed-loop optimization module, initiating a full-link quality assessment task at a configurable cycle (a full assessment is performed weekly by default, with a special assessment triggered after significant business changes). The assessment system covers three core dimensions: metadata quality, ontology quality, and knowledge base quality. Each dimension has clearly defined quantitative assessment standards, calculation methods, and passing thresholds, corresponding one-to-one with the technical nodes of the entire process from S1 to S5. The specific assessment rules are as follows: Metadata quality dimension: Corresponding to the S1 metadata collection and preprocessing stage, the core assessment indicators are completeness rate, standardization rate, and accuracy rate. The evaluation rules, which are aligned with the e-commerce product business scenario, are as follows: Completeness rate: refers to the percentage of core metadata fields (product code, category, unit price constraint, compliance qualification, field business description, etc.) that are completely filled. The calculation formula is "number of core metadata fields that have been completely filled ÷ total number of core metadata fields × 100%", and the qualified threshold is not less than 98%. Standardization rate: refers to the proportion of metadata formats that conform to enterprise data standards, including data type specifications, naming specifications, value range formats, constraint rule specifications, etc. The calculation formula is "number of metadata items that conform to the standards ÷ total number of metadata items × 100%", and the qualified threshold is not less than 99%. Accuracy: This refers to the semantic consistency between metadata descriptions and actual business meanings. It is verified through sampling comparison and pre-fine-tuned domain-wide model semantic verification. The calculation formula is "number of semantically consistent metadata samples ÷ total number of sampled samples × 100%", with a passing threshold of no less than 98%. During the evaluation process, the system automatically scans all metadata in the data catalog, verifying its completeness and standardization item by item. For randomly selected samples, the pre-fine-tuned domain-wide model is used to perform semantic consistency verification, outputting quantitative scores and anomaly details for each indicator.

[0126] Ontology quality dimension: Corresponding to the S2 ontology construction and S24 cross-domain semantic alignment stages, the core assessment indicators are business adaptability, logical consistency, and cross-domain alignment accuracy. Business Adaptability: This measures the degree of matching between ontology concepts, attributes, and relationships and the enterprise's e-commerce business scenarios. It is calculated by semantically comparing the business terms defined in the ontology with the business metadata in the data catalog and the product operation business query logs. The calculation formula is "number of ontology concepts with successful business matching ÷ total number of core ontology concepts × 100%", and the qualified threshold is not less than 95%. Logical consistency: Verify whether there are semantic conflicts and logical contradictions in the ontology, including conflicts between class equivalence and mutual exclusion, conflicts between attribute value ranges, hierarchical circular inheritance, and contradictions between axioms and rules. Call the deterministic symbolic logic AI inference engine that is reused throughout the entire process of this solution to perform consistency detection. The passing standard is no logical conflict. Cross-domain alignment accuracy: This assesses the correctness of merging and differentiating ontology concepts across different business domains. For aligned concept pairs in e-commerce business domains such as home appliances, food, and apparel, verification is performed through automated validation and expert sampling. The calculation formula is "number of correctly aligned concept pairs ÷ total number of sampled concept pairs × 100%", with a passing threshold of no less than 99%. After the assessment, an ontology quality sub-report is generated, clearly defining the scope of impact and root cause of non-compliant items.

[0127] Knowledge base quality dimensions: Corresponding to the S4 three-tier architecture business knowledge base construction phase, the core assessment metrics are accuracy, coverage, and consistency. Accuracy: refers to the correct proportion of knowledge stored in the knowledge base (product instance triples, compliant business rules, category association reasoning conclusions, etc.). It is verified through random sampling + pre-fine-tuning of the domain-wide model for verification + expert review. The calculation formula is "number of correctly verified knowledge items ÷ total number of sampled knowledge items × 100%", and the passing threshold is not less than 99%. Coverage: This refers to the degree to which the knowledge base covers the core business domains of e-commerce. It assesses the proportion of key entities, concepts, and rules in the knowledge base for core business domains such as home appliances, food, apparel, orders, fulfillment, and after-sales service. The passing standard is 100% coverage of the core business domains. Consistency: Verify whether there are contradictory knowledge in the knowledge base, including issues such as the same entity being assigned mutually exclusive attribute values, conflicting statements of the same business rule, and inference conclusions contradicting basic rules. A full scan is performed using an AI inference engine and conflict detection algorithm. The passing standard is no redundant and no conflicting knowledge.

[0128] After the comprehensive evaluation is completed, the quality control and model optimization module summarizes the quantitative indicators, anomaly details, root cause analysis, and improvement suggestions for metadata quality, ontology quality, and knowledge base quality, and generates a structured and standardized quality evaluation report, which serves as the core basis for subsequent closed-loop optimization.

[0129] S62: Feed the quality assessment results back to the data catalog system to drive the completion, optimization and standardization of the corresponding metadata, and simultaneously feed back ontology defects and knowledge conflict issues to the incremental update process of the ontology and knowledge base, so as to achieve full-link reverse optimization.

[0130] For S62, the closed-loop optimization module synchronously feeds back to the data catalog system any metadata quality anomalies identified in the S61 quality assessment report, including missing business terms, inconsistent field definitions, non-standard constraint rules, and outdated data standards. Based on this feedback, the data catalog system automatically or semi-automatically triggers a metadata completion and optimization process, forming a closed loop with the S34 metadata rectification process. Specific optimization rules include: For metadata containing missing business descriptions and terms (such as no business comments for product fields or no compliance standard descriptions for product categories), the pre-tuned domain model is invoked, and candidate business terms and comments are generated by combining field names, data samples, and related business documents. After confirmation by the data administrator, the terms and comments are automatically completed. For metadata with non-standard formats and inconsistent definitions (such as inconsistent data types in the product unit price field and inconsistent category coding rules), the format unification and standardization correction will be automatically performed in accordance with the enterprise data standards. For expired or invalid business rules and data standards, mark them as pending review and push them to the corresponding business person in charge and data owner. After review and update, synchronize them to the data directory.

[0131] This closed-loop feedback mechanism continuously improves the metadata quality of the data catalog, providing higher-quality input for subsequent ontology construction and knowledge base updates, forming a two-way empowerment closed loop of "metadata quality improvement → ontology and knowledge base quality improvement → reverse optimization of metadata governance".

[0132] S63: Based on quality assessment results, explicit user error correction feedback, and implicit business behavior feedback, automatically optimize the AI ​​agent prompts, HST-Encoder semantic encoding model parameters, and semantic alignment algorithm weights that are relied upon for ontology construction, knowledge extraction, semantic alignment, and reasoning.

[0133] For S63, the closed-loop optimization module collects two types of core optimization criteria: first, the shortcomings of the indicators in the S61 quality assessment report (such as the accuracy of ontology concept extraction not meeting the standard, the error rate of cross-domain semantic alignment being too high, and the knowledge extraction illusion rate exceeding the standard); second, explicit feedback from users during the use of the knowledge base (knowledge correction annotation, rule validity evaluation, likes / dislikes of search results) and implicit behavioral feedback (product semantic retrieval click-through rate, business rule adoption rate, question-answer matching accuracy). Based on the above criteria, automated optimization is performed on the pre-fine-tuned domain model that can be reused throughout the entire process. The specific optimization content and implementation methods are as follows: Automated optimization of prompt word engineering: For the Prompt templates of core links such as ontology concept extraction, attribute relationship recognition, unstructured knowledge extraction, and semantic alignment, based on quality assessment results and user feedback, the constraint rules, example content, and output format of the Prompt are automatically and iteratively optimized through Bayesian optimization algorithm to strengthen the strong semantic constraints of the ontology and continuously reduce the illusion rate of large models; Automated hyperparameter tuning: For core hyperparameters such as similarity weights of the BM-SA semantic alignment algorithm, semantic vector dimension of the HST-Encoder, rule confidence threshold of the AI ​​inference engine, and support / confidence threshold of the association rule mining algorithm, automated optimization is performed based on changes in quality indicators. After each optimization, the effect is verified in a gray-scale environment, and updated to the production environment after the indicators meet the standards. Lightweight Incremental Fine-Tuning: To address the shortcomings of large models in adapting to enterprise e-commerce businesses, user-annotated error correction data and high-quality business samples are collected to construct a fine-tuning dataset. Lightweight incremental fine-tuning is performed using reinforcement learning based on human feedback (RLHF) to enhance the model's semantic understanding capabilities for core e-commerce business scenarios such as product compliance, category operation, and order fulfillment, and further improve the accuracy of ontology construction and knowledge extraction.

[0134] After each optimization, the model version, optimization content, and changes in metrics are recorded, supporting version retrospection and effect comparison. Through this continuous self-iteration mechanism, the adaptability of the domain-wide model to the specific business of the enterprise is continuously improved, the illusion rate is continuously reduced, and a full-link self-optimization closed loop of "quality assessment → feedback optimization → model iteration → quality improvement" is formed.

[0135] like Figure 3 As shown, in some embodiments, this application also proposes a system for automatically constructing an ontology and business knowledge base based on a data catalog using AI, applicable to the method for automatically constructing an ontology and business knowledge base based on a data catalog in the above embodiments. The system includes: The metadata semanticization module, corresponding to step S1 in the method embodiment, is used to collect metadata from various domains in the data catalog system, obtain associated unstructured business documents, group the corresponding domains to form several domain groups, perform semantic enhancement processing on each domain group based on the unstructured business documents, and generate semantic processing results for each domain group. The semantic processing results include at least a structured metadata dataset, a metadata semantic vector library, and a mapping table between metadata and associated unstructured business documents. The ontology construction module, corresponding to step S2 in the method embodiment, communicates with the metadata semanticization module. It is used to call the domain big model for ontology recognition processing based on the semanticization processing results, and to construct the ontology model and its supporting mapping relationship and axiom rule base based on the processing results. The ontology recognition processing includes extracting ontology concepts, identifying the attributes and relationships between ontology concepts, generating ontology axioms and standardized business rules, and performing cross-domain semantic alignment and fusion of ontology concepts from different domain groups. The bidirectional mapping and verification module, corresponding to step S3 in the method embodiment, communicates with the ontology construction module and the data catalog system. It is used to establish a bidirectional traceable mapping relationship between the ontology model and the metadata of the data catalog, perform bidirectional verification processing, and form an ontology instance library and generate feedback information for optimizing metadata quality based on the verification results. The bidirectional verification processing includes consistency verification and incremental updates of the ontology model and mapping relationship based on newly added or changed metadata in the data catalog, and compliance verification of ontology instance data extracted from business data sources based on ontology axioms and rule base.

[0136] Furthermore, some embodiments also include a knowledge base construction module: corresponding to step S4 of the method embodiment, it is communicatively connected to the bidirectional mapping and verification module and the ontology construction module, and is used to construct the basic layer of the knowledge base with the ontology model as the unified semantic framework and the ontology instance library as the core data; based on the pre-fine-tuned domain large model and strong semantic constraints of the ontology, it automatically extracts unstructured knowledge such as business rules, business processes, and business cases from unstructured business documents, completes semantic alignment with ontology concepts, and then integrates them into the knowledge base to form a business knowledge layer; through a deterministic AI reasoning engine and graph mining algorithm, it performs logical reasoning and implicit knowledge mining on explicit knowledge to form a reasoning knowledge layer; finally, it completes the construction of a three-layer architecture multimodal business knowledge base, and builds a multi-dimensional intelligent indexing system with semantic vector index, full-text index, and ontology semantic index.

[0137] Incremental Iteration Module: Corresponding to step S5 in the method embodiment, it interfaces with the enterprise data catalog system, business data source, ontology construction module, and multimodal business knowledge base construction module, respectively. It is used to monitor metadata change events in the data catalog and incremental data changes in the business data source in real time through an event-driven mechanism, automatically triggering incremental updates of the ontology model and synchronous iteration of the business knowledge base; at the same time, it is responsible for the canary release, automated verification, exception rollback, and full lifecycle version management of incremental update content.

[0138] Closed-loop optimization module: Corresponding to step S6 in the method embodiment, it communicates and connects with all the aforementioned functional modules and the enterprise data catalog system. It is used to build a multi-dimensional quantitative evaluation system covering metadata quality, ontology semantic adaptability, and knowledge base quality. It regularly performs full-link automated quality evaluation and generates standardized quality evaluation reports. The quality evaluation results are fed back to the data catalog system to drive the completion, optimization, and standardization of metadata. At the same time, based on the quality evaluation results and user feedback, it performs automated optimization of prompt word engineering, algorithm hyperparameter tuning, and lightweight incremental fine-tuning on the pre-fine-tuned domain model that can be reused throughout the entire process, realizing a self-optimizing closed loop throughout the entire process.

[0139] The specific execution methods, implementation details, and technical effects of each functional module in this embodiment have been fully and thoroughly described in the foregoing method embodiments, and will not be repeated here.

[0140] Based on the embodiments described above, using the technical solution of this application to construct the ontology and business knowledge base can achieve the following beneficial technical effects: This solution achieves end-to-end AI automation from metadata semantic enhancement, automated ontology construction, multimodal knowledge base generation to full lifecycle iterative management. It shortens the implementation cycle of traditional ontology construction projects from several months to hours, effectively reduces manual annotation and expert input, and significantly lowers the professional threshold and implementation cost of enterprise-level ontology and knowledge base construction. Small and medium-sized enterprises can also quickly complete the construction of business ontology and knowledge base based on this solution.

[0141] This solution uses the enterprise data catalog as the core anchor point to build a two-way traceable mapping and real-time linkage mechanism between the ontology model and metadata and business data. At the same time, through an event-driven incremental iteration mechanism, it realizes the real-time dynamic adaptation of the ontology model and business knowledge base to business changes, transforming the traditional static and rigid ontology model into an active semantic model that can continuously iterate with business changes. Compared with traditional solutions, the consistency between ontology and business data is improved to near the ideal level, fundamentally avoiding the common industry problem of ontology being disconnected from business and "unusable after one construction".

[0142] This solution is based on the full business metadata and business context of the enterprise data catalog, and uses the ontology semantic framework as a hard boundary to strongly constrain the entire process of concept extraction, semantic understanding, and knowledge generation of the domain big model, thereby reducing the illusion rate of the big model from the root. The generated ontology model can be deeply adapted to the enterprise's own business scenarios, data asset status and management rules. The adaptability of ontology concepts to enterprise business scenarios is greatly improved, and the problem of the disconnect between general ontology and actual enterprise business and poor usability is completely solved.

[0143] This solution deeply reuses the existing metadata assets, data lineage, access control, and quality control results of the enterprise data catalog, eliminating the need to repeatedly build data governance infrastructure and significantly reducing project implementation costs. At the same time, through a full-chain quality assessment system, the quality assessment results of the ontology and knowledge base drive the completion, optimization, and standardization of the metadata of the data catalog, forming a two-way empowerment closed loop of "metadata quality improvement → ontology and knowledge base quality optimization → reverse empowerment of data governance", which simultaneously enhances the core value of the enterprise's data assets and knowledge assets.

[0144] This solution utilizes the BM-SA bidirectional mapping self-enhancing semantic alignment algorithm to achieve automated alignment and fusion of ontologies across business domains. This completely solves the problem of semantic silos and barriers across departments, systems, and business domains within enterprises, and constructs a unified enterprise-level business semantic standard. Based on this, a three-layer architecture multimodal business knowledge base can directly support full-scenario business applications such as intelligent semantic retrieval, intelligent business question answering, automated compliance auditing, intelligent product operation, and intelligent risk control, enabling rapid business value transformation of enterprise knowledge assets.

[0145] Furthermore, it should be noted that the formulas and / or expressions involved in the relevant embodiments are all dimensionless calculations. The formulas are derived from software simulations using a large amount of collected data to obtain the most recent real-world formula. The preset parameters in the formulas are set by those skilled in the art according to the actual situation. For example, if there are weighting coefficients and proportional coefficients, their values ​​are set to quantify each parameter to obtain a specific value, which facilitates subsequent comparison. Regarding the size of the weighting coefficients and proportional coefficients, it is acceptable as long as they do not affect the proportional relationship between the parameters and the quantified values.

[0146] The above embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and are not intended to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the scope of the technology disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for automatically constructing ontology and business knowledge base based on a data catalog using AI, characterized in that, The ontology construction process includes: Step S1: Collect metadata from each domain in the data catalog system and obtain the associated unstructured business documents. Group the corresponding domains to form several domain groups. Perform semantic enhancement processing on each domain group based on the unstructured business documents to generate semantic processing results for each domain group. The semantic processing results include at least a structured metadata dataset, a metadata semantic vector library, and a mapping table between metadata and associated unstructured business documents. Step S2: Based on the semantic processing results, the domain big model is called to perform ontology recognition processing, and an ontology model and its supporting mapping relationship and axiom rule base are constructed based on the processing results. The ontology recognition processing includes extracting ontology concepts, identifying the attributes and relationships between ontology concepts, generating ontology axioms and standardized business rules, and performing cross-domain semantic alignment and fusion of ontology concepts from different domain groups. Step S3: Establish a bidirectional traceable mapping relationship between the ontology model and the metadata of the data directory, perform bidirectional verification processing, and form an ontology instance library based on the verification results and generate feedback information for optimizing metadata quality. The bidirectional verification processing includes performing consistency verification and incremental updates on the ontology model and mapping relationship based on newly added or changed metadata in the data directory, and performing compliance verification on ontology instance data extracted from business data sources based on the ontology axioms and rule base.

2. The method for automatically constructing ontology and business knowledge base based on data catalog using AI according to claim 1, characterized in that, The specific steps for generating the semantic processing result in step S1 include: S11: Perform deduplication, completion, and format standardization on the collected metadata to form standardized metadata; S12: Classify and cluster the standardized metadata according to preset business domain rules to form several domain groups; S13: Based on the domain-wide model, standardized business semantic tags and semantic vectors are generated for the metadata entities in each domain group, forming a structured meta-dataset and a metadata semantic vector library; S14: Based on the standardized business semantic tags, extract the corresponding metadata business semantic supplementary information from the associated unstructured business documents, establish a mapping table between the metadata and the unstructured business documents, and provide business context constraints for subsequent ontology construction.

3. The method for automatically constructing ontology and business knowledge base based on data catalog using AI according to claim 1, characterized in that, Step S2 specifically includes: S21: Construct an AI agent, input the semantic processing result into the AI ​​agent, the AI ​​agent calls the domain big model, extracts the ontology concept of the business domain with the business semantics of the metadata as the core, and divides it into layers according to the three-level architecture of public ontology layer, domain ontology layer and scene ontology layer. S22: Identify the data attributes and object attributes of each ontology concept, and generate the domain, range, cardinality constraints of the data attributes, as well as the inheritance relationship, association relationship and semantic characteristics of the object attributes; S23: Based on the constraint rules and business rules of metadata, generate ontology logical axioms and output the rule base in SWRL form; S24: Perform semantic alignment and fusion of ontology concepts from different business domains. Based on semantic vector similarity calculation and domain-wide model semantic reasoning, distinguish between homonymous and synonymous concepts, merge synonymous concepts, generate ontology alignment mapping table, and optimize ontology concepts, attributes and axiom rules in reverse based on the alignment results. S25: Output a standardized ontology model, ontology concept-metadata mapping table, ontology alignment mapping table, and rule base, which serve as the ontology model and its associated mapping relationships and axiomatic rule base.

4. The method for automatically constructing ontology and business knowledge base based on data catalog using AI according to claim 3, characterized in that, The semantic alignment and fusion of ontology concepts from different business domains in S24 further includes: S241: Generate semantic vectors of ontology concepts and metadata through the HST-Encoder semantic encoding model, calculate the initial similarity from three dimensions: text semantics, structural topology and instance distribution, and then fuse them according to preset weights to obtain the basic comprehensive similarity. S242: Obtain historically verified mapping information as prior knowledge, introduce a time decay mechanism to calculate the historical mapping enhancement factor, and combine it with the basic comprehensive similarity to calculate the comprehensive mapping confidence. S243: Generate a dynamic threshold based on the confidence statistics of historically verified mapping pairs. Based on the comparison between the comprehensive mapping confidence and the dynamic threshold, add the mapping pairs to the ontology concept-metadata mapping table, submit them for manual review, or discard them. The annotation results of the manual review serve as positive and negative samples for model fine-tuning. S244: When a new mapping pair is successfully verified, the semantic vector of the ontology concept is updated with weights, and the semantic encoding model is incrementally fine-tuned based on the positive and negative samples that have been verified. S245: Periodically clean up mapping pairs that have been retained for longer than a preset period and whose confidence level is lower than a preset threshold.

5. The method for automatically constructing ontology and business knowledge base based on data catalog using AI according to claim 3, characterized in that, Step S3 specifically includes: S31: Based on the ontology concept-metadata mapping table, establish a bidirectional traceable mapping relationship between the ontology model and the data catalog metadata. The bidirectional traceable mapping relationship includes: mapping between ontology classes and business tables, mapping between ontology data attributes and table fields, and mapping between ontology object attributes and table associations or data lineages. The bidirectional traceable mapping relationship is then solidified into the data catalog as the core basis for triggering incremental ontology updates when metadata changes occur. S32: Based on the bidirectional traceable mapping relationship, extract ontology instance data from the business data source associated with the data directory, generate standardized RDF triples, and form an ontology instance library, which serves as the core data foundation for the subsequent construction of the business knowledge base. S33: Perform bidirectional verification processing, specifically including: based on the newly added or changed metadata in the data catalog, perform dynamic consistency verification on the ontology model and mapping relationship through the bidirectional traceable mapping relationship; if the verification fails, trigger incremental update of the ontology. Based on the ontology axioms and rule base, an AI inference engine is used to automatically verify the ontology instance data. The automatic verification includes data consistency verification, semantic integrity verification, and identification of redundant or conflicting instances. Instances with abnormal verification are handled in a graded manner according to preset rules. S34: Feedback the verification results to the data catalog system, update the quality score and anomaly label of the corresponding metadata, and generate feedback information for optimizing metadata quality.

6. The method for automatically constructing ontology and business knowledge base based on data catalog using AI according to claim 5, characterized in that, It also includes step S4, which involves constructing a multimodal business knowledge base based on the constructed ontology model. Step S4 includes: S41: Using the ontology model as the semantic framework and the ontology instance library as the core data, construct the basic layer of the knowledge base. Based on the bidirectional traceable mapping relationship between the ontology model and the metadata of the data directory, connect the basic layer with the permission system and security level system of the data directory, and assign security tags, permission tags and business domain tags to knowledge entities. S42: Based on the domain big model and ontology semantic constraints, automatically extract business rules, business processes and business cases from unstructured business documents, semantically align the extracted knowledge entities with ontology concepts, and integrate them into the base layer to form a business knowledge layer. S43: Based on the ontology axioms and rule base, use AI reasoning engine and graph mining algorithm to perform compliance reasoning on explicit knowledge in the knowledge base, mine implicit knowledge corresponding to business scenarios, and update the verified implicit knowledge in reverse to the ontology axioms and rule base to form a reasoning knowledge layer. S44: Implement hierarchical classification and management of the basic layer, business knowledge layer and reasoning knowledge layer, construct interconnected semantic vector index, full-text index and ontology semantic index, and establish a synchronous update mechanism for the knowledge base, ontology model and data catalog, and output a hierarchical enterprise-level business knowledge base.

7. The method for automatically constructing ontology and business knowledge base based on data catalog using AI according to claim 6, characterized in that, It also includes step S5 of incremental iterative update, which includes: S51: Real-time monitoring of preset metadata change events in the data catalog through an event-driven mechanism, as well as real-time acquisition of incremental data changes from business data sources, and filtering and classifying the monitored change events, filtering invalid changes and classifying update levels according to the scope of impact. S52: Automatically trigger the corresponding level of ontology incremental update based on the monitored change event. The incremental update includes adding or adjusting ontology concepts, attributes, relationships, axioms and cross-domain alignment mapping relationships, as well as synchronously updating the bidirectional mapping relationship between the ontology and the data directory. After the incremental update is completed, a consistency check is performed based on the bidirectional verification processing method in step S3. If the check fails, an automatic rollback is triggered. S53: Based on the incremental updates of the ontology and incremental business data, automatically synchronize and update the instance data, knowledge entities and corresponding indexes in the knowledge base, and extract knowledge from newly added unstructured documents for semantic alignment and fusion. S54: Perform canary releases of incremental updates, set verification rules and rollback mechanisms, and manage versions, supporting historical version backtracking and comparison.

8. The method for automatically constructing ontology and business knowledge base based on data catalog using AI according to claim 7, characterized in that, The preset metadata change events are events that affect the semantic consistency between the ontology model and the metadata, including at least one of the following: adding a data table, deleting a data table, modifying field definitions, updating business terms, adjusting the business domain, and changing the data source; incremental data changes of the business data source are obtained in real time through log capture or data change notification mechanisms.

9. The method for automatically constructing ontology and business knowledge base based on data catalog using AI according to claim 7, characterized in that, It also includes step S6, which is a closed-loop optimization of the entire process. Step S6 includes: S61: Construct a multi-dimensional quality assessment system, triggering automated assessments according to preset cycles or incremental update events to generate quality assessment results. The assessment dimensions include metadata completeness, standardization, and accuracy; business adaptability, logical consistency, and cross-domain alignment accuracy of the ontology model; and knowledge accuracy, business coverage, and semantic consistency of the business knowledge base. S62: Feed the quality assessment results back to the data catalog system to drive the completion, optimization, and standardization of corresponding metadata, and simultaneously feed back ontology defects and knowledge conflict issues to the incremental update process of the ontology and knowledge base, achieving end-to-end reverse optimization. S63: Based on the quality assessment results, explicit user error correction feedback, and implicit business behavior feedback, automatically optimize the AI ​​agent prompts, HST-Encoder semantic encoding model parameters, and semantic alignment algorithm weights that are relied upon for ontology construction, knowledge extraction, semantic alignment, and reasoning.

10. A system for automatically constructing ontology and business knowledge base based on a data catalog using AI, applicable to the method of automatically constructing ontology and business knowledge base based on a data catalog according to any one of claims 1 to 5, characterized in that, include: The metadata semanticization module is used to collect metadata from various domains in the data catalog system, obtain associated unstructured business documents, group the corresponding domains to form several domain groups, perform semantic enhancement processing on each domain group based on the unstructured business documents, and generate semantic processing results for each domain group. The semantic processing results include at least a structured metadata dataset, a metadata semantic vector library, and a mapping table between metadata and associated unstructured business documents. The ontology construction module communicates with the metadata semanticization module and is used to call the domain big model for ontology recognition processing based on the semanticization processing results, and to construct the ontology model and its supporting mapping relationship and axiom rule base based on the processing results. The ontology recognition processing includes extracting ontology concepts, identifying the attributes and relationships between ontology concepts, generating ontology axioms and standardized business rules, and performing cross-domain semantic alignment and fusion of ontology concepts from different domain groups. The bidirectional mapping and verification module communicates with the ontology construction module and the data catalog system. It is used to establish a bidirectional traceable mapping relationship between the ontology model and the metadata of the data catalog, perform bidirectional verification processing, and form an ontology instance library and generate feedback information for optimizing metadata quality based on the verification results. The bidirectional verification processing includes consistency verification and incremental updates of the ontology model and mapping relationship based on newly added or changed metadata in the data catalog, and compliance verification of ontology instance data extracted from business data sources based on the ontology axioms and rule base.

Images