A financial regulation cause-effect analysis method fusing counterfactual reasoning

Abstract

The application discloses a kind of financial regulations causal analysis methods of fusion counterfactual reasoning, it is related to computer processing technical field, the method includes the following steps: S01, obtains multimodal financial regulations data;S02, the multimodal financial regulations data is carried out knowledge enhanced semantic alignment and entity relationship joint discovery, constructs initial multimodal financial knowledge graph;S03, responds to the new financial regulations data, based on the incremental distillation strategy to the initial multimodal financial knowledge graph is dynamically updated, generates dynamic multimodal financial knowledge graph.The application is updated dynamically by the dynamic updating mechanism that hierarchical ordering, incremental distillation and differentiating weight distribution are coordinated, in combination with knowledge enhanced path reasoning and multi-granularity cross-modal entity link, the efficient dynamic evolution of financial regulations knowledge graph is realized, and the efficiency of compliance review and early warning accuracy in the scene of investment banking manuscript management and risk control is significantly improved.

Description

Technical Field

[0001] This invention relates to the field of computer processing technology, and in particular to a method for causal analysis of financial regulations that integrates counterfactual reasoning. Background Technology

[0002] With increasingly stringent financial regulatory environments and a more complex legal framework, financial institutions face significant challenges in business compliance and risk control. Traditional compliance management relies heavily on manual analysis of regulatory texts and static knowledge bases, which struggles to address the multi-source, heterogeneous, and dynamically evolving nature of regulatory data. Specifically: On the one hand, financial regulatory data usually exists in multiple modalities such as text, tables, and flowcharts. Traditional single-modal information processing technologies cannot effectively integrate these heterogeneous data, resulting in incomplete extraction of entities and relationships and difficulties in semantic integration. On the other hand, regulations are frequently revised and updated, and existing knowledge graph construction technologies are mostly static, lacking the ability to continuously learn and update dynamically, leading to outdated knowledge and inaccurate compliance judgments.

[0003] In summary, how to achieve dynamic fusion of multimodal financial regulatory data, accurate reasoning with knowledge enhancement, and efficient incremental updates of the data graph have become urgent technical problems to be solved in the current fintech field. Summary of the Invention

[0004] To address the aforementioned technical problems, the present invention employs a financial regulatory causal analysis method that integrates counterfactual reasoning, comprising the following steps: S01. Obtain multimodal financial regulatory data, wherein the multimodal financial regulatory data includes at least text modal data, tabular modal data and image modal data; S02. Perform knowledge-enhanced semantic alignment and joint discovery of entity relationships on the multimodal financial regulatory data to construct an initial multimodal financial knowledge graph; S03. In response to newly added financial regulatory data, the initial multimodal financial knowledge graph is dynamically updated based on an incremental distillation strategy to generate a dynamic multimodal financial knowledge graph. The dynamic update includes: performing hierarchical sorting on the newly added financial regulations data, dividing it into multiple learning levels based on the topological distance from the old graph, and determining the learning order of newly added triples in each level according to node centrality and relation betweenness centrality; and retaining the knowledge of the old graph through inter-layer representation distillation to obtain the updated multimodal financial knowledge graph.

[0005] Preferably, step S02, which involves knowledge-enhanced semantic alignment and joint discovery of entity relationships in the multimodal financial regulatory data, includes: The FinBERT model, OCR technology, and cross-modal attention mechanism are used to perform structured parsing and feature extraction on the text modal data, table modal data, and image modal data, respectively, to achieve semantic alignment of multimodal data; Based on knowledge graphs in the financial field, a knowledge-guided joint learning framework is constructed, which combines conditional random field constraints and multimodal association matrices to identify cross-modal entities; By using graph attention networks to perform multi-hop relationship reasoning on identified cross-modal entities, implicit cross-modal association paths can be discovered.

[0006] Based on spectral knowledge, an updated multimodal financial knowledge graph is obtained.

[0007] Preferably, the hierarchical sorting includes: Based on the breadth-first search algorithm, the layers are sorted according to the topological distance between the new triples and the old graph; The importance score of each newly added triplet is calculated using node centrality and relation betweenness centrality, and then the triplet is sorted within the layer.

[0008] Preferably, the method of preserving old spectral knowledge through interlayer representation distillation includes: If the same entity appears in multiple layers, the difference between its current representation and the historical representation of the nearest layer is constrained by inter-layer representation distillation. A distillation loss function based on KL divergence is used in combination with L1 / L2 regularization to dynamically adjust the distillation intensity during training, so as to suppress noise propagation while retaining old knowledge. Differentiated weights are assigned to distillation loss based on the node degree distribution and relation betweenness centrality in the graph topology features. Priority is given to protecting the semantic integrity of core regulatory entities with high node degree and their related relationships, and then updates are performed sequentially.

[0009] Preferably, the method also includes cross-modal entity linking of the multimodal financial regulatory map based on multi-granularity interaction: A unified feature representation method is adopted, and pre-trained ViT and FinBERT models are used to extract global and local features of visual and textual modalities, respectively. By fusing global semantic and local lexical information through text global-local interaction units, enhancing the robustness of explicit visual features through visual dual interaction units, and mining implicit semantic associations between text and vision through cross-modal fusion interaction units; Based on the modality type and data quality of the input data, the output weights of the text global-local interaction unit, the visual dual interaction unit, and the cross-modal fusion interaction unit are dynamically adjusted through scene adaptive weight allocation; A unit consistency loss function is introduced to constrain the matching process of each interactive unit, thereby achieving accurate linking and semantic alignment of cross-modal entities.

[0010] Preferably, the method also includes path reasoning with knowledge enhancement based on the dynamic multimodal financial knowledge graph, which uses a relational graph convolutional network to dynamically model entities and relations in the dynamic multimodal financial knowledge graph. Different types of relations correspond to relational embedding representations of different dimensions. The relational graph convolutional network dynamically adjusts the information transmission weight of relation types according to the semantic context of each entity to generate a low-dimensional vector representation that integrates multi-hop neighbor information. Features from different modalities are fused with the initial embeddings of graph nodes and used as input to the graph neural network. Based on the reasoning capabilities of graph neural networks, probabilistic path walking is used to automatically discover cross-modal association paths. During the path search process, a knowledge-guided path pruning strategy and path length penalty are introduced to eliminate redundant path branches that are not directly related to compliance. The comprehensive reliability score of each candidate path is calculated based on the product of the reasoning scores at each step in the path, and the complete reasoning chain of the query entity association is output.

[0011] Preferably, the knowledge-enhanced path reasoning is applied to investment banking documentation management and risk control scenarios, including: In response to a query request for investment banking working papers, the system automatically traces the consistency of financial data along the inference chain, identifies logical inconsistencies in the data in the working papers based on the verified inference chain, and matches and verifies the expert information database involved in the inference chain. In response to risk control monitoring requests, the system automatically tracks the correlation between fund flows and regulatory provisions along the inference chain and the financial industry compliance rules engine, and issues warnings for abnormal transaction behaviors.

[0012] Preferably, the dynamic update further includes: The parameters of the old graph are frozen by freezing the embedding parameters of non-new nodes in the old graph, and only the representation of new entities and relations is optimized. Unfreeze all parameters and fine-tune them together to achieve a deep integration of new and old knowledge.

[0013] The present invention has at least the following beneficial effects: The KL divergence distillation loss is adopted to balance the accurate replication of old knowledge with the rapid learning of new knowledge. Combined with L1 / L2 regularization to suppress noise propagation, the numerical stability of the distillation process is guaranteed. At the same time, the distillation loss is assigned differentiated weights based on the node degree distribution and relation betweenness centrality, prioritizing the protection of the semantic integrity of core regulatory entities with high node degrees and their related relationships.

[0014] Different types of relations correspond to different dimensions of relation embedding representation. RGCN dynamically adjusts the information transmission weight of relation types according to the semantic context of the entity. At the same time, it introduces knowledge-guided path pruning strategy and path length penalty in the GNN-based reasoning process to eliminate redundant path branches that are not directly related to compliance. The comprehensive reliability score is calculated based on the product of the reasoning scores of each step in the path before the reasoning chain is output. Attached Figure Description

[0015] 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.

[0016] Figure 1 The flowchart illustrates a financial regulatory causal analysis method that integrates counterfactual reasoning, as provided in Embodiment 1 of the present invention. Detailed Implementation

[0017] 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.

[0018] 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.

[0019] Example 1

[0020] This embodiment provides a financial regulatory causal analysis method that integrates counterfactual reasoning. The method includes the following steps: Figure 1 As shown: S01. Obtain multimodal financial regulatory data, which includes at least text modal data, tabular modal data, and image modal data.

[0021] Specifically, through pre-configured data interfaces or web crawler modules, the system connects to the official websites of financial regulatory agencies, exchange announcement platforms, and internal corporate compliance databases, collecting data categorized by regulatory type, publication date, and business area. Taking the revised Securities Law as an example, the text data obtained by the system is the complete text of the revised law; the tabular data may be a "comparison table of old and new clauses" or an "administrative penalty amount statistics table" accompanying the revision; and the image data may be a "registration system reform flowchart" from the legislative explanation or a "compliance review guidance diagram" compiled by a securities firm.

[0022] For the text modality, FinBERT (a pre-trained BERT model for the financial domain) is used for word segmentation, part-of-speech tagging, and named entity recognition to extract entities such as "financial institutions," "information disclosure obligations," and "insider trading." For the table modality, high-precision OCR (Optical Character Recognition) technology is used to convert image-formatted tables into structured data. Then, regular expressions or semantic matching models (such as Sentence-BERT) are used to establish correspondences between table headers and column values ​​and clause numbers and business terms in the text. For the image modality, in addition to OCR extraction of text from images, image classification models (such as ResNet) are used to identify chart types (such as flowcharts, line charts, and bar charts), and a cross-modal attention mechanism is used to initially establish implicit associations between image regions and text descriptions. The output of this stage is not the final knowledge graph, but rather provides aligned intermediate feature vectors with modal labels for subsequent entity linking and relationship discovery. For example, when dealing with the flowchart of the registration-based IPO system reform, the system can identify the nodes in the flowchart ("enterprise application", "exchange review" and "CSRC registration") and the flow relationships between the nodes. It will then preliminarily align this visual information with the corresponding registration-based IPO process description in the text, thus laying the foundation for multimodal fusion at the source of the data.

[0023] S02. Perform knowledge-enhanced semantic alignment and joint discovery of entity relationships on multimodal financial regulatory data to construct an initial multimodal financial knowledge graph.

[0024] Furthermore, the above embodiments involve knowledge-enhanced semantic alignment and joint discovery of entity relationships for multimodal financial regulatory data, including: FinBERT model, OCR technology and cross-modal attention mechanism are used to perform structured parsing and feature extraction on text modal data, table modal data and image modal data respectively, so as to achieve semantic alignment of multimodal data; Based on knowledge graphs in the financial field, a knowledge-guided joint learning framework is constructed, which combines conditional random field constraints and multimodal association matrices to identify cross-modal entities; By using graph attention networks to perform multi-hop relationship reasoning on identified cross-modal entities, implicit cross-modal association paths can be discovered.

[0025] Specifically, for text modalities (such as legal clauses and regulatory notices), a FinBERT model fine-tuned with financial corpora is used for context encoding, outputting semantic vectors for each word. For table modalities (such as capital adequacy ratio calculation tables and fee rate tables), high-precision OCR is used to recognize the table structure, and then regular expressions or semantic matching models are used to align the table headers and cell contents with the "clause number" and "constraints" in the text paragraphs. For image modalities (such as business process diagrams and scanned signature pages), OCR is used to extract the text from the image, and convolutional neural networks are used to detect the chart type and region. Based on this, a cross-modal attention mechanism is introduced: using text features as the query, and table features and image features as the key and value respectively, attention weights are calculated to associate numerical entities in the table with flowchart nodes in the image to the corresponding legal clause text, thus completing the semantic alignment of multimodal data.

[0026] Subsequently, the system enters the knowledge-guided joint discovery phase of entities and relationships. This phase constructs a joint learning framework that integrates two types of prior knowledge: First, knowledge graphs in the financial field; Second, there are the constraints of Conditional Random Fields (CRFs) defined by industry experts.

[0027] For aligned multimodal features, the system first generates the probability distribution of candidate entities; then, it uses a CRF model to globally constrain the entity sequence, forcing the output of label combinations that conform to financial logic. Simultaneously, the system constructs a multimodal association matrix: each row of the matrix corresponds to an entity mention in the text, and each column corresponds to a candidate entity region in a table or image. The matrix elements are calculated using cross-modal feature similarity.

[0028] After entity recognition, the system uses a Graph Attention Network (GAT) to perform multi-hop relationship reasoning to discover implicit cross-modal association paths. The GAT constructs a heterogeneous graph using identified entities as nodes and initial relationships derived from the cross-modal association matrix as edges. At each layer of the GAT, each node calculates attention coefficients for its neighboring nodes.

[0029] S03. In response to newly added financial regulatory data, the initial multimodal financial knowledge graph is dynamically updated based on an incremental distillation strategy to generate a dynamic multimodal financial knowledge graph.

[0030] Specifically, the newly added financial regulations data undergoes structured analysis to extract a set of triples containing head entities, relations, and tail entities. This set is then compared with existing knowledge in the old graph to identify truly new triples. Subsequently, the system employs a hierarchical sorting mechanism: based on a breadth-first search (BFS) algorithm, triples are divided into multiple learning levels according to the topological distance between entities in the new triples and the old graph. Triples closer to core regulatory nodes (e.g., triples directly amending a clause of the Securities Law) are assigned to higher priority levels. Within each level, importance scores are assigned to each triple by calculating node centrality (e.g., the degree of the entity in the old graph) and relation betweenness centrality (e.g., the frequency of the relation in multiple compliance paths). Triples with higher scores are prioritized for training. This ensures that new knowledge with the greatest impact on the graph structure (e.g., revised key clauses) is learned first, thus avoiding the disruption of the existing knowledge system caused by random learning order.

[0031] The system treats the old graph embeddings (i.e., the entity and relation vector representations before the update) as "teachers" and the newly added triples being trained, along with their associated old entity representations, as "students." When the same entity (e.g., "information disclosure obligor") appears in multiple learning levels, the inter-level representation distillation mechanism constrains the difference between the entity's current representation and the historical representation of the most recent layer. Specifically, it employs a distillation loss function based on KL divergence, combined with L1 / L2 regularization terms, to dynamically adjust the distillation intensity. For core entities with high node degrees and important relationships (e.g., "China Securities Regulatory Commission"), the distillation loss weight is increased, forcing the new representation to be as close as possible to the old representation. Simultaneously, the temperature parameter in the distillation loss function controls the smoothness of knowledge transfer, ensuring that new knowledge, when rapidly adapting to new provisions, does not excessively interfere with the already stable core regulatory semantics. The training process employs a two-stage strategy: In the first stage, the embedding parameters of non-new nodes in the old graph are frozen, and only the representations of new entities and relations are optimized, allowing new knowledge to find suitable embedding positions without disrupting the old structure; in the second stage, all parameters are unfrozen, and joint fine-tuning is performed to achieve deep integration of new and old knowledge, ultimately outputting an updated dynamic multimodal financial knowledge graph.

[0032] The dynamic update in the above embodiments includes: sorting the newly added financial regulations data in layers, dividing multiple learning layers according to the topological distance from the old graph, and determining the learning order of the newly added triples in each layer according to the node centrality and relation betweenness centrality; retaining the knowledge of the old graph through inter-layer representation distillation to obtain the updated multimodal financial knowledge graph.

[0033] The above hierarchical sorting includes: Based on the breadth-first search algorithm, the layers are sorted according to the topological distance between the new triples and the old graph; The importance score of each newly added triplet is calculated using node centrality and relation betweenness centrality, and then the triplet is sorted within the layer.

[0034] Secondly, in the above embodiments, knowledge of old spectra is preserved through interlayer representation distillation, including: If the same entity appears in multiple layers, the difference between its current representation and the historical representation of the nearest layer is constrained by inter-layer representation distillation. A distillation loss function based on KL divergence is used in combination with L1 / L2 regularization to dynamically adjust the distillation intensity during training, so as to suppress noise propagation while retaining old knowledge. Differentiated weights are assigned to distillation loss based on the node degree distribution and relation betweenness centrality in the graph topology features. Priority is given to protecting the semantic integrity of core regulatory entities with high node degree and their related relationships, and then updates are performed sequentially.

[0035] Specifically, during the dynamic updating of the financial regulatory knowledge graph, when a new batch of regulatory data is added (such as newly released revised provisions of the Securities Law and related regulatory cases), the system first activates the hierarchical sorting module. This module, based on the breadth-first search (BFS) algorithm, calculates the shortest topological distance between each new triple (i.e., an "entity-relationship-entity" unit, such as "securities company-constrained-new net capital rule") and existing entities in the original graph. Based on the distance, the new triples are divided into multiple learning levels: those with a distance of 1 (directly connected to core entities in the old graph) are assigned to the first level for priority learning, those with a distance of 2 are assigned to the second level, and so on. Within each level, the system further calculates the importance score for each triple: it calculates the node centrality (i.e., how many other entities it is connected to) of the head and tail entities in the triple, as well as the relation betweenness centrality (i.e., the frequency with which the relation acts as a shortest path bridge in the entire graph). The weighted sum of these three factors is used as the comprehensive importance score; triples with higher scores are learned earlier within the same level. The principle behind this is that in the financial regulatory knowledge graph, core regulations (such as the general provisions of the Securities Law) often have extremely high node degrees. Prioritizing the learning of newly added provisions that are directly related to them can ensure the stability of the semantic backbone structure and prevent marginal noise provisions from interfering with model training too early.

[0036] After completing the hierarchical sorting, the system enters the incremental distillation phase. The core idea of ​​the distillation process is to "distill" the entity representations already learned in the old graph into the training of the new model, preventing catastrophic forgetting when learning new knowledge. In specific implementation, if the same entity (e.g., "anti-money laundering obligor") appears in both the old graph and is updated again in a newly added layer, the system applies an inter-layer representation distillation constraint to that entity: minimizing the KL divergence between the entity's current new representation and its old representation saved in the most recent layer of historical training, while simultaneously using L1 / L2 regularization terms. During training, the weight coefficients of the distillation loss are dynamically adjusted; that is, when the new data is noisy, the regularization term is automatically increased to suppress overfitting, and when the new data is of high quality, the distillation intensity is appropriately reduced to absorb new knowledge. Furthermore, graph topological features are used for differentiated weight allocation: the distribution of node degree and relation betweenness centrality of all entities in the statistical graph are used to assign higher distillation protection weights to core regulatory entities with high node degree (such as the core node "information disclosure obligation" which connects hundreds of specific clauses), making their representation updates more constrained by old knowledge; while for less common entities with low degree (such as a local pilot financial rule), lower distillation weights are assigned, allowing their representations to adapt to new changes more quickly. The update order follows the principle of "core priority, hierarchical progression": high-weight entities and their relationships are processed first to ensure the integrity of the main semantics, and then low-frequency edge nodes are updated layer by layer.

[0037] Example 2

[0038] Based on the above embodiment one, this embodiment also includes cross-modal entity linking of the multimodal financial regulatory graph based on multi-granularity interaction: A unified feature representation method is adopted, and pre-trained ViT and FinBERT models are used to extract global and local features of visual and textual modalities, respectively. By fusing global semantic and local lexical information through text global-local interaction units, enhancing the robustness of explicit visual features through visual dual interaction units, and mining implicit semantic associations between text and vision through cross-modal fusion interaction units; Based on the modality type and data quality of the input data, the output weights of the text global-local interaction unit, the visual dual interaction unit, and the cross-modal fusion interaction unit are dynamically adjusted through scene adaptive weight allocation; A unit consistency loss function is introduced to constrain the matching process of each interactive unit, thereby achieving accurate linking and semantic alignment of cross-modal entities.

[0039] Specifically, the pre-trained ViT model is used to extract multi-scale features from images (such as regulatory flowcharts, financial statement charts, and equity structure diagrams) to obtain global visual features and local key region features (such as table cells and legend labels); at the same time, the FinBERT model is used to perform deep semantic encoding on text (such as regulatory clauses and investment banking document descriptions) to obtain global semantic representations and local lexical features (such as entity words and relational phrases). Based on feature alignment, three parallel interaction units were designed: the Text Global-Local Interaction Unit (TGLU) integrates the global context and local key entities of the text through an attention mechanism, solving the problem of insufficient context caused by abbreviations and ambiguities in regulatory texts; the Visual Dual Interaction Unit (VDLU) uses a bidirectional gating mechanism to jointly enhance the global layout and local visual elements (such as arrow connecting lines and numerical annotations) of the image, effectively suppressing the interference of image noise (such as blurred scans and seal occlusion); and the Cross-Modal Fusion Interaction Unit (CMFU) mines the implicit semantic relationship between text and image through cross-modal projection and gating fusion strategies, such as the correspondence between "related party transactions" described in the text and the long dashed arrow in the image. To adapt to the uneven data quality in real financial documents, this method innovatively introduces a scene-adaptive weight allocation mechanism: the system dynamically adjusts the output weights of the above three interaction units according to the modal completeness of the input data (such as whether the image is missing or the text is verbose) and the quality confidence (such as OCR recognition confidence and text grammatical fluency). For example, when processing an investment banking document containing only blurry scanned charts without explanatory text, the weight of TGLU units is automatically reduced while the weight of VDLU units is increased to enhance visual feature mining. Conversely, when dealing with a plain text regulatory document, the capabilities of VDLU units are weakened while the capabilities of TGLU units are highlighted. Finally, a unit consistency loss function is introduced to constrain the consistency of the linking results of the three interaction units for the same entity through contrastive learning. If TGLU links an entity to graph node A while VDLU links it to node B, the loss function penalizes this discrepancy, forcing multimodal evidence to converge to a unified and correct node.

[0040] Example 3

[0041] Based on the above embodiment one, this embodiment also includes path reasoning based on knowledge enhancement using a dynamic multimodal financial knowledge graph: A relational graph convolutional network is used to dynamically model entities and relations in a dynamic multimodal financial knowledge graph. Different types of relations correspond to different dimensions of relation embedding representations. The relational graph convolutional network dynamically adjusts the information transmission weights of relation types according to the semantic context of each entity, and generates a low-dimensional vector representation that integrates multi-hop neighbor information. Features from different modalities are fused with the initial embeddings of graph nodes and used as input to the graph neural network. Based on the reasoning capabilities of graph neural networks, probabilistic path walking is used to automatically discover cross-modal association paths. During the path search process, a knowledge-guided path pruning strategy and path length penalty are introduced to eliminate redundant path branches that are not directly related to compliance. The comprehensive reliability score of each candidate path is calculated based on the product of the reasoning scores at each step in the path, and the complete reasoning chain of the query entity association is output.

[0042] Furthermore, the knowledge-enhanced path reasoning in the above embodiments is applied to investment banking document management and risk control scenarios, including: In response to query requests for investment banking working papers, the system automatically traces the consistency of financial data along the inference chain, identifies logical inconsistencies in the data in the working papers based on the verified inference chain, and matches and verifies the expert information database involved in the inference chain. In response to risk control monitoring requests, the system automatically tracks the correlation between fund flows and regulatory provisions along the inference chain and the financial industry compliance rules engine, providing early warnings of abnormal transaction behaviors.

[0043] Specifically, a relational graph convolutional network (RGCN) is used to model a dynamically updated multimodal financial knowledge graph. Unlike traditional GCNs, this method assigns different dimensional relation embedding representations to different types of relations in the graph (such as "constrained by," "violation," "reference," and "involving fund flows"). During message passing, the RGCN dynamically adjusts the information transmission weights of various relations based on the semantic context of each entity node (for example, the entity "a listed company" focuses more on financial disclosure relationships in investment banking scenarios, while focusing more on fund flow relationships in risk control scenarios). For example, when processing the "violation" relationship between "listed company A" and "CSRC penalty decision," the RGCN strengthens the feature transmission along this relationship path while weakening other relations unrelated to the query (such as "registered address"). After multiple convolutions, each entity node obtains a low-dimensional vector representation that incorporates multi-hop neighbor information. This representation includes both the semantics of the regulatory text and features from cross-modal data (such as financial tables and equity structure diagrams). Next, features from different modalities (such as chart trends in images and numerical sequences in tables) are concatenated with the initial embeddings of graph nodes and used as input to the graph neural network to achieve semantic alignment with knowledge enhancement. During the inference phase, the system automatically discovers cross-modal association paths based on a pre-trained RGCN using probabilistic path walking. Specifically, starting from the query entity (such as a suspicious transaction or anomalies in financial data in a prospectus), the probability of walking along the relation edge at each step is obtained by normalizing the dot product of the current node embedding and the edge embedding using softmax. To prevent path explosion and redundant output, the system introduces a knowledge-guided path pruning strategy and a path length penalty during path search: the former uses predefined financial compliance rules (such as "the anti-money laundering rule chain passes through a maximum of three intermediate accounts" or "financial reconciliation relationships must comply with accounting standards") to prune branches that clearly do not conform to business logic; the latter applies an exponential decay penalty to excessively long paths, encouraging the inference chain to remain concise. Ultimately, the system uses the product of the reasoning scores at each step in each candidate path as the overall reliability score for that path, and outputs the complete reasoning chain with the highest score, thus providing users with an interpretable and explicit chain of related evidence.

[0044] When this path reasoning is applied to investment banking document management and risk control scenarios, its operating principle is further linked with specific business rule engines. In the investment banking document management scenario, when the system receives a query request for a company's IPO prospectus document (e.g., "verify whether the operating revenue and operating cash flow match in the past three years"), the knowledge-enhanced path reasoning engine automatically starts from the two financial indicator entities of "operating revenue" and "operating cash flow" and performs a probabilistic walk along the "reconciliation relationship" edge learned by RGCN. Since the document usually contains multiple versions of financial tables, audit adjustment explanations, and related party transaction disclosures, path reasoning can penetrate different tables and text paragraphs to trace the reconciliation chain at the original document level.

[0045] Example 4

[0046] Based on the above embodiment one, the dynamic update of this embodiment also includes: adopting a two-stage training strategy. In the first stage, the old graph parameters are frozen by freezing the embedding parameters of non-new nodes in the old graph, and only the representation of new entities and relations is optimized. In the second stage, all parameters are unfrozen and fine-tuned together to achieve deep integration of new and old knowledge.

[0047] Specifically, upon receiving new financial regulations and their corresponding triples, the system first identifies the new entities (e.g., "programmatic trading reporting obligors") and new relations (e.g., "should report") involved in the new triples. At this stage, the system freezes the embedding parameters of all non-new nodes (i.e., existing entities, such as "securities companies" and "stock exchanges") in the initial multimodal financial knowledge graph. This is achieved by setting the gradient mask of these frozen parameters to zero during backpropagation gradient calculation, preventing them from participating in parameter updates. Simultaneously, the weight matrices corresponding to old relations in the relational graph convolutional network (RGCN) are also locked. The model only randomly initializes the embedding vectors of new entities and relations, allowing them to interact with frozen old nodes during forward propagation, optimized through incremental distillation loss functions (such as the aforementioned KL divergence combined with L1 / L2 regularization). This stage involves fewer training epochs (typically 5-10 epochs), enabling new knowledge to be quickly "grafted" onto the old graph structure without disrupting the distribution of the old semantic space due to large-scale gradient updates.

[0048] After the first stage ensures that the newly added representations converge to the old graph space, the system unfreezes the embedding parameters and relation matrices of the old nodes, allowing all graph parameters to participate in joint optimization. At this point, a smaller learning rate (typically 1 / 10 of the first stage) and a smaller number of epochs (e.g., 3-5) are used to fine-tune all entities and relations using a joint loss function (incremental distillation loss + cross-modal alignment loss + compliance path inference loss). The purpose of this stage is to eliminate local semantic conflicts that may have arisen in the first stage due to the forced "insertion" of new nodes, creating a deeply integrated and continuous representation space between old and new knowledge. For example, after joint fine-tuning, the distinction between "programmatic trading reporting obligor" and existing neighboring entities such as "high-frequency trading account" becomes clearer, while the causal path weights between the two and "self-regulatory measures" are further enhanced.

[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: Acquire multimodal financial regulatory data, which includes at least textual, tabular, and image-based data. We perform knowledge-enhanced semantic alignment and joint discovery of entity relationships on multimodal financial regulatory data to construct an initial multimodal financial knowledge graph. In response to newly added financial regulatory data, the initial multimodal financial knowledge graph is dynamically updated based on an incremental distillation strategy to generate a dynamic multimodal financial knowledge graph.

[0051] 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.

[0052] 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.

[0053] Example 6

[0054] 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: Acquire multimodal financial regulatory data, which includes at least textual, tabular, and image-based data. We perform knowledge-enhanced semantic alignment and joint discovery of entity relationships on multimodal financial regulatory data to construct an initial multimodal financial knowledge graph. In response to newly added financial regulatory data, the initial multimodal financial knowledge graph is dynamically updated based on an incremental distillation strategy to generate a dynamic multimodal financial knowledge graph.

[0055] 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 causal analysis method for financial regulations that integrates counterfactual reasoning, characterized in that, Includes the following steps: S01. Obtain multimodal financial regulatory data, wherein the multimodal financial regulatory data includes at least text modal data, tabular modal data and image modal data; S02. Perform knowledge-enhanced semantic alignment and joint discovery of entity relationships on the multimodal financial regulatory data to construct an initial multimodal financial knowledge graph; S03. In response to newly added financial regulatory data, the initial multimodal financial knowledge graph is dynamically updated based on an incremental distillation strategy to generate a dynamic multimodal financial knowledge graph. The dynamic update includes: performing hierarchical sorting on the newly added financial regulations data, dividing it into multiple learning levels based on the topological distance from the old graph, and determining the learning order of newly added triples in each level according to node centrality and relation betweenness centrality; and retaining the knowledge of the old graph through inter-layer representation distillation to obtain the updated multimodal financial knowledge graph.

2. The financial regulatory causal analysis method integrating counterfactual reasoning as described in claim 1, characterized in that, Step S02, which involves knowledge-enhanced semantic alignment and joint entity relationship discovery of the multimodal financial regulatory data, includes: The FinBERT model, OCR technology, and cross-modal attention mechanism are used to perform structured parsing and feature extraction on the text modal data, table modal data, and image modal data, respectively, to achieve semantic alignment of multimodal data; Based on knowledge graphs in the financial field, a knowledge-guided joint learning framework is constructed, which combines conditional random field constraints and multimodal association matrices to identify cross-modal entities; By using graph attention networks to perform multi-hop relationship reasoning on identified cross-modal entities, implicit cross-modal association paths can be discovered.

3. The financial regulatory causal analysis method integrating counterfactual reasoning as described in claim 1, characterized in that, The hierarchical sorting includes: Based on the breadth-first search algorithm, the layers are sorted according to the topological distance between the new triples and the old graph; The importance score of each newly added triplet is calculated using node centrality and relation betweenness centrality, and then the triplet is sorted within the layer.

4. The financial regulatory causal analysis method integrating counterfactual reasoning according to claim 1, characterized in that, The method of preserving old spectral knowledge through interlayer representation distillation includes: If the same entity appears in multiple layers, the difference between its current representation and the historical representation of the nearest layer is constrained by inter-layer representation distillation. A distillation loss function based on KL divergence is used in combination with L1 / L2 regularization to dynamically adjust the distillation intensity during training, so as to suppress noise propagation while retaining old knowledge. Differentiated weights are assigned to distillation loss based on the node degree distribution and relation betweenness centrality in the graph topology features. Priority is given to protecting the semantic integrity of core regulatory entities with high node degree and their related relationships, and then updates are performed sequentially.

5. The financial regulatory causal analysis method integrating counterfactual reasoning according to claim 1, characterized in that, It also includes cross-modal entity linking based on multi-granularity interaction for the multimodal financial regulatory map, which uses pre-trained ViT and FinBERT models to extract global and local features of visual and textual modalities, respectively. By fusing global semantic and local lexical information through text global-local interaction units, enhancing the robustness of explicit visual features through visual dual interaction units, and mining implicit semantic associations between text and vision through cross-modal fusion interaction units; Based on the modality type and data quality of the input data, the output weights of the text global-local interaction unit, the visual dual interaction unit, and the cross-modal fusion interaction unit are dynamically adjusted through scene adaptive weight allocation; A unit consistency loss function is introduced to constrain the matching process of each interactive unit, thereby achieving accurate linking and semantic alignment of cross-modal entities.

6. The financial regulatory causal analysis method integrating counterfactual reasoning according to claim 1, characterized in that, It also includes path reasoning based on the aforementioned dynamic multimodal financial knowledge graph for knowledge enhancement: The entities and relationships in the dynamic multimodal financial knowledge graph are dynamically modeled using a relational graph convolutional network. Different types of relationships correspond to different dimensions of relational embedding representations. The relational graph convolutional network dynamically adjusts the information transmission weights of relation types according to the semantic context of each entity, generating a low-dimensional vector representation that integrates multi-hop neighbor information. Features from different modalities are fused with the initial embeddings of graph nodes and used as input to the graph neural network. Based on the reasoning capabilities of graph neural networks, probabilistic path walking is used to automatically discover cross-modal association paths. During the path search process, a knowledge-guided path pruning strategy and path length penalty are introduced to eliminate redundant path branches that are not directly related to compliance. The comprehensive reliability score of each candidate path is calculated based on the product of the reasoning scores at each step in the path, and the complete reasoning chain of the query entity association is output.

7. The financial regulatory causal analysis method integrating counterfactual reasoning as described in claim 6, characterized in that, The knowledge-enhanced path reasoning described above is applied to investment banking documentation management and risk control scenarios, including: In response to a query request for investment banking working papers, the system automatically traces the consistency of financial data along the inference chain, identifies logical inconsistencies in the data in the working papers based on the verified inference chain, and matches and verifies the expert information database involved in the inference chain. In response to risk control monitoring requests, the system automatically tracks the correlation between fund flows and regulatory provisions along the inference chain and the financial industry compliance rules engine, and issues warnings for abnormal transaction behaviors.

8. The financial regulatory causal analysis method integrating counterfactual reasoning according to claim 1, characterized in that, The dynamic update also includes: The parameters of the old graph are frozen by freezing the embedding parameters of non-new nodes in the old graph, and only the representation of new entities and relations is optimized. Unfreeze all parameters and fine-tune them together to achieve a deep integration of new and old knowledge.

9. 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 financial regulatory causal analysis method that incorporates counterfactual reasoning as described in any one of claims 1 to 8.

10. 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 financial regulatory causal analysis method incorporating counterfactual reasoning as described in any one of claims 1 to 8.

Images