Metadata retrieval method and device, storage medium and computer program product

Abstract

The invention discloses a metadata retrieval method and device, a storage medium and a computer program product, and relates to the technical field of data processing.The method comprises the steps that a natural language query instruction input by a user is obtained; calling a pre-constructed semantic understanding model to carry out query vectorization on the natural language query instruction to obtain a metadata query vector; performing semantic retrieval matching on the metadata query vector in a pre-constructed metadata vector library to obtain a preliminary matching result set; and judging the number of results of the preliminary matching result set, and generating a metadata retrieval result based on the number of the results. A metadata query vector is generated by calling a pre-constructed semantic understanding model, semantic retrieval matching is carried out, and precise metadata retrieval is carried out on natural language description input by a user cooperatively.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to metadata retrieval methods, devices, storage media, and computer program products. Background Technology

[0002] Traditional metadata retrieval systems primarily employ structured rules, keyword matching, or predefined query languages. Their core retrieval logic relies on users explicitly inputting metadata fields, with the system returning results through precise matching of field values. However, due to a lack of semantic understanding capabilities, traditional metadata retrieval systems cannot perform semantic association matching based on user-input natural language descriptions. Consequently, when users input natural language queries with non-fixed fields, the system cannot accurately identify the query intent and locate semantically relevant metadata, relying solely on precise keyword or field matching, thus failing to meet semantic-level metadata retrieval needs.

[0003] Therefore, how to perform accurate metadata retrieval based on natural language descriptions input by users has become a technical problem that this application urgently needs to solve.

[0004] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main objective of this application is to provide a metadata retrieval method, device, storage medium, and computer program product, aiming to solve the technical problem of how to perform accurate metadata retrieval based on natural language descriptions input by users.

[0006] To achieve the above objectives, this application proposes a metadata retrieval method, the method comprising: Obtain the user's input in natural language; The natural language query instruction is vectorized by calling a pre-built semantic understanding model to obtain a metadata query vector; Semantic retrieval and matching of the metadata query vectors are performed in a pre-built metadata vector library to obtain a preliminary matching result set; Determine the number of results in the preliminary matching result set, and generate metadata retrieval results based on the number of results.

[0007] In one embodiment, the step of invoking a pre-built semantic understanding model to vectorize the natural language query instruction and obtain a metadata query vector includes: The natural language query instructions are preprocessed to obtain standardized query instructions; The standardized query command is mapped to a fixed-dimensional high-dimensional vector space by calling a pre-built semantic understanding model to obtain an initial vector; The initial vector is converted into a unit vector, and the data type of the unit vector is adjusted to obtain the metadata query vector.

[0008] In one embodiment, the step of performing semantic retrieval and matching on the metadata query vector in a pre-built metadata vector library to obtain a preliminary matching result set further includes: Collect initial metadata from multi-source heterogeneous systems; The initial metadata is cleaned and standardized, and the logical relationships between the initial metadata are verified. If the logical relationship verification of the initial metadata passes, the business terminology table, lineage diagram, operational metadata, and sample data description in the initial metadata are integrated to obtain unified metadata; The unified metadata is converted into metadata vectors and then stored in a vector database to obtain a metadata vector library.

[0009] In one embodiment, the step of performing semantic retrieval and matching on the metadata query vector in a pre-built metadata vector library to obtain a preliminary matching result set includes: Based on the metadata query vector, an approximate nearest neighbor search is performed in the metadata vector library to obtain a candidate vector set; Calculate the cosine similarity between each metadata vector in the candidate vector set and the metadata query vector; The candidate vector set is filtered based on the cosine similarity to obtain a preliminary matching result set.

[0010] In one embodiment, the step of determining the number of results in the preliminary matching result set and generating metadata retrieval results based on the number of results includes: Determine the number of results in the preliminary matching result set; If the number of results is less than a preset lower threshold, the semantic understanding model is invoked to generate extended retrieval suggestions, the candidate vector set is re-selected based on the extended retrieval suggestions, and metadata retrieval results are generated based on the re-selected candidate vector set. If the number of results exceeds a preset upper limit threshold, the semantic understanding model is invoked to generate intelligent filtering conditions, the candidate vector set is re-filtered based on the intelligent filtering conditions, and metadata retrieval results are generated based on the re-filtered candidate vector set. If the number of results is neither less than the preset lower threshold nor greater than the upper threshold, then the semantic relevance ranking is performed on the preliminary matching result set to obtain the metadata retrieval results.

[0011] In one embodiment, the step of re-filtering the candidate vector set based on the expanded search suggestions and generating metadata retrieval results based on the re-filtered candidate vector set includes: Based on the extended search suggestions, an extended search is performed in the metadata vector database to obtain a new vector set; By merging the candidate vector set and the newly added vector set, a re-filtered candidate vector set is obtained; Calculate the cosine similarity between each metadata vector in the re-filtered candidate vector set and the metadata query vector; Return to the execution steps: filter the candidate vector set according to the cosine similarity to obtain a preliminary matching result set.

[0012] In one embodiment, the step of re-filtering the candidate vector set based on the intelligent filtering conditions and generating metadata retrieval results based on the re-filtered candidate vector set includes: The candidate vector set is filtered based on the intelligent filtering conditions to obtain a re-filtered candidate vector set. Calculate the cosine similarity between each metadata vector in the re-filtered candidate vector set and the metadata query vector; Return to the execution steps: filter the candidate vector set according to the cosine similarity to obtain a preliminary matching result set.

[0013] In addition, to achieve the above objectives, this application also proposes a metadata retrieval device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the metadata retrieval method as described above.

[0014] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and which, when executed by a processor, implements the steps of the metadata retrieval method described above.

[0015] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the metadata retrieval method described above.

[0016] One or more technical solutions proposed in this application have at least the following technical effects: First, the system acquires the user's natural language query command: breaking the dependence on fixed field input, it receives the user's unstructured natural language request, providing raw input for subsequent semantic processing. Further, a pre-built semantic understanding model is invoked to vectorize the natural language query command, obtaining a metadata query vector: utilizing the semantic embedding capability of the semantic understanding model, the user's natural language query is transformed into a high-dimensional semantic vector. This vector encodes the user's deep semantic intent, providing a unified representation space for subsequent semantic similarity matching. Second, semantic retrieval matching is performed on the metadata query vector in a pre-built metadata vector library, obtaining a preliminary matching result set: through semantic retrieval matching, semantically relevant metadata is accurately located. Finally, the number of results in the preliminary matching result set is determined, and metadata retrieval results are generated based on this number: the metadata retrieval results are adjusted according to the number of results to ensure that the final returned results both conform to the user's semantic intent and have an appropriate scale, further improving the effectiveness of the retrieval. In summary, this application, by invoking a pre-built semantic understanding model to generate metadata query vectors and performing semantic retrieval matching, collaboratively achieves accurate metadata retrieval from user-input natural language descriptions. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0018] 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, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the first embodiment of the metadata retrieval method of this application. Figure 2 This is a flowchart illustrating the second embodiment of the metadata retrieval method of this application. Figure 3 This is a flowchart illustrating the fourth embodiment of the metadata retrieval method of this application. Figure 4 A simplified flowchart illustrating the metadata retrieval method provided in this application; Figure 5 This is a schematic diagram of the module structure of the metadata retrieval device according to an embodiment of this application; Figure 6 This is a schematic diagram of the device structure of the hardware operating environment involved in the metadata retrieval method in this application embodiment.

[0020] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0021] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0022] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0023] The main solution of this application embodiment is: to obtain a natural language query instruction input by the user; to call a pre-built semantic understanding model to vectorize the natural language query instruction to obtain a metadata query vector; to perform semantic retrieval matching on the metadata query vector in a pre-built metadata vector library to obtain a preliminary matching result set; to determine the number of results in the preliminary matching result set, and to generate a metadata retrieval result based on the number of results.

[0024] In this embodiment, for ease of description, the following description uses the metadata retrieval system as the execution entity.

[0025] This application's embodiments take into account that: Traditional metadata retrieval systems mainly employ structured rules, keyword matching, or predefined query languages. Their core retrieval logic relies on users explicitly inputting metadata fields, and the system returns results through precise matching of field values. However, due to the lack of semantic understanding capabilities, traditional metadata retrieval systems cannot perform semantic association matching based on user-input natural language descriptions. Consequently, when users input natural language queries with non-fixed fields, the system cannot accurately identify the query request and locate semantically relevant metadata, and can only rely on precise matching of keywords or fields, failing to meet the semantic-level metadata retrieval needs.

[0026] Therefore, this application provides a solution. First, it obtains the user's input natural language query instruction: breaking the dependence on fixed field input, it receives the user's unstructured natural language request, providing raw input for subsequent semantic processing. Further, it calls a pre-built semantic understanding model to vectorize the natural language query instruction, obtaining a metadata query vector: utilizing the semantic embedding capability of the semantic understanding model, the user's natural language query is transformed into a high-dimensional semantic vector. This vector encodes the user's deep semantic intent, providing a unified representation space for subsequent semantic similarity matching. Second, it performs semantic retrieval matching on the metadata query vector in a pre-built metadata vector library, obtaining a preliminary matching result set: through semantic retrieval matching, semantically relevant metadata is accurately located. Finally, it determines the number of results in the preliminary matching result set and generates metadata retrieval results based on this number: adjusting the metadata retrieval results according to the number of results ensures that the final returned results both conform to the user's semantic intent and have an appropriate scale, further improving the effectiveness of the retrieval. In summary, this application, by calling a pre-built semantic understanding model to generate metadata query vectors and performing semantic retrieval matching, collaboratively achieves accurate metadata retrieval from the user's input natural language description.

[0027] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device or metadata retrieval system capable of performing the above functions. The following description uses a metadata retrieval system as an example to illustrate this embodiment and the subsequent embodiments.

[0028] Based on this, embodiments of this application provide a metadata retrieval method, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the metadata retrieval method of this application.

[0029] In this embodiment, the metadata retrieval method includes steps S10 to S40: Step S10: Obtain the natural language query command input by the user; Natural language query commands refer to commands entered by users in the form of unstructured language used in daily human communication to express their metadata query needs. They are not limited to fixed-format field inputs and can contain any semantic description content.

[0030] The purpose of obtaining natural language query commands input by users is to break the dependence of traditional metadata retrieval on fixed field inputs, receive diverse query requests from users, provide raw input for subsequent semantic processing, and thus achieve more flexible metadata retrieval.

[0031] In one possible implementation, the system supports multiple input methods to obtain natural language query instructions, including text input box input, speech-to-text input (achieved through an integrated speech recognition module), and image recognition to extract text input (using OCR technology to parse user-uploaded images containing query content), to cover user needs in different scenarios.

[0032] Step S20: Invoke the pre-built semantic understanding model to vectorize the natural language query instruction to obtain the metadata query vector; Semantic understanding models refer to artificial intelligence models with deep semantic parsing capabilities, which can transform natural language into semantic representations that machines can understand. Query vectorization refers to the process of converting natural language query commands into high-dimensional semantic vectors; metadata query vectors refer to high-dimensional vectors obtained through query vectorization, used to encode the semantic intent of user queries.

[0033] The purpose of calling a pre-built semantic understanding model to vectorize natural language query instructions is to transform the user's natural language request into a machine-computable vector form, providing a unified representation space for subsequent semantic retrieval and matching, and solving the technical bottleneck of traditional keyword matching being unable to capture deep semantics.

[0034] In one possible implementation, the semantic understanding model can be a natural language processing model, and the model can be fine-tuned by using domain-specific metadata corpus (such as a business terminology table, metadata description documents, etc.) to improve the accuracy of the vector representation of metadata-related semantics; at the same time, the dimension of the metadata query vector can be set to 768 or 1024 dimensions to balance computational efficiency and semantic expressive power.

[0035] Step S30: Perform semantic retrieval and matching on the metadata query vector in the pre-built metadata vector library to obtain a preliminary matching result set; A metadata vector library refers to a database that stores metadata semantic vectors, where the metadata vectors are transformed from metadata information (such as business terms, lineage relationships, sample data descriptions, etc.).

[0036] Semantic retrieval matching refers to the process of finding the semantically closest metadata by calculating the similarity between the metadata query vector and the metadata vector stored in the metadata vector library; the preliminary matching result set refers to the set of metadata that is semantically similar to the query vector obtained through semantic retrieval matching.

[0037] The purpose of semantic retrieval and matching of metadata query vectors in a pre-built metadata vector library is to find the metadata most relevant to the user query from a semantic level, replacing the traditional keyword exact matching and improving the accuracy of the retrieval.

[0038] In one possible implementation, semantic retrieval matching uses the cosine similarity calculation method, and the metadata vector library is built on the FAISS vector database to support efficient similarity queries. In addition, the metadata vector library will periodically synchronize with the metadata updates of the unified metadata center (such as adding tables / columns, updating business term mappings, etc.), automatically regenerating and replacing the corresponding metadata vectors to ensure data timeliness.

[0039] Step S40: Determine the number of results in the preliminary matching result set, and generate metadata retrieval results based on the number of results.

[0040] The number of results refers to the number of metadata entries contained in the initial matching result set; the metadata retrieval results refer to the final, optimized metadata query results presented to the user.

[0041] The purpose of determining the number of results in the initial matching result set and generating metadata retrieval results based on that number is to optimize the scale of the results, ensure that users can efficiently obtain the information they need, and avoid the inconvenience caused by too many or too few results.

[0042] In one possible implementation, the system presets a threshold range for the number of results (e.g., 5 to 50 results): if the number of results exceeds 50, it automatically generates filtering conditions based on metadata attributes (e.g., "Should the data owner be restricted to the user's management department?"); if the number of results is less than 5, it provides suggestions based on semantic expansion (e.g., "Should age metadata related to the user's birth date be included?"); if the number of results is within the threshold range, it directly returns the results. Furthermore, it should be noted that the system-generated filtering conditions or expanded suggestions will be personalized based on the user's historical query behavior (e.g., the data owners the user frequently focuses on) to improve the relevance of the suggestions.

[0043] This embodiment provides a metadata retrieval method. First, it obtains the user's input natural language query instruction: breaking the dependence on fixed field input, it receives the user's unstructured natural language request, providing raw input for subsequent semantic processing. Further, it calls a pre-built semantic understanding model to vectorize the natural language query instruction, obtaining a metadata query vector: utilizing the semantic embedding capability of the semantic understanding model, the user's natural language query is transformed into a high-dimensional semantic vector. This vector encodes the user's deep semantic intent, providing a unified representation space for subsequent semantic similarity matching. Second, it performs semantic retrieval matching on the metadata query vector in a pre-built metadata vector library, obtaining a preliminary matching result set: through semantic retrieval matching, semantically relevant metadata is accurately located. Finally, it determines the number of results in the preliminary matching result set and generates metadata retrieval results based on this number: adjusting the metadata retrieval results according to the number of results ensures that the final returned results both conform to the user's semantic intent and have an appropriate scale, further improving the effectiveness of the retrieval. In summary, this application, by calling a pre-built semantic understanding model to generate metadata query vectors and performing semantic retrieval matching, collaboratively achieves accurate metadata retrieval from user-input natural language descriptions.

[0044] In one feasible implementation, step S20 may include steps S21 to S23: Step S21: Perform data preprocessing on the natural language query command to obtain a standardized query command; Data preprocessing refers to the process of standardizing and denoising natural language query commands; standardized query commands refer to query commands that, after data preprocessing, have a unified format, clear content, and are adapted to the input requirements of subsequent models.

[0045] The purpose of preprocessing natural language query commands is to eliminate redundant information (such as typos and special symbols) and standardize the format (such as capitalization and punctuation) in the original query, so that the query commands conform to the input specifications of the semantic understanding model, thereby improving the accuracy and consistency of subsequent vector generation.

[0046] In one possible implementation, data preprocessing includes noise filtering (removing emojis and irrelevant brackets from the query), error correction (correcting typos through contextual semantics), word segmentation (breaking long sentences into semantic units), and format unification (converting all letters to lowercase). In addition, for query commands that are mixed in multiple languages, the system automatically identifies and converts them into a single target language to adapt to the model's single-language input requirements.

[0047] Step S22: Invoke the pre-built semantic understanding model to map the standardized query command to a fixed-dimensional high-dimensional vector space to obtain the initial vector; A fixed-dimensional high-dimensional vector space refers to a vector representation space with a preset number of dimensions (such as 768 or 1024 dimensions), where each dimension corresponds to a semantic feature of the text; the initial vector refers to the original semantic vector output by the semantic understanding model that has not undergone subsequent optimization.

[0048] The purpose of calling the semantic understanding model to map standardized query instructions to a high-dimensional vector space is to transform text-based queries into machine-computable vector forms, encode the deep semantic intent of the query, and provide a foundation for subsequent semantic retrieval and matching.

[0049] In one possible implementation, the semantic understanding model uses a large natural language processing model and is fine-tuned using domain-specific metadata corpus (such as a business terminology table and metadata description documents) to enhance the ability to capture semantics related to metadata; a fixed dimension of 768 is selected to balance semantic expressiveness and computational efficiency; in addition, the system performs validity checks on the initial vector output by the model, checking for the presence of NaN values ​​to ensure that the vector is usable.

[0050] Step S23: Convert the initial vector into a unit vector and adjust the data type of the unit vector to obtain the metadata query vector.

[0051] A unit vector is a vector with a magnitude of 1, obtained through normalization; data type refers to the storage format of vector elements; metadata query vector refers to the final vector used for subsequent semantic retrieval and matching after transformation and adjustment.

[0052] The purpose of converting the initial vector to a unit vector is to unify the vector magnitude and eliminate the influence of length on similarity calculation; the purpose of adjusting the data type is to optimize storage and computation efficiency.

[0053] In one possible implementation, the unit vector conversion uses L2 normalization (calculating the initial vector magnitude and dividing each element by the magnitude); the data type is adjusted to bfloat16 format to reduce storage usage and accelerate computation; in addition, the system performs magnitude verification on the converted unit vector (ensuring the magnitude is close to 1), and if it does not meet the requirement, it is re-normalized.

[0054] Based on the first embodiment of this application, a second embodiment of this application is proposed. In the second embodiment of this application, content that is the same as or similar to that in the first embodiment described above can be referred to the above description and will not be repeated hereafter.

[0055] Based on this, please refer to Figure 2 , Figure 2 This is a schematic flowchart of the second embodiment provided in this application. Figure 2 As shown, before step S30, the metadata retrieval method further includes steps S01 to S04: Step S01: Collect initial metadata from the multi-source heterogeneous system; Multi-source heterogeneous systems refer to multiple data systems with different sources, data structures, or storage methods, such as databases, data warehouses, and data lakes; initial metadata refers to the raw metadata information collected from multi-source heterogeneous systems that has not been cleaned and integrated.

[0056] The purpose of collecting initial metadata from multi-source heterogeneous systems is to obtain metadata covering different business scenarios and data sources, providing basic data support for the subsequent construction of a unified metadata center, and achieving the comprehensiveness and completeness of metadata.

[0057] In one possible implementation, the system adopts a multi-mode acquisition strategy, including acquiring metadata (such as SCHEMATA table and TABLES table information) from relational databases through database links, acquiring metadata from data lakes through API services, acquiring metadata from data warehouses through scheduled tasks, and adding a real-time acquisition module to synchronize frequently updated metadata in real time to ensure the timeliness of metadata.

[0058] Step S02: Perform data cleaning and data standardization on the initial metadata, and verify the logical relationships between the initial metadata. Data cleaning refers to the process of removing redundant, erroneous, or missing information from initial metadata; data standardization refers to the process of adjusting the format and standardizing the content of metadata from different sources according to unified rules; logical relationship verification refers to the process of checking whether the relationships between metadata conform to business logic or technical rules.

[0059] The purpose of cleaning and standardizing the initial metadata and verifying logical relationships is to improve the quality of the metadata, eliminate data discrepancies, ensure consistency and correctness among metadata, and lay a reliable foundation for subsequent integration steps.

[0060] In one possible implementation, data cleaning includes removing duplicate table information and supplementing missing column description information; data standardization includes unifying field names to the format of "table name_column name" and unifying data types to standard enumeration values; logical relationship verification includes checking whether foreign keys in the table correspond to existing columns and verifying whether the data flow in the lineage relationship is closed loop.

[0061] Step S03: If the logical relationship verification of the initial metadata passes, the business terminology table, lineage graph, operational metadata, and sample data description in the initial metadata are integrated to obtain unified metadata; A business terminology table refers to the collection of officially defined business terms and their mapping relationships with physical assets; a lineage graph refers to a relationship graph that records the source and destination of data; operational metadata refers to information describing the operational attributes of metadata, such as update frequency, data owner, and data quality score; and sample data descriptions refer to rich text descriptions added to key tables or columns.

[0062] Unified metadata refers to a set of metadata that is structurally consistent and has complete content, formed by integrating the above information.

[0063] If the logical relationship of the initial metadata passes verification, the purpose of integrating the business terminology table, lineage graph, operational metadata, and sample data description to obtain unified metadata is to build a centralized and unified metadata center, providing complete basic data for subsequent metadata vectorization and semantic retrieval.

[0064] In one possible implementation, the system integrates data in the order of "business terminology table, lineage diagram, operational metadata, and sample data description" to ensure that business logic takes precedence. In addition, it should be noted that metadata version control is implemented, generating a new metadata version each time an integration operation is performed and recording the version change log for easy retrospective and management.

[0065] Step S04: After converting the unified metadata into metadata vectors, store them in the vector database to obtain the metadata vector library.

[0066] A vector database refers to a database system specifically designed for storing and retrieving high-dimensional vectors; a metadata vector library refers to a collection of vector databases that store all metadata vectors.

[0067] The purpose of converting unified metadata into metadata vectors and storing them in a vector database to obtain a metadata vector library is to provide a data foundation for subsequent semantic retrieval and matching using large models, thereby upgrading from "field matching" to "semantic association matching".

[0068] In one possible implementation, the system calls the embedding interface of the natural language large model to convert the unified metadata into a 768-dimensional metadata vector, and selects FAISS as the vector database for storage. In addition, it should be noted that the system adds a vector periodic update mechanism. When the unified metadata changes, the corresponding metadata vector is automatically regenerated and the vector database is updated to ensure the consistency between the metadata vector and the unified metadata.

[0069] In this embodiment, initial metadata is collected from multi-source heterogeneous systems. The metadata quality is improved through cleaning, standardization, and logical relationship verification. This results in the integration of unified metadata, including a business terminology table and lineage relationships. The metadata is then converted into vectors and stored in a vector database, constructing a high-quality, unified metadata vector library. This provides comprehensive and accurate basic data support for subsequent semantic understanding, vector matching, and intelligent retrieval based on large models. It effectively solves the problems of scattered sources, inconsistent quality, and inability to semantically correlate traditional metadata retrieval, thereby improving the accuracy and efficiency of the overall retrieval process.

[0070] Based on the above embodiments of this application, a third embodiment of this application is proposed. In this third embodiment, content that is the same as or similar to that in the above embodiments can be referred to the above description, and will not be repeated hereafter.

[0071] In this embodiment, step S30, which involves semantic retrieval and matching of the metadata query vector in a pre-built metadata vector library to obtain a preliminary matching result set, may include steps S31 to S33: Step S31: Based on the metadata query vector, perform an approximate nearest neighbor search in the metadata vector library to obtain a candidate vector set; Approximate nearest neighbor search refers to an algorithm that quickly finds the set of vectors most similar to the target vector in a high-dimensional space; the candidate vector set refers to the set of vectors that are semantically similar to the metadata query vector obtained through approximate nearest neighbor search.

[0072] The candidate vector set refers to the set of vectors that are semantically similar to the metadata query vector, obtained through approximate nearest neighbor search. The purpose of performing approximate nearest neighbor search in the metadata vector library based on the metadata query vector is to quickly narrow down the scope of semantic matching, avoiding the inefficiency caused by performing an exact search on the entire metadata vector set, and thus improving retrieval speed.

[0073] In one possible implementation, the system employs the HNSW (Hierarchical Navigable SmallWorlds) algorithm to perform an approximate nearest neighbor search. This algorithm achieves fast retrieval of high-dimensional vectors by constructing a hierarchical graph structure, and the retrieval accuracy is close to that of an exact search.

[0074] Step S32: Calculate the cosine similarity between each metadata vector in the candidate vector set and the metadata query vector; Cosine similarity refers to the cosine value of the angle between two vectors, ranging from [-1, 1]. A larger value indicates greater semantic similarity between the vectors. The purpose of calculating the cosine similarity between each metadata vector and the metadata query vector in the candidate vector set is to quantify the semantic similarity between each candidate vector and the query vector, providing a basis for subsequent screening.

[0075] In one possible implementation, the system uses matrix operations from the BLAS library to accelerate cosine similarity calculation, converting the candidate vector set into a matrix and performing batch dot product operations with the metadata query vector, which greatly improves computational efficiency.

[0076] Step S33: Filter the candidate vector set according to the cosine similarity to obtain a preliminary matching result set.

[0077] Cosine similarity filtering refers to retaining vectors with similarity higher than a preset threshold; the preliminary matching result set refers to the set of metadata vectors that meet the basic semantic requirements after filtering.

[0078] The purpose of filtering candidate vector sets based on cosine similarity is to obtain metadata vectors that are semantically sufficiently similar to the query vector, and to filter out irrelevant vectors.

[0079] In one possible implementation, the system employs dynamic threshold filtering: based on the cosine similarity distribution of the candidate vector set, a threshold is set to 1.2 times the median similarity, ensuring a balance between the relevance and quantity of the filtering results. For example, in a specific implementation, a fixed threshold of 0.7 is set to filter out metadata vectors with a cosine similarity ≥ 0.7, resulting in a preliminary matching result set containing 30 vectors.

[0080] In this embodiment, the semantic matching range is quickly narrowed by near nearest neighbor search, the semantic similarity is quantified by calculating cosine similarity, and the candidate vector set is filtered to obtain preliminary matching results. This realizes the transformation from "full vector traversal" to "precise semantic filtering", which greatly improves the efficiency and accuracy of metadata retrieval. It lays the foundation for subsequent intelligent adjustment of the number of search results and effectively solves the problem of missed detection and false detection caused by traditional retrieval relying on complete keyword matching.

[0081] Based on the above embodiments of this application, a fourth embodiment of this application is proposed. In this fourth embodiment, content that is the same as or similar to that in the above embodiments can be referred to the above description, and will not be repeated hereafter.

[0082] refer to Figure 3 , Figure 3 This is a flowchart illustrating the fourth embodiment of this application, as shown below. Figure 3 As shown, step S40, which involves determining the number of results in the preliminary matching result set and generating metadata retrieval results based on that number of results, may include steps S41 to S44: Step S41: Determine the number of results in the preliminary matching result set; The purpose of determining the number of results in the initial matching result set is to determine the subsequent processing method (expansion, filtering, or direct sorting) to ensure that users receive an appropriate number of search results.

[0083] In one possible implementation, the system uses a counter to count the number of results and stores the result count in a temporary variable for later comparison with a preset threshold. For example, in one specific implementation, the system counts the number of vectors in the initial matching result set as 30, resulting in a result count of 30.

[0084] Step S42: If the number of results is less than a preset lower threshold, the semantic understanding model is invoked to generate extended retrieval suggestions, the candidate vector set is re-selected based on the extended retrieval suggestions, and metadata retrieval results are generated based on the re-selected candidate vector set. The preset lower threshold refers to the minimum number of results set; expanded search suggestions refer to suggestions generated by the model to expand the search scope; re-filtering refers to adjusting the search conditions based on the suggestions.

[0085] If the number of results is less than a preset lower threshold, the purpose of calling the semantic understanding model to generate extended search suggestions is to solve the problem of insufficient results and help users find more relevant metadata.

[0086] In one possible implementation, the semantic understanding model generates synonyms or related business terms based on the query semantics, such as expanding the query "birthday" to "date of birth," "birth anniversary," "age," etc. For example, in a specific implementation, if the number of results is 3, which is less than the lower threshold of 5, the model generates a suggestion: "Does it contain age metadata related to the date of birth?" The system then re-filters the candidate vector set based on the suggestion, obtaining metadata retrieval results containing 10 vectors.

[0087] Step S43: If the number of results is greater than a preset upper limit threshold, the semantic understanding model is invoked to generate intelligent filtering conditions, the candidate vector set is re-filtered based on the intelligent filtering conditions, and metadata retrieval results are generated based on the re-filtered candidate vector set. The preset upper limit threshold refers to the maximum number of results set; intelligent filtering conditions refer to the conditions generated by the model to narrow the search scope; re-filtering refers to adjusting the search results based on the conditions.

[0088] If the number of results exceeds the preset upper limit threshold, the purpose of calling the semantic understanding model to generate intelligent filtering conditions is to solve the problem of too many results and help users locate the correct information.

[0089] In one possible implementation, the model generates personalized filtering conditions based on the user's historical behavior. For example, in one specific implementation, if the number of results is 60, which is greater than the upper threshold of 50, the model generates the condition "Should the data owner be restricted to the user's management department?", and the system re-filters based on the condition to obtain metadata retrieval results containing 35 vectors.

[0090] Step S44: If the number of results is neither less than the preset lower threshold nor greater than the upper threshold, then perform semantic relevance sorting on the preliminary matching result set to obtain metadata retrieval results.

[0091] Semantic relevance ranking refers to ranking results based on factors such as semantic similarity, user behavior, and metadata importance; metadata retrieval results refer to the final results displayed to the user.

[0092] The purpose of performing semantic relevance sorting on the initial matching result set is to prioritize the display of the most relevant results, thereby improving the user experience.

[0093] In one possible implementation, the system employs a multi-dimensional weighted ranking: cosine similarity, user historical click weight, and metadata importance, to comprehensively calculate the ranking score. For example, in a specific implementation, the system performs a multi-dimensional weighted ranking on the vectors of the initial matching result set to obtain the ranked metadata retrieval results, where "user's birthday" is ranked first and "user's age" is ranked second.

[0094] In this embodiment, by determining the number of preliminary matching results, the large model is dynamically invoked to generate extended suggestions or filtering conditions, or semantic relevance ranking is performed to achieve intelligent dynamic adjustment of search results. This avoids information overload when there are too many results and solves the problem of insufficient retrieval when there are too few results. At the same time, by prioritizing the display of content that matches the user's intent through multi-dimensional ranking, the flexibility of metadata retrieval and user experience are significantly improved, meeting the accurate retrieval needs in different scenarios.

[0095] In one feasible implementation, step S42 may include steps S421 to S424: Step S421: Based on the extended search suggestions, perform an extended search in the metadata vector library to obtain a new vector set; Extended search refers to the search operation performed after adjusting the search criteria based on extended search suggestions; the new vector set refers to the new vector set obtained through extended search that is semantically similar to the metadata query vector.

[0096] The purpose of extending the search based on the extended search suggestions in the metadata vector library is to address the problem of insufficient initial matching results, supplement more metadata vectors related to the query semantics, and ensure that users can obtain sufficient search results.

[0097] In one possible implementation, the system transforms the extended retrieval suggestions into extended query vectors, and then performs an approximate nearest neighbor search in the metadata vector library to obtain a new set of vectors, thereby improving the accuracy of the extended retrieval.

[0098] Step S422: Merge the candidate vector set and the newly added vector set to obtain a re-filtered candidate vector set; Fusion refers to the operation of merging the candidate vector set and the newly added vector set and removing duplicate vectors; the candidate vector set refers to the set of similar vectors obtained through approximate nearest neighbor search; the newly added vector set refers to the new vector set obtained by expanding the search; the re-filtered candidate vector set refers to the fused set of vectors without duplicates.

[0099] The purpose of merging the candidate vector set and the newly added vector set is to integrate the old and new search results, avoid duplicate vectors occupying resources, and retain all relevant vectors.

[0100] In one possible implementation, the system employs a deduplication mechanism to calculate the similarity of vectors in two sets. If the similarity exceeds a preset threshold, it is determined to be a duplicate and removed to ensure the uniqueness of the result.

[0101] Step S423: Calculate the cosine similarity between each metadata vector in the re-filtered candidate vector set and the metadata query vector; The re-filtered candidate vector set refers to the fused set of vectors without repetition; the metadata vector refers to the vector that encodes the semantics of metadata stored in the metadata vector library; the metadata query vector refers to the vector that encodes the semantic intent of the user's query; and the cosine similarity refers to the cosine value of the angle between two vectors, used to quantify the degree of semantic similarity.

[0102] The purpose of calculating the cosine similarity between each metadata vector and the metadata query vector in the re-filtered candidate vector set is to requantify the semantic similarity between each vector and the query vector, so as to provide an accurate basis for subsequent filtering.

[0103] Step S424, return to the execution step: filter the candidate vector set according to the cosine similarity to obtain a preliminary matching result set.

[0104] The purpose of returning to the filtering step is to obtain a new preliminary matching result set based on the recalculated cosine similarity, thus solving the problem of insufficient original results.

[0105] In another feasible implementation, step S43 may include steps S431 to S433: Step S431: Filter the candidate vector set based on the intelligent filtering conditions to obtain a re-filtered candidate vector set; Intelligent filtering conditions refer to the conditions generated by the semantic understanding model to narrow down the search scope; filtering refers to excluding vectors that do not meet the requirements based on the conditions; the candidate vector set refers to the previous set of similar vectors; the re-filtered candidate vector set refers to the set of vectors that meet the conditions after filtering.

[0106] The purpose of using intelligent filtering conditions to filter candidate vector sets is to solve the problem of too many preliminary results and to accurately narrow down the search scope.

[0107] In one possible implementation, the system employs a multi-condition combined filtering mechanism, which decomposes the intelligent filtering conditions into multiple sub-conditions and filters the candidate vector set sequentially, thereby improving the accuracy of filtering.

[0108] Step S432: Calculate the cosine similarity between each metadata vector in the re-filtered candidate vector set and the metadata query vector; The re-filtered candidate vector set refers to the set of vectors that meet the filtering criteria.

[0109] The purpose of calculating the cosine similarity between each metadata vector and the metadata query vector in the re-filtered candidate vector set is to requantify the similarity between each filtered vector and the query, providing a basis for subsequent filtering.

[0110] In one possible implementation, the system uses the BLAS library for batch calculations to accelerate the calculation process of cosine similarity.

[0111] Step S433: Return to the execution step: Filter the candidate vector set according to the cosine similarity to obtain a preliminary matching result set.

[0112] The purpose of returning to the filtering step is to obtain a result set that meets the threshold based on the recalculated similarity, thus solving the problem of too many original results.

[0113] For example, to help understand the implementation flow of the metadata retrieval method obtained by combining the above embodiments, please refer to... Figure 4 , Figure 4 A simplified flowchart of a metadata retrieval method is provided, specifically: Metadata preprocessing and vector construction: Initial metadata is collected from heterogeneous systems such as data lakes and data warehouses, cleaned, standardized and verified for logical relationships, and then integrated into unified metadata. Semantic vectorization is then used to transform it into metadata vectors and stored in a vector database.

[0114] User query processing: The user inputs a natural language query, which is then parsed by the large model's semantic understanding and transformed into a metadata query vector.

[0115] Preliminary semantic matching: Calculate the cosine similarity between the query vector and the metadata vector in the vector database, and filter out a set of semantically similar metadata vectors. Extract the metadata corresponding to the similar vectors to form a preliminary matching result set.

[0116] Results Quantity Judgment and Adjustment: Too Many Results: The large model generates dynamic filtering conditions, and after confirmation with the user, filters the candidate set based on the conditions, recalculates the similarity, and filters again to obtain a simplified result; Too Few Results: The large model generates expanded search suggestions, and after confirmation with the user, expands the search scope to obtain a new vector set, merges the original candidate set, recalculates the similarity, and filters again to obtain a supplemented result.

[0117] Results sorting and display: The final results are sorted based on multiple dimensions such as semantic relevance, correlation, user usage, and asset importance, and presented to users, with traceability links provided to support further verification.

[0118] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the metadata retrieval method of this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0119] This application also provides a metadata retrieval device; please refer to [reference needed]. Figure 5 The metadata retrieval device includes: The instruction acquisition module 10 is used to acquire natural language query instructions input by the user; Vector query module 20 is used to call a pre-built semantic understanding model to vectorize the natural language query command and obtain metadata query vectors; The retrieval and matching module 30 is used to perform semantic retrieval and matching on the metadata query vector in a pre-built metadata vector library to obtain a preliminary matching result set; The result adjustment module 40 is used to determine the number of results in the preliminary matching result set and generate metadata retrieval results based on the number of results.

[0120] The metadata retrieval device provided in this application, employing the metadata retrieval method in the above embodiments, can solve the technical problem of metadata retrieval. Compared with the prior art, the beneficial effects of the metadata retrieval device provided in this application are the same as those of the metadata retrieval method provided in the above embodiments, and other technical features in the metadata retrieval device are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0121] This application provides a metadata retrieval device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the metadata retrieval method in Embodiment 1 above.

[0122] The following is for reference. Figure 6The diagram illustrates a structural schematic of a metadata retrieval device suitable for implementing embodiments of this application. The metadata retrieval device in these embodiments may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 6 The metadata retrieval device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0123] like Figure 6 As shown, the metadata retrieval device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in the read-only memory 1002 or a program loaded from the storage device 1003 into the random access memory 1004. The random access memory 1004 also stores various programs and data required for the operation of the metadata retrieval device. The processing unit 1001, the read-only memory 1002, and the random access memory 1004 are interconnected via a bus 1005. An input / output interface 1006 is also connected to the bus. Typically, the following systems can be connected to the input / output interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the metadata retrieval device to communicate wirelessly or wiredly with other devices to exchange data. While the figure shows metadata retrieval devices with various systems, it should be understood that implementing or having all of the systems shown is not required. More or fewer systems may be implemented alternatively.

[0124] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0125] The metadata retrieval device provided in this application, employing the metadata retrieval method in the above embodiments, can solve the technical problem of metadata retrieval. Compared with the prior art, the beneficial effects of the metadata retrieval device provided in this application are the same as those of the metadata retrieval method provided in the above embodiments, and other technical features in this metadata retrieval device are the same as those disclosed in the method of the previous embodiment, and will not be repeated here.

[0126] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0127] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0128] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the metadata retrieval method in the above embodiments.

[0129] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0130] The aforementioned computer-readable storage medium may be included in the metadata retrieval device; or it may exist independently and not be assembled into the metadata retrieval device.

[0131] The aforementioned computer-readable storage medium carries one or more programs. When the aforementioned one or more programs are executed by the metadata retrieval device, the metadata retrieval device causes the following: to acquire a natural language query instruction input by a user; to call a pre-built semantic understanding model to vectorize the natural language query instruction into a metadata query vector; to perform semantic retrieval matching on the metadata query vector in a pre-built metadata vector library to obtain a preliminary matching result set; to determine the number of results in the preliminary matching result set; and to generate a metadata retrieval result based on the number of results.

[0132] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0133] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0134] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0135] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described metadata retrieval method, thereby solving the technical problem of metadata retrieval. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the metadata retrieval method provided in the above embodiments, and will not be repeated here.

[0136] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the metadata retrieval method described above.

[0137] The computer program product provided in this application can solve the technical problem of metadata retrieval. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the metadata retrieval method provided in the above embodiments, and will not be repeated here.

[0138] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A metadata retrieval method, characterized in that, The metadata retrieval method includes: Obtain the user's input in natural language; The natural language query instruction is vectorized by calling a pre-built semantic understanding model to obtain a metadata query vector; Semantic retrieval and matching of the metadata query vectors are performed in a pre-built metadata vector library to obtain a preliminary matching result set; Determine the number of results in the preliminary matching result set, and generate metadata retrieval results based on the number of results.

2. The metadata retrieval method as described in claim 1, characterized in that, The step of calling a pre-built semantic understanding model to vectorize the natural language query instruction and obtain the metadata query vector includes: The natural language query instructions are preprocessed to obtain standardized query instructions; The standardized query command is mapped to a fixed-dimensional high-dimensional vector space by calling a pre-built semantic understanding model to obtain an initial vector; The initial vector is converted into a unit vector, and the data type of the unit vector is adjusted to obtain the metadata query vector.

3. The metadata retrieval method as described in claim 1, characterized in that, Before the step of performing semantic retrieval and matching on the metadata query vector in the pre-built metadata vector library to obtain a preliminary matching result set, the method further includes: Collect initial metadata from multi-source heterogeneous systems; The initial metadata is cleaned and standardized, and the logical relationships between the initial metadata are verified. If the logical relationship verification of the initial metadata passes, the business terminology table, lineage diagram, operational metadata, and sample data description in the initial metadata are integrated to obtain unified metadata; The unified metadata is converted into metadata vectors and then stored in a vector database to obtain a metadata vector library.

4. The metadata retrieval method as described in claim 1, characterized in that, The step of performing semantic retrieval and matching on the metadata query vector in a pre-built metadata vector library to obtain a preliminary matching result set includes: Based on the metadata query vector, an approximate nearest neighbor search is performed in the metadata vector library to obtain a candidate vector set; Calculate the cosine similarity between each metadata vector in the candidate vector set and the metadata query vector; The candidate vector set is filtered based on the cosine similarity to obtain a preliminary matching result set.

5. The metadata retrieval method as described in claim 4, characterized in that, The step of determining the number of results in the preliminary matching result set and generating metadata retrieval results based on the number of results includes: Determine the number of results in the preliminary matching result set; If the number of results is less than a preset lower threshold, the semantic understanding model is invoked to generate extended retrieval suggestions, the candidate vector set is re-selected based on the extended retrieval suggestions, and metadata retrieval results are generated based on the re-selected candidate vector set. If the number of results exceeds a preset upper limit threshold, the semantic understanding model is invoked to generate intelligent filtering conditions, the candidate vector set is re-filtered based on the intelligent filtering conditions, and metadata retrieval results are generated based on the re-filtered candidate vector set. If the number of results is neither less than the preset lower threshold nor greater than the upper threshold, then the semantic relevance ranking is performed on the preliminary matching result set to obtain the metadata retrieval results.

6. The metadata retrieval method as described in claim 5, characterized in that, The step of re-filtering the candidate vector set based on the expanded search suggestions and generating metadata retrieval results based on the re-filtered candidate vector set includes: Based on the extended search suggestions, an extended search is performed in the metadata vector database to obtain a new vector set; By merging the candidate vector set and the newly added vector set, a re-filtered candidate vector set is obtained; Calculate the cosine similarity between each metadata vector in the re-filtered candidate vector set and the metadata query vector; Return to the execution steps: filter the candidate vector set according to the cosine similarity to obtain a preliminary matching result set.

7. The metadata retrieval method as described in claim 5, characterized in that, The step of re-filtering the candidate vector set based on the intelligent filtering conditions and generating metadata retrieval results based on the re-filtered candidate vector set includes: The candidate vector set is filtered based on the intelligent filtering conditions to obtain a re-filtered candidate vector set. Calculate the cosine similarity between each metadata vector in the re-filtered candidate vector set and the metadata query vector; Return to the execution steps: filter the candidate vector set according to the cosine similarity to obtain a preliminary matching result set.

8. A metadata retrieval device, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the metadata retrieval method as described in any one of claims 1 to 7.

9. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the metadata retrieval method as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the steps of the metadata retrieval method as described in any one of claims 1 to 7.

Images