Multimodal financial regulation atlas dynamic construction updating method

Abstract

The application discloses a kind of multi-modal financial regulations atlas dynamic construction updating method, it is related to computer processing technical field, including the following steps: obtaining multi-modal financial regulations data, the multi-modal financial regulations data at least includes text modal data, table modal data and image modal data;The knowledge-enhanced multi-granularity interactive cross-modal entity linking is executed to the multi-modal financial regulations data, to identify and link the financial regulations entity in different modal data, generate initial financial regulations atlas.The application aligns the regulations entity in financial text, table and image by multi-granularity interactive cross-modal entity linking and field adaptation mechanism, improves cross-modal semantic fusion precision, and based on the incremental distillation strategy of timeliness perception, combined with hierarchical sorting and version attenuation constraint, effectively suppresses catastrophic forgetting in knowledge update, ensures that new and old regulations are logically rigorous.

Description

Technical Field

[0001] This invention relates to the field of computer processing technology, and in particular to a method for dynamically constructing and updating a multimodal financial regulatory map. Background Technology

[0002] With the increasing complexity of the financial regulatory system and the frequent revisions of regulations, financial institutions face the challenge of processing massive amounts of multi-source, heterogeneous regulatory data in business scenarios such as compliance review, risk control, and regulatory reporting. Financial regulatory data not only includes plain text provisions and explanations but also exists extensively in multimodal formats such as regulatory reports with complex row and column relationships and scanned images with signatures and annotations. These different modalities of data collectively constitute a complete semantic expression of regulations. For example, textual provisions define compliance obligations, tables refine quantitative indicator thresholds, and images preserve the original document's format and validity indicators. However, traditional information processing methods typically only perform keyword searches or rule matching for a single text modality, severing the inherent semantic connections between cross-modal data. This makes it difficult to fully capture implicit constraints and chains of evidence during automated compliance analysis.

[0003] In existing technologies, some solutions have attempted to extract triples from text to construct a knowledge graph using natural language processing techniques, while a few studies have addressed information extraction from images or tables. However, these methods often employ a shallow "identify first, then align" strategy, meaning they extract entities independently within each modality and then perform simple matching based on rules, lacking multi-granular interaction and deep fusion at the feature level. On one hand, numerical constraints in tables and qualitative descriptions in text often have implicit conditional dependencies, making it impossible for simple string matching to establish cross-modal semantic links. On the other hand, financial regulations have a rigorous hierarchical structure and a complex terminology system, and existing cross-modal alignment models do not incorporate domain prior knowledge, resulting in limited accuracy in entity linking and the introduction of numerous noisy relationships into the initial knowledge graph, weakening the reliability of the knowledge base. Summary of the Invention

[0004] To address the aforementioned technical problems, the present invention provides a method for dynamically constructing and updating a multimodal financial regulatory map, comprising the following steps: Acquire multimodal financial regulatory data, which includes at least text modal data, tabular modal data, and image modal data; The multimodal financial regulatory data is subjected to knowledge-enhanced multi-granularity interactive cross-modal entity linking to identify and link financial regulatory entities in different modalities, generating an initial financial regulatory map. When new financial regulatory data is acquired, the initial financial regulatory map is dynamically updated based on an incremental distillation strategy to obtain an updated financial regulatory map.

[0005] Preferably, the implementation of knowledge-enhanced multi-granularity interactive cross-modal entity linking on the multimodal financial regulatory data includes: The text modal data is encoded using a pre-trained model in the financial field, and features are extracted from the table modal data and the image modal data using an OCR model to obtain a first feature set. By using text global-local interaction units and cross-modal fusion interaction units, multi-granularity interaction processing is performed on the encoding and the first feature set to generate a cross-modal soft matching probability distribution. Cross-modal feature fusion is performed based on the cross-modal soft matching probability distribution to obtain cross-modal links of financial regulatory entities; Furthermore, before performing the cross-modal feature fusion, the multimodal financial regulatory data is filtered for noise based on the cross-modal soft matching probability distribution to remove entity mention pairs with cross-modal confidence levels below a preset threshold.

[0006] Preferably, the global semantics and local lexical information of the text are fused, wherein the global semantics are obtained through the overall semantic encoding of the legal provisions, and the local lexical information is obtained through the context-enhanced encoding of financial terminology keywords in the legal provisions, and the two are fused by an adaptive gating mechanism. The cross-modal fusion interaction unit mines the implicit semantic association between text and vision through cross-modal projection and gating fusion strategies. The visual features include table structure features extracted from table modal data and document layout features extracted from image modal data.

[0007] Preferably, the incremental distillation strategy dynamically updates the initial financial regulatory map to obtain an updated financial regulatory map, including: Based on the breadth-first search algorithm, the newly added triples in the new financial regulations data are sorted between layers according to the topological distance between the newly added triples and the initial financial regulations graph. Then, the importance scores of each triple are calculated by node centrality and relation betweenness centrality to sort them within the layer, resulting in a hierarchical sorting result. Among them, the calculation of node centrality introduces the timeliness weight factor of the regulations clauses, and assigns higher centrality weight to the entity nodes corresponding to the latest revised regulations. Based on the hierarchical sorting results and the dynamically adjusted distillation loss function, the knowledge distillation of the newly added triples is integrated into the initial financial regulatory map. If an entity appears in multiple levels, the current representation and historical representation of the entity are constrained by the dynamically adjusted distillation loss function. The distillation loss function combines L1 and L2 regularization terms and dynamically adjusts the L2 regularization strength through the timeliness decay coefficient, so that the knowledge forgetting of older version entities is subject to stricter constraints. A two-stage training strategy is adopted. In the first stage, the parameters of the initial financial regulatory map are frozen and only the representation of the newly added triples is optimized. In the second stage, all parameters are fine-tuned to integrate new and old knowledge and generate an updated financial regulatory map. Furthermore, prior to the first stage, the newly added triples are subject to financial regulatory constraints verification, and new triples that do not conform to the logical structure of financial regulations are eliminated. The logical structure of financial regulations includes the hierarchical relationship between regulatory clauses and the constraint of referential integrity.

[0008] Preferably, the method further includes: after generating the initial financial regulatory map or the updated financial regulatory map, constructing a hybrid index structure based on a unified vector space, wherein the hybrid index structure includes a hierarchical navigation graph HNSW for approximate nearest neighbor search and an inverted index for keyword matching, and the unified vector space is jointly constructed by multimodal representations of financial regulatory entities, clauses and their relationship types; Furthermore, during the dynamic update process, the affected index entries in the hybrid index structure are updated synchronously based on the hierarchical sorting results to maintain the consistency between the index and the map data.

[0009] Preferably, the method further includes: performing consistency management on cross-modal financial regulatory graph data update operations based on ACID transaction principles; capturing data change events through an event-driven architecture, locating affected subgraphs using graph differencing algorithms, and performing incremental updates only on the affected subgraphs; and after the second stage of the two-stage training is completed, performing version consistency verification on the updated financial regulatory graph to verify the matching of the revision version number of the regulatory clauses with the version number of the associated references, and triggering a rollback operation if the verification fails.

[0010] Preferably, the method further includes: using the updated financial regulatory graph as a knowledge retrieval enhancement knowledge source, in response to user queries, retrieving relevant regulatory clauses and evidence chains from the knowledge source, and inputting them into a large language model to generate compliance analysis results; Furthermore, traceability information is embedded in the compliance analysis results. This traceability information includes at least the original text of the cited regulatory clauses, the identifiers of the associated graph nodes, and the version timestamp, in order to verify the traceability of the generated results.

[0011] The present invention has at least the following beneficial effects: By using a dynamic graph construction and update method for multimodal financial regulatory data, this approach effectively solves the problems of insufficient entity semantic alignment accuracy, catastrophic forgetting during graph updates, and chaotic integration of old and new regulatory versions when processing cross-modal and time-sensitive financial regulatory data.

[0012] This not only ensures the retrieval efficiency and concurrency consistency of massive multimodal regulatory data in high-frequency update scenarios, but also achieves accurate location and incremental correction of affected subgraphs through graph differencing algorithms, greatly reducing system maintenance overhead. Finally, by using the updated graph as the knowledge source for retrieval enhancement generation, a complete traceability link with version timestamps is embedded in the generated results.

[0013] Not only does it utilize breadth-first search combined with node betweenness centrality to perform hierarchical sorting of newly added triples, but more importantly, it innovatively introduces a regulatory timeliness weight factor and decay coefficient into the distillation loss function. This allows the graph to consciously retain the core representation of the latest revised regulations during the update process, while constraining the interference of old knowledge, thus enabling rapid replacement during the fusion of new and old knowledge. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a flowchart of a method for dynamically constructing and updating a multimodal financial regulatory map, provided in Embodiment 1 of the present invention. Figure 2 This is a flowchart illustrating the process architecture of a method for dynamically constructing and updating a multimodal financial regulatory map, as provided in Embodiment 1 of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0018] Example 1

[0019] This embodiment provides a method for dynamically constructing and updating a multimodal financial regulatory map. The method includes the following steps: Figure 1 and Figure 2 As shown: Acquire multimodal financial regulatory data, which includes at least textual, tabular, and image-based data. Specifically, the acquisition of text modal data mainly targets plain text documents such as regulatory documents, window guidance opinions, judicial interpretations, and industry self-regulatory guidelines published on the official websites of regulatory agencies. (Distributed web crawlers and timed synchronization scripts are deployed to obtain the corresponding data, including HTML, PDF, or DOCX format files, which are crawled to the local server).

[0020] The acquisition of tabular modal data focuses on the content presented in the financial regulatory system in the form of structured values, classification standards or constraints. Typical sources include capital adequacy ratio calculation tables, liquidity coverage ratio stress test templates, and related party transaction disclosure format guidelines issued by regulatory agencies along with the main text of the regulations.

[0021] Image modal data acquisition primarily targets non-machine-readable regulatory documents in the form of scanned copies, faxes, or stamped photocopies. This type of data is particularly common in cross-border regulatory agreements, historical regulatory retrospectives, and internal compliance approval processes. The acquisition process begins with scanning the paper documents using a document scanner, a scanner, or a mobile terminal at an optical character recognition (OCR) resolution of at least 300 DPI, generating lossless compressed images in PNG or TIFF format. Subsequently, the process relies on a cascaded visual processing pipeline: first, Canny edge detection and Hough transform algorithms are used to correct tilt and denoise the background of the scanned image, enhancing the contrast between text and background; then, an OCR engine based on CRNN (Convolutional Recurrent Neural Network) and attention mechanisms is deployed to recognize printed and handwritten text line by line, decoding pixel information into Unicode character sequences.

[0022] Furthermore, knowledge-enhanced multi-granularity interactive cross-modal entity linking is performed on multimodal financial regulatory data to identify and link financial regulatory entities in different modalities, generating an initial financial regulatory map; The aforementioned processing method for "enhanced knowledge of multimodal financial regulatory data enforcement, multi-granularity interactive cross-modal entity linking" includes the following steps: Text modal data is encoded using a pre-trained model in the financial field, and features are extracted from table modal data and image modal data using an OCR model to obtain the first feature set. By using text global-local interaction units and cross-modal fusion interaction units, multi-granularity interaction processing is performed on the encoding and the first feature set to generate a cross-modal soft matching probability distribution. Cross-modal feature fusion is performed based on cross-modal soft matching probability distribution to obtain cross-modal links of financial regulatory entities; Furthermore, before performing cross-modal feature fusion, noise data filtering is performed on the multimodal financial regulatory data based on the cross-modal soft matching probability distribution to remove entity mention pairs with cross-modal confidence scores below a preset threshold.

[0023] Furthermore, the text is used to fuse global semantics and local lexical information, where global semantics are obtained through the overall semantic encoding of the legal provisions, and local lexical information is obtained through context-enhanced encoding of financial terminology keywords in the legal provisions. The two are then fused together using an adaptive gating mechanism. The cross-modal fusion interaction unit mines the implicit semantic association between text and vision through cross-modal projection and gating fusion strategies. The visual features include table structure features extracted from table modal data and document layout features extracted from image modal data.

[0024] Specifically, for text modal data, a language model pre-trained on a large scale on financial corpora is used for encoding. This model can map the professional terminology, sentence structure, and contextual semantics in legal provisions into vector representations containing deep financial semantics. For tabular and image modal data, an optical character recognition model is used for feature extraction. For tabular data, not only is the text content within cells extracted, but also the row and column positions and the table header hierarchy are retained as structural features. For image data, in addition to extracting the text content, the document's layout features are further captured, such as paragraph columns, signature positions, and annotation areas. The above encoding and extraction results together constitute the first feature set, serving as the basic input for subsequent cross-modal interactive processing.

[0025] After obtaining the first set of features, this invention refines the semantic features of the text modality through a text global-local interaction unit. The design of this unit is based on the dual importance of macro-level logic and micro-level wording in financial regulatory texts: the overall expression of the regulatory clauses defines their legal scope and normative purpose, while the financial terminology keywords embedded in the clauses carry precise indicator definitions and compliance thresholds. To this end, the text global-local interaction unit first encodes the complete context of the regulatory clauses through a global encoding branch, obtaining the overall semantic representation of the clauses within the entire regulatory framework; simultaneously, it identifies the financial terminology keywords appearing in the clauses through a local enhancement branch, and performs window-style enhancement encoding on the qualifiers and modifiers before and after these keywords, generating a refined local representation at the terminology level. Subsequently, an adaptive gating mechanism is introduced to dynamically weight and fuse global semantic and local lexical information. This gating mechanism uses a learnable parameter network to automatically adjust the fusion ratio of global and local information based on the characteristics of the current clause content: for principled and general clauses, the model tends to assign higher weights to global semantics; while for clauses containing specific numerical standards or defining terms, the model automatically increases the contribution of local lexical information. The text features after this fusion process not only retain the overall meaning of the legal provisions but also accurately characterize the semantic boundaries of key terms, providing a high-quality single-modal representation foundation for subsequent semantic alignment with the visual modality.

[0026] Meanwhile, the cross-modal fusion interaction unit is responsible for establishing deep semantic connections between textual and visual features. This unit first maps feature vectors from the textual modality and those from the table and image modalities to a shared semantic space of the same dimension through a cross-modal projection layer, eliminating differences in distribution scale and representation dimension between features from different modalities. Based on this, a gated fusion strategy is employed to interactively aggregate textual and visual features: specifically, for each pair of candidate textual and visual entity mentions, the model calculates their cosine similarity in the shared space as an initial matching score. Simultaneously, an attention mechanism is introduced to calculate the weighted aggregation result of textual features on visual features and the symmetrically weighted aggregation result of visual features on textual features. The gated network adaptively outputs a fusion gating coefficient based on the information saturation of the two aggregation results, controlling the mixed weights of textual and visual features in the final cross-modal representation. Through iterative calculations using multiple stacked cross-modal fusion interaction units, the model ultimately generates a cross-modal soft-matching probability distribution for each pair of cross-modal entity mentions. This distribution quantifies the probability that a legal entity in the text and a candidate region in the image or table point to the same objective entity.

[0027] After generating the cross-modal soft-match probability distribution, this invention introduces a noise data filtering step to improve the quality of the graph construction before performing the final feature fusion and entity linking. Specifically, a confidence threshold is preset, and only entity mention pairs with cross-modal soft-match probabilities higher than this threshold are retained to proceed to the next stage, while low-confidence pairs below the threshold are treated as noise and discarded. This filtering mechanism effectively suppresses erroneous links introduced by OCR recognition errors, table layout ambiguities, or poor image quality, ensuring that only semantically closely related and reliable cross-modal entity pairs can enter the subsequent fusion and linking process, thereby guaranteeing the accuracy and purity of the initial financial regulatory graph from the source.

[0028] High-quality entity mention pairs retained after noise filtering enter the cross-modal feature fusion stage to generate the final cross-modal entity links. In this stage, the system weights and fuses the text entity representations with their corresponding visual region representations using matching weights calculated previously, forming a unified multimodal entity embedding vector. This embedding vector carries both the semantic definition from the text description and incorporates numerical constraint information from tables or layout effectiveness features from images. Subsequently, based on this fused vector, nearest neighbor retrieval or link prediction is performed in an existing financial regulatory entity database to identify and merge the same financial regulatory entity across different modalities.

[0029] Furthermore, when new financial regulatory data is acquired, the initial financial regulatory map is dynamically updated based on an incremental distillation strategy to obtain an updated financial regulatory map.

[0030] The processing method of "dynamically updating the initial financial regulatory map using the incremental distillation strategy to obtain the updated financial regulatory map" in the above embodiments includes the following steps: Based on the breadth-first search algorithm, the newly added triples in the new financial regulations data are sorted between layers according to the topological distance between the newly added triples and the initial financial regulations graph. The importance scores of each triple are calculated by node centrality and relation betweenness centrality to sort them within the layer, resulting in a hierarchical sorting result. Among them, the calculation of node centrality introduces the timeliness weight factor of the regulations clauses, and assigns higher centrality weight to the entity nodes corresponding to the latest revised regulations. Based on the hierarchical sorting results and the dynamically adjusted distillation loss function, the knowledge distillation of the newly added triples is integrated into the initial financial regulatory map. If an entity appears in multiple levels, the current representation and historical representation of the entity are constrained by the dynamically adjusted distillation loss function. The distillation loss function combines L1 and L2 regularization terms and dynamically adjusts the L2 regularization strength through the timeliness decay coefficient, so that the knowledge forgetting of older version entities is more strictly constrained. A two-stage training strategy is adopted. In the first stage, the parameters of the initial financial regulatory map are frozen and only the representation of newly added triples is optimized. In the second stage, all parameters are fine-tuned to integrate new and old knowledge and generate an updated financial regulatory map. Furthermore, prior to the first phase, the newly added triples are subject to financial regulatory constraints verification, and new triples that do not conform to the logical structure of financial regulations are eliminated. The logical structure of financial regulations includes the hierarchical relationship between regulatory clauses and the constraints on the integrity of reference.

[0031] Specifically, the system uses existing regulatory entities in the initial graph as anchor points and employs a breadth-first search algorithm to calculate the topological distance between each new triplet and the existing graph structure, i.e., the shortest hop count from the new entity to the core regulatory clause network. Triples with closer distances and more direct associations are assigned to higher-priority levels to ensure the stability of the graph's core framework during updates. Within the same level, a secondary sort is performed based on node centrality and relation betweenness centrality, calculating the importance score of each triplet to the overall knowledge structure. The calculation of node centrality does not solely rely on graph topology indicators but incorporates a timeliness weight factor reflecting the unique attributes of financial regulations. For example, the centrality weight of entity nodes corresponding to the latest revised regulatory clauses is proactively increased to ensure that new knowledge representing currently effective legal norms participates in distillation and fusion first, while triples associated with historical versions or repealed clauses have their processing priority reduced accordingly. Thus, the system generates a hierarchical sorting result that reflects both topological affinity and the importance of regulatory timeliness.

[0032] When the same entity appears repeatedly in multiple update levels, a semantic shift occurs between the entity's embedded representation in the historical state and the new representation learned in the current batch. To prevent new knowledge from overwriting the effective components of old knowledge, the distillation loss function enforces a certain degree of alignment between the current representation and the historical representation. This loss function is composed of a combination of L1 and L2 regularization terms, where the L1 term promotes the preservation of sparsity differences between the old and new representations, and the L2 term smooths the magnitude of the representation shift. For the specific needs of the timeliness of financial regulations, the L2 regularization strength is not fixed but dynamically scaled using a timeliness decay coefficient related to the time difference. Specifically, the further back in time the regulations associated with the entity took effect, the smaller the decay coefficient and the higher the corresponding L2 regularization strength, thus imposing a stronger preservation constraint on the representation of older versions of knowledge and inhibiting its easy rewriting during updates; conversely, for the most recently effective regulatory entities, the decay coefficient approaches a large value, and the L2 constraint is correspondingly relaxed, allowing new knowledge to be integrated into the representation space with greater freedom.

[0033] More importantly, before the first phase of training begins, the system has specifically set up a compliance checkpoint in the field of financial regulations. For the candidate triples initially extracted from the newly added multimodal data, the system will call predefined regulatory logic structure rules to perform compliance screening.

[0034] Example 2

[0035] After completing the above embodiment one, that is, after generating the initial financial regulatory map or the updated financial regulatory map, a hybrid index structure based on a unified vector space is constructed. The hybrid index structure includes a hierarchical navigation graph HNSW for approximate nearest neighbor search and an inverted index for keyword matching. The unified vector space is jointly constructed by the multimodal representation of financial regulatory entities, clauses and their relationship types. Furthermore, during the dynamic update process, the affected index entries in the hybrid index structure are updated synchronously based on the hierarchical sorting results to maintain the consistency between the index and the map data.

[0036] Specifically, in the implementation process, the financial regulatory entity nodes that have completed cross-modal entity links and their associated clause text content are first mapped into dense vector representations using the aforementioned pre-trained financial domain model and multimodal encoder. Simultaneously, the relation types in the graph are also mapped to the same vector space through a relation embedding layer, ensuring that entity vectors, clause content vectors, and relation vectors maintain geometric correlation within a unified semantic manifold.

[0037] Based on the construction of the unified vector space described above, the HNSW index in the hybrid index structure takes over the semantic similarity retrieval task. HNSW is a graph-based multi-layer approximate nearest neighbor search structure. Its operating principle involves constructing a hierarchical, navigable small-world network. During retrieval, it quickly jumps from the top-level sparse graph to the neighborhood of the query vector, then descends layer by layer to the bottom-level dense graph for fine-grained searching, thus achieving high-precision vector retrieval with logarithmic time complexity. Specifically, each node in the HNSW layer is a vectorized representation of the entities and terms in the graph within the multimodal unified vector space, and the edges between nodes are constructed based on the cosine similarity or Euclidean distance between vectors. When a user inputs a natural language query, the query is first converted into a query vector by the same encoder. The HNSW index can then retrieve a list of candidate entities semantically closest to the query from millions of entity vectors within tens of milliseconds.

[0038] Meanwhile, the inverted index in the hybrid index structure handles the needs for precise term matching and structured field filtering. This is achieved by using the unique identifier of each financial regulatory entity and clause as a document ID, constructing an inverted list of all associated keywords to form a mapping dictionary. This mapping dictionary is then used as the filtering scope for HNSW vector retrieval, thus ensuring both semantic matching accuracy and the rigor of the business logic in the search results.

[0039] Furthermore, maintaining consistency between the index and the graph data is crucial during the dynamic updating of the financial regulatory graph; otherwise, it will lead to problems such as lagging search results or referencing outdated versions of regulations. During the first and second phases of incremental update training, the system uses a graph differencing algorithm to capture changes in entity nodes and their neighborhood subgraphs affected by newly added triples. For entity nodes whose representations are updated at the vector space level (i.e., nodes whose embedding vectors have significantly shifted due to distillation learning), the system does not perform a full reconstruction of the HNSW index, but instead employs a local re-insertion strategy. Specifically, based on the importance score of the entity node in the hierarchical ranking results, if it is a core hub node, the recalculation of the neighborhood connections and local index repair of its HNSW layer are triggered; if it is an edge leaf node, only the node's own vector storage and the corresponding keyword entries in the inverted index are updated. At this point, the system identifies the entity as a node with high centrality based on the hierarchical ranking results, and then triggers the reconstruction operation of the local neighborhood edges related to this node in the HNSW index, synchronously updating the version timestamp field associated with it in the inverted index. For marginal clause entities that are only indirectly referenced by the revised content, the system performs only lightweight vector replacement and inverted record appending.

[0040] Example 3

[0041] During the completion of Embodiments 1 and 2 above, consistency management is performed on cross-modal financial regulatory graph data update operations based on the ACID transaction principle; data change events are captured through an event-driven architecture, the affected subgraphs are located using the graph difference algorithm, and incremental updates are performed only on the affected subgraphs; and after the second stage of the two-stage training is completed, version consistency verification is performed on the updated financial regulatory graph to verify the matching of the revision version number of the regulatory clauses with the version number of the associated references. If the verification fails, a rollback operation is triggered.

[0042] Specifically, in the operation and maintenance of the financial regulatory knowledge graph, each modification operation to the graph (such as adding regulatory clause entities, modifying the reference relationships between clauses, marking the repeal status of clauses, etc.) is regarded as an independent database transaction. This solution strictly follows the ACID transaction principle to manage the consistency of cross-modal update operations. Specifically, atomicity ensures that each set of related modifications either takes full effect or takes no effect at all, avoiding the intermediate state where only the main text of the regulation is updated while the relationship edge between it and the appendix table is missed; consistency ensures that the updated graph state always meets the preset financial regulatory logical constraints.

[0043] When an external system imports new financial regulation data packages or generates revision instructions through manual annotation, the system generates a corresponding "regulatory change event" and publishes it to the message bus. The graph update engine, acting as the event consumer, does not scan and reconstruct the entire graph upon receiving the change event. Instead, it calls a graph differencing algorithm to accurately quantify the scope of the change's impact. Specifically, it first performs subgraph matching in the existing graph based on the new triplet information carried by the event, using breadth-first search to identify the set of neighboring nodes reachable from the changed node via referencing edges or hierarchical dependent edges. Then, by comparing the hash difference between the adjacency matrices of the old and new graph structures, it quickly locates the minimum connected subgraph where the topology has changed or attributes have been updated. This process concentrates computational resources only within the boundaries of the affected subgraph, greatly reducing the complexity of the update operation.

[0044] Example 4

[0045] Based on the above embodiment one, the updated financial regulatory graph is used as a knowledge retrieval enhancement knowledge source to respond to user queries, retrieve relevant regulatory clauses and evidence chains from the knowledge source, and input them into the large language model to generate compliance analysis results. Furthermore, traceability information is embedded in the compliance analysis results. This traceability information includes at least the original text of the cited regulatory clauses, the identifiers of the related graph nodes, and the version timestamp, in order to verify the traceability of the generated results.

[0046] Specifically, once the updated financial regulatory graph is constructed, the system uses it as a structured knowledge retrieval enhancement knowledge source. This knowledge source differs from traditional unstructured document libraries; its core advantage lies in the fact that entities, relationships, and attributes within the graph all carry explicit semantic tags, version identifiers, and cross-modal reference paths. Responding to user query requests input through the interactive interface, the system first initiates a hybrid index retrieval process. Furthermore, the system automatically arranges the evidence chain and assembles the context based on the relationship paths in the graph. The system traverses these paths, extracting the original regulatory clauses from the text modality, specific financial indicator data from the table modality, and the signature validity information of relevant attachments from the image modality, and converts them into structured contextual prompt text according to logical hierarchy.

[0047] After constructing the chain of evidence, the system inputs this structured context along with the user's original query into the large language model. The large language model no longer relies on its own potentially outdated or ambiguous parameterized knowledge, but instead performs compliance analysis and reasoning strictly based on the accurate original legal texts and quantitative data provided by the system. The model output will be an analysis report containing clear conclusions, original legal citations, and logical deduction steps.

[0048] To ensure strong traceability of the generated results, the system simultaneously performs embedded annotation of traceability information during the compliance analysis results generation stage. This process is not simply adding references at the end of the document, but rather anchoring fine-grained information within the text stream generated by the model for each segment involving regulatory arguments.

[0049] Example 5

[0050] This invention provides a non-transitory computer-readable storage medium storing at least one instruction or at least one program segment, which is loaded and executed by a processor to implement the following steps: Obtain multimodal financial regulatory data.

[0051] Perform knowledge-enhanced, multi-granularity, interactive cross-modal entity linking on multimodal financial regulatory data to identify and link financial regulatory entities in different modalities, generating an initial financial regulatory map.

[0052] When new financial regulatory data is acquired, the initial financial regulatory map is dynamically updated based on the incremental distillation strategy to obtain the updated financial regulatory map.

[0053] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0054] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0055] Example 6

[0056] This invention provides an electronic device, including a processor and a memory, wherein the memory stores at least one instruction or at least one program segment, and the at least one instruction or the at least one program segment is loaded and executed by the processor to implement the following steps: Obtain multimodal financial regulatory data.

[0057] Perform knowledge-enhanced, multi-granularity, interactive cross-modal entity linking on multimodal financial regulatory data to identify and link financial regulatory entities in different modalities, generating an initial financial regulatory map.

[0058] When new financial regulatory data is acquired, the initial financial regulatory map is dynamically updated based on the incremental distillation strategy to obtain the updated financial regulatory map.

[0059] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for dynamically constructing and updating a multimodal financial regulatory map, characterized in that, Includes the following steps: Acquire multimodal financial regulatory data, which includes at least text modal data, tabular modal data, and image modal data; The multimodal financial regulatory data is subjected to knowledge-enhanced multi-granularity interactive cross-modal entity linking to identify and link financial regulatory entities in different modalities, generating an initial financial regulatory map. When new financial regulatory data is acquired, the initial financial regulatory map is dynamically updated based on an incremental distillation strategy to obtain an updated financial regulatory map.

2. The method for dynamically constructing and updating a multimodal financial regulatory map according to claim 1, characterized in that, The implementation of knowledge-enhanced multi-granularity interactive cross-modal entity links on the multimodal financial regulatory data includes: The text modal data is encoded using a pre-trained model in the financial field, and features are extracted from the table modal data and the image modal data using an OCR model to obtain a first feature set. By using text global-local interaction units and cross-modal fusion interaction units, multi-granularity interaction processing is performed on the encoding and the first feature set to generate a cross-modal soft matching probability distribution. Cross-modal feature fusion is performed based on the cross-modal soft matching probability distribution to obtain cross-modal links of financial regulatory entities; Furthermore, before performing the cross-modal feature fusion, the multimodal financial regulatory data is filtered for noise based on the cross-modal soft matching probability distribution to remove entity mention pairs with cross-modal confidence levels below a preset threshold.

3. The method for dynamically constructing and updating a multimodal financial regulatory map according to claim 2, characterized in that, The global semantics and local lexical information of the text are used to fuse the text. The global semantics are obtained through the overall semantic encoding of the legal provisions, and the local lexical information is obtained through the context-enhanced encoding of financial terminology keywords in the legal provisions. The two are then fused by an adaptive gating mechanism. The cross-modal fusion interaction unit mines the implicit semantic association between text and vision through cross-modal projection and gating fusion strategies. The visual features include table structure features extracted from table modal data and document layout features extracted from image modal data.

4. The method for dynamically constructing and updating a multimodal financial regulatory map according to claim 1, characterized in that, The incremental distillation strategy dynamically updates the initial financial regulatory map to obtain an updated financial regulatory map, including: Based on the breadth-first search algorithm, the newly added triples in the new financial regulations data are sorted between layers according to the topological distance between the newly added triples and the initial financial regulations graph. Then, the importance scores of each triple are calculated by node centrality and relation betweenness centrality to sort them within the layer, resulting in a hierarchical sorting result. Among them, the calculation of node centrality introduces the timeliness weight factor of the regulations clauses, and assigns higher centrality weight to the entity nodes corresponding to the latest revised regulations. Based on the hierarchical sorting results and the dynamically adjusted distillation loss function, the knowledge distillation of the newly added triples is integrated into the initial financial regulatory map. If an entity appears in multiple levels, the current representation and historical representation of the entity are constrained by the dynamically adjusted distillation loss function. The distillation loss function combines L1 and L2 regularization terms and dynamically adjusts the L2 regularization strength through the timeliness decay coefficient, so that the knowledge forgetting of older version entities is subject to stricter constraints. A two-stage training strategy is adopted. In the first stage, the parameters of the initial financial regulatory map are frozen and only the representation of the newly added triples is optimized. In the second stage, all parameters are fine-tuned to integrate new and old knowledge and generate an updated financial regulatory map. Furthermore, prior to the first stage, the newly added triples are subject to financial regulatory constraints verification, and new triples that do not conform to the logical structure of financial regulations are eliminated. The logical structure of financial regulations includes the hierarchical relationship between regulatory clauses and the constraint of referential integrity.

5. The method for dynamically constructing and updating a multimodal financial regulatory map according to claim 4, characterized in that, The method further includes: after generating the initial financial regulatory map or the updated financial regulatory map, constructing a hybrid index structure based on a unified vector space, wherein the hybrid index structure includes a hierarchical navigation graph HNSW for approximate nearest neighbor search and an inverted index for keyword matching, and the unified vector space is jointly constructed by multimodal representations of financial regulatory entities, clauses and their relationship types; Furthermore, during the dynamic update process, the affected index entries in the hybrid index structure are updated synchronously based on the hierarchical sorting results to maintain the consistency between the index and the map data.

6. The method for dynamically constructing and updating a multimodal financial regulatory map according to claim 4, characterized in that, The method further includes: performing consistency management on cross-modal financial regulatory graph data update operations based on ACID transaction principles; capturing data change events through an event-driven architecture, locating affected subgraphs using graph differencing algorithms, and performing incremental updates only on the affected subgraphs; and after the second stage of the two-stage training is completed, performing version consistency verification on the updated financial regulatory graph to verify the matching of the revision version number of the regulatory clauses with the version number of the associated references, and triggering a rollback operation if the verification fails.

7. The method for dynamically constructing and updating a multimodal financial regulatory map according to claim 1, characterized in that, The method further includes: using the updated financial regulatory graph as a knowledge retrieval enhancement knowledge source, in response to user queries, retrieving relevant regulatory clauses and evidence chains from the knowledge source, and inputting them into a large language model to generate compliance analysis results; Furthermore, traceability information is embedded in the compliance analysis results. This traceability information includes at least the original text of the cited regulatory clauses, the identifiers of the associated graph nodes, and the version timestamp, in order to verify the traceability of the generated results.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method for dynamically constructing and updating the multimodal financial regulatory map as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method for dynamically constructing and updating a multimodal financial regulatory map as described in any one of claims 1 to 7.

Images