A data retrieval method and device based on meta-model semantic analysis

Abstract

The application discloses a data retrieval method and device based on meta-model semantic analysis. The data in the meta-model is preprocessed. The preprocessed data is subjected to semantic analysis. The meta-model is optimized according to the semantic analysis result, and an optimized meta-model is obtained. The data to be retrieved is identified, and key information is obtained. The key information is converted by using the metadata in the optimized meta-model, and a query instruction is obtained. The data corresponding to the query instruction is filtered from a database, and is determined as a retrieval result of the data to be retrieved. The semantic information of the deep layer of the data is mined through preprocessing and semantic analysis, so that the internal correlation between the data is understood. On this basis, the meta-model is optimized, the expression ability of the meta-model to the data structure and the relationship is enhanced, and a more accurate query instruction is generated. Finally, the key information is converted into a retrieval condition by means of the optimized meta-model, accurate data query is realized, and the result accuracy is significantly improved.

Description

Technical Field

[0001] This application relates to the field of information retrieval technology, and in particular to a data retrieval method and apparatus based on meta-model semantic analysis. Background Technology

[0002] With the rapid development of big data, how to efficiently and accurately retrieve and share massive amounts of data has become a key challenge in supporting decision-making and promoting digital transformation.

[0003] Existing data retrieval methods primarily rely on simple ranking algorithms. These methods compare keywords with entries in a database to return results relevant to the query. While this approach can provide basic retrieval functionality in some cases, simple ranking algorithms can only sort based on static matching rules, ignoring potential correlations and contextual information between data, resulting in often low accuracy of the retrieved results. Summary of the Invention

[0004] This application provides a data retrieval method and apparatus based on meta-model semantic analysis, with the aim of improving the accuracy of the retrieved results.

[0005] To achieve the above objectives, this application provides the following technical solution:

[0006] The first aspect of this application provides a data retrieval method based on meta-model semantic analysis, including:

[0007] Obtain the data to be retrieved and the meta-model; the meta-model includes the data and basic information about the data.

[0008] The data in the meta-model is preprocessed to obtain preprocessed data;

[0009] Semantic analysis is performed on the preprocessed data to obtain semantic analysis results;

[0010] The meta-model is optimized based on the semantic analysis results to obtain the optimized meta-model;

[0011] The data to be retrieved is identified to obtain key information; the key information includes at least intent and scene parameter information.

[0012] The key information is transformed using the metadata in the optimized meta-model to obtain a query instruction;

[0013] The database is used to filter out data that corresponds to the query command, and this data is then identified as the retrieval result for the data to be retrieved.

[0014] Optionally, the process of obtaining the metamodel includes:

[0015] Instance layer is built based on user data;

[0016] The model layer is constructed based on the first metadata; the first metadata is data describing the instance layer.

[0017] A meta-model layer is constructed based on the second metadata; the second metadata is data describing the structure and semantics of the first metadata.

[0018] A meta-meta-model layer is constructed based on third metadata; the third metadata is data describing the structure and semantics of the second metadata.

[0019] A metamodel is constructed based on the instance layer, the model layer, the metamodel layer, and the metametamodel layer.

[0020] Optionally, performing semantic analysis on the preprocessed data to obtain semantic analysis results includes:

[0021] The preprocessed data is semantically annotated to obtain annotated data;

[0022] The annotated data is semantically aligned to obtain aligned data;

[0023] A semantic network is constructed based on the aligned data;

[0024] Semantic conflict detection is performed on the semantic network to obtain semantic analysis results.

[0025] Optionally, after performing semantic analysis on the preprocessed data to obtain the semantic analysis results, the method further includes:

[0026] The semantic analysis results are compared with the standard dataset;

[0027] When the semantic analysis result is consistent with the standard dataset, the rule engine is used to perform conflict detection on the semantic analysis result;

[0028] If there are no conflicting semantics in the semantic analysis results, then check whether the semantic analysis results meet the business requirements;

[0029] If the semantic analysis result meets the business requirements, then the semantic analysis result is determined to have passed verification;

[0030] If the semantic analysis result does not meet the business requirements, then the semantic analysis result is determined to have failed verification.

[0031] Optionally, optimizing the meta-model based on the semantic analysis results to obtain an optimized meta-model includes:

[0032] Collect information on usage issues and new data issues related to the meta-model.

[0033] Key issues are extracted from the semantic analysis results;

[0034] The key issues, usage issues, and new data issues are sorted to determine the priority of each issue;

[0035] The meta-model is optimized sequentially according to the problem, from highest to lowest priority, to obtain the optimized meta-model.

[0036] Optionally, the step of identifying the data to be retrieved to obtain key information includes:

[0037] The data to be retrieved is classified to obtain the business domain to which the data belongs;

[0038] Identify the relevance between the data to be retrieved and historical dialogues to obtain a complete question context containing contextual information;

[0039] Based on the aforementioned business domain and the complete problem context, semantic understanding instructions are constructed;

[0040] Filter similar question examples corresponding to the business domain from the knowledge base;

[0041] The semantic understanding instruction and the similar question examples are input into the large model to obtain the intent;

[0042] Scene recognition is performed on the data to be retrieved to obtain scene parameter information;

[0043] The intent and the scenario parameter information are identified as key information.

[0044] Optionally, the step of performing scene recognition on the data to be retrieved to obtain scene parameter information includes:

[0045] The data to be retrieved is subjected to scene identification to obtain the business scene to which the data to be retrieved belongs;

[0046] The business scenarios are divided into one or more query task scenarios;

[0047] Based on all query task scenarios, extract the parameter information contained in the data to be retrieved, and use the parameter information to update the word slots;

[0048] When the slot is missing the parameter information required for the query task scenario, a completion prompt is generated based on the completion information template.

[0049] When new parameter information is received based on the completion prompt information, the parameter information and the new parameter information are determined as scene parameter information.

[0050] A second aspect of this application provides a data retrieval device based on meta-model semantic analysis, comprising:

[0051] An acquisition unit is used to acquire the data to be retrieved and the meta-model; the meta-model includes the data and its basic information.

[0052] The preprocessing unit is used to preprocess the data in the meta-model to obtain preprocessed data;

[0053] The analysis unit is used to perform semantic analysis on the preprocessed data to obtain semantic analysis results.

[0054] An optimization unit is used to optimize the meta-model based on the semantic analysis results to obtain an optimized meta-model.

[0055] The identification unit is used to identify the data to be retrieved and obtain key information; the key information includes at least intent and scene parameter information.

[0056] The conversion unit is used to convert the key information using the metadata in the optimized meta-model to obtain a query instruction;

[0057] The retrieval unit is used to filter out data corresponding to the query instruction from the database and determine it as the retrieval result of the data to be retrieved.

[0058] Optionally, the acquisition unit includes:

[0059] The instance layer construction subunit is used to build the instance layer based on user data;

[0060] A model layer construction subunit is used to construct a model layer based on first metadata; the first metadata is data describing the instance layer.

[0061] The metamodel layer construction subunit is used to construct the metamodel layer based on the second metadata; the second metadata is data describing the structure and semantics of the first metadata.

[0062] The meta-meta-model layer construction subunit is used to construct the meta-meta-model layer based on the third metadata; the third metadata is data that describes the structure and semantics of the second metadata.

[0063] The metamodel construction subunit is used to construct a metamodel based on the instance layer, the model layer, the metamodel layer, and the metametamodel layer.

[0064] Optionally, the analysis unit is specifically used for:

[0065] The preprocessed data is semantically annotated to obtain annotated data;

[0066] The annotated data is semantically aligned to obtain aligned data;

[0067] A semantic network is constructed based on the aligned data;

[0068] Semantic conflict detection is performed on the semantic network to obtain semantic analysis results.

[0069] The technical solution provided in this application preprocesses the data in the metamodel; performs semantic analysis on the preprocessed data; optimizes the metamodel based on the semantic analysis results to obtain an optimized metamodel; identifies the data to be retrieved to obtain key information; transforms the key information using metadata in the optimized metamodel to obtain query instructions; and filters data corresponding to the query instructions from the database and determines them as the retrieval results for the data to be retrieved. By preprocessing and semantic analysis, the deep semantic information of the data is mined, thereby understanding the inherent relationships between data. Based on this, the metamodel is optimized to enhance its ability to express data structures and relationships, thus generating more accurate query instructions. Finally, the optimized metamodel is used to convert key information into retrieval conditions, achieving accurate data retrieval and significantly improving the accuracy of the results. Attached Figure Description

[0070] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0071] Figure 1 A flowchart illustrating a data retrieval method based on meta-model semantic analysis provided in this application embodiment;

[0072] Figure 2 An example diagram of metamodel storage provided in an embodiment of this application;

[0073] Figure 3 A schematic diagram of a four-layer structure of a meta-model provided in an embodiment of this application;

[0074] Figure 4 A schematic diagram illustrating an instance layer construction process provided in an embodiment of this application;

[0075] Figure 5 This application provides a schematic diagram of a metamodel layer establishment process.

[0076] Figure 6 A schematic diagram illustrating the process of establishing a meta-model layer is provided in this embodiment of the application.

[0077] Figure 7 A schematic diagram of a semantic analysis process provided in an embodiment of this application;

[0078] Figure 8 A schematic diagram of a data preprocessing flow provided in an embodiment of this application;

[0079] Figure 9 A schematic diagram of a data formatting process provided in an embodiment of this application;

[0080] Figure 10 A schematic diagram illustrating intent recognition provided for an embodiment of this application;

[0081] Figure 11 A schematic diagram of a scene parameter recognition process provided in an embodiment of this application;

[0082] Figure 12 A schematic diagram illustrating the execution of an intent, provided for an embodiment of this application;

[0083] Figure 13 A schematic diagram of a data retrieval process provided in an embodiment of this application;

[0084] Figure 14 A schematic diagram illustrating an operation and maintenance management method provided in an embodiment of this application;

[0085] Figure 15 A schematic diagram of a data retrieval process provided in an embodiment of this application;

[0086] Figure 16 This is a schematic diagram of the architecture of a data retrieval device based on meta-model semantic analysis provided in an embodiment of this application. Detailed Implementation

[0087] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0088] In this application, the terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0089] like Figure 1 The flowchart shown is a data retrieval method based on meta-model semantic analysis provided in an embodiment of this application, including the following steps:

[0090] S101: Obtain the data to be retrieved and the meta-model.

[0091] The metamodel includes data and its basic information. The basic information includes the data's structure, attributes, and relationships.

[0092] Specifically, the data to be retrieved is the question entered by the user. For example, querying for XX type models and sorting them according to professional type.

[0093] It's important to note that a metamodel is a high-level abstraction for modeling data or information; it defines the structure, attributes, and relationships of the data. In data retrieval and sharing, metamodel building techniques are used to create a unified description and representation of data.

[0094] After clarifying the metadata strategy and gaining a deep understanding of business needs, the primary design step is to create the data architecture for the metamodel repository. This step requires setting different levels of metamodel access permissions based on actual needs. Typically, high-level conceptual models focus on describing the overall relationships between systems, while low-level metamodels are more detailed, specifically depicting the definition of each attribute, the model's constituent elements, and its processing flow. See also Figure 2 The diagram shown is an example of a metamodel storage, which may be in a hierarchical or list format. It visually illustrates how models at different levels (such as high-level conceptual models and low-level detailed models) are categorized, stored, and managed in the metamodel repository, thereby concretizing the "hierarchical access permission" design mentioned above.

[0095] The metamodel construction method based on domain ontology establishes a metamodel by formally defining the concepts, attributes, and relationships within the domain. As a structured representation of domain knowledge, domain ontology provides a unified framework for understanding terms and concepts; therefore, this type of metamodel can accurately reflect the data structure and semantic relationships within the domain.

[0096] A metamodel is essentially a "model of models," defining the basic elements needed to build a model within a specific domain. A typical metamodel structure can be divided into four layers: instance layer, model layer, metamodel layer, and metametamodel layer. These four layers exhibit a progressive relationship from concrete to abstract, with each layer being an abstract and specification description of the layer below. For a detailed explanation of the four-layer structure of a metamodel, please refer to [link to relevant documentation]. Figure 3 The content shown.

[0097] Optionally, in another embodiment of this application, the specific implementation of obtaining the meta-model includes processes A1 to A5.

[0098] A1: Build an instance layer based on user data.

[0099] The instance layer consists of data that is intended to be described, typically user data. The instance layer is primarily used to describe detailed information within the information domain.

[0100] For details, see Figure 4 The diagram illustrates an instance layer construction process. First, instance elements must be clearly defined in terms of element type, attributes, and constraints. Based on this, relationships between instances are established according to one-to-one, one-to-many, or many-to-many relationship patterns. Semantic Web technologies (such as RDF and OWL) are then used to formally describe these relationships, thus constructing an instance relationship network with semantic expressive capabilities. Furthermore, a reasonable instance structure needs to be designed to enhance the hierarchy and readability of the data. Specifically, instance elements (such as users, orders, and products) can be organized hierarchically according to business logic to form a hierarchical structure, and instance relationships can be presented in a network form to intuitively reflect their business connections. Finally, an evaluation and optimization mechanism should be established to continuously verify and optimize the instance layer through data validation, performance optimization, and user feedback.

[0101] A2: Build the model layer based on the first metadata.

[0102] The first metadata describes the instance layer. That is, it describes the information layer, which consists of user data and is an instance of the model layer.

[0103] Specifically, the model layer is an "abstract language" (i.e., a language without specific syntax or symbols) defined to describe the information layer.

[0104] It is important to note that data modeling, as a fundamental aspect of data governance, is a systematic task that requires long-term advancement and continuous improvement. Data models comprehensively analyze data requirements using structured languages ​​and employ standardized design methods to uniformly classify and hierarchically model all data, ensuring that all data definitions conform to standard requirements and clearly describe their basic, management, business, and technical attributes. Based on this, primary metadata, as "data about data," describes data from different dimensions, forming various metadata types. Metadata management covers the entire process of creation, storage, integration, and control, providing support for applications such as data catalogs and data lineage analysis. Reference data, as the basis for data classification and cataloging, such as network classifications and provincial / municipal codes, is managed uniformly to enhance readability and maintainability, preventing data silos caused by inconsistent system standards, and thus supporting data integration, sharing, and analytical applications. Master data, as core data that needs to be shared and repeatedly referenced across systems, such as network resource data, is managed primarily to ensure consistency of master data values ​​across systems.

[0105] Furthermore, indicator data is data processed and calculated according to business rules. Its management involves uniformly defining aspects such as business meaning, generation system, calculation formula, and value range to achieve comprehensive consistency in name, description, formula, source, and reference, thereby ensuring data standardization and comparability. Specifically, the indicator data definition template can be found in Table 1.

[0106] Table 1

[0107]

[0108] A3: Construct a meta-model layer based on the second metadata.

[0109] The second metadata is data that describes the structure and semantics of the first metadata.

[0110] Specifically, the meta-model layer is an "abstract language" defined to describe the model layer, and it is a further abstract description of the model layer.

[0111] It should be noted that, see Figure 5This diagram illustrates a process for building a metamodel layer. The metamodel construction process begins with clarifying its basic concepts, including identifying key elements such as classes, attributes, relationships, and operations. The characteristics of each element (e.g., data type, value range, and constraints) are defined, and relationships, inheritance, and dependencies between elements are established to form a complete metamodel structure. Next, the hierarchical structure of the metamodel should be designed, dividing it into layers such as meta-metamodel (M3), metamodel (M2), model (M1), and user object (M0) based on modeling requirements. The relationships and mapping rules between each layer are clearly defined to ensure the consistency and scalability of the overall structure. In the description and representation phase, a suitable modeling language (such as UML or MOF) is selected to define the metamodel's metadata, covering information such as element names, types, and descriptions. A graphical representation should be constructed as much as possible to improve intuitive understanding and communication efficiency. Finally, the metamodel is validated through consistency checks and usability testing to ensure it meets business needs and has practical effectiveness. Continuous iteration and optimization based on feedback ensure that it can adapt to business development and technological evolution, maintaining its advanced nature and practicality.

[0112] A4: Construct a meta-meta-model layer based on third-party metadata.

[0113] The third metadata is data that describes the structure, semantics, and attributes of the second metadata.

[0114] Specifically, the meta-meta-model layer is an "abstract language" defined to describe meta-models. The definition of a meta-meta-model is more abstract and concise than that of a meta-model. A meta-meta-model can define multiple meta-models, and each meta-model can be associated with multiple meta-meta-models. For specific process setup, refer to the meta-model layer, i.e. Figure 6 The diagram shows the process of establishing the meta-model layer.

[0115] It's important to note that the four-layer metamodel architecture, developed by the Object Management Group (OMG), is the core foundation of the Unified Modeling Language (UML) architecture. This architecture not only precisely defines the semantics of complex models but also recursively applies semantics to different layers to construct the overall semantic structure. This provides architectural support for the expansion of the UML metamodel and promotes its integration with other standard implementations based on the same architecture. On the other hand, data schema-based metamodel construction methods start from the structure of the data itself, such as database schemas and XML schemas. By parsing and abstracting these schemas, they extract the data structure and attribute information to construct the corresponding metamodel.

[0116] A5: Construct a metamodel based on the instance layer, model layer, metamodel layer, and metametamodel layer.

[0117] S102: Perform data preprocessing on the data in the meta-model to obtain preprocessed data.

[0118] Specifically, the data in the meta-model is cleaned, standardized, and formatted to obtain preprocessed data, thereby ensuring the accuracy and consistency of the data.

[0119] S103: Perform semantic analysis on the preprocessed data to obtain the semantic analysis results.

[0120] The semantic analysis results indicate the semantic meaning of the data.

[0121] Specifically, semantic analysis techniques are used to perform semantic analysis on the preprocessed data. Semantic analysis techniques mainly include natural language processing-based semantic analysis, knowledge graph-based semantic analysis, and machine learning-based semantic analysis.

[0122] Natural language processing (NLP) techniques are used to analyze and understand the deeper meaning of natural language, thereby extracting semantic information from the data. The process typically includes a series of steps such as word segmentation, part-of-speech tagging, syntactic analysis, and semantic role labeling, aiming to interpret the text layer by layer and ultimately achieve a precise understanding of the data's semantics.

[0123] Knowledge graphs are a structured method of knowledge representation that organizes entities, attributes, and their relationships in the form of a graph, thereby enabling efficient management and utilization of knowledge. Semantic analysis based on knowledge graphs works by mining and utilizing the defined entity and relationship networks within the graph to perform semantic association and reasoning, thus revealing the semantic information behind the data.

[0124] Semantic analysis models are trained using machine learning algorithms to automatically parse the semantics of data. The core of this approach is to enable the model to learn and extract key feature patterns and complex semantic relationships from massive amounts of data, thereby allowing the model to apply these learned rules to the semantic understanding and analysis of new data.

[0125] Optionally, in another embodiment of this application, the specific implementation of step S103 includes processes B1 to B4.

[0126] B1: Perform semantic annotation on the preprocessed data to obtain annotated data.

[0127] It is understandable that semantic labels or annotations are applied to the preprocessed data to clarify its meaning and purpose.

[0128] Optionally, semantic annotation can be performed on the preprocessed data using manual annotation or automated tools.

[0129] B2: Perform semantic alignment on the labeled data to obtain aligned data.

[0130] Optionally, standard mapping and text parsing methods can be used to semantically align the labeled data, thereby anchoring the data definition to obtain aligned data.

[0131] B3: Construct a semantic network based on the aligned data.

[0132] It is understandable that by deriving the influence relationships between aligned data, a complete semantic network is constructed using rule matching and the model for relationships not explicitly defined in the meta-model.

[0133] B4: Perform semantic conflict detection on the semantic network to obtain semantic analysis results.

[0134] Optionally, semantic conflict detection can be performed on the semantic network using rule-based verification and cross-layer comparison methods to obtain semantic analysis results.

[0135] For details on the implementation flow of processes B1 to B4, please refer to [link / reference]. Figure 7 The content shown. In Figure 7 First, analysis is performed based on ontology, rule-based methods, and machine learning. Then, the data is preprocessed, semantically labeled, and semantically reasoned to obtain semantic analysis results. The semantic analysis results are then validated, and if the validation is successful, the meta-model is optimized and improved using the semantic analysis results.

[0136] Optionally, after step S103, the semantic analysis results need to be verified to ensure their accuracy and reliability. Therefore, another embodiment of this application provides a method for verifying semantic analysis results, including processes C1 to C5.

[0137] C1: Compare the semantic analysis results with the standard dataset.

[0138] This involves extracting standard semantic samples from a standard dataset, using a comparison tool to compare the semantic analysis results with the standard dataset, and calculating the matching degree between the two.

[0139] Specifically, the matching degree is calculated as follows: Matching degree = (Number of samples whose semantic analysis results match the standard semantic samples / Total number of validation samples) × 100%.

[0140] It should be noted that erroneous samples are manually reviewed to confirm whether the error is due to an analysis error or an omission in the standard dataset, and the review results are recorded.

[0141] In some examples, if the semantic analysis results are inconsistent with the standard dataset, it is determined that the semantic analysis results have not passed validation.

[0142] C2: When the semantic analysis results are consistent with the standard dataset, the rule engine is used to perform conflict detection on the semantic analysis results.

[0143] Specifically, a semantic rule base is first built, and the rule engine is called to input the semantic analysis results into the semantic rule base to perform conflict detection. If a semantic conflict is detected, it is sorted according to the scope of impact (for example, cross-data source conflicts have higher priority than single data source conflicts). Semantic conflicts are repaired in order of priority from high to low. After repair, the rule engine is used again to perform conflict detection on the semantic analysis results to ensure that there are no semantic conflicts.

[0144] C3: If there are no conflicting semantics in the semantic analysis results, then check whether the semantic analysis results meet the business requirements.

[0145] If the semantic analysis results meet the business requirements, then process C4 is executed; if the semantic analysis results do not meet the business requirements, then process C5 is executed.

[0146] Among these, business requirements are the conditions that must be met to achieve the goal. For example...

[0147] Specifically, the process for verifying whether the semantic analysis results meet business requirements is as follows: establish application scenarios; each application scenario includes input requirements, expected results, and test cases; input the semantic analysis results into the application scenarios and execute the test cases; count the number of searches that meet the expected results and calculate the search accuracy; when the number of errors due to semantic inconsistency is 0, it indicates that the semantic analysis results meet business requirements.

[0148] C4: The semantic analysis results have been verified.

[0149] C5: The semantic analysis results have failed validation.

[0150] In addition, during the verification process, it is necessary to collect intuitive feedback from users regarding the "convenience of semantic analysis results in supporting business operations," such as specific issues raised by users, such as "whether key data is missed in the search results" or "whether additional manual intervention is required to adjust the semantics when sharing across systems." Subsequently, for the issues identified in these feedbacks, the key aspects of semantic analysis are traced back and analyzed, with a focus on optimizing the terminology mapping and relationship definition rules. After optimization, the system is redeployed and scenario testing is re-executed to form a closed-loop iterative mechanism that continuously improves the semantic analysis model's ability to support actual business needs.

[0151] S104: Optimize the meta-model based on the semantic analysis results to obtain the optimized meta-model.

[0152] Understandably, when semantic analysis results reveal problems in testing or real-world applications, such as identifying semantic inconsistencies in the meaning of terms in different contexts, or discovering that the metamodel cannot describe certain business concepts or relationships (missing attributes or relationships), these issues can be used as a basis for precise correction of the metamodel.

[0153] It's important to note that semantic analysis results can be directly used to support data governance and business decisions. For example, based on the clear definitions of data concepts in the metamodel, unified data standards, data quality rules, and even global governance strategies can be formulated. Furthermore, by clarifying and unifying the semantic information of metamodels across different systems, data interoperability can be fundamentally enhanced, enabling seamless integration and efficient sharing of cross-system data. However, business environments and data ecosystems are constantly evolving, and the semantic information of the metamodel must also be dynamically updated accordingly. Therefore, a continuous monitoring and maintenance mechanism must be established to regularly review and optimize the semantic definition of the metamodel to ensure that it always accurately reflects business realities and maintains its timeliness and authority throughout the entire data lifecycle.

[0154] Furthermore, in the data integration process, the first step is data collection. Original documents scattered across Word documents, Elasticsearch systems, enterprise cloud storage, and personal computers are aggregated and centrally stored, with corresponding storage standards and management requirements established, significantly improving the breadth and richness of data collection. The next step is data cleaning, where a standardized processing workflow is built and an intelligent processing mechanism is introduced, focusing on improving the accuracy and completeness of the corpus. Simultaneously, AI technology is used to automatically extract and label the raw data, effectively reducing repetitive and low-quality manual operations. For a detailed diagram of the data preprocessing workflow, please refer to [link to relevant documentation]. Figure 8 The content shown.

[0155] Finally, data formatting is performed. Using the DDL statements of the database tables as input, and combining a large model and metadata knowledge base approach, missing core metadata items are automatically generated and existing inaccurate definitions are corrected. It is estimated that this can reduce metadata maintenance workload by more than 50%, comprehensively achieving the standardization and high-quality governance of data resources. For details on the data formatting process, please refer to [link to relevant documentation]. Figure 9 The content shown.

[0156] Optionally, in another embodiment of this application, the specific implementation of step S104 includes processes D1 to D4.

[0157] D1: Issues related to the use of the meta-model and the addition of new data.

[0158] For example, new data issues include whether the metamodel covers new data sources and whether the metamodel has the associated definitions for inventory data and alert rules.

[0159] D2: Extract key issues from the semantic analysis results.

[0160] Among them, the key issues are semantic conflict type and semantic missing type.

[0161] In addition, it is necessary to record the key issues, the types of key issues, the meta-model level at which the key issues are located (such as instance layer, model layer), and the scope of impact (such as single data source / entire system).

[0162] D3: Sort the key issues, usage issues, and new data issues to determine the priority of each issue.

[0163] It is understandable that key issues, usage issues, and new data issues should be prioritized, with high-impact, low-cost issues addressed first to avoid wasting resources. For example, the issue priorities can be seen in Table 2.

[0164] Table 2

[0165]

[0166] In addition, domain experts can score key issues, usage issues, and new data issues to determine the priority of each issue.

[0167] D4: Optimize the meta-model in descending order of priority according to the problem to obtain the optimized meta-model.

[0168] Specifically, based on different types of problems, different optimization schemes are constructed in conjunction with the hierarchical characteristics of the meta-model. The meta-model is optimized according to the optimization scheme corresponding to the problem in order of priority from high to low. At the same time, the relevant components of the meta-model (such as repositories and tools) are updated to avoid execution deviations.

[0169] For example, optimization scheme 1 is semantic accuracy optimization (resolving conflicts / missing information), optimization scheme 2 is structural adaptability optimization (adjusting hierarchy / relationships), (1) hierarchical structure adjustment: adding subdivided domain data and adjusting the redundancy of hierarchical elements, (2) element relationship expansion: if business needs add new associations, add semantic relationships in the meta-model layer and clarify the relationship logic, optimization scheme 3 is application effectiveness optimization: retrieval scenario optimization and sharing scenario optimization.

[0170] It should be noted that during the execution of the metamodel optimization scheme, a layered implementation mechanism should be adopted according to the priority of the issues. For P0 level urgent issues (such as semantic conflicts), the data architect takes the lead in directly correcting the attributes or rules of the conflicting elements in the metamodel repository (such as changing the data type of "Customer ID" in System B from integer to string), and simultaneously updating the validation rules of the semantic annotation tool. At the same time, the relevant system teams (such as the data sharing interface development team) are notified to ensure collaborative adaptation. For P1 / P2 level important or routine issues (such as structural adjustments and relationship expansion), the modeling team adjusts the element and hierarchical relationships through the metamodel modeling tool and submits them for review by domain experts. The development team then expands the association rules in the semantic reasoning engine accordingly and conducts unit tests to ensure that the relationship logic is accurately identified and executed.

[0171] Furthermore, after the metamodel optimization is completed, its design document needs to be revised simultaneously, recording the optimization content, executors, and timelines in detail to ensure all changes are traceable. The next step is verification and closed-loop management, employing a three-pronged mechanism of data verification, business testing, and long-term monitoring to ensure optimization effectiveness and prevent new problems: First, the semantic analysis verification process is reused to compare key indicators before and after optimization to quantify the improvement; second, the actual application effect of the metamodel is tested in simulated business scenarios to confirm its supporting capabilities; finally, a metamodel monitoring dashboard is built to continuously track core indicators such as "number of new semantic conflicts" and "business scenario adaptation rate." Once an anomaly is detected, a new round of optimization is triggered, thus forming a continuously iterating and self-improving closed-loop management system.

[0172] In some examples, during the data annotation phase, the platform supports both template import and interface input to quickly complete detailed data annotation of the metamodel, ensuring effective maintenance of its information. Secondly, in terms of metadata modeling, the platform supports management of target database tables via DSL import or manual input, and has the ability to manage API input and output parameters. It is currently compatible with the modeling needs of various relational and columnar databases. Furthermore, in business model construction, the platform supports defining the relationships between indicators, data tables, and APIs, and can configure preprocessing of query results, such as rounding and percentage conversion. Finally, it also provides a recommendation metamodel setting function, which can complete the initial recommendation settings and support flexible configuration of the underlying recommendation algorithm, thereby improving modeling efficiency and intelligence.

[0173] S105: Identify the data to be retrieved to obtain key information.

[0174] The key information includes at least intent and scenario parameter information.

[0175] Optionally, the specific implementation process of step S105 is as follows: classify the data to be retrieved to obtain the business domain to which the data to be retrieved belongs; identify the relevance between the data to be retrieved and historical dialogues to obtain a complete question context containing contextual information; construct semantic understanding instructions based on the business domain and the complete question context; select similar question examples corresponding to the business domain from the knowledge base; input the semantic understanding instructions and similar question examples into the large model to obtain the intent; perform scene recognition on the data to be retrieved to obtain scene parameter information; and determine the intent and scene parameter information as key information.

[0176] It should be noted that in the intent recognition task, cue engineering and knowledge retrieval techniques are comprehensively used to improve the classification performance of the large model. Specifically, cue engineering enhances the model's ability to understand historical context in multi-turn dialogues, achieving more accurate intent discrimination; at the same time, a knowledge retrieval mechanism based on vector matching is introduced to build a dynamically updated few-shot example library, enabling the model to obtain reference information highly relevant to the current question in real time. This significantly improves the accuracy and overall adaptability of intent recognition while effectively controlling token consumption.

[0177] To better illustrate the intent recognition task, the following explanation will be provided in conjunction with specific application scenarios. (See [link to relevant documentation]). Figure 10 This is a schematic diagram illustrating intent recognition. Figure 10 In this process, the user inputs a question (i.e., the data to be retrieved) as "Query XXX model". First, the relevance of the data to be retrieved to historical dialogues is identified to obtain a complete question context containing contextual information. The data to be retrieved is then classified to determine whether it belongs to the business domain of "model query" or "metric query". Based on the business domain and the complete question context, a semantic understanding instruction is constructed. Similar question examples corresponding to the business domain are selected from the knowledge base, such as "Does the XXX related model have a certain number of customers?". The semantic understanding instruction and similar question examples are then input into the large model to obtain the intent.

[0178] Optionally, the specific implementation process for scene identification of the data to be retrieved to obtain scene parameter information is as follows: scene identification of the data to be retrieved to obtain the business scene to which the data to be retrieved belongs; the business scene is divided to obtain one or more query task scenes; based on all query task scenes, the parameter information contained in the data to be retrieved is extracted, and the slots are updated using the parameter information; when the slots are missing the parameter information required by the query task scene, completion prompt information is generated based on the completion information template; when new parameter information is received based on the completion prompt information, the parameter information and the new parameter information are determined as scene parameter information.

[0179] Understandably, this involves identifying the specific interactive context of the data to be retrieved and then finely segmenting it into scenarios. Based on this, parameter information is extracted and the corresponding slots are dynamically updated. If incomplete information is detected, a large language model guides the user to interactively complete the information, ensuring that the parameters required for the data query are complete and the structure is intact. Furthermore, the system utilizes the agent's tool selection capabilities to automatically match the most suitable processing tool (i.e., scenario) from a pre-set toolset (corresponding to the segmented scenario set) to accurately and efficiently resolve the specific sub-questions contained in the current question-and-answer session.

[0180] In some examples, within network domain data sharing scenarios, the query task scenarios are first identified and subdivided, such as "a municipal branch office querying 5G base station operation data for a certain area" or "a provincial company sharing monthly network congestion rate data for the entire province." Parameter information, including data granularity, anonymization requirements, and authorization scope, is extracted from these scenarios. A hierarchical semantic structure of "main slot + sub-slot" is then constructed based on the network domain meta-model to achieve accurate mapping and semantic alignment between parameter information and slots. To adapt to business changes, a dynamic slot update mechanism is established, forming a closed-loop management system of "trigger-verification-update." Simultaneously, network domain-specific test cases are used to continuously verify the accuracy of parameter extraction and slot adaptability, ensuring that the data sharing process is executed efficiently and compliantly under multiple constraints, including security, authorization, and formatting.

[0181] For example, Figure 11 The diagram illustrates a scenario parameter identification process. The data to be retrieved is "the XX type of model to be queried and sorted according to the order of professional type". The process identifies the business scenario to which the data to be retrieved belongs and divides the business scenario into time and location indicator Q&A scenario, network element indicator Q&A scenario, and location information indicator Q&A scenario. After extracting the parameter information in the query task scenario and updating the slots, the user is guided to complete the information. The completion prompt is "Please complete the time and location information. For example, you can try asking me to query all XX type models in 24 years".

[0182] E7: Identify intent and scene parameter information as key information.

[0183] S106: Use the metadata in the optimized meta-model to transform key information and obtain query instructions.

[0184] Among them, query commands include, but are not limited to, SQL query commands.

[0185] Understandably, the key information is first converted into a standard format, and then metadata is used to annotate the key information in order to realize the functions of natural language to SQL query commands and natural language to API (NL2API).

[0186] In some examples, to reduce the uncertainty of NL2SQL output, the online knowledge data question answering system introduces metadata to uniformly describe the relationships between metrics, data tables, fields, and APIs. By constructing an intermediate layer, a large model is used as an "auxiliary" component for semantic understanding and transformation: the large model leverages its advantages in natural language understanding to extract key information from user questions; then, based on the mapping rules defined in the metadata, this information is accurately converted into standard SQL query instructions, ultimately executing database access and data retrieval. This design ensures both the flexibility of semantic parsing and the structured accuracy of query generation through metadata constraints.

[0187] For details, see Figure 12 This is a schematic diagram illustrating an intention to execute. Figure 12 In the process, the system receives the data to be retrieved (i.e., the user-related model built by the sub-system in 2025), extracts slot data from the data to be retrieved, including scene data extraction, time data extraction and indicator information extraction, obtains key information, and uses the metadata in the optimized meta-model to transform the key information to obtain the query instruction.

[0188] S107: Filter out the data corresponding to the query command from the database and determine it as the retrieval result of the data to be retrieved.

[0189] It's worth noting that the automatic generation of chart data is achieved by leveraging the NL2Code capabilities of the large model. Based on the LangGraph framework, a multi-agent collaborative model is employed for data analysis. A built-in routing mechanism controls interaction conditions and processes, allowing each agent to autonomously select its next interaction partner within a preset range, ensuring a flexible and controllable collaborative process. In typical application scenarios, it can simulate multiple roles such as data experts, network optimization engineers, and operations managers collaboratively analyzing and discussing query results, ultimately generating high-quality analytical conclusions. This model not only significantly improves the accuracy of the analysis process and the reliability of the results but also allows users to customize discussion topics according to their needs, further enhancing the practical value of data analysis.

[0190] For details, see Figure 13 This paper illustrates a data retrieval process that filters data from a database corresponding to a query command, obtains the query results (i.e., the retrieval results for the data to be retrieved), and can generate list and chart data based on the query results before presenting them to the user. The retrieval results for the data to be retrieved are analyzed using multi-agent analytics.

[0191] In some examples, based on five pre-defined core scenarios and 23 sub-scenarios, combined with 10 question generation rules, 101 types of test sampling data, and 424 question permutations, a rich and comprehensive test dataset can be automatically constructed. In the automatic test case execution phase, both manual triggering and automatic scheduling modes are supported. Especially after adding new business metrics or system version updates, tests can be quickly executed to verify the accuracy and stability of the data Q&A function. In the automatic test case analysis phase, test results can be visualized through charts and reports, and specific reason analyses can be provided for failed tests, thereby accurately locating the root cause of the problem and effectively supporting subsequent optimization.

[0192] It's important to emphasize that during use, user question and answer logs and feedback are automatically collected. Operations personnel regularly review user activity and mark unsatisfactory responses. This marked data is then imported into the automated test suite to facilitate subsequent optimization and verification, thus achieving a closed-loop operational process. For details, please refer to... Figure 14 This diagram illustrates an operation and maintenance management approach.

[0193] For a better explanation of the contents shown in steps S101 to S107, see [link to documentation]. Figure 15 The diagram illustrates a data retrieval process. First, a meta-model semantic analysis framework is constructed, and then the meta-model is used to perform data retrieval. The data retrieval process includes data preprocessing, data preparation, and the construction of the semantic analysis framework.

[0194] It should be noted that, through the content shown in steps S101 to S107, the embodiments of this application have the following beneficial effects:

[0195] 1. Employing advanced data extraction, transformation, and loading (ETL) processes, it can automatically extract data from various heterogeneous data sources (including relational databases, non-relational databases, file systems, and web services), and ensure data consistency and accuracy through intelligent cleaning and transformation. Building upon this, it innovatively introduces a meta-model as a semantic mediator, achieving unified mapping and integration of different data sources at the semantic level, providing reliable support for subsequent data analysis and retrieval.

[0196] 2. To effectively address the interoperability challenges arising from semantic heterogeneity between systems, an ontology-based approach was introduced to construct an ontology and meta-model system encompassing domain concepts and their relationships. This system not only clearly defines the structure and attributes of data but also profoundly reveals the inherent semantic connections and business logic between data. Through ontology mapping and semantic matching algorithms, the system achieves semantic-level interoperability of data and functions across systems, thereby significantly improving the accuracy and intelligence of data retrieval.

[0197] 3. To meet a wide range of system integration needs, standardized system interfaces were designed and a middleware architecture was built. The interfaces follow common communication protocols such as RESTful API and SOAP, which can easily connect to various business systems and data sources; the middleware layer is responsible for data routing, format conversion and caching, effectively reducing the coupling between systems and comprehensively enhancing the scalability and maintainability of the overall architecture.

[0198] like Figure 16 The diagram shown is an architectural schematic of a data retrieval device based on meta-model semantic analysis provided in an embodiment of this application. The data retrieval device includes: an acquisition unit 100, a preprocessing unit 200, an analysis unit 300, an optimization unit 400, an identification unit 500, a conversion unit 600, and a retrieval unit 700.

[0199] The acquisition unit 100 is used to acquire the data to be retrieved and the meta-model; the meta-model includes the data and the basic information of the data.

[0200] Acquisition unit 100 includes:

[0201] The instance layer construction subunit is used to build the instance layer based on user data.

[0202] The model layer construction subunit is used to build the model layer based on the first metadata; the first metadata is the data describing the instance layer.

[0203] The metamodel layer construction subunit is used to build the metamodel layer based on the second metadata; the second metadata is data that describes the structure and semantics of the first metadata.

[0204] The meta-meta-model layer construction subunit is used to construct the meta-meta-model layer based on the third metadata; the third metadata is data that describes the structure and semantics of the second metadata.

[0205] The metamodel building subunit is used to build a metamodel based on the instance layer, model layer, metamodel layer, and metametamodel layer.

[0206] The preprocessing unit 200 is used to preprocess the data in the meta-model to obtain preprocessed data.

[0207] Analysis unit 300 is used to perform semantic analysis on the preprocessed data and obtain semantic analysis results.

[0208] The analysis unit 300 is specifically used for: semantically labeling the preprocessed data to obtain labeled data; semantically aligning the labeled data to obtain aligned data; constructing a semantic network based on the aligned data; and performing semantic conflict detection on the semantic network to obtain semantic analysis results.

[0209] The optimization unit 400 is used to optimize the meta-model based on the semantic analysis results to obtain the optimized meta-model.

[0210] The optimization unit 400 is specifically used for: collecting usage issues and new data issues of the meta-model; extracting key issues from the semantic analysis results; sorting the key issues, usage issues, and new data issues to obtain the priority of each issue; and optimizing the meta-model in descending order of priority to obtain the optimized meta-model.

[0211] The identification unit 500 is used to identify the data to be retrieved and obtain key information; the key information includes at least intent and scene parameter information.

[0212] The identification unit 500 includes:

[0213] The classification subunit is used to classify the data to be retrieved to obtain the business domain to which the data belongs.

[0214] The first identification subunit is used to identify the relevance between the data to be retrieved and historical dialogues, thereby obtaining a complete question context containing contextual information.

[0215] Construct sub-units to build semantic understanding instructions based on business domains and complete problem contexts.

[0216] The filtering sub-unit is used to filter similar question examples that correspond to the business domain from the knowledge base.

[0217] The input sub-unit is used to input semantic understanding instructions and similar question examples into the large model to obtain the intent.

[0218] The second identification subunit is used to perform scene identification on the data to be retrieved and obtain scene parameter information.

[0219] The second identification subunit is specifically used for: performing scene identification on the data to be retrieved to obtain the business scene to which the data to be retrieved belongs; dividing the business scene to obtain one or more query task scenes; extracting parameter information contained in the data to be retrieved based on all query task scenes, and updating the slots using the parameter information; when the slots lack the parameter information required by the query task scene, generating completion prompt information based on the completion information template; and when receiving new parameter information based on the completion prompt information, determining the parameter information and the new parameter information as scene parameter information.

[0220] The information determination subunit is used to determine the intent and scene parameter information as key information.

[0221] The transformation unit 600 is used to transform key information using metadata in the optimized meta-model to obtain query instructions.

[0222] The retrieval unit 700 is used to filter out data corresponding to the query command from the database and determine the retrieval results as the data to be retrieved.

[0223] In summary, by preprocessing and semantic analysis to mine deep semantic information from data, we can understand the inherent relationships between data points. Based on this, the meta-model is optimized to enhance its ability to express data structures and relationships, thereby generating more precise query instructions. Finally, the optimized meta-model is used to transform key information into search conditions, achieving accurate data queries and significantly improving the accuracy of results.

[0224] Combination Figure 16 The data retrieval device also includes, as shown, the following:

[0225] The comparison unit is used to compare the semantic analysis results with a standard dataset.

[0226] The detection unit is used to perform conflict detection on the semantic analysis results when the semantic analysis results are consistent with the standard dataset, using the rule engine.

[0227] The verification unit is used to verify whether the semantic analysis results meet the business requirements if there are no conflicting semantics in the semantic analysis results.

[0228] The first determining unit is used to determine that the semantic analysis result passes verification if the semantic analysis result meets the business requirements.

[0229] The second determining unit is used to determine that the semantic analysis result has not passed the verification if the semantic analysis result does not meet the business requirements.

[0230] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. In particular, for system or system embodiments, since they are fundamentally similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The systems and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. Components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0231] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0232] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A data retrieval method based on meta-model semantic analysis, characterized in that, include: Obtain the data to be retrieved and the meta-model; The meta-model includes data and basic information about the data; The data in the meta-model is preprocessed to obtain preprocessed data; Semantic analysis is performed on the preprocessed data to obtain semantic analysis results; The semantic analysis results are compared with the standard dataset; When the semantic analysis result is consistent with the standard dataset, the rule engine is used to perform conflict detection on the semantic analysis result; If there are no conflicting semantics in the semantic analysis results, then check whether the semantic analysis results meet the business requirements; If the semantic analysis result meets the business requirements, then the semantic analysis result is determined to have passed verification; If the semantic analysis result does not meet the business requirements, then the semantic analysis result is determined to have failed verification. The meta-model is optimized based on the semantic analysis results to obtain the optimized meta-model; The data to be retrieved is identified to obtain key information; The key information includes at least intent and scene parameter information; The key information is transformed using the metadata in the optimized meta-model to obtain a query instruction; The database is used to filter out data that corresponds to the query command, and this data is determined as the retrieval result for the data to be retrieved. The step of performing semantic analysis on the preprocessed data to obtain semantic analysis results includes: semantically labeling the preprocessed data to obtain labeled data; semantically aligning the labeled data to obtain aligned data; constructing a semantic network based on the aligned data; and performing semantic conflict detection on the semantic network to obtain semantic analysis results. The step of optimizing the meta-model based on the semantic analysis results to obtain an optimized meta-model includes: collecting usage issues and new data issues of the meta-model; extracting key issues from the semantic analysis results; sorting the key issues, usage issues, and new data issues to obtain the priority of each issue; and optimizing the meta-model sequentially according to the issues in descending order of priority to obtain the optimized meta-model.

2. The method according to claim 1, characterized in that, The process of obtaining the metamodel includes: Instance layer is built based on user data; The model layer is constructed based on the first metadata; the first metadata is data describing the instance layer. A meta-model layer is constructed based on the second metadata; the second metadata is data describing the structure and semantics of the first metadata. A meta-meta-model layer is constructed based on third metadata; the third metadata is data describing the structure and semantics of the second metadata. A metamodel is constructed based on the instance layer, the model layer, the metamodel layer, and the metametamodel layer.

3. The method according to claim 1, characterized in that, The process of identifying the data to be retrieved to obtain key information includes: The data to be retrieved is classified to obtain the business domain to which the data to be retrieved belongs; Identify the relevance between the data to be retrieved and historical dialogues to obtain a complete question context containing contextual information; Based on the aforementioned business domain and the complete problem context, semantic understanding instructions are constructed; Filter similar question examples corresponding to the business domain from the knowledge base; The semantic understanding instruction and the similar question examples are input into the large model to obtain the intent; Scene recognition is performed on the data to be retrieved to obtain scene parameter information; The intent and the scenario parameter information are identified as key information.

4. The method according to claim 3, characterized in that, The step of performing scene recognition on the data to be retrieved to obtain scene parameter information includes: The data to be retrieved is subjected to scene identification to obtain the business scene to which the data to be retrieved belongs; The business scenarios are divided into one or more query task scenarios; Based on all query task scenarios, extract the parameter information contained in the data to be retrieved, and use the parameter information to update the word slots; When the slot is missing the parameter information required for the query task scenario, a completion prompt is generated based on the completion information template. When new parameter information is received based on the completion prompt information, the parameter information and the new parameter information are determined as scene parameter information.

5. A data retrieval device based on meta-model semantic analysis, characterized in that, include: The acquisition unit is used to acquire the data to be retrieved and the meta-model. The meta-model includes data and basic information about the data; The preprocessing unit is used to preprocess the data in the meta-model to obtain preprocessed data; The analysis unit is used to perform semantic analysis on the preprocessed data to obtain semantic analysis results. The comparison unit is used to compare the semantic analysis results with a standard dataset; The detection unit is used to perform conflict detection on the semantic analysis result using a rule engine when the semantic analysis result is consistent with the standard dataset. The verification unit is used to verify whether the semantic analysis result meets the business requirements if there are no conflicting semantics in the semantic analysis result. The first determining unit is configured to determine that the semantic analysis result passes verification if the semantic analysis result meets the business requirements; The second determining unit is used to determine that the semantic analysis result has failed verification if the semantic analysis result does not meet the business requirements; An optimization unit is used to optimize the meta-model based on the semantic analysis results to obtain an optimized meta-model. The identification unit is used to identify the data to be retrieved and obtain key information; The key information includes at least intent and scene parameter information; The conversion unit is used to convert the key information using the metadata in the optimized meta-model to obtain a query instruction; The retrieval unit is used to filter out data corresponding to the query instruction from the database and determine it as the retrieval result of the data to be retrieved; Specifically, the analysis unit is used to: semantically annotate the preprocessed data to obtain annotated data; semantically align the annotated data to obtain aligned data; construct a semantic network based on the aligned data; and perform semantic conflict detection on the semantic network to obtain semantic analysis results. Specifically, the optimization unit is used to: collect usage issues and new data issues of the meta-model; extract key issues from the semantic analysis results; sort the key issues, usage issues, and new data issues to obtain the priority of each issue; and optimize the meta-model according to the issues in descending order of priority to obtain the optimized meta-model.

6. The apparatus according to claim 5, characterized in that, The acquisition unit includes: The instance layer construction subunit is used to build the instance layer based on user data; A model layer construction subunit is used to construct a model layer based on first metadata; the first metadata is data describing the instance layer. The metamodel layer construction subunit is used to construct the metamodel layer based on the second metadata; the second metadata is data describing the structure and semantics of the first metadata. The meta-meta-model layer construction subunit is used to construct the meta-meta-model layer based on the third metadata; the third metadata is data that describes the structure and semantics of the second metadata. The metamodel construction subunit is used to construct a metamodel based on the instance layer, the model layer, the metamodel layer, and the metametamodel layer.

Images