A Metamodel-Driven Two-Way Mapping Method and System for Financial Data Standards
By using a meta-model-driven approach to structurally decompose financial data and construct a dual-network topology, the problems of low mapping accuracy and consistency failure in the standard mapping of financial data are solved. This enables efficient and dynamic maintenance and adaptive updates of mapping relationships, thereby improving the stability and adaptability of the system.
Patent Information
- Application Number
- CN202611156358.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-31
- Publication Date
- 2026-11-17
AI Technical Summary
Financial data standard mapping suffers from low mapping accuracy, low efficiency, and difficulty in scaling up. Existing mapping schemes cannot dynamically adapt to updates in financial standards and adjustments in business systems, leading to a gradual loss of standard consistency.
A meta-model-driven approach is adopted to structurally decompose multi-source heterogeneous financial data, construct a dual-network heterogeneous topology of forward and reverse mapping networks, and realize the dynamic maintenance and adaptive updating of mapping relationships through dynamic bidirectional consistency analysis and adaptive reconstruction operation.
It improves mapping accuracy and semantic ambiguity handling capabilities, ensures consistent maintenance and adaptive evolution of mapping relationships, guarantees system stability and engineering availability, and adapts to the actual needs of financial data governance scenarios.
Smart Images

Figure CN123115107A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of financial data processing, and more specifically, to a method and system for bidirectional mapping of financial data standards based on a meta-model driven approach. Background Technology
[0002] As the digital transformation of the financial industry deepens, financial institutions have accumulated a wide range of business data in diverse formats. Problems such as inconsistent data standards and definitions across different business systems are becoming increasingly prominent. To meet the needs of regulatory reporting, data sharing, and business collaboration, mapping heterogeneous business data to a unified financial data standard has become a crucial aspect of data governance.
[0003] Currently, financial data standard mapping mainly relies on two methods: First, manual benchmarking, where data governance personnel manually establish the correspondence between business fields and standard data elements based on business experience. While this method can achieve high accuracy in certain areas, it is inefficient and difficult to scale. Second, automated mapping methods based on keyword matching or similarity calculations. Although these methods improve mapping efficiency, their accuracy often fails to meet actual business needs due to differences in field naming conventions and semantic ambiguities. Furthermore, existing mapping schemes generally use static mapping tables or one-way mapping relationships. Once financial standards are updated or business systems are adjusted, existing mapping relationships are difficult to automatically adapt and maintain synchronously, causing standard consistency to gradually become invalid over time.
[0004] Therefore, how to achieve dynamic maintenance and adaptive updating of mapping relationships while ensuring mapping accuracy is a technical challenge that urgently needs to be solved in the field of financial data standard mapping. Summary of the Invention
[0005] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, the present invention provides a meta-model-driven bidirectional mapping method for financial data standards, the method comprising:
[0006] Multi-source heterogeneous financial data is decomposed into a structured dataset at the meta-model dimension. Based on the decomposition results, standard intra-data elements within the financial standard and standard-business data elements between the financial standard and the business system data dictionary are pre-established to form a structured dataset with related relationships.
[0007] Based on the structured dataset, a dual-network heterogeneous topology containing a forward mapping network and a reverse mapping network is constructed. Forward mapping edges and corresponding forward mapping weights are configured for the forward mapping network, and reverse mapping edges and corresponding reverse mapping weights are configured for the reverse mapping network. At the same time, a consistency bridge edge is established between the forward mapping network and the reverse mapping network to apply consistency tolerance constraints to the forward mapping weights and the reverse mapping weights.
[0008] Dynamic bidirectional consistency analysis is performed on the heterogeneous dual-network topology. The dynamic bidirectional consistency analysis includes: calculating structural consistency components in the structural dimension of the heterogeneous dual-network topology, calculating semantic consistency components in the semantic dimension, and tracking the temporal change trajectory of bidirectional consistency indices over a continuous time window; calculating the bidirectional consistency index of each mapping edge based on the structural consistency components, the semantic consistency components, and the temporal change trajectory, and detecting mapping drift and bidirectional conflicts based on the temporal change trend of the bidirectional consistency indices to identify precursors of critical mismatch.
[0009] Based on the analysis results of the dynamic bidirectional consistency analysis, an adaptive reconstruction operation is performed on the heterogeneous topology of the dual network. The adaptive reconstruction operation includes pruning failed mapping edges, growing mapping edges for newly emerging mapping requirements, reconnecting mapping edges with bidirectional conflicts, and implementing protection strategies for core mapping edges that meet preset protection conditions.
[0010] Multidimensional feedback data is collected throughout the entire mapping process, and based on this multidimensional feedback data, the disassembly rules of the structured decomposition, the structural parameters of the heterogeneous dual-network topology, the detection threshold of the dynamic bidirectional consistency analysis, and the triggering conditions of the adaptive reconstruction operation are adjusted in a closed loop to achieve continuous system evolution.
[0011] Furthermore, the present invention also provides a meta-model-driven bidirectional mapping system for financial data standards, which is used to store and execute the above-described methods.
[0012] Based on the above aspects, the present invention has at least the following beneficial effects compared to the prior art:
[0013] (i) Significant improvement in mapping accuracy and semantic ambiguity handling capabilities
[0014] This invention extracts naming semantic primitives, type semantic primitives, context semantic primitives, and relational semantic primitives from data elements and fuses them into a joint semantic representation. This comprehensively represents the semantic information of data elements from multiple orthogonal dimensions, effectively overcoming mapping errors caused by non-standard field naming or homonyms in traditional keyword matching methods. When the similarity of the joint semantic representation falls into the fuzzy range, a joint disambiguation mechanism performs cross-validation across the dimensions of each semantic primitive, significantly improving the accuracy of semantic relationship determination. Simultaneously, by tracking the semantic drift of data elements in different business scenarios through context anchor vectors, the mapping relationship can dynamically adapt to changes in business semantics, further ensuring mapping accuracy.
[0015] (ii) Consistency maintenance and adaptive evolution capability of bidirectional mapping relationship
[0016] This invention achieves quantitative constraints and automatic calibration of bidirectional mapping relationships by constructing forward and reverse mapping networks and establishing consistency bridging edges. When the forward and reverse mapping weights deviate from the consistency tolerance threshold, the system automatically performs a bidirectional consistency calibration operation to ensure that the mapping relationship remains bidirectionally symmetrical. Simultaneously, through a feedback-driven "adjustment-verification-calibration" closed-loop strategy, the system can continuously optimize mapping weights based on explicit feedback (manual review results) and implicit feedback (actual usage frequency and adoption rate), and dynamically adjust the tolerance threshold in conjunction with the overall distribution of consistency indicators. This enables the system to achieve an adaptive balance between feedback-driven optimization and consistency constraint maintenance, effectively solving the technical challenge that existing static mapping schemes cannot continuously evolve with the version of financial standards and business changes.
[0017] (III) System stability and engineering availability
[0018] This invention employs a multi-level protection mechanism to implement differentiated protection for core mapping edges, ensuring that critical mapping relationships, such as those for regulatory reporting, are not mistakenly deleted by automatic pruning operations. An event queue mechanism decouples attribute change-triggered re-verification from the main loop, preventing arbitrary interruptions to the main process by event-driven operations and ensuring system stability and control flow clarity. A cold-start strategy addresses the issue of an empty mapping network during initial system deployment, giving the solution complete engineering usability. These mechanisms collectively ensure the stable and reliable operation of this invention in real-world financial data governance scenarios, demonstrating promising prospects for industrial application. Attached Figure Description
[0019] Figure 1 This is a flowchart illustrating the method provided in an embodiment of the present invention;
[0020] Figure 2 This is a flowchart illustrating the dual-network topology construction and consistency calibration process in this invention;
[0021] Figure 3 This is a module architecture diagram of the system of the present invention. Detailed Implementation
[0022] like Figures 1-3 As shown.
[0023] This technical solution provides a meta-model-driven bidirectional mapping method and system for financial data standards. The technical details are explained in detail below using a specific application scenario.
[0024] I. Application Scenario Description
[0025] This invention can be applied to data standardization and governance scenarios in various financial institutions. The following is an example of a specific implementation scenario: A national commercial bank plans to map data from multiple business systems, including its core business system, credit management system, and anti-money laundering system, to the ISO 20022 financial data standard to meet regulatory reporting requirements and internal data sharing needs.
[0026] The bank's core business system stores customer information, account information, transaction records, and other data. Its data dictionary contains numerous business fields, such as "cust_nm" (customer name), "id_num" (ID number), and "acct_bal" (account balance) in the customer information table, and "txn_dt" (transaction date), "txn_amt" (transaction amount), and "txn_cd" (transaction code) in the transaction record table. The naming conventions for these fields are inconsistent; some fields use English abbreviations, while others use Chinese pinyin abbreviations, and the same business meaning may have different field names in different systems. The ISO 20022 standard library defines standardized data elements, such as "CustomerName," "IdentificationNumber," "Amount," and "Date."
[0027] In traditional manual mapping methods, data governance personnel need to compare business fields with standard data elements one by one to establish a mapping relationship table. This bank has over 2,000 data tables and tens of thousands of business fields, making manual mapping extremely labor-intensive. Furthermore, when the ISO 20022 standard is upgraded from the 2019 version to the 2022 version, all existing mapping relationships need to be re-verified. The technical solution of this invention is intended to solve the above problems.
[0028] II. Structured Decomposition and Pre-construction of Related Elements
[0029] This method first performs a structured decomposition of multi-source heterogeneous financial data. Taking the customer information table of the aforementioned core business system as an example, the system reads the table's metadata information, including field names, data types, field lengths, and constraints. For the "cust_nm" field, the system extracts four semantic primitives.
[0030] The extraction process of the naming semantic primitives is as follows: The system performs lexical analysis on the field name "cust_nm", identifying that "cust" is an abbreviation for "customer" and "nm" is an abbreviation for "name", thus restoring "cust_nm" to the core business term "customer name". The type semantic primitive generates a structural feature vector based on the information that the field's data type is character and its length is 50 characters. The context semantic primitive generates a business scenario label based on the information that the field belongs to the "customer information table", belongs to the "core business system", and is used to identify customer identity in the business process. The relational semantic primitive extracts a potential relationship graph based on the fact that the field has a co-occurrence relationship with the "id_num" field in the database schema and a dependency relationship with the "acct_bal" field through the customer ID.
[0031] The system inputs the four primitives mentioned above into four parallel semantic encoders to generate four independent semantic fingerprint vectors, and then concatenates these four vectors into a joint semantic representation. This joint semantic representation is a multi-dimensional comprehensive representation of the semantic information of the data element.
[0032] The structured dataset is stored as a graph. Each data element in the dataset is a node, and node attributes include the data element's name, data type, system, semantic primitive vector, and joint semantic representation. Candidate associations between data elements are represented as labeled edges, each carrying a comprehensive association confidence value. This graph-structured dataset serves as the input data source for subsequent topology construction.
[0033] When the system processes the "cust_nm" field and the "CustomerName" data element in the ISO 20022 standard library, it calculates the cosine similarity of their joint semantic representations. If this similarity falls within a preset fuzzy range, such as between 0.55 and 0.85, the system triggers a joint disambiguation mechanism. The upper and lower bounds of this fuzzy range can be configured according to the strictness of semantic judgment in the business scenario. In this embodiment, the lower bound of 0.55 indicates that the two data elements are basically unrelated semantically when the value is below this, and no disambiguation needs to be triggered; the upper bound of 0.85 indicates that the two data elements are highly consistent semantically when the value is above this, and can be directly judged as a match; the fuzzy range between the two requires comprehensive judgment using the joint disambiguation mechanism. This mechanism calculates the pairwise similarity of the two data elements in the naming semantic primitive dimension, type semantic primitive dimension, context semantic primitive dimension, and relational semantic primitive dimension, respectively, to obtain a four-dimensional similarity vector. This vector is then input into the disambiguation classifier, which comprehensively judges the semantic relationship between the two. In this example, the disambiguation classifier outputs a "synonym" determination result, confirming that "cust_nm" and "CustomerName" have the same semantic meaning.
[0034] The system maintains a context anchor vector for each data element. This vector records the business context distribution of the data element during its use in historical mapping processes. For example, when the "cust_nm" field is mapped multiple times to the "CustomerName" standard data element in a personal business scenario, the weight of the "personal business" dimension in its context anchor vector gradually accumulates. When the system processes a data element for the first time, its context anchor vector is initialized to empty; as mapping operations accumulate, the vector is gradually improved. The system calculates the semantic drift of the same data element between different mapping times based on the joint semantic representation. For example, if the "cust_nm" field is frequently mapped to "BeneficiaryName" (payee name) instead of "CustomerName" on a certain day, the system calculates the difference between its joint semantic representation and the historical average. If this difference exceeds a preset drift threshold, a semantic drift alarm is triggered. In this embodiment, the preset drift threshold is set to 0.15. The meaning of this threshold is: when the Euclidean distance between the joint semantic representation of a data element and its historical average exceeds 0.15, the system determines that the semantics of the data element has changed significantly and an alarm needs to be triggered to remind data governance personnel to pay attention to whether the mapping relationship of the data element is still valid.
[0035] The system uses joint semantic representation and context anchoring vectors to differentiate and label the decomposed financial data elements into three layers: core, extension, and adaptation. The core layer consists of immutable basic data types and common business elements in financial standards, such as "Amount" and "Date," whose meanings remain consistent across different business scenarios. The extension layer comprises variable elements supporting diverse business scenarios; for example, "Party" refers to the payer or payee in a payment scenario, and the borrower or guarantor in a credit scenario. The adaptation layer consists of elements that meet the personalized needs of specific institutions, such as the bank's unique internal product codes. The system dynamically adjusts the labeling hierarchy based on semantic drift and mapping usage—when the semantics of an extension-layer data element stabilizes over long-term use, the system automatically upgrades it to the core layer; when a core-layer data element frequently exhibits semantic drift in a specific business scenario, the system automatically downgrades it to the extension or adaptation layer.
[0036] III. Construction of Heterogeneous Dual-Network Topology
[0037] After the structured decomposition is completed, the system constructs a dual-network heterogeneous topology containing a forward mapping network and a reverse mapping network. The nodes in this topology include three types: standard nodes, service nodes, and meta-model nodes.
[0038] Standard nodes correspond to data elements in the ISO 20022 standard library. For example, standard node S_001 corresponds to the "CustomerName" data element, S_002 corresponds to the "IdentificationNumber" data element, and S_003 corresponds to the "Amount" data element. Business nodes correspond to fields in the business system. For example, business node B_001 corresponds to the "cust_nm" field in the core business system, B_002 corresponds to the "id_num" field, and B_003 corresponds to the "txn_amt" field. Metamodel nodes correspond to abstract concepts in the metamodel layer. For example, M_001 corresponds to the abstract concept of "customer information," and M_002 corresponds to the abstract concept of "transaction information."
[0039] Standard nodes inherit from metamodel nodes, and business nodes also inherit from metamodel nodes. Specifically, S_001 (CustomerName) and B_001 (cust_nm) both inherit from M_001 (customer information), therefore S_001 and B_001 share the same semantic base. Multiple standard nodes inheriting from the same metamodel node, as well as multiple business nodes, also share a semantic base. For example, S_001 (CustomerName) and S_004 (BeneficiaryName) both inherit from M_001, and they are semantically related; B_001 (cust_nm) and B_005 (cstmr_nm) both inherit from M_001, and both represent customer names.
[0040] The system defines forward mapping edges as directed edges from business nodes to standard nodes. For example, B_001→S_001 means that the "cust_nm" field can be mapped to the "CustomerName" standard data element. The system defines reverse mapping edges as directed edges from standard nodes to business nodes. For example, S_001→B_001 means that the "CustomerName" standard data element can be reverse mapped to the "cust_nm" field. The system defines semantic association edges as synonymous, near-synonymous, and hierarchical semantic relationships between standard nodes or business nodes. For example, there is a near-synonymous semantic association edge between S_001 (CustomerName) and S_004 (BeneficiaryName). The system defines constraint edges as mutually exclusive and dependency constraints in financial business rules, such as the constraint that "transaction amount" must be a positive number.
[0041] The initial forward mapping weight of the forward mapping edge is determined based on a comprehensive score of name similarity, semantic similarity, and structural similarity. Name similarity is calculated based on the literal matching degree of field names. Semantic similarity is calculated based on named semantic primitives and contextual semantic primitives—for example, "cust_nm" and "CustomerName" are highly matched in named semantic primitives (both represent customer names) and also highly matched in contextual semantic primitives (both are used in customer information scenarios), thus their semantic similarity is high. Structural similarity is calculated based on type semantic primitives—"cust_nm" is a character type with a length of 50, and "CustomerName" is also a character type with a length of 70; they are structurally similar, thus their structural similarity is high. Based on the scores of the above three dimensions, the system assigns an initial forward mapping weight of 0.92 to B_001→S_001.
[0042] The initial reverse mapping weight of the reverse mapping edge is set to the mirror value of the corresponding forward mapping weight. In this example, the initial reverse mapping weight of S_001→B_001 is also 0.92. The system also assigns a preset adjustable redundancy to the reverse mapping edge, allowing the reverse mapping weight to deviate slightly from the forward mapping weight, in order to adapt to the actual situation in real business scenarios where the forward and reverse mappings are not completely symmetrical.
[0043] The system establishes a consistency bridge edge between the forward and reverse mapping networks. This bridge edge constrains the absolute value of the weight difference between corresponding forward and reverse mapping edges to not exceed a preset consistency tolerance threshold. In this example, the preset consistency tolerance threshold is 0.05. If at a certain moment the forward mapping weight of B_001→S_001 is 0.92, and the reverse mapping weight of S_001→B_001 is 0.88, the absolute value of the difference is 0.04, which does not exceed the tolerance threshold of 0.05, and the system does not intervene. If the reverse mapping weight further decreases to 0.86, the absolute value of the difference reaches 0.06, exceeding the tolerance threshold, and the system triggers a bidirectional consistency calibration operation. This calibration operation replaces the original weights with the weighted average of the forward mapping weight of 0.92 and the reverse mapping weight of 0.86. Assuming that the weighting coefficients for both the forward and reverse weights are 0.5, the calibrated forward and reverse mapping weights are both (0.92+0.86) / 2=0.89, thus satisfying the tolerance threshold again.
[0044] The consistency tolerance threshold is a dynamically maintained constraint. After each change in mapping weights, the system re-verifies and re-executes the aforementioned calibration operation. When the first standard node pointed to by any forward mapping edge in the forward mapping network and the first business node pointed to by the corresponding reverse mapping edge in the reverse mapping network share a common semantic basis through meta-model nodes, the system tightens the consistency tolerance threshold between that forward and reverse mapping edge by a preset margin. In this embodiment, the tightening margin is 40%, i.e., the tolerance threshold is adjusted from 0.05 to 0.03. When there is no common semantic basis between the standard node and the business node, the system relaxes the consistency tolerance threshold by a preset margin. In this embodiment, the relaxation margin is 60%, i.e., the tolerance threshold is adjusted from 0.05 to 0.08.
[0045] IV. Dynamic Bidirectional Consistency Analysis
[0046] After the heterogeneous dual-network topology is constructed, the system performs dynamic bidirectional consistency analysis on this topology. This analysis includes three dimensions: structural consistency components in the structural dimension, semantic consistency components in the semantic dimension, and temporal consistency components in the time dimension.
[0047] The system first calculates the bidirectional consistency index between each forward mapping edge and its corresponding reverse mapping edge based on the forward mapping weights and reverse mapping weights. This index quantifies the degree of mapping symmetry between the forward and reverse mapping networks. For the mapping edge pair B_001 and S_001, the bidirectional consistency index is calculated as: C = 1 - |wf - wr| / (wf + wr), where wf is the forward mapping weight and wr is the reverse mapping weight. If wf = 0.92 and wr = 0.89, then C = 1 - |0.92 - 0.89| / (0.92 + 0.89) = 1 - 0.03 / 1.81 ≈ 0.983. The closer the C value is to 1, the higher the bidirectional consistency.
[0048] Based on the aforementioned bidirectional consistency index, the system calculates structural consistency, semantic consistency, and temporal consistency components. The structural consistency component reflects the symmetry of the dual-network topology at the structural level, calculated by statistically analyzing the proportion and evenness of the distribution of corresponding edges in the forward and reverse mapping networks. The semantic consistency component reflects the semantic consistency of the mapping relationship, calculated by comparing the matching degree of semantic primitives involved in the forward and reverse mappings. The temporal consistency component reflects the stability of the consistency index across different time slices, calculated by comparing the change in the consistency index between the current and previous times. The system weights and merges these three components into a comprehensive consistency score. In this embodiment, the weighting coefficient for the structural consistency component is 0.25, the weighting coefficient for the semantic consistency component is 0.45, and the weighting coefficient for the temporal consistency component is 0.30. The semantic consistency component has the highest weight because the degree of matching at the semantic level is the most crucial factor in the standard mapping of financial data; the temporal consistency component has the next highest weight to reflect the system's focus on the stability of the mapping relationship over time; the structural consistency component has a relatively low weight.
[0049] The system extracts the context vector of the current business context from the context anchor vector of the data element. This context vector contains the business scenario identifier corresponding to the current mapping operation and the field usage pattern under that scenario. For example, when the system processes a mapping request for a personal account opening business, the value of the "business scenario" dimension in the context vector is "personal account opening," and the value of the "field usage pattern" dimension is "mainly customer information fields." The system uses this context vector as a conditional parameter for the forward mapping weight and the reverse mapping weight. By weighting the context vector with the initial values of the forward mapping weight and the reverse mapping weight, the system ensures that the mapping weight between the same business node and the same standard node has different weight values in different business contexts. For example, in the "personal account opening" scenario, the weight of "cust_nm" mapping to "CustomerName" is increased to 0.95; while in the "cross-border remittance" scenario, "cust_nm" may be more likely to map to "BeneficiaryName," and the weight allocation is adjusted accordingly.
[0050] When the context vector changes, the system recalculates the bidirectional consistency index of the affected mapping edges and marks the consistency fluctuations caused by the context change. The change in mapping weights caused by the context vector is the object of the consistency tolerance constraint. After recalculating the bidirectional consistency index, if the absolute value of the difference between the forward mapping weight and the reverse mapping weight exceeds the consistency tolerance threshold, the aforementioned bidirectional consistency calibration operation is triggered, so that the mapping weights meet the tolerance threshold again after completing the context adaptation.
[0051] The system tracks the change trajectory of the bidirectional consistency index of each mapped edge within a continuous time window. A 30-day time window is set, and the system records the bidirectional consistency index of each mapped edge daily. When the decrease in the bidirectional consistency index of a mapped edge relative to the previous moment exceeds a preset single-step decrease threshold (e.g., a sudden drop from 0.98 to 0.85, a daily decrease of 0.13, exceeding the single-step decrease threshold of 0.10), the system marks the mapped edge as a consistency-sensitive edge. When the cumulative decrease in the bidirectional consistency index of a mapped edge within a preset cumulative window exceeds a preset cumulative decrease threshold (e.g., a decrease from 0.98 to 0.82 over 7 consecutive days, a cumulative decrease of 0.16, exceeding the cumulative decrease threshold of 0.15), the system marks the mapped edge as a consistency-declining edge and triggers a consistency repair warning.
[0052] The system simultaneously monitors the overall connectivity index of the heterogeneous dual-network topology. This overall connectivity index includes the coupling strength between the forward mapping network and the reverse mapping network. The coupling strength is calculated based on the number of consistent bridging edges and the mapping density between the nodes on both sides of a consistent bridging edge. If the number of consistent bridging edges is 150 and the mapping density between the nodes on both sides (i.e., the ratio of the actual number of mapped edges to the possible number of mapped edges) is 0.75, the coupling strength can be expressed as the product or weighted sum of the two. When the coupling strength is lower than a preset coupling strength threshold, the system triggers a dual-network coupling structure warning.
[0053] When the proportion of consistency decay edges exceeds a preset threshold, or the coupling strength is lower than a preset coupling strength threshold, the system determines that the heterogeneous dual-network topology has entered a critical mismatch precursor state and outputs a reconstruction trigger signal. For example, if there are a total of 1000 mapping edges in the system, and 120 of them have consistency decay edges, the proportion is 12%, which exceeds the 10% proportion threshold, the system determines that it has entered a critical mismatch precursor state.
[0054] After completing the structured decomposition described above, the system then enters the association pre-construction phase. The core task of this phase is to pre-establish potential association channels between data elements, laying the foundation for subsequent topology construction. Association pre-construction is divided into three progressive stages.
[0055] The first stage involves pre-building static structural relationships. The system iterates through the foreign key dependencies in the database schema. For example, the transaction log table in the core business system contains a "cust_id" field, which acts as a foreign key pointing to the primary key "id" in the customer information table. Based on this, the system establishes a first candidate relationship between the "cust_id" field and the "id" field. The system also analyzes the co-occurrence relationships of data elements in the database tables—the three fields "cust_nm", "id_num", and "acct_bal" in the customer information table are frequently referenced simultaneously in the same query statement, and the system establishes a second candidate relationship between them. The system also establishes a third candidate relationship based on the matching patterns of field naming prefixes and suffixes; for example, all fields prefixed with "txn_" ("txn_dt", "txn_amt", "txn_cd") have naming-level relationships.
[0056] The second stage involves the pre-construction of semantic evolution associations. The system acquires the naming evolution trajectory of financial data elements throughout historical version evolution. This trajectory records multiple naming expressions of the same financial concept in different standard versions or different business systems. For example, in the evolution of the ISO 20022 standard from version 2013 to version 2019 and then to version 2022, the concept of "customer name" has successively used multiple naming forms such as "Nm", "CustomerName", and "PartyName". Based on these evolution trajectories, the system establishes cross-version semantic association edges for multiple data elements with different names but traceable to the same financial concept, as a fourth candidate association edge. For example, even if a field in a certain business system is named "cust_nm", while the corresponding data element in the standard library has been updated to "PartyName", the system can still identify the association between the two through the naming evolution trajectory.
[0057] The third stage is the pre-construction of business flow associations. The system establishes cross-system association edges based on the flow path of data elements in upstream and downstream financial business systems. For example, data from a customer account opening transaction flows from the core business system to the credit management system and then to the anti-money laundering system. The flow path of the customer's name field between these systems forms a complete business flow. The system identifies this flow path and establishes cross-system association edges between relevant fields distributed across different systems, as the fifth candidate association edges.
[0058] For the five types of candidate association edges mentioned above, the system assigns corresponding confidence levels. Static confidence reflects the reliability of establishing associations based on the static structure of the database; its value is higher when the foreign key relationship is clear and complete, and lower when the foreign key relationship is missing or ambiguous. Co-occurrence confidence reflects the reliability of establishing associations based on co-occurrence relationships within the same table; its value increases when the co-occurrence frequency is high. Named matching confidence reflects the reliability of establishing associations based on naming pattern matching; its value increases when multiple fields share the same naming prefix or suffix. Semantic evolution confidence reflects the reliability of establishing associations based on naming evolution trajectories; its value increases when the evolution chain is complete and continuous. Business flow confidence reflects the reliability of establishing associations based on business flows; its value increases when the flow path has been verified multiple times by actual business processes.
[0059] The system weights and merges the five confidence levels mentioned above into a single comprehensive association confidence level. The weighting coefficients can be dynamically adjusted based on the business scenario—in existing system upgrade projects, the weights of static confidence and named matching confidence can be appropriately increased; in cross-system data integration projects, the weight of business flow confidence can be appropriately increased. When the comprehensive association confidence level exceeds a preset association threshold, the system confirms the corresponding candidate association edge as a valid association edge and marks it as a semantic association edge to be added. When at least three candidate association edges exist between any two data elements, it indicates that there is multi-dimensional association evidence between these two data elements. The system automatically increases the weighted fusion weight of the comprehensive association confidence level—specifically, by multiplying the comprehensive association confidence level by a preset multi-source corroboration coefficient, which is greater than 1, making it easier for the comprehensive association confidence level to exceed the preset association threshold.
[0060] In the subsequent construction of the dual-network heterogeneous topology, the system will formally add the aforementioned semantically related edges to the topology as semantically related edges.
[0061] After the structured decomposition is completed, the system enters the dual-network heterogeneous topology construction phase. The node definition of the topology involves three levels. The standard node layer corresponds to the data elements in the ISO 20022 standard library, with each standard data element corresponding to one standard node. The business node layer corresponds to the fields in the business system's data dictionary, with each field corresponding to one business node. The metamodel node layer corresponds to the abstract concepts of the metamodel layer; these abstract concepts are a high-level summary of the common semantics of standard nodes and business nodes.
[0062] The inheritance relationships between nodes form the vertical structure of the topology. Standard nodes inherit from metamodel nodes—for example, the "CustomerName" standard node inherits from the "Customer Information" metamodel node. Business nodes also inherit from metamodel nodes—the "cust_nm" business node also inherits from the "Customer Information" metamodel node. Because both inherit from the same metamodel node, S_001 and B_001 share the same semantic base. This semantic base means that they are semantically homologous, and even if there are differences in their literal names, the system can still recognize their inherent semantic relationship. Multiple standard nodes inheriting from the same metamodel node, as well as multiple business nodes, also share a semantic base—for example, both "CustomerName" and "BeneficiaryName" inherit from the "Customer Information" metamodel node, and they are semantically related.
[0063] In addition to semantic inheritance edges, the topology contains various other edge types. Semantic inheritance edges are directed edges from standard nodes to metamodel nodes and from business nodes to metamodel nodes. Based on these semantic inheritance edges, the system establishes a bidirectional index path between standard nodes and business nodes. This bidirectional index path has two directions: starting from any standard node, tracing along the semantic inheritance edge to the metamodel node, and then descending along the reverse semantic inheritance edge of the metamodel node to all business nodes inheriting from the same metamodel node, forming the first direction index from standard nodes to business nodes; starting from any business node, tracing along the semantic inheritance edge to the metamodel node, and then descending along the reverse semantic inheritance edge of the metamodel node to all standard nodes inheriting from the same metamodel node, forming the second direction index from business nodes to standard nodes.
[0064] Forward mapping edges are directed edges from business nodes to standard nodes. Reverse mapping edges are directed edges from standard nodes to business nodes. Semantic association edges are edges that express synonymous, near-synonymous, or hierarchical semantic relationships between standard nodes or between business nodes. Constraint edges are directed edges that express mutual exclusion and dependency constraints in financial business rules.
[0065] The initial forward mapping weight of a forward mapping edge is determined by a combination of three similarity dimensions. The first dimension is name similarity, calculated based on the literal matching of field names. The second dimension is semantic similarity, calculated based on named semantic primitives and contextual semantic primitives—this semantic similarity reflects the degree of matching between two data elements at the level of business meaning and business scenario. The third dimension is structural similarity, calculated based on type semantic primitives—this structural similarity reflects the degree of matching between two data elements at the level of structural features such as data type, field length, and code value constraints. Based on the scores of these three dimensions, the system assigns an initial weight between 0 and 1 to each forward mapping edge.
[0066] The initial reverse mapping weight of the reverse mapping edge is set to the mirror value of the corresponding forward mapping weight. However, considering that the symmetry of forward and reverse mapping is not absolute in real business scenarios, the system presets an adjustable redundancy for each reverse mapping edge. This redundancy allows the reverse mapping weight to deviate slightly from the forward mapping weight, in order to adapt to scenarios such as multiple business fields being mapped to the same standard field.
[0067] The system establishes a consistency bridge edge between the forward and reverse mapping networks. Each consistency bridge edge corresponds to a pair of forward and reverse mapping edges. The function of this bridge edge is to constrain the absolute value of the weight difference between the two edges to not exceed a preset consistency tolerance threshold. This tolerance threshold is a dynamically maintained constraint—after each change in mapping weights, the system re-verifies this constraint and, if necessary, re-executes a bidirectional consistency calibration operation. When the absolute value of the weight difference between the forward and reverse mapping edges exceeds the tolerance threshold, the system triggers a calibration operation. This calibration operation replaces the original weights with a weighted average of the forward and reverse mapping weights, ensuring that both re-meet the tolerance threshold after calibration.
[0068] The weighted averages involved in the calibration operation are calculated as follows: w_f_new = (α × w_f + β × w_r) / (α + β), w_r_new = (α × w_f + β × w_r) / (α + β). Where w_f is the forward mapping weight before calibration, w_r is the reverse mapping weight before calibration, α is the weighting coefficient of the forward weights, and β is the weighting coefficient of the reverse weights. In this embodiment, α and β can both be 1, i.e., a simple arithmetic average. In other embodiments, the ratio of α and β can be adjusted according to actual business needs to make the calibrated weights more biased towards either the forward or reverse direction.
[0069] The tolerance threshold value differs between mapping pairs that share a common semantic basis and those that do not. When the first standard node pointed to by the forward mapping edge and the first business node pointed to by the reverse mapping edge trace back to a common semantic basis through the metamodel node, the system tightens the consistency tolerance threshold of that mapping pair by a preset margin. For example, for S_001 and B_001, which share the semantic basis of "customer information," the tolerance threshold is set to 0.03. When there is no common semantic basis between the standard node and the business node, the system relaxes the tolerance threshold by a preset margin. For example, a heterogeneous mapping pair across business domains that does not share a common semantic basis has a tolerance threshold set to 0.08 to accommodate the semantic differences that inevitably exist in mapping scenarios across semantic bases.
[0070] After the heterogeneous dual-network topology is constructed, the system continuously performs dynamic bidirectional consistency analysis on this topology. This analysis is not a one-time static check, but is repeatedly executed periodically or triggered by events during system operation to monitor the health status of the mapping relationship in real time.
[0071] Dynamic bidirectional consistency analysis first calculates the bidirectional consistency index between each forward mapping edge and its corresponding reverse mapping edge. For the aforementioned mapping pair B_001 (cust_nm) and S_001 (CustomerName), assuming the current forward mapping weight is 0.92 and the reverse mapping weight is 0.89, the system calculates the bidirectional consistency index as follows: C = 1 - |wf - wr| / (wf + wr). Here, wf represents the forward mapping weight, and wr represents the reverse mapping weight. Substituting the values, we get C = 1 - |0.92-0.89| / (0.92+0.89) = 1 - 0.03 / 1.81 ≈ 0.983. This index quantifies the degree of symmetry between the forward and reverse mappings; the closer the C value is to 1, the higher the bidirectional consistency.
[0072] Based on the aforementioned bidirectional consistency index, the system calculates structural consistency, semantic consistency, and temporal consistency components. The structural consistency component reflects the degree of symmetry of the dual-network topology at the structural level. The system counts the total number of mapping edges in the forward mapping network and the total number of mapping edges in the reverse mapping network, calculating their ratio. Simultaneously, it counts the degree distribution of each node in both the forward and reverse networks, obtaining the structural consistency component by comparing the similarity between the forward and reverse degree distributions. The semantic consistency component reflects the consistency of the mapping relationship at the semantic level. The system compares the degree of matching between the semantic primitive combinations upon which the forward mapping depends and the semantic primitive combinations upon which the reverse mapping depends. The temporal consistency component reflects the stability of the consistency index across different time slices. The system compares the change in the bidirectional consistency index between the current time and the previous time; the smaller the change, the higher the temporal consistency component. The system weights and fuses these three components into a comprehensive consistency score.
[0073] The system extracts the context vector of the current business context from the context anchor vector of the data element. The context anchor vector is established for each data element during the structured decomposition phase and is continuously maintained, recording the business context distribution of that data element in the historical mapping process. The current business context includes the business scenario identifier corresponding to the current mapping operation and the field usage pattern under that scenario. For example, when the system processes a mapping request for a personal account opening business, the business scenario dimension of the context vector is "personal account opening," and the field usage pattern dimension is "primarily customer information fields." The system uses this context vector as a conditional parameter for the forward mapping weight and the reverse mapping weight. By weighting the context vector with the initial values of the forward mapping weight and the reverse mapping weight, the mapping weight between the same business node and the same standard node has different weight values under different business contexts. Specifically, in the scenario of personal account opening, the activation value of the customer information dimension in the context vector is higher, and the system increases the positive weight of mapping "cust_nm" to "CustomerName" from 0.92 to 0.95; while in the scenario of cross-border remittance, the activation value of the transaction participant dimension in the context vector is higher, and "cust_nm" may be more inclined to map to "BeneficiaryName", and the weight allocation is adjusted accordingly.
[0074] When the context vector changes, the system recalculates the bidirectional consistency index of the affected mapping edges and marks the consistency fluctuations caused by the context change. The change in mapping weights caused by the context vector is the target of the consistency tolerance constraint. After recalculating the bidirectional consistency index, if the absolute value of the difference between the forward and reverse mapping weights exceeds the consistency tolerance threshold, the system triggers the aforementioned bidirectional consistency calibration operation, ensuring that the mapping weights meet the tolerance threshold again after context adaptation. For example, after a context switch, the forward weight of B_001→S_001 becomes 0.95, while the reverse weight of S_001→B_001 remains 0.89, with an absolute difference of 0.06, exceeding the tolerance threshold of 0.03. The system immediately performs a calibration operation, adjusting both to a weighted average of 0.92.
[0075] The system tracks the change trajectory of the bidirectional consistency index of each mapping edge within a continuous time window. A 30-day time window is set, and the system collects and records the bidirectional consistency index of each mapping edge during daily off-peak hours. When the bidirectional consistency index of a mapping edge decreases by more than a preset single-step decrease threshold relative to the previous day, the system marks that mapping edge as a consistency-sensitive edge. For example, if the bidirectional consistency index of S_003 (Amount) and B_003 (txn_amt) drops sharply from 0.97 to 0.85 on a certain day, a single-day decrease of 0.12 exceeding the preset threshold of 0.10, this mapping edge is marked as a consistency-sensitive edge. When the cumulative decrease of the bidirectional consistency index of a mapping edge within a preset cumulative window exceeds a preset cumulative decrease threshold, the system marks that mapping edge as a consistency decay edge and triggers a consistency repair warning. For example, if the bidirectional consistency index between B_003 and S_003 drops from 0.97 to 0.80 within 7 consecutive days, with a cumulative decrease of 0.17 exceeding the preset threshold of 0.15, the system will generate a repair warning notification to alert data governance personnel to pay attention to the anomaly in the mapping relationship.
[0076] The system simultaneously monitors the overall connectivity index of the heterogeneous dual-network topology. This overall connectivity index includes the coupling strength between the forward mapping network and the reverse mapping network. The coupling strength is calculated as follows:
[0077] κ = N_bridge × ρ_map
[0078] Where κ represents the coupling strength, N_bridge represents the number of consistent bridging edges, and ρ_map represents the mapping density between the nodes on both sides. The mapping density ρ_map is calculated as follows: ρ_map = N_actual / (N_source × N_target), where N_actual is the total number of actual positive and negative mapping edges, N_source is the total number of business nodes, and N_target is the total number of standard nodes.
[0079] The mapping density is the ratio of the actual number of mapping edges to the theoretical maximum number of mapping edges. If the number of consistent bridging edges is 150 and the mapping density between the nodes on both sides is 0.75, the coupling strength can be expressed as the product of 150 and 0.75. When the coupling strength is lower than the preset coupling strength threshold, the system triggers a dual-network coupling structure warning.
[0080] When the proportion of consistency-decaying edges exceeds a preset threshold, or the coupling strength falls below a preset coupling strength threshold, the system determines that the heterogeneous dual-network topology has entered a critical mismatch precursor state and outputs a reconstruction trigger signal. For example, if there are a total of 1000 mapped edges in the system, and 130 edges are currently marked as consistency-decaying, representing 13%, which exceeds the preset threshold of 10%, the system determines that it has entered a critical mismatch precursor state. After the reconstruction trigger signal is output, the system immediately enters the adaptive reconstruction phase.
[0081] The adaptive reconstruction operation is initiated upon receiving a reconstruction trigger signal. This operation first implements differentiated protection strategies for mapping edges in the heterogeneous dual-network topology. The system classifies the protection level into three levels based on the importance of the mapping edges. Mapping edges within the Level 1 protection zone are those corresponding to standard nodes commonly referenced by more than five business systems, core mapping edges involved in regulatory reporting, and mapping edges corresponding to specific financial standard versions. For example, the mapping edge corresponding to the "TransactionId" data element used by the People's Bank of China's large-value payment system in the ISO 20022 standard is a Level 1 protected object, and the system prohibits any automatic pruning operations on it. Mapping edges within the Level 2 protection zone are those frequently queried in the past 30 days, with "frequently queried" defined as exceeding 100 queries. For mapping edges within the Level 2 protection zone, a consistency impact assessment is required before pruning—the system simulates deleting the mapping edge, calculates the number of directly and indirectly related nodes affected, and assesses the proportion of nodes belonging to the Level 1 protection zone among the affected nodes. Pruning is only allowed when the assessment result is lower than a preset impact threshold. In this embodiment, the preset impact threshold is set to 5%. The meaning of this threshold is as follows: if, after simulating the deletion of a mapping edge, the proportion of nodes belonging to the first-level protection zone among the affected nodes exceeds 5%, then the pruning operation is not allowed; if this proportion does not exceed 5%, then pruning is allowed. Mapping edges within the third-level protection zone are all mapping edges other than those in the first-level and second-level protection zones, and the system directly performs pruning operations on mapping edges within the third-level protection zone.
[0082] The specific triggering conditions for pruning operations are related to the bidirectional consistency index. When the bidirectional consistency index of any mapped edge is continuously below a preset low threshold for a duration exceeding a preset number of detection cycles, the system performs the corresponding pruning operation based on the protection level to which the mapped edge belongs. The preset low threshold is set to 0.60, and the preset number of detection cycles is set to 5 consecutive detections. If the bidirectional consistency index of a mapped edge in a Level 3 protected area is below 0.60 in 5 consecutive detections, the system directly marks the mapped edge as a candidate for pruning and performs a deletion operation. If the mapped edge belongs to a Level 2 protected area, the system first triggers a consistency impact range assessment and performs pruning only if the assessment results allow. If the mapped edge belongs to a Level 1 protected area, the system maintains the mapped edge unchanged and records the abnormal event.
[0083] The growth operation is triggered when a financial standard version update is detected or a new field is added to a business system. When the ISO 20022 standard was upgraded from version 2019 to version 2022, several data elements were added to the standard library. For each new standard node, the system traverses all existing business nodes and calculates the candidate mapping confidence score based on a comprehensive score of name similarity, semantic similarity, and structural similarity. The calculation method for this comprehensive score is consistent with the calculation method for the initial forward mapping weight. When the candidate mapping confidence score exceeds a preset growth threshold, the system automatically generates the corresponding candidate forward mapping edge and candidate reverse mapping edge and adds them to the dual-network heterogeneous topology. The preset growth threshold is set to 0.75. For example, the ISO 20022 2022 version adds the "GreenIndicator" data element to identify green financial transactions. The system calculates and finds that the candidate mapping confidence score between the "env_tag" field of a certain business system and this data element reaches 0.82, exceeding the growth threshold of 0.75. The system automatically generates the corresponding forward and reverse mapping edges and adds them to the topology.
[0084] Reconnection operations handle bidirectional conflict scenarios. In real-world financial business scenarios, it is common for a standard data element to correspond to multiple business fields. For example, the "CustomerName" data element in the ISO 20022 standard may simultaneously correspond to the "cust_nm" field in the core business system and the "cstmr_full_nm" field in the credit management system, both of which are valid mappings. When the system detects that any business node points to the first standard node via a forward mapping edge, but the first standard node does not have any reverse mapping edges pointing to that business node, the system determines that a mapping to be verified is missing and triggers a reverse mapping edge completion verification. The specific operation of this completion verification is as follows: calculate the business semantic similarity between the business node and each business node pointed to by all existing reverse mapping edges of the first standard node. If all similarities are lower than a preset many-to-one effective threshold, the system determines that there is a bidirectional conflict; if there is at least one business node whose business semantic similarity with the business node is not lower than the preset many-to-one effective threshold, the system determines that there is a valid many-to-one mapping relationship between the business node and the first standard node, and no reconnection operation is triggered.
[0085] The above judgment logic is illustrated with an example. Standard node S_001 (CustomerName) already has a reverse mapping edge pointing to business node B_001 (cust_nm). Now, business node B_005 (cstmr_full_nm) is detected pointing to S_001 via a forward mapping edge, but S_001 does not have a reverse mapping edge pointing to B_005. The system triggers reverse mapping edge completion verification, calculates the business semantic similarity between B_005 and B_001. Both represent customer names and are semantically highly similar, with a similarity of 0.95, which is not lower than the preset many-to-one effective threshold of 0.70. The system determines that a valid many-to-one mapping relationship exists between B_005 and S_001 and does not trigger a reconnection operation. Conversely, if a business node B_020 (txn_rmk) points to S_001 via a forward mapping edge, but the business semantic similarity between B_020 and B_001 is only 0.35, which is lower than the 0.70 threshold, the system determines that a bidirectional conflict has occurred.
[0086] When a bidirectional conflict is detected, the system triggers a reconnection operation. The reconnection operation works as follows: the set of reverse mapping edges of the first standard node is updated to include the reverse mapping edges of the service node, so that the first standard node is simultaneously reverse-mapped to multiple service nodes, including the service node in question. Taking the aforementioned conflict between B_020 and S_001 as an example, the system updates the set of reverse mapping edges of S_001 to include both B_001 and B_020, so that S_001 is simultaneously reverse-mapped to both service nodes B_001 and B_020. The system then performs a bidirectional consistency calibration operation on the newly added reverse mapping edges to ensure that they meet tolerance constraints with their corresponding forward mapping edges.
[0087] After the structured decomposition is completed, the system enters the association pre-construction stage. This stage is independent of the subsequent topology construction process, and its task is to pre-establish candidate association channels between data elements to provide input for topology construction.
[0088] The pre-construction of associations is executed sequentially in three stages. The first stage is the pre-construction of static structural associations. The system traverses the entire database schema and identifies three types of static clues. The first type of clue is foreign key dependency—the system scans the foreign key definitions of all data tables and finds that the "cust_id" field in the transaction log table points to the "id" field in the customer information table. Based on this, a first candidate association edge is established between the two data elements "cust_id" and "id". The second type of clue is co-occurrence relationship—the system analysis finds that the three fields "cust_nm", "id_num", and "acct_bal" are frequently queried simultaneously in the customer information table, with the proportion of references to these three fields in the same query reaching 75%. Based on this, second candidate association edges are established between them respectively. The third type of clue is naming pattern matching—the system identifies all fields with the prefix "txn_", including "txn_dt", "txn_amt", and "txn_cd", and establishes a third candidate association edge between them based on the naming pattern of shared prefixes.
[0089] The second phase involves the pre-construction of semantic evolution associations. The system accesses a historical version database of financial standards to obtain the evolution trajectory of data element naming across different versions. Taking the ISO 20022 standard as an example, during its evolution from version 2013 to 2019 and then to 2022, the concept of "customer name" successively used three naming forms: "Nm", "CustomerName", and "PartyName". The system also retrieves the evolution records of internal field names from the bank's historical data dictionary. In the bank's system, "customer name" has successively used multiple forms such as "cust_name", "cstmr_nm", and "customer_nm". The system integrates the standard evolution trajectory and the internal evolution trajectory to form a complete naming evolution chain. Based on this, cross-version semantic association edges are established for data elements with different names but traceable to the same financial concept, serving as the fourth candidate association edge.
[0090] The third stage involves the pre-construction of business flow associations. The system extracts complete business flow logs from the bank's core business system, tracing the flow path of data elements between upstream and downstream systems. Taking a personal account opening transaction as an example, customer name data originates from the customer information table in the core business system, flows through the data exchange platform to the borrower information table in the credit management system, and then flows through the regulatory reporting module to the suspicious transaction report table in the anti-money laundering system. The system identifies this complete flow path and establishes cross-system association edges between customer name-related fields distributed across the three systems, serving as the fifth candidate association edge.
[0091] For the five types of candidate association edges mentioned above, the system assigns corresponding confidence scores. The static confidence score assignment rules are as follows: candidate association edges based on foreign key dependencies are assigned a score of 0.90, candidate association edges based on co-occurrence relationships within the same table are assigned a score of 0.70, and candidate association edges based on named pattern matching are assigned a score of 0.80. When multiple clues point to the same pair of data elements, the highest value is taken. Co-occurrence confidence is calculated based on co-occurrence frequency, with a score of 0.75 when the co-occurrence frequency is 75%. Named matching confidence is calculated based on the completeness of prefix or suffix matching, with a score of 0.85 for a complete match and 0.60 for a partial match. Semantic evolution confidence is calculated based on the completeness of the evolution chain, with a score of 0.95 when the chain length reaches 5 or more versions, and 0.80 when the chain length is 3 to 4 versions. The confidence level of the business flow is calculated based on the actual number of business verifications traversed by the flow path. When the number of verifications exceeds 10, the value is assigned to 0.90, and when the number of verifications is between 5 and 10, the value is assigned to 0.75.
[0092] The system weights and fuses the above five confidence levels into a comprehensive correlation confidence level. The weighted fusion calculation method is as follows:
[0093] C_total = λ_static × C_static + λ_cooccur × C_cooccur + λ_naming ×C_naming + λ_evolution × C_evolution + λ_flow × C_flow
[0094] Where C_total is the overall association confidence score, C_static is the static confidence score, C_cooccur is the co-occurrence confidence score, C_naming is the named matching confidence score, C_evolution is the semantic evolution confidence score, and C_flow is the business flow confidence score; λ_static, λ_cooccur, λ_naming, λ_evolution, and λ_flow are the corresponding weighting coefficients, and the sum of the five weighting coefficients is 1. In this embodiment, the five weighting coefficients are 0.25, 0.15, 0.15, 0.25, and 0.20, respectively, with the static confidence score and the semantic evolution confidence score having the highest weights because the static structure of the database and the semantic evolution history are the most reliable evidence of association.
[0095] When the overall association confidence exceeds the preset association threshold, the system confirms the corresponding candidate association edge as a valid association edge and marks it as a semantic association edge to be added. In this embodiment, the preset association threshold is set to 0.70. Taking the "CustomerName" data element in ISO20022 2019 and the "cust_nm" field in the bank's core business system as examples, their static confidence is 0.90, co-occurrence confidence is 0.75, naming matching confidence is 0.85, semantic evolution confidence is 0.80, business flow confidence is 0.90, and the weighted fusion overall association confidence is:
[0096] 0.90×0.25+0.75×0.15+0.85×0.15+0.80×0.25+0.90×0.20 = 0.225+0.1125+0.1275+0.20+0.18 = 0.845, which exceeds the association threshold of 0.70. Therefore, this candidate association edge is marked as a semantic association edge to be added.
[0097] When at least three candidate association edges exist between any two data elements, it indicates that there is multi-dimensional association evidence between these two data elements. The system automatically increases the judgment weight of the comprehensive association confidence under the condition of multi-source corroboration. The specific increase method is to multiply the comprehensive association confidence by a preset multi-source corroboration coefficient. In this embodiment, the multi-source corroboration coefficient is 1.15. Taking "CustomerName" and "cust_nm" as examples, there are three candidate association edges between them: static structural association (foreign key dependency), semantic evolution association (cross-version semantic evolution), and business flow association (upstream and downstream flow path). The comprehensive association confidence of 0.845 is increased to 0.972 after multiplying by 1.15, which significantly exceeds the association threshold of 0.70. This increase operation gives association edges from multiple independent evidence channels a higher priority in subsequent topology construction.
[0098] During the subsequent construction of the dual-network heterogeneous topology, the system will formally add all the aforementioned semantically related edges to the topology as semantically related edges. Semantically related edges are synonymous, near-synonymous, and hierarchical semantic relationships between standard nodes or between business nodes.
[0099] The feedback loop phase is executed in parallel during the continuous operation of the system. The system collects and maps multi-dimensional feedback data throughout the entire process.
[0100] The first type is explicit feedback, consisting of manual review of the mapping results by data governance personnel. The bank's data governance team conducts weekly sampling reviews of the system-generated mapping results. When a governance personnel confirms a mapping edge is correct, the system records a positive confirmation signal; when a governance personnel corrects the mapping target of a mapping edge, the system records a correction signal; when a governance personnel rejects a mapping edge, the system records a rejection signal. The system assigns +1 to confirmation signals, 0 to correction signals, and -1 to rejection signals, and accumulates the review score for each mapping edge.
[0101] The second type is implicit feedback, which consists of the frequency with which the mapped edge is actually queried and the acceptance rate of the mapping results by the business system. The system records the query log for each mapped edge during runtime. For example, a certain mapped edge was queried 1500 times by the business system in a given month, with 1400 of the returned results being directly used by the business system without triggering manual review, resulting in an acceptance rate of 93.3%. The system records the usage frequency of this mapped edge as 1500 times and the acceptance rate as 93.3%.
[0102] The third type is consistency feedback, which consists of the time-series trend of the bidirectional consistency index. The system calculates the bidirectional consistency index of each mapped edge weekly and compares it with the value of the previous week to generate a time-series trend curve.
[0103] The fourth category is external feedback, consisting of official version change notifications for financial standards and updates to regulatory rules. The system subscribes to standard update notifications issued by the ISO organization. When a new version of the standard is released, the system automatically receives the change list, including a list of newly added data elements, a list of modified data elements, and a list of obsolete data elements.
[0104] The above four types of feedback data employ a closed-loop coordination strategy of "adjustment-verification-calibration" for adjusting each link of the system.
[0105] First, the positive mapping weights are adjusted based on implicit feedback. Taking a mapping edge B_001→S_001 as an example, this mapping edge was queried 1500 times in the current month, with an acceptance rate of 93.3%. The system uses the product of usage frequency and acceptance rate as the adjustment amount; 1500 multiplied by 93.3% yields an adjustment amount of 1400. The system adjusts the positive mapping weight from 0.92 to 0.92 plus the normalized increment of the adjustment amount. The normalization method is to divide the adjustment amount by a preset normalization base of 10000, with an increment of approximately 0.14, resulting in an adjusted positive weight of 0.94.
[0106] The system adjusts the reverse mapping weight based on explicit feedback. Taking a mapping edge S_001→B_001 as an example, this edge received 3 approvals, 0 corrections, and 0 rejections in the current month, with an approval score of +3. The system adjusts the reverse mapping weight from 0.89 to 0.89 plus the normalized increment of the approval score. The normalization method is to divide the approval score by a preset normalization base of 10, with an increment of approximately 0.03. The adjusted reverse weight is 0.92.
[0107] After each adjustment, the system recalculates the bidirectional consistency index of the mapping edges corresponding to the adjusted weights. If the absolute value of the difference between the adjusted forward mapping weights and the reverse mapping weights does not exceed the consistency tolerance threshold, the system confirms that the adjustment has taken effect. If the absolute value of the difference exceeds the tolerance threshold, the system triggers a bidirectional consistency calibration operation, replacing the original weights with the weighted average of the forward and reverse mapping weights, so that both weights meet the tolerance threshold again after adjustment.
[0108] Taking the aforementioned adjustment as an example, if the positive weight is adjusted to 0.94 and the negative weight is adjusted to 0.92, the absolute value of the difference is 0.02, which does not exceed the tolerance threshold of 0.03, and the adjustment takes effect. If the positive weight is adjusted to 0.97 and the negative weight remains at 0.89, the absolute value of the difference, 0.08, exceeds the tolerance threshold of 0.03, and the system triggers a calibration operation, adjusting both to (0.97+0.89) / 2=0.93. After calibration, the tolerance threshold is met again.
[0109] The system dynamically adjusts the tolerance threshold of the consistency tolerance constraint based on the overall distribution of the bidirectional consistency index. The system calculates the bidirectional consistency index distribution of all mapped edges monthly. If the average value of the overall distribution increases from 0.92 to 0.96, it indicates an increase in the overall consistency level of the system, and the system tightens the tolerance threshold from 0.05 to 0.04, requiring higher precision bidirectional symmetry. If the average value of the overall distribution decreases from 0.92 to 0.85, it indicates a decrease in the overall consistency level of the system, and the system relaxes the tolerance threshold from 0.05 to 0.07, providing the system with more fault tolerance.
[0110] The system distributes multidimensional feedback data across three dimensions. When distributed by business system dimension, feedback signals from core business systems only affect the mapping edges associated with that core business system, preventing feedback errors from the credit management system from impacting the mapping relationships of core business systems. When distributed by financial standard version dimension, feedback signals corresponding to the ISO 20022 2019 version only affect all mapping edges associated with that version, without propagation across versions. When distributed by user operation dimension, the review operations of a specific data governance personnel only affect the mapping edges operated by that personnel. This three-dimensional distribution mechanism ensures accurate attribution of feedback signals and prevents the spread of errors.
[0111] The system also adjusts the decomposition rules in the structured decomposition process based on feedback data. The system monitors the trigger frequency and accuracy of each semantic primitive dimension in the joint disambiguation mechanism. After three months of continuous monitoring, it was found that the disambiguation trigger frequency for the named semantic primitive dimension was 30%, with 15% of its judgments being manually corrected; the disambiguation trigger frequency for the context semantic primitive dimension was 10%, with 5% of its judgments being manually corrected. Since the correction rate of the context semantic primitive is significantly lower than that of the named semantic primitive, it indicates that the context semantic primitive has stronger semantic discrimination capabilities. The system increases the fusion weight of the context semantic primitive from 0.25 to 0.35 and decreases the fusion weight of the named semantic primitive from 0.30 to 0.20, making subsequent decomposition more biased towards the business context dimension.
[0112] The system also adjusts the triggering conditions for the adaptive reconstruction process based on feedback data. The system monitors the actual effect of pruning operations—whether any valid mapping edges were mistakenly pruned among the pruned edges. After three months of continuous operation, statistics showed that the system pruned a total of 120 mapping edges, of which 8 were confirmed as valid mapping edges and restored during subsequent business verification, resulting in a false pruning rate of 6.7%. This false pruning rate exceeded the preset 5% threshold, so the system automatically increased the preset low threshold used for pruning operations from 0.60 to 0.65, making future pruning operations more cautious. Simultaneously, statistics revealed that the system missed pruning 30 mapping edges—these edges met the original pruning conditions but were not pruned and were subsequently confirmed as invalid, resulting in a missed pruning rate of 25%. The missed shearing rate exceeded the preset 10% missed shearing rate threshold. However, increasing the low threshold would further increase the missed shearing rate. Therefore, the system adopted a compromise strategy—maintaining the preset low threshold at 0.60, but shortening the number of detection cycles from 5 consecutive times to 3 consecutive times, making the system more sensitive to persistent low consistency mapping.
[0113] Through the closed-loop adjustment of the above multi-dimensional feedback, the system continuously optimizes the parameters of each link during operation, achieving the self-evolution goal of continuously improving mapping accuracy while ensuring bidirectional consistency.
[0114] In addition to forward mapping edges, reverse mapping edges, semantic association edges, and constraint edges, heterogeneous dual-network topologies also include semantic inheritance edges. Semantic inheritance edges are directed edges from standard nodes to metamodel nodes, and directed edges from business nodes to metamodel nodes. Taking the aforementioned bank as an example, the standard node S_001 corresponds to the "CustomerName" data element in the ISO 20022 standard, and the metamodel node M_001 corresponds to the abstract concept of "customer information." The system establishes a semantic inheritance edge between S_001 and M_001, pointing from S_001 to M_001, indicating that S_001 inherits from M_001. Similarly, the business node B_001 corresponds to the "cust_nm" field in the core business system, which is also inherited from M_001. The system establishes a semantic inheritance edge between B_001 and M_001, pointing from B_001 to M_001. Standard nodes and business nodes share a semantic basis by inheriting from the same metamodel node.
[0115] Based on the semantic inheritance edges mentioned above, the system establishes a bidirectional index path between standard nodes and business nodes. This bidirectional index path is divided into two directions.
[0116] The first-direction index starts from the standard node. Taking S_001 (CustomerName) as an example, the system traces along the semantic inheritance edge from S_001 to M_001 to the metamodel node M_001 (customer information), and then descends along the reverse semantic inheritance edge of M_001 to find all business nodes that inherit from M_001, including B_001 (cust_nm), B_005 (cstmr_full_nm), and B_008 (party_name). The system records this index path, forming the first-direction index from S_001 to B_001, B_005, and B_008. The establishment of this index path enables the system to quickly locate all business nodes affected by a specific standard node.
[0117] The second-direction index starts from the business node. Taking B_001 (cust_nm) as an example, the system traces along the semantic inheritance edge from B_001 to M_001 to the metamodel node M_001, and then descends along the reverse semantic inheritance edge of M_001 to find all standard nodes that inherit from M_001, including S_001 (CustomerName) and S_004 (BeneficiaryName). The system records this index path, forming the second-direction index from B_001 to S_001 and S_004. The establishment of this index path enables the system to quickly locate all candidate standard nodes corresponding to a specific business node.
[0118] Bidirectional indexing plays a crucial role in standard version updates and business system changes. When the ISO 20022 standard is upgraded from version 2019 to version 2022, the attribute of the standard node S_001 (CustomerName) changes—the definition of this data element is expanded from "the customer's full name" to "the customer's complete name, including last name, first name, and middle name." Upon detecting this attribute change event, the system does not immediately perform revalidation but instead records the attribute change event in the event queue. The event queue is a first-in, first-out data structure; each event entry contains the event type (attribute change), the event occurrence time, the node identifiers involved in the change, and a detailed description of the change. The system uses the first direction index to locate all business nodes affected by this standard node attribute change, including B_001, B_005, and B_008, marking these business nodes and all their associated mapping edges as pending validation, while simultaneously recording the execution position of the current main loop.
[0119] The main loop refers to the loop process consisting of the five steps defined in claim 1. When an attribute change occurs, the system may be executing any step in the main loop. For example, if the system is currently executing the dynamic bidirectional consistency analysis step, the system records the current execution position as "Step 3 - In Progress", and then continues to complete the remaining work of the current step without being interrupted by the attribute change.
[0120] When the field attribute of any business node changes, the system adopts the same processing strategy. For example, when the bank's core business system upgrades its data dictionary, expanding the length of the "cust_nm" field from 50 characters to 200 characters, the system detects this field attribute change but does not immediately perform a re-validation. Instead, it records the field attribute change event in the event queue and locates all standard nodes affected by the business node's field attribute change along the second direction index, including S_001 and S_004. The system marks these standard nodes and their associated mapping edges as pending validation, while simultaneously recording the execution position of the current main loop.
[0121] After the current step in the main loop completes, the system checks if the event queue is empty. If the event queue is empty, the system proceeds to the next step in the main loop. If the event queue is not empty, the system pauses proceeding to the next step in the main loop and prioritizes performing consistency rechecks on all nodes in the pending verification state and their associated mapping edges.
[0122] Taking the scenario of upgrading from ISO 20022 standard version 2019 to version 2022 as an example, after the current main loop step is completed, the system detects a "CustomerName attribute change" event in the event queue and then initiates a consistency re-verification process. The specific steps of this re-verification process are as follows: First, the bidirectional consistency index of the affected mapping edges is recalculated. For the B_001→S_001 mapping edge, since the definition of S_001 has changed, the system re-extracts the semantic primitives of S_001 based on the updated standard definition, then recalculates the forward mapping weight and reverse mapping weight between B_001 and S_001, and then recalculates the bidirectional consistency index accordingly. The same operation is performed for the B_005→S_001 and B_008→S_001 mapping edges.
[0123] After the consistency recheck is completed, the system clears the corresponding processed events in the event queue, removes the pending verification status, and determines the subsequent path based on the recheck result. If the recheck result indicates a bidirectional consistency error—for example, after the definition of S_001 changes, the forward mapping weight between B_001 and S_001 decreases from 0.92 to 0.76, and the reverse weight decreases from 0.89 to 0.71, with an absolute difference of 0.05, exceeding the tolerance threshold of 0.03—the system triggers the adaptive refactoring operation described in claim 5. After the refactoring operation is completed, the system resumes execution of the next step of the current main loop. If the recheck result indicates that bidirectional consistency is normal, the system directly resumes execution of the next step of the current main loop.
[0124] This event queue mechanism ensures that the processing of attribute changes will not interrupt the normal execution of the main loop, avoiding uncontrollable loop nesting and timing conflicts.
[0125] When the system is first deployed, the initial state of the heterogeneous dual-network topology is empty, with no nodes or edges. At this time, the system needs to execute a cold start strategy.
[0126] During the cold start phase, the system traverses the entire data dictionary of the bank's core business system, creating a corresponding business node for each business field. Simultaneously, it traverses the ISO 20022 standard library, creating a corresponding standard node for each standard data element. The system calculates the initial mapping confidence between each pair of business nodes and standard nodes based on the joint semantic representation. The initial mapping confidence is calculated as a weighted fusion value of the similarity of the four semantic primitives in their respective dimensions. Taking business node B_001 (cust_nm) and standard node S_001 (CustomerName) as an example, the system calculates the similarity between their naming semantic primitives to be 0.95 (both representing customer names), the type semantic primitive similarity to be 0.90 (both character-type), the context semantic primitive similarity to be 0.85 (both used in customer information scenarios), and the relational semantic primitive similarity to be 0.70 (cust_nm and id_num co-occur in the same table, and CustomerName and IdentificationNumber are associated in the standard library). The four similarities were fused using the same weighting coefficients as in the structured decomposition stage, resulting in an initial mapping confidence score of 0.89.
[0127] The cold start strategy uses a cold start threshold as the criterion for establishing initial mapping edges. The cold start threshold is higher than the preset growth threshold in the normal operation mode. In this embodiment, the preset growth threshold in the normal operation mode is 0.75, and the cold start threshold is set to 0.85. The higher threshold ensures that the initial mapping edges established in the cold start phase have higher credibility, reducing the workload of subsequent manual review.
[0128] When the initial mapping confidence exceeds the cold start threshold, the system generates the corresponding initial forward mapping edge and initial reverse mapping edge. The initial mapping confidence of B_001 and S_001 is 0.89, which exceeds the cold start threshold of 0.85. The system generates a forward mapping edge from B_001 to S_001 and a reverse mapping edge from S_001 to B_001, and marks these two mapping edges as pending review and adds them to the dual-network heterogeneous topology.
[0129] For mapping edges that are pending review, the system suspends pruning and reconnection operations. This means that these mapping edges will not be automatically pruned or reconnected until they are manually reviewed and confirmed, in order to avoid the initial mapping relationships during the cold start phase being mistakenly manipulated by the system.
[0130] The bank's data governance team reviews each mapping edge in the pending review stage. The review process involves the system generating a review workbench page that lists all mapping edges awaiting review, along with their corresponding business fields, standard data elements, and initial mapping confidence levels. Governance personnel verify the correctness of each mapping relationship and can choose to confirm, correct, or reject. Confirmation indicates the mapping edge is correct, correction indicates the mapping target needs adjustment, and rejection indicates the mapping relationship is invalid.
[0131] After the administrator confirms that the B_001→S_001 mapping edge is correct, the system converts the mapping edge into a formal mapping edge and removes the pause status of the pruning and reconnection operations for that mapping edge. If the administrator corrects a mapping edge, the system regenerates the mapping edge according to the corrected mapping target and converts it into a formal mapping edge. If the administrator rejects a mapping edge, the system deletes the mapping edge from the topology.
[0132] As the review process continues, the number of formal mapping edges in the heterogeneous dual-network topology gradually increases. When the number of formal mapping edges reaches the preset cold start completion threshold, the system determines that the cold start is complete and switches to the normal operating mode. In this embodiment, the preset cold start completion threshold is 1000 formal mapping edges. This threshold is set according to the size of the bank's business data dictionary—the core business system involves approximately 1200 core fields related to regulatory reporting and data sharing, and 1000 mapping edges cover 83% of these core fields, which is sufficient to support daily business operations. After the system switches to the normal operating mode, the adaptive reconstruction operation uses a preset growth threshold instead of the cold start threshold, and the system begins to automatically perform pruning, growth, and reconnection operations, while the feedback loop begins to operate.
[0133] The aforementioned method steps can be executed by a meta-model-driven bidirectional mapping system for financial data standards. This system comprises five functional modules, each corresponding to one of the five core steps in the aforementioned method. The specific implementation methods of each module are described below in conjunction with the aforementioned banking application scenario.
[0134] The structured decomposition module is responsible for performing structured decomposition of multi-source heterogeneous financial data at the meta-model dimension, and pre-establishing intra-standard associations between data elements within the financial standard and standard-business associations between financial standard data elements and business system data dictionaries based on the decomposition results.
[0135] This module takes as input the data dictionaries of the bank's core business system, credit management system, and anti-money laundering system, as well as the structured description files of the ISO 20022 standard library. It reads the table structure definitions, field definitions, constraint definitions, and foreign key definitions from each data dictionary and uses a metamodel parsing engine to uniformly parse the attributes of each data element. The metamodel parsing engine maps the descriptive information of each data element to a unified metamodel structure according to a predefined metamodel framework. Specifically, this includes normalizing and converting different representations of data types in different systems (e.g., mapping VARCHAR in DB2 and VARCHAR2 in Oracle to string types within a unified framework). The engine outputs a unified metamodel dataset, in which each data element is assigned a globally unique identifier.
[0136] The structured decomposition module incorporates a semantic primitive extractor. This extractor automatically analyzes the name, data type, table structure, system location, and relationships with other data elements for each data element. Based on the aforementioned four-fold semantic primitive extraction rules, the semantic primitive extractor sequentially generates named semantic primitives, type semantic primitives, context semantic primitives, and relational semantic primitives. These four primitives are then input into a semantic encoder to generate a joint semantic representation. Specifically, the semantic primitive extractor is implemented as follows: named semantic primitives are generated using a word segmenter and stemmer based on a financial domain dictionary; type semantic primitives are generated using a feature encoder based on data type; context semantic primitives are generated using embedding vectors based on business scenario tags; and relational semantic primitives are generated using a relational encoder based on a graph attention mechanism.
[0137] This module also includes a pre-built association submodule. This submodule performs the aforementioned three-stage progressive association pre-building: traversing foreign key dependencies in the database schema to generate the first candidate association edge; analyzing co-occurrence frequency and field naming patterns to generate the second and third candidate association edges; accessing the historical version database of financial standards to obtain naming evolution trajectories to generate the fourth candidate association edge; and extracting upstream and downstream flow paths from business flow logs to generate the fifth candidate association edge. The pre-built association submodule assigns corresponding confidence levels to the five types of candidate association edges, calculates the comprehensive association confidence level through weighted fusion, and marks candidate association edges exceeding a preset association threshold as semantic association edges to be added. This submodule outputs a structured dataset, stored in a graph structure, containing all data element nodes and their joint semantic representation, context anchor vectors, hierarchical labeling information, and a list of semantic association edges to be added.
[0138] The dual-network modeling module constructs a heterogeneous dual-network topology containing forward mapping networks and reverse mapping networks based on the structured dataset output by the structured decomposition module.
[0139] This module first creates three types of nodes. Standard nodes read the definition, attributes, and metamodel inheritance relationships of each standard data element from the structured description file of the ISO 20022 standard library, creating a standard node for each standard data element in the topology. Business nodes read the definition and attributes of each field from the data dictionary of each business system, creating a business node for each field in the topology. Metamodel nodes read the list of abstract concepts in the metamodel layer from the preset metamodel framework, creating a metamodel node for each abstract concept in the topology. After node creation, the dual-network modeling module establishes semantic inheritance edges based on the inheritance relationship information in the structured dataset—for each standard node, it determines its corresponding metamodel node and creates directed edges from the standard node to the metamodel node; for each business node, it determines its corresponding metamodel node and creates directed edges from the business node to the metamodel node. After node creation and semantic inheritance edge establishment are completed, the module, based on the list of semantic association edges to be added, formally adds each semantic association edge to the topology as a semantic association edge.
[0140] This module also includes a mapping edge initialization submodule. This submodule calculates name similarity, semantic similarity based on naming semantic primitives and context semantic primitives, and structural similarity based on type semantic primitives for each business node and each standard node. It then weights and fuses these three similarities into a comprehensive score, which serves as the initial forward mapping weight for the forward mapping edge. Simultaneously, this submodule generates corresponding reverse mapping edges, sets their initial reverse mapping weights to the mirror values of the forward mapping weights, and configures a preset adjustable redundancy parameter for each reverse mapping edge. This submodule establishes a consistency bridging edge between the forward and reverse mapping networks. Each bridging edge is associated with a pair of corresponding forward and reverse mapping edges, and its consistency tolerance threshold is recorded. This module outputs a complete dual-network heterogeneous topology data structure, including a set of nodes, a set of edges, and the weight and constraint parameters for each edge.
[0141] The dynamic analysis module performs dynamic bidirectional consistency analysis on a heterogeneous dual-network topology. This module includes a consistency calculation engine, a context adapter, and an alert trigger.
[0142] The consistency calculation engine continuously calculates a bidirectional consistency metric for each mapping edge. Based on the forward and reverse mapping weights, the engine calculates the bidirectional consistency metric according to the aforementioned formula. The consistency calculation engine also simultaneously calculates structural consistency, semantic consistency, and temporal consistency components, and weights and fuses these three components into a comprehensive consistency score.
[0143] The context adapter extracts the context vector of the current business context from the context anchor vector of the data element. The context adapter listens for mapping requests from the business system, extracts the business scenario identifier and field usage pattern from the requests, and generates the corresponding context vector. The context adapter weights this vector with the initial values of the mapping weights to generate the context-adapted weight values. When the context vector changes, the context adapter recalculates the bidirectional consistency metrics of the affected mapping edges and checks whether a consistency calibration operation has been triggered.
[0144] The early warning trigger tracks the change trajectory of the bidirectional consistency index of each mapped edge within a continuous time window. It maintains a time-series database storing daily snapshots of the mapped edge status. The trigger periodically calculates the single-step decline rate and cumulative decline rate, marking mapped edges exceeding thresholds as consistency-sensitive or consistency-declining edges and triggering corresponding repair warnings. Simultaneously, the trigger monitors the overall connectivity index of the heterogeneous dual-network topology, calculating the coupling strength between the forward and reverse mapped networks. When the proportion of consistency-declining edges exceeds a preset proportion threshold or the coupling strength falls below a preset coupling strength threshold, the trigger determines that the topology has entered a critical mismatch precursor state and outputs a reconstruction trigger signal.
[0145] The adaptive reconfiguration module starts upon receiving a reconfiguration trigger signal. This module includes a protection policy executor, a pruning executor, a growth executor, and a reconnection executor.
[0146] The protection policy executor configures a multi-level protection mechanism for the heterogeneous dual-network topology. This executor scans the attributes of all mapped edges in the topology, identifying mapped edges within the first-level protection zone (mapped edges corresponding to standard nodes commonly referenced by more than a preset number of business systems, core mapped edges involved in regulatory reporting, and mapped edges corresponding to specific financial standard versions), mapped edges within the second-level protection zone (mapped edges frequently queried within a preset time period), and mapped edges within the third-level protection zone (all other mapped edges). The protection policy executor sets different pruning permissions for each level of protection zone.
[0147] The pruning actuator monitors the bidirectional consistency index of each mapped edge. When the bidirectional consistency index of a mapped edge remains below a preset low threshold for an extended period exceeding a preset number of detection cycles, the pruning actuator performs the corresponding pruning operation based on the protection level of that mapped edge. For mapped edges in Level 3 protected areas, pruning and deletion are performed directly; for mapped edges in Level 2 protected areas, a consistency impact range assessment is first triggered, and pruning is only performed if the assessment results allow; pruning is prohibited for mapped edges in Level 1 protected areas.
[0148] The growth executor detects events such as updates to financial standard versions and the addition of new fields in the business system. When a new node is detected, the growth executor calculates the confidence of candidate mappings between the new node and existing nodes. When the confidence exceeds a preset growth threshold, it automatically generates corresponding candidate forward and reverse mapping edges and adds them to the topology.
[0149] The reconnection executor detects bidirectional conflict scenarios. When it detects that a business node points to a standard node via a forward mapping edge, but the standard node does not have a reverse mapping edge pointing to the business node, the reconnection executor triggers a reverse mapping edge completion verification. It determines whether a valid many-to-one mapping relationship exists by calculating the business semantic similarity. If a conflict is found, the reconnection executor updates the set of reverse mapping edges of the standard node to include the business node, so that the standard node is simultaneously reverse-mapped to multiple business nodes.
[0150] The feedback evolution module collects multi-dimensional feedback data throughout the entire mapping process and adjusts the disassembly rules of the structured disassembly module, the structural parameters of the dual-network modeling module, the detection threshold of the dynamic analysis module, and the triggering conditions of the adaptive reconstruction module in a closed loop based on the feedback data.
[0151] The feedback evolution module comprises four collection sub-modules. The explicit feedback collection sub-module provides a manual review interface, recording the review actions (confirmation, correction, or rejection) of data governance personnel and converting them into scoring signals. The implicit feedback collection sub-module monitors the query logs of the business system for mapped edges, calculating the query count and adoption rate for each mapped edge. The consistency feedback collection sub-module reads the time-series trends of bidirectional consistency indicators from the time-series database of the dynamic analysis module. The external feedback collection sub-module subscribes to the notification service officially released by financial standards, automatically receiving version change and regulatory rule update information.
[0152] The feedback evolution module also includes an adjustment executor. This executor employs a closed-loop coordination strategy of "adjustment-verification-calibration"—adjusting the forward mapping weights based on implicit feedback and the reverse mapping weights based on explicit feedback. After each adjustment, the bidirectional consistency index is recalculated; if it exceeds the tolerance, a calibration operation is triggered. The adjustment executor dynamically adjusts the consistency tolerance threshold based on the overall distribution of the bidirectional consistency index. Furthermore, the adjustment executor adjusts the fusion weights of each semantic primitive in the structured decomposition module based on the disambiguation trigger frequency and correction rate, and adjusts the pruning threshold and detection cycle number of the adaptive reconstruction module based on the pruning error rate and missed pruning rate. The adjustment executor accurately attributes feedback data according to the business system dimension, financial standard version dimension, and user operation dimension, ensuring that feedback signals are allocated to the corresponding mapping edge set.
[0153] The five modules described above work together to implement a meta-model-driven bidirectional mapping system for financial data standards. Deployed within the bank's data governance platform, this system operates as an independent microservice, interacting with core business systems, credit management systems, anti-money laundering systems, and the ISO 20022 standard library via application programming interfaces (APIs). Data transfer between modules employs a message queue mechanism for asynchronous decoupling, ensuring that each module can be independently expanded and maintained.
[0154] The above-described specific implementation describes the entire process of a complete metamodel-driven bidirectional mapping system for financial data standards, from cold start to regular operation and continuous evolution. This process begins with a cold start strategy for the initial deployment of the spatiotemporal topology. Topology initialization is completed through the establishment of high-confidence initial mapping edges and manual review and confirmation. Subsequently, the system enters a regular operation mode. In this mode, structured decomposition, dual-network construction, dynamic analysis, adaptive reconstruction, and feedback loop constitute a continuously operating five-order loop. The association pre-construction stage provides high-quality candidate association edge inputs for topology construction. The event queue mechanism ensures that changes in external attributes are processed promptly without interfering with the main loop. The feedback loop drives continuous optimization of parameters in each stage of the system. The above five modules can be deployed as independent microservices, communicating through application programming interfaces (APIs). The system runs in a hardware environment equipped with processors, memory, and network interfaces, interacting with business systems through data exchange interfaces. This specific implementation is merely a preferred embodiment of the present invention. Those skilled in the art can make various modifications and variations after reading this specification without departing from the spirit and scope of the invention. The scope of protection of the present invention is defined by the appended claims.
Claims
1. A meta-model-driven bidirectional mapping method for financial data standards, characterized in that, Includes the following steps: Multi-source heterogeneous financial data is decomposed into a structured dataset at the meta-model dimension. Based on the decomposition results, standard intra-data elements within the financial standard and standard-business data elements between the financial standard and the business system data dictionary are pre-established to form a structured dataset with related relationships. Based on the structured dataset, a dual-network heterogeneous topology containing a forward mapping network and a reverse mapping network is constructed. Forward mapping edges and corresponding forward mapping weights are configured for the forward mapping network, and reverse mapping edges and corresponding reverse mapping weights are configured for the reverse mapping network. At the same time, a consistency bridge edge is established between the forward mapping network and the reverse mapping network to apply consistency tolerance constraints to the forward mapping weights and the reverse mapping weights. Dynamic bidirectional consistency analysis is performed on the heterogeneous dual-network topology. The dynamic bidirectional consistency analysis includes: calculating structural consistency components in the structural dimension of the heterogeneous dual-network topology, calculating semantic consistency components in the semantic dimension, and tracking the temporal change trajectory of bidirectional consistency indices over a continuous time window; calculating the bidirectional consistency index of each mapping edge based on the structural consistency components, the semantic consistency components, and the temporal change trajectory, and detecting mapping drift and bidirectional conflicts based on the temporal change trend of the bidirectional consistency indices to identify precursors of critical mismatch. Based on the analysis results of the dynamic bidirectional consistency analysis, an adaptive reconstruction operation is performed on the heterogeneous topology of the dual network. The adaptive reconstruction operation includes pruning failed mapping edges, growing mapping edges for newly emerging mapping requirements, reconnecting mapping edges with bidirectional conflicts, and implementing protection strategies for core mapping edges that meet preset protection conditions. Multidimensional feedback data is collected throughout the entire mapping process, and based on this multidimensional feedback data, the disassembly rules of the structured decomposition, the structural parameters of the heterogeneous dual-network topology, the detection threshold of the dynamic bidirectional consistency analysis, and the triggering conditions of the adaptive reconstruction operation are adjusted in a closed loop to achieve continuous system evolution.
2. The method for bidirectional mapping of financial data standards based on meta-model driven according to claim 1, characterized in that, The structured decomposition of multi-source heterogeneous financial data at the meta-model dimension specifically includes: For each data element in the multi-source heterogeneous financial data, four semantic primitives are extracted, including naming semantic primitives, type semantic primitives, context semantic primitives and relational semantic primitives, and corresponding semantic representations are generated respectively. The semantic representations are then merged into a joint semantic representation. The naming semantic primitive is used to represent the core business vocabulary information of the data element, the type semantic primitive is used to represent the structural feature information of the data element, the context semantic primitive is used to represent the business scenario information of the data element, and the relational semantic primitive is used to represent the potential association information between data elements. When the similarity of the joint semantic representation of any two data elements falls within a preset fuzzy range, a joint disambiguation mechanism is triggered. The joint disambiguation mechanism comprehensively determines the semantic relationship between the two data elements based on the similarity of the two data elements in each semantic primitive dimension. A context anchor vector is maintained for each data element. The context anchor vector records the business context distribution of the data element in the historical mapping process. The semantic drift of the data element between different mapping times is calculated based on the joint semantic representation. When the semantic drift exceeds a preset drift threshold, a semantic drift alarm is triggered. Based on the joint semantic representation and the context anchoring vector, the decomposed financial data elements are differentiated and labeled according to the core layer, the extension layer and the adaptation layer. The core layer consists of immutable basic data types and general business elements in financial standards, the extension layer consists of variable elements that support diverse business scenarios, and the adaptation layer consists of adaptation elements that meet the personalized needs of specific institutions. Furthermore, the hierarchy of the differentiated tags is dynamically adjusted based on the semantic drift and mapping usage.
3. The method for bidirectional mapping of financial data standards based on meta-model driven according to claim 2, characterized in that, The construction of the dual-network heterogeneous topology, which includes a forward mapping network and a reverse mapping network, specifically includes: Standard nodes, business nodes, and metamodel nodes are defined as node types in the heterogeneous dual-network topology. The standard node corresponds to a data element in the financial standard library, the business node corresponds to a field in the business system data dictionary, and the metamodel node corresponds to an abstract concept in the metamodel layer. The standard node inherits from the metamodel node, and the business node inherits from the metamodel node. Multiple standard nodes and multiple business nodes that inherit from the same metamodel node have a shared semantic basis. The forward mapping edge is defined as a directed edge from the business node to the standard node; the reverse mapping edge is defined as a directed edge from the standard node to the business node; the semantic association edge is defined as the synonym, near-synonym, and hierarchical semantic relationship edge between the standard nodes or between the business nodes; and the constraint edge is defined as the mutual exclusion constraint and dependency constraint relationship edge in the financial business rules. The initial forward mapping weight of the forward mapping edge is determined based on a comprehensive score of name similarity, semantic similarity calculated from the naming semantic primitive and the context semantic primitive, and structural similarity calculated from the type semantic primitive. The initial reverse mapping weight of the reverse mapping edge is set as the mirror value of the forward mapping weight, and the reverse mapping edge is given a preset adjustable redundancy to allow bidirectional non-perfect symmetry. A consistency bridge edge is established between the forward mapping network and the reverse mapping network. The consistency bridge edge is used to constrain the absolute value of the weight difference between corresponding forward mapping edges and reverse mapping edges to not exceed a preset consistency tolerance threshold. The consistency tolerance threshold is a dynamically maintained constraint condition, which is re-verified and re-executed after each change in mapping weights. When the absolute value of the weight difference between the forward mapping edge and the reverse mapping edge exceeds the consistency tolerance threshold, a bidirectional consistency calibration operation is triggered. The bidirectional consistency calibration operation is as follows: the weighted average of the forward mapping weights and the reverse mapping weights replaces the original weights, so that the forward mapping weights and the reverse mapping weights satisfy the consistency tolerance threshold again after calibration. Furthermore, when the first standard node pointed to by any forward mapping edge in the forward mapping network and the first business node pointed to by the corresponding reverse mapping edge in the reverse mapping network are traced back to a common semantic base through the meta-model node, the consistency tolerance threshold between the forward mapping edge and the reverse mapping edge is tightened by a preset amount; when there is no common semantic base between the standard node and the business node, the consistency tolerance threshold is widened by a preset amount to adapt to heterogeneous mapping scenarios across semantic bases.
4. The method for bidirectional mapping of financial data standards based on meta-model driven according to claim 2, characterized in that, The dynamic bidirectional consistency analysis specifically includes: Based on the forward mapping weights and the reverse mapping weights, a bidirectional consistency index is calculated between each forward mapping edge and its corresponding reverse mapping edge. The bidirectional consistency index is used to quantify the degree of mapping symmetry between the forward mapping network and the reverse mapping network. Based on the bidirectional consistency index, the structural consistency component, semantic consistency component, and temporal consistency component of the heterogeneous dual-network topology in the structural dimension, the semantic consistency component, and the temporal consistency component in different time slices are calculated respectively. The structural consistency component, the semantic consistency component, and the temporal consistency component are then weighted and fused into a comprehensive consistency score. The context vector of the current business context is extracted from the context anchor vector. The context anchor vector records the business context distribution of data elements in the historical mapping process. The current business context is the business scenario identifier corresponding to the current mapping operation and the field usage mode under that scenario. The context vector is used as the condition parameter of the forward mapping weight and the reverse mapping weight. By weighting the context vector with the initial values of the forward mapping weight and the reverse mapping weight respectively, the mapping weight between the same business node and the same standard node has different weight values under different business contexts. When the context vector changes, the bidirectional consistency index of the affected mapping edge is recalculated, and the consistency fluctuation caused by the context change is marked. The mapping weight change caused by the context vector is the object of the consistency tolerance constraint. After recalculating the bidirectional consistency index, if the absolute value of the difference between the forward mapping weight and the reverse mapping weight exceeds the consistency tolerance threshold, the bidirectional consistency calibration operation described in claim 3 is triggered, so that the mapping weight meets the consistency tolerance threshold again after completing the context adaptation. Track the change trajectory of the bidirectional consistency index of each mapping edge within a continuous time window. When the decrease of the bidirectional consistency index relative to the previous moment exceeds a preset single-step decrease threshold, mark the mapping edge as a consistency sensitive edge. When the cumulative decrease of the bidirectional consistency index within a preset cumulative window exceeds a preset cumulative decrease threshold, mark the mapping edge as a consistency decay edge and trigger a consistency repair warning. The overall connectivity index of the heterogeneous dual-network topology is monitored. The overall connectivity index includes the coupling strength between the forward mapping network and the reverse mapping network. The coupling strength is calculated based on the number of consistent bridging edges and the mapping density between the nodes on both sides of the consistent bridging edge. When the coupling strength is lower than a preset coupling strength threshold, a dual-network coupling structure warning is triggered. When the proportion of the consistency decay edge exceeds a preset proportion threshold, or the coupling strength is lower than the preset coupling strength threshold, the dual-network heterogeneous topology is determined to have entered a critical mismatch precursor state, and a reconstruction trigger signal is output.
5. The method for bidirectional mapping of financial data standards based on meta-model driven according to claim 1, characterized in that, The pruning, growth, reconnection, and protection strategies implemented in the adaptive reconstruction operation specifically include: A multi-level protection mechanism is configured for the heterogeneous dual-network topology. This mechanism divides the protection level into Level 1, Level 2, and Level 3 protection zones according to the importance of the mapping edges. Mapping edges within the Level 1 protection zone are those corresponding to standard nodes commonly referenced by more than a preset number of business systems, core mapping edges involved in regulatory reporting, and mapping edges corresponding to specific financial standard versions. No automatic pruning operations are allowed on mapping edges within the Level 1 protection zone. Mapping edges within the Level 2 protection zone are those frequently queried within a preset time period. Before performing pruning operations on mapping edges within the Level 2 protection zone, a consistency impact assessment is required, and pruning is only allowed if the assessment result is below a preset impact threshold. Mapping edges within the Level 3 protection zone are all mapping edges other than those in the Level 1 and Level 2 protection zones. Pruning operations are directly performed on mapping edges within the Level 3 protection zone. When the bidirectional consistency index of any mapping edge is continuously lower than the preset low threshold and the duration exceeds the preset number of detection cycles, the corresponding pruning operation is performed according to the protection level to which the mapping edge belongs. When a financial standard version update or a new field is added to the business system, the confidence score of the candidate mapping is calculated based on the comprehensive score of name similarity, semantic similarity and structural similarity between the new node and the existing node. When the confidence score of the candidate mapping exceeds the preset growth threshold, the corresponding candidate forward mapping edge and candidate reverse mapping edge are automatically generated and added to the dual-network heterogeneous topology. When it is detected that any service node points to the first standard node via the forward mapping edge, and the first standard node does not have any reverse mapping edge pointing to the arbitrary service node, it is determined that a mapping to be verified is missing, and reverse mapping edge completion verification is triggered. The reverse mapping edge completion verification is as follows: calculate the business semantic similarity between the arbitrary service node and each service node pointed to by all existing reverse mapping edges of the first standard node. If all similarities are lower than a preset many-to-one effective threshold, it is determined to be the bidirectional conflict. If there is at least one service node whose business semantic similarity with the arbitrary service node is not lower than the preset many-to-one effective threshold, it is determined that there is a valid many-to-one mapping relationship between the arbitrary service node and the first standard node, and the reconnection operation is not triggered. When the bidirectional conflict is determined to occur, the reconnection operation is triggered. The reconnection operation is as follows: the set of reverse mapping edges of the first standard node is updated to include the reverse mapping edges of the arbitrary business node, so that the first standard node is simultaneously reverse mapped to multiple business nodes including the arbitrary business node.
6. The method for bidirectional mapping of financial data standards based on meta-model driven according to claim 1, characterized in that, The process involves collecting multi-dimensional feedback data throughout the entire mapping process, and then adjusting the structured decomposition rules, the structural parameters of the heterogeneous dual-network topology, the detection threshold of the dynamic bidirectional consistency analysis, and the triggering conditions of the adaptive reconstruction operation in a closed loop based on this multi-dimensional feedback data. Specifically, this includes: The multidimensional feedback data includes: explicit feedback consisting of the results of manual benchmarking and review; implicit feedback consisting of the frequency of actual querying and use of the mapping edge and the adoption rate of the mapping results by the business system; consistency feedback consisting of the time-series change trend of the bidirectional consistency indicators; and external feedback consisting of official version change notices of financial standards and updates to regulatory rules. The adjustment adopts a closed-loop coordination strategy of "adjustment-verification-calibration": the forward mapping weight is adjusted according to the product of usage frequency and adoption rate in the implicit feedback, and the reverse mapping weight is adjusted according to the manual confirmation operation in the explicit feedback; after each adjustment, the bidirectional consistency index of the mapping edge corresponding to the adjusted weight is recalculated. If the absolute value of the difference between the adjusted forward mapping weight and the reverse mapping weight does not exceed the consistency tolerance threshold, the adjustment is confirmed to be effective; if the absolute value of the difference between the adjusted forward mapping weight and the reverse mapping weight exceeds the consistency tolerance threshold, the bidirectional consistency calibration operation described in claim 3 is triggered, and the weighted average of the forward mapping weight and the reverse mapping weight is used to replace the original weight, so that the adjusted forward mapping weight and the reverse mapping weight meet the consistency tolerance threshold again. Furthermore, the tolerance threshold of the consistency tolerance constraint is dynamically adjusted according to the overall distribution of the bidirectional consistency index. When the overall consistency level of the system increases, the tolerance threshold is tightened, and when the overall consistency level of the system decreases, the tolerance threshold is relaxed, so as to achieve an adaptive balance between feedback-driven optimization and consistency constraint maintenance. Furthermore, the multidimensional feedback data is allocated to the mapping edges associated with the corresponding business system according to the business system dimension, to all mapping edges associated with the corresponding version according to the financial standard version dimension, and to the operation mapping edges according to the user operation dimension.
7. The method for bidirectional mapping of financial data standards based on meta-model driven according to claim 1, characterized in that, The pre-established intra-standard associations between data elements within the financial standard and the standard-business associations between the financial standard data elements and the business system data dictionary specifically include: A three-stage progressive association pre-construction is performed, which includes a static structure association pre-construction stage, a semantic evolution association pre-construction stage, and a business flow association pre-construction stage; In the static structure association pre-construction stage, a first candidate association edge is established based on the foreign key dependency relationship of data elements in the database schema, a second candidate association edge is established based on the co-occurrence relationship of data elements in the database table, and a third candidate association edge is established based on the matching pattern of the naming prefix and suffix of data element fields. In the semantic evolution association pre-construction stage, the naming evolution trajectory of financial data element names in the historical version evolution process is obtained. The naming evolution trajectory records multiple naming expressions of the same financial concept in different standard versions or different business systems. Based on the naming evolution trajectory, cross-version semantic association edges are established for multiple data elements with different names but traced back to the same financial concept, as the fourth candidate association edge. In the business flow association pre-construction stage, cross-system association edges are established based on the flow path of data elements in the upstream and downstream systems of financial business, serving as the fifth candidate association edge; The first candidate association edge, the second candidate association edge, the third candidate association edge, the fourth candidate association edge, and the fifth candidate association edge are respectively assigned corresponding static confidence, co-occurrence confidence, naming matching confidence, semantic evolution confidence, and business flow confidence. The static confidence, co-occurrence confidence, naming matching confidence, semantic evolution confidence, and business flow confidence are then weighted and merged into a comprehensive association confidence. When the overall association confidence exceeds the preset association threshold, the corresponding candidate association edge is confirmed and marked as a semantic association edge to be added; Furthermore, when there are at least three candidate association edges between any two data elements, the weighted fusion weight of the comprehensive association confidence is automatically increased, including: multiplying the comprehensive association confidence by a preset multi-source verification coefficient, or adding the comprehensive association confidence to a preset multi-source verification score, so that the comprehensive association confidence is more likely to exceed the preset association threshold after adjustment. In the subsequent construction of the dual-network heterogeneous topology, the semantic association edges to be added are formally added to the dual-network heterogeneous topology as semantic association edges.
8. The method for bidirectional mapping of financial data standards based on meta-model driven according to claim 3, characterized in that, The heterogeneous dual-network topology also includes semantic inheritance edges, which are directed edges from the standard node to the meta-model node and directed edges from the service node to the meta-model node. The standard node and the service node share the semantic basis by inheriting from the same meta-model node. Based on the semantic inheritance edge, a bidirectional index path is established between the standard node and the business node. The bidirectional index path is as follows: starting from any standard node, tracing along the semantic inheritance edge to the metamodel node, and then descending along the reverse semantic inheritance edge of the metamodel node to all business nodes that inherit from the same metamodel node, forming a first directional index from the standard node to the business node; starting from any business node, tracing along the semantic inheritance edge to the metamodel node, and then descending along the reverse semantic inheritance edge of the metamodel node to all standard nodes that inherit from the same metamodel node, forming a second directional index from the business node to the standard node. When the attribute of any standard node changes, instead of immediately performing a re-verification, the attribute change event is recorded in the event queue, and all business nodes affected by the attribute change of the standard node are located along the first direction. All business nodes and their associated mapping edges are marked as pending verification, and the execution position of the current main loop is recorded. When the field attribute of any business node changes, the re-validation is not performed immediately. Instead, the field attribute change event is recorded in the event queue, and all standard nodes affected by the field attribute change of the business node are located along the second direction index. All standard nodes and their associated mapping edges are marked as pending validation, and the execution position of the current main loop is recorded. After the current step of the current main loop is completed, check if the event queue is empty. If the event queue is not empty, pause the next step of the main loop and prioritize performing consistency re-verification on all nodes of the state to be verified and their associated mapping edges. After the consistency re-verification is completed, clear the corresponding processed events in the event queue, clear the state to be verified, and determine the subsequent path based on the re-verification result. If the re-verification result indicates that there is a bidirectional consistency error, trigger the adaptive reconstruction operation described in claim 5, and then resume the execution of the next step of the current main loop. If the re-verification result indicates that the bidirectional consistency is normal, then the execution of the next step of the current main loop will resume directly.
9. The method for bidirectional mapping of financial data standards based on meta-model driven according to claim 2, characterized in that, The initial state of the dual-network heterogeneous topology is empty, and the method also includes a cold start strategy when the system is first deployed: Under the cold start strategy, the initial mapping confidence between each business node and each standard node is calculated based on the joint semantic representation. The initial mapping confidence is a weighted fusion value of the similarity of the named semantic primitive, the type semantic primitive, the context semantic primitive, and the relational semantic primitive in their respective dimensions. The cold start strategy uses a cold start threshold as the criterion for establishing the initial mapping edge. The cold start threshold is higher than the preset growth threshold in the normal operation mode. When the initial mapping confidence exceeds the cold start threshold, the corresponding initial forward mapping edge and initial reverse mapping edge are generated, and the initial forward mapping edge and the initial reverse mapping edge are marked as pending review and added to the dual-network heterogeneous topology. For the mapped edges that are in the pending review state, suspend the pruning operation and the reconnection operation, and wait for manual review and confirmation; After manual review and confirmation, the mapping edge in the pending review state is converted into a formal mapping edge, and the paused state of the pruning operation and the reconnection operation is lifted. When the number of formally mapped edges in the heterogeneous dual-network topology reaches the preset cold start completion threshold, the system is determined to have completed the cold start and switches to the normal operation mode; in the normal operation mode, the adaptive reconstruction operation uses the preset growth threshold instead of the cold start threshold.
10. A meta-model-driven bidirectional mapping system for financial data standards, characterized in that, The system employs the meta-model-driven bidirectional mapping method for financial data standards as described in any one of claims 1 to 9, wherein the system comprises: The structured decomposition module is used to perform structured decomposition of multi-source heterogeneous financial data at the meta-model dimension, and pre-establish standard intra-standard associations between data elements within the financial standard and standard-business associations between the financial standard data elements and the business system data dictionary based on the decomposition results, so as to form a structured dataset with association relationships. The dual-network modeling module is used to construct a dual-network heterogeneous topology containing a forward mapping network and a reverse mapping network based on the structured dataset, and to configure forward mapping edges and corresponding forward mapping weights for the forward mapping network, and to configure reverse mapping edges and corresponding reverse mapping weights for the reverse mapping network. At the same time, a consistency bridge edge is established between the forward mapping network and the reverse mapping network to apply consistency tolerance constraints to the forward mapping weights and the reverse mapping weights. A dynamic analysis module is used to perform dynamic bidirectional consistency analysis on the heterogeneous dual-network topology. The dynamic bidirectional consistency analysis includes: calculating structural consistency components in the structural dimension of the heterogeneous dual-network topology, calculating semantic consistency components in the semantic dimension, and tracking the temporal change trajectory of bidirectional consistency indicators in a continuous time window; calculating the bidirectional consistency indicator of each mapping edge based on the structural consistency components, the semantic consistency components, and the temporal change trajectory, and detecting mapping drift and bidirectional conflicts based on the temporal change trend of the bidirectional consistency indicators to identify precursors of critical mismatch. The adaptive reconstruction module is used to perform adaptive reconstruction operations on the heterogeneous topology of the dual network based on the analysis results of the dynamic bidirectional consistency analysis. The adaptive reconstruction operations include pruning failed mapping edges, growing mapping edges for newly emerging mapping requirements, reconnecting mapping edges with bidirectional conflicts, and implementing protection strategies for core mapping edges that meet preset protection conditions. The feedback evolution module is used to collect multi-dimensional feedback data throughout the mapping process, and to make closed-loop adjustments to the disassembly rules of the structured disassembly module, the structural parameters of the dual-network modeling module, the detection threshold of the dynamic analysis module, and the triggering conditions of the adaptive reconstruction module based on the multi-dimensional feedback data, so as to achieve continuous evolution of the system.
