Cross-system heterogeneous metadata intelligent collection method and system based on LLM

Abstract

The present invention discloses a method and system for intelligently collecting heterogeneous metadata across systems based on LLM, comprising the following steps: S1, collecting and preprocessing heterogeneous system data to construct a heterogeneous graph; S2, parsing nodes using a large language model, concatenating structural features, and forming an initial node representation; S3, constructing a heterogeneous graph transformer, performing multi-layer propagation and low-rank decomposition, and outputting a final node representation; S4, generating field alignment, interface mapping, and dependency paths, and planning the collection path and sequence; S5, executing the collection task, converting the structured format, writing to the metadata center, and performing closed-loop optimization. This invention achieves intelligent collection, semantic parsing, and automated integration of heterogeneous metadata across systems, significantly improving the efficiency and intelligence of metadata management in a multi-source data environment.

Description

Technical Field

[0001] The present invention relates to the field of data management technology, and in particular to a method and system for intelligently collecting heterogeneous metadata across systems based on LLM. Background Art

[0002] Against the backdrop of increasing complexity in information systems and increasing data heterogeneity, enterprises and organizations often need to integrate data systems from multiple sources, such as relational databases, non-relational databases, big data platforms, and distributed storage. These heterogeneous systems exhibit significant differences in metadata structure, data types, semantic expressions, and interface protocols, making the collection and management of cross-system metadata an extremely arduous task. Traditional data integration methods often rely on manually configured collection scripts or the use of fixed templates for adaptation, and are unable to flexibly cope with the frequent changes and complex semantic differences in source systems. Some current ETL-based tools lack the ability to fine-tune semantic understanding and structural recognition when processing complex semi-structured and unstructured data, making it difficult to support the needs of dynamic metadata management.

[0003] With the development of artificial intelligence, large language models have demonstrated superior performance in semantic parsing and language understanding. Graph neural networks, especially heterogeneous graph neural architectures, have also been widely used in modeling complex, multi-type relationships. However, existing technologies still suffer from the following common flaws: First, while large language models can support semantic parsing, they lack a deep integration mechanism with system context and structural features when dealing with complex system metadata. Second, graph neural networks are susceptible to interference from redundant paths when processing high-dimensional features and multi-class edge nodes, resulting in semantic degradation and information redundancy, making it difficult to output accurate node representations. Third, metadata collection paths and scheduling strategies are often set based on manual experience, lacking an explainable semantic matching mechanism and structural stability assessment, making automated planning impossible. Fourth, the lack of a closed-loop optimization mechanism for feedback on the collection process prevents adaptive adjustment of models and strategies.

[0004] Therefore, how to provide a cross-system heterogeneous metadata intelligent collection method and system based on LLM is a problem that those skilled in the art urgently need to solve. Summary of the Invention

[0005] One objective of the present invention is to propose a method and system for intelligently collecting heterogeneous metadata across systems based on LLM. This method fully integrates the semantic understanding capabilities of large language models with the structural modeling advantages of heterogeneous graph neural networks, constructing a multi-layer information dissemination network under a low-rank optimization mechanism. Combined with field alignment, interface mapping, and path planning strategies, it enables automatic metadata recognition, semantic parsing, path planning, and dynamic collection for structured, semi-structured, and unstructured data. This method boasts a high degree of automation, strong semantic accuracy, strong cross-system adaptability, and robust sustainable optimization capabilities, significantly improving metadata integration efficiency and intelligent management in a multi-source heterogeneous system environment.

[0006] The LLM-based cross-system heterogeneous metadata intelligent collection method according to an embodiment of the present invention includes the following steps:

[0007] S1. Collect structured, semi-structured, and unstructured data from multiple heterogeneous information systems and pre-process them to construct a heterogeneous graph containing nodes and edges.

[0008] S2. Based on the constructed heterogeneous graph, the structural attribute information of each node is extracted to generate a node structural feature vector. The original description information of each node is semantically parsed using a large language model to generate a semantic representation vector. This is then embedded and concatenated with the node structural feature vector to form the initial node representation.

[0009] S3. Construct a heterogeneous graph Transformer network, input the initial node representation, set the node type mapping and edge type attention mechanism, perform multi-layer information propagation, calculate the attention weight and message matrix in each layer, perform low-rank decomposition on the attention weight and message matrix, introduce a rank-aware masking mechanism in the multi-layer information propagation process, combine the semantic representation vector with the edge connection strength in the heterogeneous graph, filter redundant paths, and output the final node representation;

[0010] S4. Based on the final node representation, generate field alignment relationships, interface mapping tables, and dependency paths, plan metadata collection paths, and determine the cross-system collection sequence and scope based on confidence levels, and determine metadata collection tasks.

[0011] S5. Execute metadata collection tasks, convert the raw metadata entries collected by each heterogeneous system into a unified structured format, and write them into the metadata center. Dynamically update the prompt template of the large language model and the rank parameter of the heterogeneous graph transformer to build a sustainable and optimized closed-loop process for heterogeneous metadata collection.

[0012] Optionally, the structured, semi-structured and unstructured data specifically include database table structures, log files, API documents, data directories and other source system metadata information.

[0013] Optionally, the preprocessing of the structured, semi-structured and unstructured data specifically includes format parsing, field extraction, annotation cleaning, key-value pair extraction, structure mapping and encoding normalization operations.

[0014] Optionally, the S2 specifically includes:

[0015] S21. For each node in the constructed heterogeneous graph, extract the associated original description information, where the original description information includes the field name, data type, data table or interface name, annotation content, and corresponding source system identifier, to form an original metadata text set;

[0016] S22. Organize the original metadata text collection into a natural language input sequence according to a preset structured input template, input it into a large language model based on the BERT structure and fine-tuned by the domain task, perform semantic understanding and context encoding operations, and obtain the semantic representation vector of the node;

[0017] S23. For each node, extract corresponding structural information features, including the encoding format of the field name, the data type encoding, the system type encoding, and the logical level or call depth of the node in the system, and construct a node structural feature vector;

[0018] S24, concatenating the semantic representation vector and the node structure feature vector to construct a combined vector;

[0019] S25. Use the combined vector as the input representation of the node in the heterogeneous graph Transformer network to form the initial representation of the node.

[0020] Optionally, the S3 specifically includes:

[0021] S31. Build a heterogeneous graph Transformer network and set the initial node representation as input;

[0022] S32. Define a node type mapping function to encode the types of all nodes in the heterogeneous graph into node type vectors as the information propagation condition of the heterogeneous graph Transformer network. The node type mapping function is a type encoding function constructed based on the system source, data mode, and entity category of the node.

[0023] S33. Construct an edge type attention mechanism, set an independent attention weight matrix for each type of edge, and perform differentiated attention calculation based on edge type attributes;

[0024] S34. Construct the heterogeneous graph Transformer network as a stacked network framework with multiple layers of propagation structure. In each propagation layer, the node representation output by the previous layer is input. The node type vector and the edge type attention weight are combined. Based on the adjacency relationship in the heterogeneous graph, the neighbor node representations of each node are weighted aggregated to generate a message mapping matrix corresponding to the edge type, and the node representation update operation is performed.

[0025] S35. In each propagation layer, perform a low-rank approximation operation on the attention weight matrix and the message mapping matrix generated by the current layer, concatenate the attention weight matrix and the message mapping matrix into a joint representation matrix, and use a low-rank decomposition operation to decompose them into the product of two low-dimensional dense matrices to generate a low-rank decomposition result matrix;

[0026] S36. Based on low-rank decomposition, a rank-aware gating mechanism is introduced to construct a regulation function that integrates rank contribution, attention significance and edge connection strength. Constructing gating factors , set the gating threshold function based on the gating factor, perform the gating screening operation, and construct the structural path mask matrix:

[0027] ;

[0028] in, is the matrix trace function, is the path attention score, is the edge connection strength, is the upper limit of the rank corresponding to the node type, To adjust the parameters, is the Sigmoid activation function, For the transmission path The corresponding left singular matrix representation in the low-rank decomposition;

[0029] S37, based on gating factors , the path attention score of each propagation path in the heterogeneous graph Transformer network Perform weighted adjustments and use the weighted results to replace the path attention scores, retaining only valid paths with significant gating factor values ​​and filtering out the propagation information of low-correlation edges;

[0030] S38. According to the distribution of node types in the heterogeneous graph Transformer network, the node type is set by integrating semantic entropy, structural centrality within the type and difference between types. The rank upper bound of :

[0031] ;

[0032] in, Node type The information entropy of Node type The average degree centrality of Node type Jensen-Shannon divergence between the distribution of non-type nodes, is the regulating factor, is the global basic constant of rank, and represents the probability of embedding vector distribution, is the set of representation vectors of all nodes corresponding to node type t;

[0033] S39. Perform a rank truncation operation on the low-rank decomposition result matrix according to the rank upper limit corresponding to each node type, and retain the rank dimension subspace corresponding to each type;

[0034] S310: Perform element-by-element multiplication of the low-rank decomposition result matrix after rank truncation and the structural path mask matrix to construct a node representation matrix with path selectivity and rank constraint characteristics, which is used as the output node representation of the current layer and input to the next propagation layer as the node initial input;

[0035] S311. After all propagation layers are completed, the representation of the nodes in the final layer is output as the final representation of the nodes in the heterogeneous graph Transformer network.

[0036] Optionally, the S4 specifically includes:

[0037] S41. Receive the final representation of the node output by the heterogeneous graph Transformer network and set the feature matching function ,in and The final representation of the nodes representing the field nodes in the two heterogeneous systems is:

[0038] ;

[0039] in, is the weighting parameter, represents the semantic cosine similarity, Representation node and Structural dependency matrix in heterogeneous graph Transformer network, Represents a vector norm;

[0040] S42, similarity score calculated based on feature matching function, using set threshold Filter candidate field node pairs to generate a field alignment relationship set ;

[0041] S43. Combine the field alignment relationship set with the node connection relationship in the heterogeneous graph Transformer network to construct the interface candidate pair set and introduce the interface mapping function , map the field alignment relationship to the parameter docking relationship between system interfaces to form an interface mapping table:

[0042] ;

[0043] in, and They are the source system and target system interfaces respectively. and is the weight coefficient, represents the semantic similarity of nodes, Indicates the interface structure similarity, Indicates when the function reaches its maximum value right;

[0044] S44. Based on the interface mapping table and the heterogeneous graph Transformer network, execute the path extraction algorithm to generate the field collection path between field nodes, construct a path triple sequence containing fields, interfaces, and call relationships, and form a dependency path graph;

[0045] S45. Comprehensively evaluate each field collection path based on the field matching degree, interface mapping consistency, and structural path stability of each path in the dependency path graph, and grade them based on the overall credibility, outputting a preliminary field collection grade score.

[0046] S46. Set the confidence level threshold based on the initial score of the field collection level , divide the field collection paths into three levels: high, medium, and low, and formulate corresponding cross-system collection priorities and path coverage, and generate an initial list of field collection tasks;

[0047] S47. Introduce a structural semantic consensus mechanism to perform consensus aggregation operations on redundant field nodes that appear in the initial list of field collection tasks, integrate the path context features of the field nodes in the path triple sequence, and the mapping deviation between the interface mapping function output and the interface mapping table to reconstruct the field alignment relationship;

[0048] S48. Based on the reconstructed field alignment relationship, interface mapping table and dependency path map, the metadata collection task is finally determined, and a multi-dimensional collection instruction set including field pairs, interface mapping pairs, path triplets and confidence levels is generated.

[0049] Optionally, the S5 specifically includes:

[0050] S51. Based on the determined metadata collection task, connect to multiple heterogeneous information systems in sequence, execute the collection operation indicated by the field pair, interface mapping pair, and path triplet, and obtain the original metadata entry;

[0051] S52. Load the corresponding metadata collection plug-in according to the system type of each heterogeneous information system, perform structural analysis on the original metadata entry, and extract field content, interface parameters, and path context information;

[0052] S53, converting the collected original metadata entries into a structured format according to the set field mapping rules, semantic tag templates and system structure specifications, and constructing metadata structure units with consistent system sources and complete semantics;

[0053] S54, writing the metadata structure unit converted into a structured format into the metadata center, and automatically triggering data consistency verification, redundancy elimination, and update marking;

[0054] S55. Dynamically update the prompt template of the large language model based on field semantic deviations, interface execution anomalies, and path coverage changes during the collection process;

[0055] S56. Based on the information propagation effect of each layer of the heterogeneous graph Transformer, the change in attention distribution and the feedback of low-rank decomposition performance during the collection process, the rank parameter of the heterogeneous graph Transformer is updated to build a sustainable and optimized closed-loop process for heterogeneous metadata collection.

[0056] The LLM-based cross-system heterogeneous metadata intelligent collection system according to an embodiment of the present invention includes the following modules:

[0057] Data acquisition and preprocessing module, used to collect structured, semi-structured and unstructured data, complete preprocessing and construct heterogeneous graphs;

[0058] The semantic parsing module is used to call the large language model to parse the node description information, generate the semantic representation vector, and embed it with the structural feature to form the initial node representation;

[0059] The graph modeling module is used to build a heterogeneous graph Transformer network, input the initial node representation, perform multi-layer information propagation and low-rank decomposition, and output the final node representation;

[0060] The path planning module is used to generate field alignment relationships, interface mapping tables, and dependency path graphs based on the final node representation, plan the collection path, and determine the collection tasks;

[0061] The task execution module is used to execute the collection task, load the collection plug-in to obtain the original metadata entries, convert them into structured metadata, generate standardized collection results and write them into the metadata center;

[0062] An intelligent optimization module is used to dynamically update the large language model prompt template based on the standardized collection results, and optimize the rank parameter of the heterogeneous graph Transformer network based on propagation feedback.

[0063] The beneficial effects of the present invention are:

[0064] This paper significantly enhances the intelligence and automation capabilities of cross-system heterogeneous metadata management by constructing a metadata collection method that integrates a large language model with a heterogeneous graph Transformer network. Compared to existing methods, this method not only supports unified modeling and parsing of structured, semi-structured, and unstructured data, but also achieves efficient information dissemination and redundant path suppression by constructing heterogeneous graphs, generating initial node representations, introducing node and edge type attention mechanisms, and combining low-rank approximation with rank-aware control mechanisms, thereby enhancing the precision and efficiency of semantic modeling.

[0065] The present invention effectively supports the dynamic planning and confidence assessment of the collection path by constructing field alignment relationships, interface mapping tables, and dependency path graphs, and further improves the stability and accuracy of multi-source field mapping in combination with the structural semantic consensus mechanism. Finally, in the process of executing the collection task, with the help of the structural unified conversion module and the prompt parameter optimization mechanism, the dynamic adjustment of the large language model prompt template and the Transformer rank parameter is realized, and a metadata collection process with closed-loop optimization capabilities is constructed, which has strong robustness, adaptability, and scalability. This method can be widely applied to enterprise-level data governance and intelligent data asset management scenarios in multi-system, multi-format, and multi-structure data environments, and has significant practical value and promotion prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0067] Figure 1 This is a flow chart of the LLM-based cross-system heterogeneous metadata intelligent collection method proposed in the present invention;

[0068] Figure 2 This is a structural diagram of the LLM-based cross-system heterogeneous metadata intelligent collection system proposed in the present invention. DETAILED DESCRIPTION

[0069] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0070] refer to Figure 1 The LLM-based cross-system heterogeneous metadata intelligent collection method includes the following steps:

[0071] S1. Collect structured, semi-structured, and unstructured data from multiple heterogeneous information systems and pre-process them to construct a heterogeneous graph containing nodes and edges.

[0072] S2. Based on the constructed heterogeneous graph, the structural attribute information of each node is extracted to generate a node structural feature vector. The original description information of each node is semantically parsed using a large language model to generate a semantic representation vector. This is then embedded and concatenated with the node structural feature vector to form the initial node representation.

[0073] S3. Construct a heterogeneous graph Transformer network, input the initial node representation, set the node type mapping and edge type attention mechanism, perform multi-layer information propagation, calculate the attention weight and message matrix in each layer, perform low-rank decomposition on the attention weight and message matrix, introduce a rank-aware masking mechanism in the multi-layer information propagation process, combine the semantic representation vector with the edge connection strength in the heterogeneous graph, filter redundant paths, and output the final node representation;

[0074] S4. Based on the final node representation, generate field alignment relationships, interface mapping tables, and dependency paths, plan metadata collection paths, and determine the cross-system collection sequence and scope based on confidence levels, and determine metadata collection tasks.

[0075] S5. Execute metadata collection tasks, convert the raw metadata entries collected by each heterogeneous system into a unified structured format, and write them into the metadata center. Dynamically update the prompt template of the large language model and the rank parameter of the heterogeneous graph transformer to build a sustainable and optimized closed-loop process for heterogeneous metadata collection.

[0076] This paper proposes a cross-system, heterogeneous metadata intelligent collection method based on a large language model and a heterogeneous graph Transformer network, significantly improving the automated parsing and unified management capabilities of metadata in heterogeneous data environments. By constructing a heterogeneous graph consisting of nodes and edges, it effectively integrates complex information structures from structured, semi-structured, and unstructured data sources, enhancing the semantic and structural linkage between data. A large language model is introduced for node semantic parsing, combined with structural embedding to form the initial node representation, achieving a fusion of semantic and structural information. Multi-layer information propagation is performed through a heterogeneous graph Transformer network, and low-rank decomposition and rank-aware masking mechanisms are introduced to improve the model's information extraction efficiency and representation accuracy, while suppressing interference from redundant paths. Based on the final node representation, the present invention generates field alignment relationships, interface mapping tables, and dependency paths, forming a metadata collection task planning strategy with a confidence grading mechanism. Through the structured transformation of collection results and the dynamic updating of model parameters, closed-loop optimization of the collection process is achieved. This method boasts strong structural uniformity, high semantic adaptation accuracy, excellent automatic adjustment capabilities, and good scalability, making it suitable for intelligent data governance and metadata integration scenarios in multi-source system environments.

[0077] In this embodiment, the structured, semi-structured and unstructured data specifically include database table structures, log files, API documents, data directories and other source system metadata information.

[0078] In this embodiment, the preprocessing of the structured, semi-structured and unstructured data specifically includes format parsing, field extraction, annotation cleaning, key-value pair extraction, structure mapping and encoding normalization operations. Format parsing refers to identifying and parsing the content structure according to the data source type; field extraction refers to extracting field names, data values ​​and their attribute information from data records; annotation cleaning is used to remove irrelevant content, format noise or redundant annotations in field descriptions to make text descriptions more standardized; key-value pair extraction mainly identifies and structures the "attribute-value" information implicit in unstructured text; structure mapping is to establish an association between the extracted fields and their corresponding business entities or interface structures to form nodes and edges in the graph structure; encoding normalization operations are used to unify field naming rules, data type expressions and encoding formats to ensure the semantic and structural consistency and alignment of data between different systems.

[0079] In this embodiment, S2 specifically includes:

[0080] S21. For each node in the constructed heterogeneous graph, extract the associated original description information, where the original description information includes the field name, data type, data table or interface name, annotation content, and corresponding source system identifier, to form an original metadata text set;

[0081] S22. Organize the original metadata text set into a natural language input sequence according to a preset structured input template, input it into a large language model based on the BERT structure and fine-tuned by domain tasks, perform semantic understanding and context encoding operations, and obtain the semantic representation vector of the node. The execution of semantic understanding and context encoding operations specifically refers to inputting the original metadata text sequence organized according to the structured input template into a large language model based on BERT and fine-tuned by metadata tasks in a specific domain. The model first performs word segmentation and embedding on the input text, and encodes each field, data type, interface name, comment, etc. into a high-dimensional vector. Then, the model uses a multi-layer self-attention mechanism to deeply explore the semantic associations and contextual relationships between various parts of the text, and captures and aggregates the contextual dependencies of words, phrases, and even entire records. Finally, the large language model outputs a semantic representation vector that integrates the business meaning of the node itself and its contextual relationship;

[0082] S23. For each node, extract corresponding structural information features, including the encoding format of the field name, the data type encoding, the system type encoding, and the logical level or call depth of the node in the system, and construct a node structural feature vector;

[0083] S24, concatenating the semantic representation vector and the node structure feature vector to construct a combined vector;

[0084] S25. Use the combined vector as the input representation of the node in the heterogeneous graph Transformer network to form the initial representation of the node.

[0085] This paper proposes a method for generating initial node representations based on the fusion of a large language model and structural features for semantic understanding and structural feature modeling of nodes in heterogeneous information systems. This method significantly improves the semantic parsing accuracy and structural adaptability of metadata. By extracting multidimensional information such as field names, data types, interface names, and annotations to form a raw metadata text set, the method comprehensively captures the business context and contextual semantics of the node. A domain-tuned large language model based on the BERT architecture is introduced for natural language processing, enabling a deep understanding of domain terms and their contextual relationships, generating representation vectors with high semantic expressiveness. Structural information such as field encoding, system type, and logical hierarchy is integrated into a structural feature vector, which is then concatenated with the semantic representation vector to form a combined vector, achieving a deep fusion of node semantics and structural features. The resulting initial node representation not only possesses context-sensitive semantic recognition capabilities but also takes into account the differential expression of system hierarchical structure and logical position, providing a high-quality input foundation for multi-layer information propagation and precise modeling in heterogeneous graph Transformer networks. This method offers the advantages of accurate representation, strong structural awareness, and good transferability, providing key support for unified modeling and high-precision parsing of multi-source heterogeneous metadata.

[0086] In this embodiment, S3 specifically includes:

[0087] S31. Build a heterogeneous graph Transformer network and set the initial node representation as input;

[0088] S32. Define a node type mapping function to encode the types of all nodes in the heterogeneous graph into node type vectors as the information propagation condition of the heterogeneous graph Transformer network. The node type mapping function is a type encoding function constructed based on the system source, data mode, and entity category of the node.

[0089] S33. Construct an edge type attention mechanism, set an independent attention weight matrix for each type of edge, and perform differentiated attention calculation based on edge type attributes;

[0090] S34. Construct the heterogeneous graph Transformer network as a stacked network framework containing a multi-layer propagation structure, input the node representation of the previous layer output in each propagation layer, combine the node type vector and the edge type attention weight, perform weighted aggregation on the neighbor node representation of each node according to the adjacency relationship in the heterogeneous graph, generate a message mapping matrix corresponding to the edge type, and perform a node representation update operation. The above operation specifically refers to performing weighted aggregation on the neighbor node representation of each node according to the adjacency relationship in the heterogeneous graph and generating a message mapping matrix corresponding to the edge type in each propagation layer of the heterogeneous graph Transformer network. Based on the adjacency relationships of the current heterogeneous graph, all direct neighbor nodes of each target node are identified. For edges of the same or different types, pre-set or dynamically calculated edge type attention weights are applied to weight the representation vectors of these neighbor nodes. Each target node will weight the feature vectors of all its neighbor nodes according to the type, structure and edge properties of all its neighbor nodes, combined with the edge type attention weights, and finally form a message vector for this node and edge type. All such weighted results are collected and arranged according to edge type, forming the message mapping matrix of this propagation layer.

[0091] S35. In each propagation layer, perform a low-rank approximation operation on the attention weight matrix and the message mapping matrix generated by the current layer, concatenate the attention weight matrix and the message mapping matrix into a joint representation matrix, and use a low-rank decomposition operation to decompose them into the product of two low-dimensional dense matrices to generate a low-rank decomposition result matrix;

[0092] S36. Based on low-rank decomposition, a rank-aware gating mechanism is introduced to construct a regulation function that integrates rank contribution, attention significance and edge connection strength. Constructing gating factors , set the gating threshold function based on the gating factor, perform the gating screening operation, and construct the structural path mask matrix:

[0093] ;

[0094] in, is the matrix trace function, is the path attention score, is the edge connection strength, is the upper limit of the rank corresponding to the node type, To adjust the parameters, is the Sigmoid activation function, For the transmission path The corresponding left singular matrix representation in the low-rank decomposition;

[0095] Gating Factor Its core purpose is to dynamically adjust the retention and suppression of different paths during information propagation to improve the accuracy of path selection and the efficiency of network representation. The gating factor combines three key factors: the rank contribution of a path in the low-rank decomposition result, which is measured by calculating the strength of the path's structural projection in the singular matrix to measure its impact on the low-rank feature space; the attention weight corresponding to the path, which represents the attention and significance of the edge in the heterogeneous graph structure; and the edge connection strength, which reflects the structural closeness or frequency of interaction between two nodes. These three factors are multiplied by the adjustment weights, then weighted and summed, and normalized using a nonlinear activation function to generate the gating factor. The resulting gating factor reflects the comprehensive role of a path in semantic contribution, structural stability, and sparsity control, providing a precise basis for path screening and information fusion. This approach suppresses redundant propagation paths while maintaining effective information transfer, improving the model's robustness and expressiveness in complex heterogeneous graph data.

[0096] S37, based on gating factors , the path attention score of each propagation path in the heterogeneous graph Transformer network Perform weighted adjustment and replace the path attention score with the weighted result, retaining only the effective paths with significant gating factor values ​​and filtering the propagation information of low-correlation edges. The path attention score of each propagation path in the heterogeneous graph Transformer network is calculated. Specifically, weighted adjustment means that in each propagation layer, the system calculates the corresponding gating factor for each path. The gating factor reflects the comprehensive importance of the path in multiple dimensions such as structural contribution, attention significance, and edge connection strength. The original attention score of each path is multiplied and weighted by the gating factor of the path to obtain the weighted new attention score.

[0097] S38. According to the distribution of node types in the heterogeneous graph Transformer network, the node type is set by integrating semantic entropy, structural centrality within the type and difference between types. The rank upper bound of :

[0098] ;

[0099] in, Node type The information entropy of Node type The average degree centrality of Node type Jensen-Shannon divergence between the distribution of non-type nodes, is the regulating factor, is the global basic constant of rank, and represents the probability of embedding vector distribution, is the set of representation vectors of all nodes corresponding to node type t;

[0100] Node Type The rank upper bound of Its core function is to assign differentiated rank constraints to different node types during the low-rank approximation process, thereby improving the model's adaptability and representation accuracy for multi-type heterogeneous data. The rank upper bound comprehensively considers three structural and semantic features: First, the node type's information entropy measures the distribution uncertainty of that type of node in the semantic space. Higher information entropy indicates more complex semantics carried by that type of node and greater required representation capability. Second, the average degree centrality of the node type in the structural graph reflects the structural importance of that type of node in the system. Higher centrality indicates more extensive connections and a greater contribution to information dissemination. Third, the Jensen-Shannon divergence between the distribution of that node type and other types of nodes measures the differences in semantic or structural distribution between different types of nodes. Greater differences indicate greater independence of that type and the need for a higher-dimensional representation space to maintain its unique characteristics. These three features are weighted and superimposed on the global base rank to determine the rank upper bound for each node type. This is used to guide the construction of each type's subspace during the low-rank approximation operation, achieving a more reasonable and accurate rank assignment mechanism.

[0101] S39. According to the rank upper limit corresponding to each node type, a rank truncation operation is performed on the low-rank decomposition result matrix to retain the rank dimension subspace corresponding to each type. Specifically, the rank truncation operation is performed on the low-rank decomposition result matrix to retain only the first several important principal components or feature dimensions of the feature subspace corresponding to each type of node in the result matrix obtained by the low-rank decomposition according to the rank upper limit set in advance for each node type, and the remaining high-dimensional or low-contribution features are discarded or set to zero;

[0102] S310. Perform element-by-element multiplication on the low-rank decomposition result matrix after rank truncation and the structural path mask matrix to construct a node representation matrix with path selectivity and rank constraint characteristics, which is used as the output node representation of the current layer and input to the next propagation layer as the initial input of the node. The construction of the node representation matrix with path selectivity and rank constraint characteristics specifically refers to first performing a rank truncation operation on the low-rank decomposition result matrix obtained in the previous layer according to the node type in each propagation layer, and only retaining the valid feature subspace within the rank upper limit range of each type of node; then multiplying the rank-truncated matrix with the structural path mask matrix one by one according to the corresponding positions of the elements, retaining the path information marked as valid in the structural path mask, and setting the invalid or gated path information to zero. The final node representation matrix has two characteristics at the same time: on the one hand, it only contains high-contribution low-dimensional features retained after dynamic rank allocation, realizing adaptive control of the expression ability of different node types; on the other hand, it only retains the information on the effective propagation path screened by the gating mechanism in the current propagation layer, filtering out irrelevant or low-value structural paths;

[0103] S311. After all propagation layers are completed, the representation of the nodes in the final layer is output as the final representation of the nodes in the heterogeneous graph Transformer network.

[0104] This paper constructs a heterogeneous graph Transformer network that integrates node type awareness and edge type attention mechanisms, combining a multi-layer propagation structure with a low-rank approximation optimization strategy to efficiently model the complex semantic and structural relationships between nodes in heterogeneous information systems. During the construction phase, heterogeneous features are explicitly modeled through a node type mapping function and an edge type attention mechanism, enabling each type of node and edge to have differentiated representation capabilities during propagation. The introduction of a low-rank approximation decomposition operation effectively reduces the redundancy and resource overhead of propagation computation while preserving the key structural characteristics of the representation matrix. Combining a rank-aware gating mechanism with a structural path mask matrix, information propagation is focused on high-contribution paths, significantly improving the utilization of effective edges and filtering out noisy paths. Furthermore, a type-level dynamic rank control strategy adaptively sets the rank upper bound for each node type based on semantic entropy and structural difference information, constructing a streamlined and highly discriminative node representation subspace. The resulting output node representation combines type awareness, structural selectivity, and rank constraints, providing a high-quality semantic structure representation for subsequent tasks such as field alignment and path planning. It offers the benefits of efficient propagation, high modeling accuracy, and strong cross-system generalization.

[0105] In this embodiment, the S4 specifically includes:

[0106] S41. Receive the final representation of the node output by the heterogeneous graph Transformer network and set the feature matching function ,in and The final representation of the nodes representing the field nodes in the two heterogeneous systems is:

[0107] ;

[0108] in, is the weighting parameter, represents the semantic cosine similarity, Representation node and Structural dependency matrix in heterogeneous graph Transformer network, Represents a vector norm;

[0109] Feature matching function This feature matching function measures the degree of match between field nodes in two heterogeneous systems. It comprehensively considers semantic similarity and structural dependency information, constructing a field matching scoring function that incorporates both semantic and structural features. The feature matching function first calculates the semantic cosine similarity between the two node representations to reflect whether the two fields are expressed identically in the semantic space. The closer the semantics, the higher the cosine similarity, indicating more consistent meaning. Secondly, it considers the product of the norms of the two node representation vectors and their inner product after projection through the structural dependency matrix. This component reflects the propagation path, connectivity, and upstream and downstream dependencies of the two fields in the heterogeneous graph structure. A balancing factor controls the weighting of the two components of the overall function. This allows for greater reliance on structural information when the systems differ significantly or when semantics are ambiguous, while greater reliance on semantic similarity when the structures are identical but the semantics differ significantly. This function effectively identifies pairs of fields in heterogeneous systems that have the same meaning but different names, or different structures but equivalent functionality.

[0110] S42, similarity score calculated based on feature matching function, using set threshold Filter candidate field node pairs to generate a field alignment relationship set , the threshold value is set Screening candidate field node pairs specifically means that after calculating the feature matching similarity scores of all candidate field node pairs, a judgment threshold is pre-set, and all field node pairs with similarity scores higher than the threshold are retained, and it is considered that they can achieve effective alignment in terms of semantics and structure; while node pairs with similarity scores lower than the threshold are eliminated or do not participate in the alignment. This screening operation ensures that only field node pairs with high semantic consistency and structural relevance are included in the field alignment relationship set, effectively excluding invalid field pairs with semantic inconsistency, structural incompatibility or unclear mapping relationships, providing a high-quality and accurate field alignment foundation for interface mapping, path planning and metadata collection.

[0111] S43. Combine the field alignment relationship set with the node connection relationship in the heterogeneous graph Transformer network to construct the interface candidate pair set and introduce the interface mapping function , map the field alignment relationship to the parameter docking relationship between system interfaces to form an interface mapping table:

[0112] ;

[0113] in, and They are the source system and target system interfaces respectively. and is the weight coefficient, represents the semantic similarity of nodes, Indicates the interface structure similarity, Indicates when the function reaches its maximum value right;

[0114] Interface mapping function This method is used to construct mapping relationships between system interfaces. Its core concept is to select, from all possible combinations of source and target interfaces, the pair of interfaces with the best semantic and structural match as the bridge for field connection. It combines two pieces of information by weighted fusion: the semantic similarity between field nodes, which measures whether two fields have consistent business meanings; and the structural similarity between interfaces, which reflects the consistency between the two interfaces in parameter form, call path, or data organization. By weighted summing these two similarities, it comprehensively assesses the suitability of a given field pair for mapping to a specific set of interface pairs. Finally, a maximum function is used to select the interface pair with the highest matching score among all possible pairs as the interface connection relationship for the field between the two systems. This method significantly improves the efficiency and accuracy of automatic interface matching between heterogeneous systems, and is particularly applicable to scenarios with inconsistent interface naming, different parameter order, or complex interface nesting. It provides a high-confidence, structurally consistent interface connection foundation for generating interface mapping tables, planning dependency paths, and executing metadata collection tasks.

[0115] S44. Based on the interface mapping table and the heterogeneous graph Transformer network, a path extraction algorithm is executed to generate a field collection path between field nodes, and a sequence of path triples containing fields, interfaces, and call relationships is constructed to form a dependency path graph. The execution of the path extraction algorithm based on the interface mapping table and the heterogeneous graph Transformer network specifically refers to using the final node representation and structured interface mapping table output by the heterogeneous graph Transformer network to embed semantically aligned and structurally accessible field nodes, interface nodes, and their associations in different systems into the same graph model. The system automatically performs a heuristic search based on the interface mapping relationship and the connection pattern, edge weight, and attention distribution of the nodes in the heterogeneous graph. It systematically starts from the source field node and searches for a data transmission path that can achieve cross-system collection along the interface mapping and the actual calls or dependencies in the graph. The heuristic search not only focuses on the semantic and structural connectivity between field nodes, but also dynamically considers factors such as interface call order, path dependency depth, and node type constraints.

[0116] S45. Comprehensively evaluate each field collection path based on the field matching degree, interface mapping consistency, and structural path stability corresponding to each path in the dependency path graph, and grade them according to the overall credibility, and output a preliminary field collection grade score. The comprehensive evaluation of each field collection path specifically refers to the system examining multiple key indicators such as the semantic matching degree of the field pairs involved in the path, the consistency of the interface mapping, and the stability of the path structure for each field collection path in the dependency path graph. After quantitatively scoring each indicator, a comprehensive score for each collection path is formed through weighted aggregation or fusion strategy. The field matching degree is mainly calculated based on the semantic similarity between node representations, the interface mapping consistency is evaluated based on the matching results in the interface mapping function and the interface mapping table, and the structural path stability is comprehensively determined based on factors such as path length, path branch complexity, and historical collection success rate. After all scores are aggregated, the system divides each field collection path into different levels of high, medium, and low according to the preset credibility grading standard, and outputs them as the basis for the collection task priority and scheduling order;

[0117] S46. Set the confidence level threshold based on the initial score of the field collection level , divide the field collection paths into three levels: high, medium, and low, and formulate corresponding cross-system collection priorities and path coverage, and generate an initial list of field collection tasks;

[0118] S47. Introduce a structural semantic consensus mechanism to perform consensus aggregation operations on redundant field nodes appearing in the initial list of field collection tasks, integrate the path context features of the field nodes in the path triple sequence, the mapping deviation between the interface mapping function output and the interface mapping table, and reconstruct the field alignment relationship. The consensus aggregation operation on redundant field nodes appearing in the initial list of field collection tasks specifically refers to merging, clustering or deduplicating field nodes with similar names, similar semantics or overlapping business functions in multiple collection paths or field alignment results based on their context information in the path triple sequence, the semantic and structural features output by the interface mapping function, and the mapping deviation in the interface mapping table. During the consensus aggregation process, the system automatically calculates the path correlation, interface consistency and structural position of similar field nodes, determines whether they actually have the same business meaning or collection target under different systems and different paths, and unifies their field alignment relationships and merges them into a single collection target, eliminating the risk of repeated collection or conflicting mapping;

[0119] S48. Based on the reconstructed field alignment relationship, interface mapping table and dependency path map, the metadata collection task is finally determined, and a multi-dimensional collection instruction set including field pairs, interface mapping pairs, path triplets and confidence levels is generated.

[0120] This paper achieves precise planning and task list generation for cross-system metadata collection tasks by constructing field alignment relationships, interface mapping tables, and dependency path graphs based on the final node representation. During feature alignment, a feature matching function is established by combining semantic similarity and structural dependency information. This ensures that the alignment relationships between field nodes not only maintain semantic consistency but also accurately transfer the structural logic between the original systems. The interface mapping function design integrates both semantic and structural similarity, effectively extending the field alignment results to interface-level mapping relationships and constructing a stable interface mapping table. Furthermore, a path extraction algorithm is introduced to generate path triplets consisting of fields, interfaces, and call relationships, forming a traceable and computable dependency path graph. Through a comprehensive assessment of path matching, interface consistency, and structural stability, collection paths are assigned differentiated trust levels, providing support for hierarchical task scheduling. Furthermore, a structural semantic consensus mechanism is incorporated to aggregate and correct redundancies and ambiguities in field alignment and interface mapping, enhancing the accuracy and stability of collection tasks. Ultimately, a multidimensional collection instruction set is formed, enabling the efficient generation and organization of cross-system, multi-type field-level metadata collection tasks, with significant advantages such as high recognition accuracy, rational task planning, and stable structural matching.

[0121] In this embodiment, the S5 specifically includes:

[0122] S51. Based on the determined metadata collection task, connect to multiple heterogeneous information systems in sequence, execute the collection operation indicated by the field pair, interface mapping pair, and path triplet, and obtain the original metadata entry;

[0123] S52. Load the corresponding metadata collection plug-in based on the system type of each heterogeneous information system, perform structural analysis on the original metadata entries, and extract field content, interface parameters, and path context information. Specifically, the structural analysis of the original metadata entries refers to using the corresponding collection plug-in or adapter based on the system type of the heterogeneous information system to parse and process the collected original metadata entries according to a preset data schema, automatically extracting and separating the field names, field values, data types, table or interface names, hierarchical relationships, interface parameters, field comments, primary and foreign key information, and restoring the structured hierarchy and contextual dependencies in the source system.

[0124] S53, converting the collected original metadata entries into a structured format according to the set field mapping rules, semantic tag templates and system structure specifications, and constructing metadata structure units with consistent system sources and complete semantics;

[0125] S54, writing the metadata structure unit converted into a structured format into the metadata center, and automatically triggering data consistency verification, redundancy elimination, and update marking;

[0126] S55. Dynamically update the prompt template of the large language model based on field semantic deviations, interface execution anomalies, and path coverage changes during the collection process;

[0127] S56. Based on the information propagation effect of each layer of the heterogeneous graph Transformer, the change in attention distribution and the feedback of low-rank decomposition performance during the collection process, the rank parameter of the heterogeneous graph Transformer is updated to build a sustainable and optimized closed-loop process for heterogeneous metadata collection.

[0128] By developing a detailed metadata collection process, this paper establishes an automated metadata acquisition and structured processing mechanism for multiple heterogeneous information systems. This method first connects heterogeneous systems one by one based on pre-defined field pairs, interface mapping pairs, and path triplets, executing precise collection operations, thereby improving the accuracy and efficiency of cross-system collection tasks. The system automatically loads corresponding collection plug-ins based on the type characteristics of each source system, performs structural analysis on the acquired raw metadata entries, and accurately extracts field content, interface parameters, and path context information. The raw entries are converted into a structured format using unified field mapping rules, semantic tag templates, and system structure specifications, generating metadata units with unified semantics and consistent structure, which are then written to the metadata center. This process also simultaneously triggers data consistency checks, redundancy elimination, and update marking, ensuring the accuracy and timeliness of the central data. During the collection process, the system dynamically monitors field semantic deviations, interface anomalies, and path coverage changes to continuously adjust and optimize the prompt templates of the large language model, enhancing semantic parsing capabilities. The system dynamically updates the rank parameters of the heterogeneous graph Transformer network based on the propagation effects, attention changes, and low-rank decomposition performance feedback during the collection process, ensuring that the model structure adapts to the dynamic characteristics of actual data distribution. The overall process forms a continuously optimized closed-loop metadata collection system, significantly improving the consistency of semantic understanding and metadata integration efficiency across heterogeneous systems, and possessing excellent scalability and intelligence.

[0129] refer to Figure 2 The LLM-based cross-system heterogeneous metadata intelligent collection system includes the following modules:

[0130] Data acquisition and preprocessing module, used to collect structured, semi-structured and unstructured data, complete preprocessing and construct heterogeneous graphs;

[0131] The semantic parsing module is used to call the large language model to parse the node description information, generate the semantic representation vector, and embed it with the structural feature to form the initial node representation;

[0132] The graph modeling module is used to build a heterogeneous graph Transformer network, input the initial node representation, perform multi-layer information propagation and low-rank decomposition, and output the final node representation;

[0133] The path planning module is used to generate field alignment relationships, interface mapping tables, and dependency path graphs based on the final node representation, plan the collection path, and determine the collection tasks;

[0134] The task execution module is used to execute the collection task, load the collection plug-in to obtain the original metadata entries, convert them into structured metadata, generate standardized collection results and write them into the metadata center;

[0135] An intelligent optimization module is used to dynamically update the large language model prompt template based on the standardized collection results, and optimize the rank parameter of the heterogeneous graph Transformer network based on propagation feedback.

[0136] Example 1:

[0137] To verify the feasibility of the present invention in practice, it was applied to a city information center. The information center was faced with the task of integrating metadata between a large number of government subsystems, including a population management system, a social security business system, a tax information system, and other heterogeneous data sources. These systems were built by different manufacturers, with inconsistent data standards, large differences in field definitions, and poor interface compatibility, posing a huge challenge to information integration and data governance. In particular, when developing a unified city-wide public service data directory, problems arose, such as difficult field alignment, complex interface dependency paths, low metadata collection efficiency, and unstable collection content.

[0138] To address these issues, the platform deployed the LLM-based, cross-system, heterogeneous metadata intelligent collection method and system proposed in this paper. The project was deployed in March 2025 in the information center's data center, and lasted two weeks for testing and collection optimization.

[0139] During the system initialization phase, the platform first collected various types of data, including data structure information, interface definition documents, system logs, and database annotations from 11 business systems. This data included structured field definition tables, semi-structured Swagger interface documents, and unstructured documentation and API logs, totaling 2.8GB. The system automatically constructed this information into a heterogeneous graph structure consisting of 11,743 nodes and 45,926 edges. The system then used a BERT model fine-tuned with government data to perform semantic parsing, generating high-quality initial node representations.

[0140] By constructing an improved heterogeneous graph Transformer network and introducing a node type-specific multi-channel propagation mechanism and a low-rank decomposition structure compression mechanism, the platform efficiently models the field node representations in each business system, generates high-confidence field alignment relationships and interface mapping tables, and constructs a dependency path graph with a total of 1,125 path triplets.

[0141] The system uses a field semantic cosine similarity of no less than 0.85 and an interface structure similarity of no less than 0.75 as screening thresholds. This improved the field alignment match rate to 93.6%, the stability of the interface mapping table by 22.4%, and the overall field collection confidence level to an average of 0.91. Ultimately, 326 valid field pairs and 87 interface mapping pairs were identified. The collection path length was kept between 3 and 7, and the collection of priority fields accounted for 62.8%.

[0142] After completing the collection task, the platform writes the structured metadata entries to the government metadata center and automatically completes data consistency verification and redundancy elimination. The system also dynamically adjusts the large language model prompt template and Transformer network rank parameters based on feedback from the collection. After three rounds of optimization, the accuracy of node semantic representation increased by approximately 6.1%, and the stability of path selection increased by 11.7%.

[0143] Table 1 Evaluation table of field alignment and collection effects of heterogeneous systems

[0144]

[0145] Table 1 shows the evaluation results of ten sets of field mappings and collections from the three major government subsystems of population, social security, and taxation to the data aggregation platform. Overall, the semantic matching degree remained above 0.87, with core primary key and time attributes such as ID number and date of birth reaching 0.93–0.95. This indicates that after semantic parsing and structural alignment using a large language model, key information has achieved highly consistent business meaning across systems. Interface mapping confidence levels generally ranged from 0.81–0.92, indicating that differences in interface structure have been effectively converged through the mapping function. In particular, the population system fields achieved the highest consistency at the interface level, demonstrating optimal compatibility between the population system and the aggregation platform interface specifications. Path stability scores were generally above 0.83, with significantly higher path stability for high-priority fields. This demonstrates that, after constructing a dependency path graph and filtering with low-rank gating, data flows for core fields are collected via stable pathways, reducing the risk of task failures or redundant calls.

[0146] From the perspective of priority distribution, high-level fields account for 60%, focusing on high-value information such as identity, household registration, time and amount; mid-level fields are mainly business status or type fields, which still retain a high degree of matching and stability, but have slight differences in interface definition or path depth, and need to pay attention to dynamic updates in subsequent collection tasks. Combining the three indicators, it can be seen that semantic matching is positively correlated with interface confidence, and high semantic similarity is often accompanied by higher interface compatibility; while path stability is affected by system topology and call chain complexity, and together with interface confidence, it determines the collection priority. This batch of field mapping results verifies the effectiveness of the method of the present invention in achieving high-precision field alignment, interface mapping and path planning in heterogeneous system scenarios.

[0147] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. The cross-system heterogeneous metadata intelligent collection method based on LLM is characterized by: The steps include: S1. Collect structured, semi-structured, and unstructured data from multiple heterogeneous information systems and pre-process them to construct a heterogeneous graph containing nodes and edges. S2. Based on the constructed heterogeneous graph, the structural attribute information of each node is extracted to generate a node structural feature vector. The original description information of each node is semantically parsed using a large language model to generate a semantic representation vector. This is then embedded and concatenated with the node structural feature vector to form the initial node representation. S3. Construct a heterogeneous graph Transformer network, input the initial node representation, set the node type mapping and edge type attention mechanism, perform multi-layer information propagation, calculate the attention weight and message matrix in each layer, perform low-rank decomposition on the attention weight and message matrix, introduce a rank-aware masking mechanism in the multi-layer information propagation process, combine the semantic representation vector with the edge connection strength in the heterogeneous graph, filter redundant paths, and output the final node representation; S4. Based on the final node representation, generate field alignment relationships, interface mapping tables, and dependency paths, plan metadata collection paths, and determine the cross-system collection sequence and scope based on confidence levels, and determine metadata collection tasks. S5. Perform metadata collection tasks, convert the raw metadata items collected by various heterogeneous systems into a unified structured format, write them to the metadata center, dynamically update the prompt templates of the large language model and the rank parameters of the heterogeneous graph transformer, and build a sustainable and optimized closed-loop process for heterogeneous metadata collection; The structured, semi-structured and unstructured data specifically include database table structures, log files, API documents, data directories and other source system metadata information.

2. The LLM-based cross-system heterogeneous metadata intelligent collection method according to claim 1 is characterized in that: The preprocessing of structured, semi-structured and unstructured data specifically includes format parsing, field extraction, annotation cleaning, key-value pair extraction, structure mapping and encoding normalization operations.

3. The LLM-based cross-system heterogeneous metadata intelligent collection method according to claim 1 is characterized in that: The S2 specifically includes: S21. For each node in the constructed heterogeneous graph, extract the associated original description information, where the original description information includes the field name, data type, data table or interface name, annotation content, and corresponding source system identifier, to form an original metadata text set; S22. Organize the original metadata text collection into a natural language input sequence according to a preset structured input template, input it into a large language model based on the BERT structure and fine-tuned by the domain task, perform semantic understanding and context encoding operations, and obtain the semantic representation vector of the node; S23. For each node, extract corresponding structural information features, including the encoding format of the field name, the data type encoding, the system type encoding, and the logical level or call depth of the node in the system, and construct a node structural feature vector; S24, concatenating the semantic representation vector and the node structure feature vector to construct a combined vector; S25. Use the combined vector as the input representation of the node in the heterogeneous graph Transformer network to form the initial representation of the node.

4. The LLM-based cross-system heterogeneous metadata intelligent collection method according to claim 1 is characterized in that: The S3 specifically includes: S31. Build a heterogeneous graph Transformer network and set the initial node representation as input; S32. Define a node type mapping function to encode the types of all nodes in the heterogeneous graph into node type vectors as the information propagation condition of the heterogeneous graph Transformer network. The node type mapping function is a type encoding function constructed based on the system source, data mode, and entity category of the node. S33. Construct an edge type attention mechanism, set an independent attention weight matrix for each type of edge, and perform differentiated attention calculation based on edge type attributes; S34. Construct the heterogeneous graph Transformer network as a stacked network framework with multiple layers of propagation structure. In each propagation layer, the node representation output by the previous layer is input. The node type vector and the edge type attention weight are combined. Based on the adjacency relationship in the heterogeneous graph, the neighbor node representations of each node are weighted aggregated to generate a message mapping matrix corresponding to the edge type, and the node representation update operation is performed. S35. In each propagation layer, perform a low-rank approximation operation on the attention weight matrix and the message mapping matrix generated by the current layer, concatenate the attention weight matrix and the message mapping matrix into a joint representation matrix, and use a low-rank decomposition operation to decompose them into the product of two low-dimensional dense matrices to generate a low-rank decomposition result matrix; S36. Based on low-rank decomposition, a rank-aware gating mechanism is introduced to construct a regulation function that integrates rank contribution, attention significance and edge connection strength. Constructing gating factors , set the gating threshold function based on the gating factor, perform the gating screening operation, and construct the structural path mask matrix; S37, based on gating factors , the path attention score of each propagation path in the heterogeneous graph Transformer network Perform weighted adjustments and use the weighted results to replace the path attention scores, retaining only valid paths with significant gating factor values ​​and filtering out the propagation information of low-correlation edges; S38. According to the distribution of node types in the heterogeneous graph Transformer network, the node type is set by integrating semantic entropy, structural centrality within the type and difference between types. The rank upper bound of ; S39. Perform a rank truncation operation on the low-rank decomposition result matrix according to the rank upper limit corresponding to each node type, and retain the rank dimension subspace corresponding to each type; S310: Perform element-by-element multiplication of the low-rank decomposition result matrix after rank truncation and the structural path mask matrix to construct a node representation matrix with path selectivity and rank constraint characteristics, which is used as the output node representation of the current layer and input to the next propagation layer as the node initial input; S311. After all propagation layers are completed, the representation of the nodes in the final layer is output as the final representation of the nodes in the heterogeneous graph Transformer network.

5. The LLM-based cross-system heterogeneous metadata intelligent collection method according to claim 1 is characterized in that: The S4 specifically includes: S41. Receive the final representation of the node output by the heterogeneous graph Transformer network and set the feature matching function ,in and The final representation of the nodes representing the field nodes in the two heterogeneous systems; S42, similarity score calculated based on feature matching function, using set threshold Filter candidate field node pairs to generate a field alignment relationship set ; S43. Combine the field alignment relationship set with the node connection relationship in the heterogeneous graph Transformer network to construct the interface candidate pair set and introduce the interface mapping function ,map the field alignment relationship to the parameter docking relationship between system interfaces to form an interface mapping table; S44. Based on the interface mapping table and the heterogeneous graph Transformer network, execute the path extraction algorithm to generate the field collection path between field nodes, construct a path triple sequence containing fields, interfaces, and call relationships, and form a dependency path graph; S45. Comprehensively evaluate each field collection path based on the field matching degree, interface mapping consistency, and structural path stability of each path in the dependency path graph, and grade them based on the overall credibility, outputting a preliminary field collection grade score. S46. Set the confidence level threshold based on the initial score of the field collection level , divide the field collection paths into three levels: high, medium, and low, and formulate corresponding cross-system collection priorities and path coverage, and generate an initial list of field collection tasks; S47. Introduce a structural semantic consensus mechanism to perform consensus aggregation operations on redundant field nodes that appear in the initial list of field collection tasks, integrate the path context features of the field nodes in the path triple sequence, and the mapping deviation between the interface mapping function output and the interface mapping table to reconstruct the field alignment relationship; S48. Based on the reconstructed field alignment relationship, interface mapping table and dependency path map, the metadata collection task is finally determined, and a multi-dimensional collection instruction set including field pairs, interface mapping pairs, path triplets and confidence levels is generated.

6. The LLM-based cross-system heterogeneous metadata intelligent collection method according to claim 1 is characterized in that: The S5 specifically includes: S51. Based on the determined metadata collection task, connect to multiple heterogeneous information systems in sequence, execute the collection operation indicated by the field pair, interface mapping pair, and path triplet, and obtain the original metadata entry; S52. Load the corresponding metadata collection plug-in according to the system type of each heterogeneous information system, perform structural analysis on the original metadata entry, and extract field content, interface parameters, and path context information; S53, converting the collected original metadata entries into a structured format according to the set field mapping rules, semantic tag templates and system structure specifications, and constructing metadata structure units with consistent system sources and complete semantics; S54, writing the metadata structure unit converted into a structured format into the metadata center, and automatically triggering data consistency verification, redundancy elimination, and update marking; S55. Dynamically update the prompt template of the large language model based on field semantic deviations, interface execution anomalies, and path coverage changes during the collection process; S56. Based on the information propagation effect of each layer of the heterogeneous graph Transformer, the change in attention distribution and the feedback of low-rank decomposition performance during the collection process, the rank parameter of the heterogeneous graph Transformer is updated to build a sustainable and optimized closed-loop process for heterogeneous metadata collection.

7. The LLM-based cross-system heterogeneous metadata intelligent collection system is applied to the LLM-based cross-system heterogeneous metadata intelligent collection method according to any one of claims 1 to 6, characterized in that: Includes the following modules: Data acquisition and preprocessing module, used to collect structured, semi-structured and unstructured data, complete preprocessing and construct heterogeneous graphs; The semantic parsing module is used to call the large language model to parse the node description information, generate the semantic representation vector, and embed it with the structural feature to form the initial node representation; The graph modeling module is used to build a heterogeneous graph Transformer network, input the initial node representation, perform multi-layer information propagation and low-rank decomposition, and output the final node representation; The path planning module is used to generate field alignment relationships, interface mapping tables, and dependency path graphs based on the final node representation, plan the collection path, and determine the collection tasks; The task execution module is used to execute the collection task, load the collection plug-in to obtain the original metadata entries, convert them into structured metadata, generate standardized collection results and write them into the metadata center; An intelligent optimization module is used to dynamically update the large language model prompt template based on the standardized collection results, and optimize the rank parameter of the heterogeneous graph Transformer network based on propagation feedback.

Images