Risk identification method, device, equipment, medium and program product

Abstract

The embodiment of the specification discloses a risk identification method, device, equipment, medium and program product, which can improve the accuracy and speed of risk identification, thereby effectively coping with massive and rapidly changing regulatory rules. Among them, the method comprises: acquiring an event description text of a target business event; performing semantic alignment on an event semantic vector corresponding to the event description text and a regulatory semantic vector corresponding to a regulatory rule text to obtain a target regulatory semantic element associated with the target business event; reasoning the target regulatory semantic element based on a dynamic knowledge graph to determine at least one target regulatory rule; the dynamic knowledge graph is updated with the update of the regulatory dynamics, and the regulatory dynamics include the update dynamics of the regulatory rule text; and performing multi-modal verification on the target business event according to the at least one target regulatory rule to obtain a risk identification result of the target business event.

Description

Technical Field

[0001] This specification relates to the field of natural language processing technology, and in particular to a risk identification method, apparatus, device, medium, and program product. Background Technology

[0002] With the increasing complexity of online businesses and the continuous strengthening of regulatory oversight, higher demands are being placed on the accuracy and speed of business risk identification. Currently, related technologies typically require professionals to manually identify risks in business events based on regulatory rules periodically issued by regulatory authorities. There is a need to provide a more accurate and faster risk identification solution to cope with the massive and rapidly changing regulatory rules. Summary of the Invention

[0003] This specification provides a risk identification method, apparatus, device, medium, and program product that can improve the accuracy and speed of risk identification, thereby effectively addressing a large number of rapidly changing regulatory rules.

[0004] Firstly, embodiments of this specification provide a risk identification method, including: Obtain the event description text of the target business event; Semantically align the event semantic vector corresponding to the event description text with the regulatory semantic vector corresponding to the regulatory rule text to obtain the target regulatory semantic elements associated with the target business event. Reasoning is performed on the semantic elements of target supervision based on dynamic knowledge graph to determine at least one target supervision rule; the dynamic knowledge graph is updated with the updates of supervision dynamics, including the updates of the supervision rule text. Multimodal verification of the target business event is performed based on at least one target regulatory rule to obtain the risk identification results of the target business event.

[0005] In one possible implementation, the dynamic knowledge graph includes multiple nodes and edges connecting the nodes; the multiple nodes include regulatory entity nodes, regulatory condition nodes, and regulatory rule nodes; each edge is used to represent the semantic relationship between the nodes; Based on dynamic knowledge graphs, reasoning is performed on the semantic elements of target supervision to determine at least one target supervision rule, including: Based on the semantic elements of the target regulatory process and the event context information of the target business events, the associated subgraphs of the target business events are determined from the dynamic knowledge graph; Identify at least one target node in the associated subgraph that is associated with the target regulatory semantic element; Perform a traversal search on the associated subgraph to obtain at least one reasoning path that starts with the target node and ends with the regulatory rule node; Based on each reasoning path, at least one target regulatory rule is determined.

[0006] In one possible implementation, determining at least one target node in the association subgraph associated with the target regulatory semantic element includes: Perform string matching between the node description information of the node to be matched in the associated subgraph and the event description information of the target business event; Perform vector similarity matching between the node semantic vector of the node to be matched in the associated subgraph and the event semantic vector of the target business event; The node attribute information of the node to be matched in the associated subgraph is matched with the event context information of the target business event. Based on the matching results of string matching, vector similarity matching, and context matching, at least one target node associated with the target regulatory semantic element is selected from the association subgraph.

[0007] In one possible implementation, each edge has an edge weight determined based on the strength of semantic relationships; each regulatory rule node has a comprehensive weight determined based on multiple rule weight factors. Based on each reasoning path, at least one target regulatory rule is determined, including: Each inference path is scored based on the edge weights of the edges in each inference path and the combined weights of the regulatory rule nodes in each inference path. Based on the scoring results of each reasoning path, select at least one target reasoning path from at least one reasoning path; Based on at least one target reasoning path, determine at least one target regulatory rule.

[0008] In one possible implementation, multimodal verification of the target business event is performed based on at least one target regulatory rule to obtain the risk identification result of the target business event, including: Based on at least one target regulatory rule, multiple different modal verifications are performed on the target business event to obtain multimodal verification results; the multiple different modal verifications include at least two of the following: text semantic verification, numerical logic verification, graph structure verification, and temporal behavior verification; The multimodal verification results are fused to obtain the risk identification results of the target business event.

[0009] In one possible implementation, the above method also includes: Obtain the text of the regulatory rules; Semantic parsing of regulatory rule text yields multiple regulatory semantic elements; these elements include regulatory entities, regulatory conditions, and regulatory rules. An initial knowledge graph is constructed based on multiple regulatory semantic elements. The initial knowledge graph includes multiple nodes and edges connecting the nodes. The multiple nodes include regulatory entity nodes, regulatory condition nodes, and regulatory rule nodes. Each edge is used to represent the semantic relationship between the nodes. The comprehensive weight of each regulatory rule node is determined based on multiple rule weight factors corresponding to each regulatory rule node. These multiple rule weight factors include rule effectiveness level weight factor, rule application frequency weight factor, rule false alarm rate weight factor, and rule importance weight factor. The combined weights of each regulatory rule node are linked to the initial knowledge graph to form a dynamic knowledge graph.

[0010] In one possible implementation, after performing multimodal verification on the target business event according to at least one target regulatory rule and obtaining the risk identification result of the target business event, the method further includes: If the risk identification results of the target business event indicate that the target business event is a high-risk business event, a risk warning message for the target business event is generated.

[0011] In one possible implementation, after performing multimodal verification on the target business event according to at least one target regulatory rule and obtaining the risk identification result of the target business event, the method further includes: Based on the risk identification results, adjust the rule weight factors corresponding to the relevant regulatory rule nodes in the dynamic knowledge graph; The overall weight of the corresponding regulatory rule nodes is updated based on the adjusted rule weight factors to form an updated dynamic knowledge graph.

[0012] Secondly, embodiments of this specification provide a risk identification device, including: The acquisition module is used to acquire the event description text of the target business event; The mapping module is used to semantically align the event semantic vector corresponding to the event description text with the regulatory semantic vector corresponding to the regulatory rule text to obtain the target regulatory semantic elements associated with the target business event. The reasoning module is used to reason about the target regulatory semantic elements based on the dynamic knowledge graph to determine at least one target regulatory rule; the dynamic knowledge graph is updated with the regulatory dynamics, which include the update dynamics of the regulatory rule text. The verification module is used to perform multimodal verification of target business events according to at least one target regulatory rule, and obtain the risk identification results of the target business events.

[0013] Thirdly, embodiments of this specification provide an electronic device, including: a processor and a memory; the memory stores a computer program, and the processor executes the computer program to implement the method steps provided in the first aspect of embodiments of this specification.

[0014] Fourthly, embodiments of this specification provide a computer storage medium storing multiple instructions adapted for loading by a processor and executing the method steps provided in the first aspect of embodiments of this specification.

[0015] Fifthly, embodiments of this specification provide a computer program product, including a computer program; when the computer program is executed by a processor, it implements the method steps provided in the first aspect of embodiments of this specification.

[0016] The aforementioned risk identification methods, devices, electronic equipment, computer storage media, and computer program products, in the event understanding stage, acquire the event description text of the target business event, automatically align the event semantic vector corresponding to the event description text with the regulatory semantic vector corresponding to the regulatory rule text, accurately extract the target regulatory semantic elements associated with the target business event in the regulatory context, realize the automated dynamic mapping between event description and regulatory rules, and ensure that the extracted regulatory semantic elements can simultaneously reflect the specific characteristics of the business scenario and the applicable logic of the regulatory rules, thereby quickly providing accurate and reliable semantic input for subsequent reasoning and verification; in the reasoning and verification stage, based on the dynamic knowledge graph updated with regulatory dynamics (such as the dynamic updates of regulatory rule text), the obtained target regulatory semantic elements are automatically reasoned to obtain specific target regulatory rules applicable to the current business event, and the target business event is verified in a multimodal manner according to the target regulatory rules to obtain the risk identification result of the target business event, thereby effectively suppressing risk misidentification caused by single-modal verification through automated reasoning and verification. The entire risk identification process achieves rapid and accurate element extraction through dynamic alignment of event semantics and regulatory semantics, and completes efficient reasoning and robust judgment by combining dynamic knowledge graphs and multimodal verification. The entire automated risk identification process improves the speed and accuracy of risk identification, thereby effectively dealing with massive and rapidly changing regulatory rules. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this specification, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1A schematic diagram illustrating the application environment of a risk identification method provided as an exemplary embodiment of this specification; Figure 2 A flowchart illustrating a risk identification method provided as an exemplary embodiment of this specification; Figure 3 A flowchart illustrating another risk identification method provided as an exemplary embodiment of this specification; Figure 4 A flowchart illustrating yet another risk identification method provided as an exemplary embodiment of this specification; Figure 5 A flowchart illustrating another risk identification method provided as an exemplary embodiment of this specification; Figure 6 A schematic diagram of the architecture of a risk identification method provided for an exemplary embodiment of this specification; Figure 7 A schematic diagram of the structure of a risk identification device provided for an exemplary embodiment of this specification; Figure 8 This is a schematic diagram of the structure of an electronic device provided as an exemplary embodiment of this specification. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this specification clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this specification.

[0020] In the description of this specification, it should be understood that the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Those skilled in the art can understand the specific meaning of these terms in this specification based on the specific circumstances. Furthermore, in the description of this specification, unless otherwise stated, "multiple" means two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship.

[0021] The risk identification method provided in the embodiments of this specification can be applied to, for example, Figure 1 In the application environment shown, the business terminal 10 communicates with the server 20 corresponding to the risk identification platform via a communication network. The data storage system may include, but is not limited to, dynamic knowledge graphs that the server 20 needs to process. The data storage system can be integrated on the server 20, or it can be placed in the cloud or on other network servers.

[0022] In some possible embodiments, in response to a user's risk identification operation, the business terminal 10 sends a risk identification request for a target business event to the server 20. In response to the risk identification request, the server 20 obtains the event description text of the target business event, semantically aligns the event semantic vector corresponding to the event description text with the regulatory semantic vector corresponding to the regulatory rule text to obtain the target regulatory semantic elements associated with the target business event, infers the target regulatory semantic elements based on a dynamic knowledge graph in the data storage system, determines at least one target regulatory rule, the dynamic knowledge graph being updated with regulatory dynamics, including the update dynamics of the regulatory rule text, and performs multimodal verification on the target business event according to at least one target regulatory rule to obtain the risk identification result of the target business event.

[0023] Understandably, the business terminal 10 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, etc. The server 20 can be implemented using a standalone server or a server cluster consisting of multiple servers.

[0024] It is worth noting that the aforementioned risk identification platform is adaptable to various risk identification scenarios. For example: In the scenario of medical insurance audit risk identification, the risk identification platform can pre-build a dynamic knowledge graph based on medical insurance regulatory rules. The dynamic knowledge graph is updated with the updates of medical insurance regulatory dynamics, including the updates of the medical insurance regulatory rule text. When a target medical treatment reimbursement event (such as an inpatient reimbursement event) is received, the risk identification platform semantically aligns the event description text of the target medical treatment reimbursement event with the regulatory semantic vector corresponding to the medical insurance regulatory rule text to obtain the target regulatory semantic elements associated with the target medical treatment reimbursement event. Based on the dynamic knowledge graph, the platform infers the target regulatory semantic elements to determine at least one target regulatory rule. Then, based on at least one target regulatory rule, the platform performs multimodal verification on the target medical treatment reimbursement event to obtain its risk identification results, thereby achieving automatic compliance screening of medical insurance reimbursement events.

[0025] In industrial safety risk identification scenarios, the risk identification platform can pre-build a dynamic knowledge graph based on production supervision rules. The dynamic knowledge graph is updated with the dynamic updates of production supervision, including the updates of the production supervision rule text. When a target industrial operation application event (such as a hot work application event) is received, the event description text of the target industrial operation application event is semantically aligned with the supervision semantic vector corresponding to the production supervision rule text to obtain the target supervision semantic elements associated with the target industrial operation application event. Based on the dynamic knowledge graph, reasoning is performed on the target supervision semantic elements to determine at least one target supervision rule. And based on at least one target supervision rule, multimodal verification is performed on the target industrial operation application event to obtain its risk identification results, thereby realizing automatic hidden danger investigation of industrial operation application events.

[0026] In the context of financial compliance risk identification, the risk identification platform can pre-construct a dynamic knowledge graph based on financial regulatory rules. This dynamic knowledge graph updates with the dynamic changes in financial regulation, including the updates to the text of financial regulatory rules. When a target financial business event (such as a credit business event) is received, the platform semantically aligns the event description text with the corresponding regulatory semantic vector of the financial regulatory rule text to obtain the target regulatory semantic elements associated with the target financial business event. Based on the dynamic knowledge graph, the platform infers the target regulatory semantic elements to determine at least one target regulatory rule. Then, it performs multimodal verification on the target financial business event according to at least one target regulatory rule to obtain its risk identification results, thereby achieving compliance assessment of the financial event.

[0027] In one embodiment, such as Figure 2 As shown, a risk identification method is provided. Taking the application of this method to a server as an example, the method includes the following steps: S202: Obtain the event description text of the target business event.

[0028] The target business event is the business event to be identified. The target business event and its corresponding event description text vary depending on the risk identification scenario. For example, in a medical insurance review risk identification scenario, the target business event could be a target medical reimbursement event; in an industrial safety risk identification scenario, the target business event could be a target industrial operation application event; and in a financial compliance risk identification scenario, the target business event could be a target financial business event.

[0029] Optionally, the server can either directly receive event description text in natural language format uploaded by the business terminal, or generate event description text based on the original business data. For example, the server can generate event description text based on the original business data using a large language model. For instance, in a medical insurance review risk identification scenario, the server receives the reimbursement document for the target medical reimbursement event uploaded by the business terminal, and extracts fields such as patient information, diagnosis description, drug name, and cost details from the reimbursement document using a large language model to form the event description text for the target medical reimbursement event. In an industrial safety risk identification scenario, the server receives the application form for the target industrial operation application event uploaded by the business terminal, and extracts fields such as operation type, operation location, and operator qualifications from the application form using a large language model to form the event description text for the target industrial operation application event. In a financial compliance risk identification scenario, the server receives the event log of the target financial business event uploaded by the business terminal, and extracts fields such as transaction subject, transaction object, transaction amount, and transaction time from the event log using a large language model to form the event description text for the target financial business event.

[0030] S204: Semantically align the event semantic vector corresponding to the event description text with the regulatory semantic vector corresponding to the regulatory rule text to obtain the target regulatory semantic elements associated with the target business event.

[0031] Optionally, during system initialization or when regulatory rules are updated, the server performs semantic parsing and vectorization encoding on the regulatory rule text to extract structured regulatory semantic elements. For each regulatory semantic element, a regulatory semantic vector is generated and stored in the regulatory rule feature library. After obtaining the event description text of the target business event, the server performs vectorization encoding on the event description text to obtain an event semantic vector. Then, by calculating the similarity between the event semantic vector and each pre-stored regulatory semantic vector in the regulatory rule feature library, the event semantic vector is semantically aligned with the regulatory semantic vector. Next, regulatory semantic vectors with similarity higher than a preset threshold are used as target regulatory semantic vectors. Finally, the structured regulatory semantic elements corresponding to the target regulatory semantic vector are used as target regulatory semantic elements. Understandably, for regulatory rule text that has not yet been cached in the regulatory rule features when the event description text is obtained, the server can perform online semantic parsing and vectorization encoding as needed to obtain regulatory semantic vectors for semantic alignment with the event semantic vectors. The encoding results are stored in the regulatory rule feature library for reuse, thereby achieving real-time perception and adaptation to changes in regulatory trends.

[0032] For example, the structured regulatory semantic elements mentioned above are six-tuples including subject, action, object, scenario, time, and condition. For instance, in the scenario of medical insurance audit risk identification, the server obtains the event description text of the target medical treatment reimbursement event, "Patient A was hospitalized at a hospital on November 15, 2025, and received one imported coronary stent with a unit price of 8,000 yuan." By aligning the event semantic vector corresponding to the above event description text with the regulatory semantic vectors pre-stored in the regulatory rule feature library, the target regulatory semantic elements associated with the target medical treatment reimbursement event are extracted: {Subject: Patient A, Action: Received, Object: Imported coronary stent, Scenario: Hospitalization, Time: 2025-11-15, Condition: {Quantity: 1, Unit Price: 8,000 yuan}}. In the industrial safety risk identification scenario, the server obtains the event description text of the target industrial operation application event, "Operator B performed equipment E maintenance work in the closed space of the chemical plant reaction workshop on December 3, 2025". By aligning the event semantic vector corresponding to the above event description text with the regulatory semantic vectors pre-stored in the regulatory rule feature library, the target regulatory semantic elements associated with the target business event are extracted: {Subject: Operator B, Action: Perform equipment maintenance, Object: Equipment E, Scenario: Closed space of chemical plant reaction workshop, Time: 2025-12-03, Condition: {Whether gas detection is performed: No}. In the scenario of financial compliance risk identification, the server obtains the event description text of the target financial business event: "Customer C transferred RMB 5.2 million to offshore company F account via mobile banking on January 10, 2026". By aligning the event semantic vector corresponding to the above event description text with the regulatory semantic vectors pre-stored in the regulatory rule feature library, the target regulatory semantic elements associated with the target business event are extracted: {Subject: Customer C, Action: Cross-border transfer, Object: Offshore company F account, Scenario: Mobile banking channel, Time: 2026-01-10, Condition: {Amount: RMB 5.2 million, Whether transaction background materials are provided: No}}.

[0033] In this embodiment, after obtaining the event description text of the target business event, the server automatically performs semantic alignment between the event semantic vector corresponding to the target business event and the regulatory semantic vector corresponding to the regulatory rule text. This accurately extracts the target regulatory semantic elements associated with the target business event in the regulatory context, realizing automated dynamic mapping between the event description and the regulatory rules. This ensures that the extracted regulatory semantic elements can simultaneously reflect the specific characteristics of the business scenario and the applicable logic of the regulatory rules, thereby quickly providing accurate and reliable semantic input for subsequent reasoning and verification.

[0034] S206: Based on dynamic knowledge graphs, reason about the semantic elements of target supervision to determine at least one target supervision rule.

[0035] The dynamic knowledge graph updates in response to regulatory dynamics, including but not limited to the addition, revision, or repeal of regulatory rules. For example, when a new regulatory rule is added, nodes and edges can be dynamically inserted into the graph; when an existing regulatory rule is revised, nodes and edges can be dynamically adjusted; and when an existing regulatory rule is repealed, the corresponding nodes and edges can be deleted. Furthermore, the dynamic knowledge graph includes multiple independent, callable subgraphs, each containing multiple nodes and edges connecting them. These nodes include regulatory entity nodes, regulatory condition nodes, and regulatory rule nodes; the edges connecting the nodes represent the semantic relationships between them.

[0036] Optionally, the server first determines the association subgraph of the target business event from the dynamic knowledge graph based on the event context information (such as business subject attribute information, business spatiotemporal attribute information, and business domain attribute information) of the target regulatory semantic elements and the target business event; then, in the association subgraph, it determines at least one candidate node associated with the target regulatory semantic elements; next, it performs a traversal search on the association subgraph to obtain at least one reasoning path with the candidate node as the starting node and the regulatory rule node as the ending node; finally, based on each reasoning path, it determines at least one target regulatory rule.

[0037] Specifically, each edge in the dynamic knowledge graph has an edge weight determined based on the strength of semantic relationships, and each regulatory rule node in the dynamic knowledge graph has a comprehensive weight determined based on multiple rule weight factors (such as rule effectiveness level weight factor, rule application frequency weight factor, rule false alarm rate weight factor, and rule importance weight factor). The process by which the server determines at least one target regulatory rule based on each inference path is as follows: Based on the node representation of the regulatory rule node and the event representation of the target business event in each inference path, the server selects at least one candidate inference path from at least one inference path that starts with a candidate node and ends with a regulatory rule node; then, based on the edge weight of each candidate inference path and the comprehensive weight of the regulatory rule nodes contained in each candidate inference path, each candidate inference path is scored; next, based on the scoring results of each candidate inference path, at least one target inference path with a higher score is selected from at least one candidate inference path; finally, at least one target regulatory rule is determined based on at least one target inference path.

[0038] S208: Perform multimodal verification of the target business event based on at least one target regulatory rule to obtain the risk identification result of the target business event.

[0039] Optionally, the server performs multiple different modal verifications on the target business event according to at least one target regulatory rule, such as performing at least two different modal verifications among text semantic verification, numerical logic verification, graph structure verification, and temporal behavior verification, to obtain multimodal verification results. The multimodal verification results are then fused to obtain the risk identification results of the target business event.

[0040] Specifically, text semantic verification is used to determine whether the event description text of the target business event matches the target regulatory rules at the semantic level. In the embodiments of this specification, the server can process the text files of the target business event through a natural language processing model (such as BERT) and output the text semantic verification results. Numerical logic verification is used to verify whether the numerical attributes in the target business event meet the quantitative constraints of the target regulatory rules. In the embodiments of this specification, the server can process the numerical files of the target business event through a DSL (Domain-Specific Language) interpreter and an SQL engine and output the numerical logic verification results. Graph structure verification is used to detect hidden risk paths in the target business event that are not explicitly declared but violate regulatory rules based on the network relationship graph of the target business event. In the embodiments of this specification, the server can process the relationship network and / or entity relationship graph of the target business event through a graph neural network model or a link prediction model and output the graph structure verification results. Temporal behavior verification is used to analyze whether there are regulatory evasion patterns in the time series data of target business events. In the embodiments of this specification, the server can process the time series data of target business events through an LSTM (Long Short-Term Memory) anomaly detection model or a GAN (Generative Adversarial Network) reconstruction error model and output the temporal behavior verification results.

[0041] Understandably, all of the verification results mentioned above can be verification scores. For example, text semantic verification results can be text semantic verification scores, numerical logic verification results can be numerical logic verification scores, graph structure verification results can be graph structure verification scores, and temporal behavior verification results can be temporal behavior verification scores, etc. The server can ultimately perform weighted fusion processing on the verification scores of multiple different modalities to obtain a comprehensive risk score for the target business event, and obtain the risk identification result of the target business event based on the comprehensive risk score of the target business event and the preset risk threshold.

[0042] It is worth noting that the above is merely a preferred embodiment of the embodiments described in this specification. In practical applications, the server can directly determine the target business event as high-risk when the verification score of any modality exceeds the corresponding preset risk threshold; alternatively, weighted voting or other integrated strategies can be used to synthesize the verification results of each modality to obtain the risk identification result of the target business event.

[0043] For example, consider a financial compliance risk identification scenario: If the target regulatory rule requires that "loan contracts must clearly stipulate repayment methods and liabilities for breach of contract," the server uses a natural language processing model (such as BERT) to identify whether key fields such as "penalty for breach of contract" and "repayment plan" exist in text files such as contracts, announcements, and emails related to the target financial business event. Based on the identification results, the server outputs a text semantic verification score. If the target regulatory rule stipulates that "single loans exceeding 500,000 yuan must be reported," the server uses a natural language processing model to extract key elements from the target regulatory rule. Then, it uses a DSL interpreter to convert the extracted key elements into predefined compliance DSL expressions, and converts these DSL expressions into SQL query statements applicable to the target data source (such as the database table corresponding to the target business event). The SQL engine then executes the query in the target data source, and finally, the server outputs a numerical logic verification score based on the query results. If the target regulatory rule requires "prohibiting repeated credit extensions to related enterprises," the server, based on the entity relationship network such as equity graphs and guarantee circles in the dynamic knowledge graph, uses graph neural network models or link prediction models to analyze the implicit connection paths between the target enterprise and the already credited enterprises. For example, it checks for the existence of joint actual controllers, cross-shareholdings, or mutual guarantees. If an undeclared but risk-transmission path is detected, a hidden connection risk is identified, and a graph structure verification score is output. If the target regulatory rule requires "prohibiting customers from circumventing large-amount reporting obligations through frequent small transactions," the server, through a time-series behavior verification module, uses time-series data such as transaction flows and approval logs, employing LSTM anomaly detection models or GAN reconstruction error models to identify the customer's behavior pattern of splitting multiple transactions within a short period. For example, if the same customer initiates three remittances, each less than $50,000, to the same overseas recipient within 24 hours, accumulating to near the foreign exchange limit, the LSTM model can identify this behavior as exhibiting typical characteristics of "split transactions," thereby triggering a risk warning and outputting a time-series behavior verification score.

[0044] Finally, the server performs weighted fusion processing on the aforementioned text semantic verification score, numerical logic verification score, graph structure verification score, and temporal behavior verification score to obtain a comprehensive risk score for the target business event. If the comprehensive risk score of the target business event is greater than a first preset risk threshold, the server determines the target business event as a high-risk business event and generates a risk warning message. If the comprehensive risk score of the target business event is greater than a second preset risk threshold but less than or equal to the first preset risk threshold, the server determines the target business event as a medium-risk business event and generates compliance recommendation information. If the comprehensive risk score of the target business event is less than or equal to the second preset risk threshold, the target business event is determined as a low-risk business event and compliance recommendation information is generated.

[0045] In this embodiment, the server automatically infers the obtained target regulatory semantic elements based on a dynamic knowledge graph to obtain specific target regulatory rules applicable to the current business event. Based on the target regulatory rules, the server performs multimodal verification on the target business event to obtain the risk identification result of the target business event. Thus, through automated inference verification, the server effectively suppresses the risk misidentification caused by single-modal verification.

[0046] In the aforementioned risk identification method, during the event understanding phase, the server acquires the event description text of the target business event and automatically aligns the event semantic vector corresponding to the event description text with the regulatory semantic vector corresponding to the regulatory rule text. This accurately extracts the target regulatory semantic elements associated with the target business event in the regulatory context, achieving automated dynamic mapping between the event description and regulatory rules. This ensures that the extracted regulatory semantic elements can simultaneously reflect the specific characteristics of the business scenario and the applicable logic of the regulatory rules, thus providing accurate and reliable semantic input for subsequent reasoning and verification. During the reasoning and verification phase, the server automatically reasones about the acquired target regulatory semantic elements based on a dynamic knowledge graph that is updated with regulatory dynamics (such as the dynamic updates of regulatory rule texts). This yields specific target regulatory rules applicable to the current business event. Multimodal verification is then performed on the target business event based on the target regulatory rules to obtain the risk identification results of the target business event. Thus, through automated reasoning and verification, the risk misidentification caused by single-modal verification is effectively suppressed. By employing the aforementioned risk identification method, rapid and accurate element extraction is achieved through dynamic alignment of event semantics and regulatory semantics. Combined with dynamic knowledge graphs and multimodal verification, efficient reasoning and robust judgment are completed. The entire automated risk identification process improves the speed and accuracy of risk identification, thereby effectively addressing massive and rapidly changing regulatory rules.

[0047] In one embodiment, such as Figure 3As shown, another risk identification method is provided. Taking the application of this method to a server as an example, the method includes the following steps: S302: Obtain the text of the regulatory rules.

[0048] Among them, regulatory rule texts are normative texts used to constrain or assess the compliance of target business events. The regulatory rule texts vary depending on the risk identification scenario. For example, in the scenario of medical insurance audit risk identification, the regulatory rule text may be medical insurance regulatory rules, such as rules on the use of medical insurance funds, or texts on medical insurance reimbursement restrictions; in the scenario of industrial safety risk identification, the regulatory rule text may be production regulatory rules, such as texts on production operation specifications, or texts on production license management regulations; in the scenario of financial compliance risk identification, the regulatory rule text may be financial regulatory rules, such as texts on credit management specifications, or texts on large-value transaction management specifications.

[0049] Optionally, the server receives regulatory rule files in PDF, Word, or HTML format uploaded by the business terminal, extracts the original text content from the regulatory rule files using a document parsing tool, and performs noise filtering and text segmentation on the original text content to generate structured regulatory rule text. Understandably, the regulatory rule text can consist of multiple independent regulatory constraint statements, serving as input for subsequent semantic vectorization and knowledge graph construction. For example, the regulatory rule text could be "A single cash transaction exceeding 500,000 yuan requires the submission of a large transaction report."

[0050] S304: Perform semantic parsing on the regulatory rule text to obtain multiple regulatory semantic elements; these multiple regulatory semantic elements include regulatory entities, regulatory conditions, and regulatory rules.

[0051] Optionally, the server performs semantic parsing on the regulatory rule text using a large language model to obtain multiple regulatory semantic elements.

[0052] The regulatory entity includes both the regulatory body that performs the regulation and the regulated object. In the scenario of identifying risks in medical insurance audits, the regulatory body could be a medical insurance regulatory agency, and the regulated object could be reimbursed drugs, insured individuals, etc. In the scenario of identifying industrial safety risks, the regulatory body could be an industrial safety production regulatory agency, and the regulated object could be the type of industrial operation, the type of industrial equipment, the industrial workers, etc. In the scenario of identifying financial compliance risks, the regulatory body could be a financial regulatory agency, and the regulated object could be financial transaction products, financial customer categories, etc.

[0053] Regulatory conditions are constraints used to limit the scope of regulatory application or the preconditions for regulatory triggering. In the scenario of medical insurance audit risk identification, regulatory conditions can be restrictive thresholds or logical rules related to medical treatment behavior, drug use, and cost structure, such as "single outpatient cost exceeding 2,000 yuan" or "patient age under 18 years old." In the scenario of industrial safety risk identification, regulatory conditions can be safety thresholds or safety types associated with industrial operation attributes, personnel qualifications, and environmental parameters, such as "continuous operation time not exceeding 8 hours" or "operation type is hot work." In the scenario of financial compliance risk identification, regulatory conditions can be numerical or logical constraints on transaction characteristics, customer attributes, and financial products, such as "single loan amount exceeding 500,000 yuan" or "wealth management products are sold to compliant investors."

[0054] Regulatory rules are regulatory provisions or normative units with independent compliance judgment significance. In the scenario of identifying risks in medical insurance audits, regulatory rules can be medical insurance regulatory rule provisions and their corresponding provision numbers, or medical insurance regulatory rule notices and their corresponding notice numbers. In the scenario of identifying industrial safety risks, regulatory rules can be production regulatory rule provisions and their corresponding provision numbers, or production regulatory rule notices and their corresponding notice numbers. In the scenario of identifying financial compliance risks, regulatory rules can be financial regulatory rule provisions and their corresponding provision numbers, or financial regulatory rule notices and their corresponding notice numbers.

[0055] S306: Construct an initial knowledge graph based on multiple regulatory semantic elements.

[0056] The initial knowledge graph consists of multiple nodes and edges connecting them. The nodes include regulatory entity nodes, regulatory condition nodes, and regulatory rule nodes; the edges represent the semantic relationships between the nodes.

[0057] Optionally, the server creates regulatory entity nodes based on regulatory entities among multiple regulatory semantic elements; creates regulatory condition nodes based on regulatory conditions among multiple regulatory semantic elements; and creates regulatory rule nodes based on regulatory rules among multiple regulatory semantic elements. Subsequently, based on the logical association between each regulatory semantic element, edges connecting each node are established. Understandably, the semantic relationship types corresponding to the above edges may be, but are not limited to, applicable, required, belonging, referenced, violated, etc. Finally, attribute fields are added to the nodes and / or edges to enhance semantic expressiveness. For example, the attribute information of the regulatory rule node includes, but is not limited to, the effective date of the regulatory rule, the scope of application of the regulatory rule, the penalty intensity of the regulatory rule, the historical hit count of the regulatory rule, and the historical false alarm rate of the regulatory rule.

[0058] It is worth noting that each edge has an edge weight determined based on the strength of the semantic relationship. Specifically, the weight of an edge representing a "reference" semantic relationship type can be determined based on the reference strength. For example, an edge representing a "direct reference" semantic relationship type has a higher weight, which can be assigned a value of 0.9; an edge representing a "direct reference" semantic relationship type has a lower weight, which can be assigned a value of 0.3. The weight of an edge representing an "applicable" semantic relationship type can be determined based on the historical verification fit of the nodes it connects to. The weight of an edge representing a "conflict" semantic relationship type can be determined based on the severity of the conflict of the nodes it connects to. Understandably, edge weights can come from static rule settings or can be dynamically adjusted based on events or feedback. For example, if a certain semantic relationship type appears frequently in the risk identification process recently, its weight will increase. In addition, each regulatory rule node has a comprehensive weight determined based on multiple rule weight factors, which will be detailed later and will not be repeated here.

[0059] S308: Determine the comprehensive weight of each regulatory rule node based on multiple rule weight factors corresponding to each regulatory rule node.

[0060] Optionally, the server determines the comprehensive weight of each regulatory rule node based on multiple rule weight factors corresponding to each regulatory rule node using a preset comprehensive weight calculation formula. These multiple rule weight factors include a rule effectiveness level weight factor, a rule application frequency weight factor, a rule false alarm rate weight factor, and a rule importance weight factor. Specifically, the comprehensive weight calculation formula is as follows: (1) in, This represents the overall weight of a regulatory rule node, which is used to characterize the dynamic importance of the regulatory rule corresponding to that node in the current regulatory environment. This is the rule effectiveness level weight factor, which is used to characterize the effectiveness level of regulatory rules. Its value range is [0,1]. For example, if the regulatory rule corresponding to a certain regulatory rule node comes from a legal provision, then the effectiveness level weight factor corresponding to that node can be 1; if the regulatory rule corresponding to a certain regulatory rule node comes from a regulatory notice, then the effectiveness level weight factor corresponding to that node can be 0.5. This is the application frequency weighting factor for the rule, which reflects the recent activity level of the regulatory rule. It is calculated based on the normalized number of times the regulatory rule is invoked within the sliding time window. The false alarm rate weighting factor is used to measure the reliability of a rule. It is determined based on the proportion of historical risk identification results where the rule was triggered but the identification result was incorrect after manual review. This is a weighting factor for the importance of rules, used to quantify the intensity of attention that regulatory agencies pay to relevant regulatory rules. , , , This is a global hyperparameter used to adjust the contribution ratio of each rule weight factor in the comprehensive weight calculation formula.

[0061] Understandably, during the system initialization phase, due to the lack of historical risk identification data, the aforementioned rule weight factors and global hyperparameters can be assigned initial values ​​by domain experts based on experience. As the system continues to run, the rule weight factors can be dynamically updated based on historical risk identification data, and the global hyperparameters can be adaptively optimized through machine learning algorithms. For example, the server can use a supervised learning algorithm to optimize the global hyperparameters by fitting the contribution of each rule weight factor to risk identification through logistic regression or a linear model based on historical risk identification data that has been manually reviewed; the server can also use a reinforcement learning algorithm to explore the global hyperparameters through policy gradients based on a reward function that optimizes the false positive rate; in addition, the server can also use a Bayesian optimization algorithm to search for the global hyperparameters that maximize the accuracy of the overall risk identification results in a black-box environment.

[0062] S310: Associate the comprehensive weights of each regulatory rule node with the initial knowledge graph to form a dynamic knowledge graph.

[0063] Optionally, after calculating the comprehensive weight of each regulatory rule node based on the effectiveness level weight factor, rule application frequency weight factor, rule false alarm rate weight factor, and rule importance weight factor, the server associates the comprehensive weight of each regulatory rule node with the initial knowledge graph constructed based on multiple regulatory semantic elements, thereby forming a dynamic knowledge graph that can evolve dynamically with changes in external regulatory trends and internal risk identification feedback.

[0064] In this embodiment, by parsing unstructured regulatory rule text into three categories of regulatory semantic elements—regulatory entities, regulatory conditions, and regulatory rules—and constructing an initial knowledge graph based on these three categories, lengthy and complex regulatory rules can be transformed into structured and computable knowledge representations. Furthermore, by determining the comprehensive weight of each regulatory rule node based on multiple rule weight factors and associating this comprehensive weight with the initial knowledge graph, a dynamic knowledge graph that can evolve dynamically with changes in external regulatory trends and internal risk identification feedback can be formed. This provides weighted regulatory rule basis for subsequent multimodal risk verification, effectively improving the accuracy and efficiency of risk identification.

[0065] S312: Obtain the event description text of the target business event.

[0066] Specifically, S312 is the same as S202, and will not be repeated here.

[0067] S314: Semantically align the event semantic vector corresponding to the event description text with the regulatory semantic vector corresponding to the regulatory rule text to obtain the target regulatory semantic elements associated with the target business event.

[0068] Specifically, S314 is the same as S204, and will not be repeated here.

[0069] S316: Based on dynamic knowledge graphs, reason about the semantic elements of target supervision to determine at least one target supervision rule.

[0070] Specifically, S316 is the same as S206, and will not be repeated here.

[0071] S318: Perform multimodal verification of the target business event based on at least one target regulatory rule to obtain the risk identification result of the target business event.

[0072] Specifically, S318 is the same as S206, and will not be repeated here.

[0073] S320: Determine whether the risk identification result of the target business event indicates that the target business event is a high-risk business event. If yes, proceed to S322; otherwise, proceed to S324.

[0074] Optionally, the server determines whether the risk identification result of the target business event indicates that the target business event is a high-risk business event based on the comprehensive risk score of the target business event. If the risk identification result of the target business event indicates that the target business event is a high-risk business event, the server generates risk warning information for the target business event; if the risk identification result of the target business event indicates that the target business event is a non-high-risk business event (such as a medium-risk or low-risk business event), the server generates compliance recommendation information for the target business event.

[0075] S322: Generate risk warning information for the target business event.

[0076] Among them, risk warning information is an alert output generated by the server when it determines that a target business event is a high-risk business event. This alert serves to warn business personnel of a clear violation risk associated with the target business event and provides risk details and emergency response guidelines. Risk warning information includes alerts about compliance risks in the target business event, descriptions of the compliance risks, and guidance on handling the compliance risks. Specifically, compliance risk alerts include the level and type of compliance risk; the description of the compliance risk includes the regulatory rules triggered by the compliance risk, the regulatory entities involved in the compliance risk, the regulatory conditions involved in the compliance risk, and the multimodal verification results of the compliance risk; the guidance on handling the compliance risk is generated based on the type of compliance risk and the business scenario, and may include suggestions such as "recommend manual review," "block the transaction," or "generate and submit a large transaction report."

[0077] S324: Generate compliance recommendation information for the target business event.

[0078] The compliance advice information is a guiding output generated by the server when it determines that the target business event is not a high-risk business event (e.g., it is a medium-risk or low-risk business event). This information assists business personnel in optimizing operations and preventing potential compliance risks. The compliance advice information includes alerts about compliance risks in the target business event, descriptions of the compliance risks, and suggestions for handling them. Specifically, the compliance risk alerts include the level and type of compliance risk; the descriptions include the regulatory rules triggering the compliance risk, the regulatory entities involved, the regulatory conditions involved, and the multimodal verification results of the compliance risk; the guidance on handling compliance risks is generated based on the type of compliance risk and the business scenario, and may include suggestions such as "suggesting supplementary due diligence materials" or "suggesting the generation of a transaction report."

[0079] In this embodiment, the server determines the risk level of a target business event based on the risk identification results and generates risk warning information or compliance suggestion information accordingly, thus achieving differentiated responses to business events with different risk levels. Specifically, when the risk identification results of a target business event indicate that it is a high-risk event, the server generates risk warning information that includes a compliance risk alert, a multi-dimensional description of the compliance risk, and guidance on handling the compliance risk, thereby achieving a risk alert with strong interpretability and operational guidance. When the risk identification results of a target business event indicate that it is a medium-risk or low-risk event, the server generates compliance suggestion information that includes a compliance risk prompt, a description of the compliance risk, and suggestions on handling the compliance risk, thereby guiding business personnel to proactively improve operational details and prevent potential risks. The entire process effectively improves the pertinence of risk response and the efficiency of risk handling.

[0080] S326: Based on the risk identification results, adjust the rule weight factors corresponding to the relevant regulatory rule nodes in the dynamic knowledge graph.

[0081] Optionally, as the system continues to run, the server can use the risk identification results of the target business event as feedback signals to continuously adjust the rule weight factors of the relevant regulatory rule nodes in the dynamic knowledge graph. Understandably, the multiple rule weight factors corresponding to each regulatory rule node in the dynamic knowledge graph include rule effectiveness level weight factors, rule application frequency weight factors, rule false alarm rate weight factors, and rule importance weight factors. The server can count the number of times each regulatory rule node hits the target business event within a preset time window at preset time intervals, and adjust the rule application frequency weight factor of the relevant regulatory rule node based on the number of times it is invoked. Furthermore, the server can also combine the manual review results of the risk identification results by domain experts, count the number of false identifications of the target business events associated with each regulatory rule node within a preset time window at preset time intervals, and update the rule false alarm rate weight factor of the relevant regulatory rule node based on the number of false identifications of the associated target business events.

[0082] It is worth noting that as the system continues to run, the server can also continuously adjust the rule weight factors of relevant regulatory rule nodes in the dynamic knowledge graph based on external regulatory dynamics. For example, the system can adjust the rule importance weight factors of relevant regulatory rule nodes by continuously monitoring the regulatory attention of external regulatory agencies (such as medical insurance regulatory agencies) or media organizations; or, the server can also adjust the effectiveness level weight factors of relevant regulatory rule nodes based on the dynamic updates of the regulatory rule text.

[0083] S328: Update the comprehensive weight of the corresponding regulatory rule nodes based on the adjusted rule weight factors to form an updated dynamic knowledge graph.

[0084] Optionally, the server performs weight rebalancing on the relevant regulatory rule nodes based on the adjusted rule weight factor and according to the aforementioned formula (1), that is, recalculates the comprehensive weight of the relevant regulatory rule nodes; then, the recalculated comprehensive weight is written into the weight attribute field of the corresponding regulatory rule node in the dynamic knowledge graph to complete the incremental update of the graph and form the updated dynamic knowledge graph.

[0085] In this embodiment, the server uses the risk identification result as a feedback signal to dynamically adjust the rule weight factors of relevant regulatory rule nodes, and recalculates the comprehensive weight based on the updated weight factors, thereby updating the dynamic knowledge graph. In subsequent risk identification, multimodal verification and rule matching can be performed based on the updated dynamic knowledge graph, so that regulatory rules with high sensitivity, high attention, or high reliability receive higher decision weights, thereby improving the accuracy of risk identification. At the same time, the incremental update mechanism of the graph avoids the overhead of full reconstruction, ensuring the continuity, stability, and efficiency of risk identification.

[0086] In one embodiment, such as Figure 4 As shown, another risk identification method is provided. Taking the application of this method to a server as an example, the method includes the following steps: S402: Obtain the event description text of the target business event.

[0087] Specifically, S402 is the same as S202, and will not be repeated here.

[0088] S404: Semantically align the event semantic vector corresponding to the event description text with the regulatory semantic vector corresponding to the regulatory rule text to obtain the target regulatory semantic elements associated with the target business event.

[0089] Specifically, S404 is the same as S204, and will not be repeated here.

[0090] S406: Based on the target regulatory semantic elements and the event context information of the target business events, determine the association subgraph of the target business events from the dynamic knowledge graph.

[0091] The event context information includes at least one of the following: business entity attribute information, business spatiotemporal attribute information, and business domain attribute information. The aforementioned business entity attribute information includes, but is not limited to, institution type and customer credit rating; the aforementioned business spatiotemporal attribute information includes, but is not limited to, the time of the business event and the region to which the business event belongs; the aforementioned business domain attribute information includes, but is not limited to, business type.

[0092] Understandably, a dynamic knowledge graph comprises multiple independent, callable subgraphs, each containing multiple nodes and edges connecting them. These nodes include regulatory entity nodes, regulatory condition nodes, and regulatory rule nodes; the edges represent the semantic relationships between the nodes. Furthermore, each subgraph corresponds to a specific regulatory theme. For example, in the scenario of medical insurance review risk identification, the dynamic knowledge graph might include subgraphs such as "Drug Use Supervision" and "Insured Person Identity Supervision"; in the scenario of industrial safety risk identification, it might include subgraphs such as "Hot Work Supervision" and "Special Equipment Operation"; and in the scenario of financial compliance risk identification, it might include subgraphs such as "Loan Application Supervision" and "Customer Due Diligence Supervision".

[0093] Optionally, the server can determine the regulatory topic corresponding to the target business event based on the target regulatory semantic elements obtained from semantic alignment and the context information of the target business event. Based on the regulatory topic corresponding to the target business event, the server can determine the association subgraph of the target business event from the dynamic knowledge graph. Specifically, the server processes the target regulatory semantic elements and event context information in the form of six tuples through a preset classification model, mapping the target business event to the corresponding regulatory topic; then, it searches for the subgraph associated with the regulatory topic in the dynamic knowledge graph, which serves as the association subgraph of the target business event.

[0094] S408: Identify at least one target node in the association subgraph that is associated with the target regulatory semantic element.

[0095] Optionally, the server filters at least one target node associated with the target regulatory semantic element from the association subgraph based on the node representations of each node to be matched in the association subgraph and the event representation of the target business event. The node to be matched is either a regulatory entity node or a regulatory condition node; the node representation includes node description information, node semantic vector, and node attribute information; and the event representation includes event description information, event semantic vector, and event context information.

[0096] In one embodiment, the server determines at least one target node in the association subgraph that is associated with the target regulatory semantic element, including: performing string matching between the node description information of the node to be matched in the association subgraph and the event description information of the target business event; performing vector similarity matching between the node semantic vector of the node to be matched in the association subgraph and the event semantic vector of the target business event; performing context matching between the node attribute information of the node to be matched in the association subgraph and the event context information of the target business event; and filtering out at least one target node associated with the target regulatory semantic element from the association subgraph based on the matching results of string matching, vector similarity matching, and context matching.

[0097] String matching refers to directly comparing the literal text of the node description information of the node to be matched with the event description information of the target event, which is used to capture display and deterministic keywords. Optionally, the server can perform keyword matching on the node description information of the node to be matched and the event description information in the associated subgraph, and calculate the keyword similarity between the two (such as Jaccard similarity). If the node description information and the event description information have an intersection on a preset keyword set and the keyword similarity exceeds a first preset threshold, then the string matching is determined to be successful.

[0098] Vector similarity matching involves encoding the node description information of the node to be matched and the event description information of the target business event into semantic vectors using a semantic encoding model, and then calculating the semantic similarity between the two. This measure assesses the closeness of the node description and event description in the semantic space, overcoming literal differences and achieving semantic-level generalized matching. Optionally, the server pre-encodes the node description information of the node to be matched and the event description information of the target business event into semantic vectors using a BERT model, and then calculates the cosine similarity between the two semantic vectors. If the cosine similarity is higher than a second preset threshold, the vector similarity match is considered successful.

[0099] Context matching refers to the logical comparison of the structured attribute information of the node to be matched with the event context information of the target business event. It is used to determine whether the target business event meets the applicable conditions of the node to be matched. Optionally, the server inputs the structured attribute information of the node to be matched and the event context information of the target business event into the large language model, so that the large language model outputs the context matching degree. If the context matching degree output by the large language model is higher than a third preset threshold, the context matching is determined to be successful.

[0100] Understandably, for any node to be matched, the server may determine whether it is a target node associated with the target regulatory semantic element using any of the following methods: Method 1: After completing string matching, vector similarity matching, and context matching, a weighted fusion strategy is used to combine the three matching results and calculate the comprehensive matching score for the node to be matched. The weights of string matching, vector similarity matching, and context matching in the weighted fusion decrease sequentially. Then, based on the comprehensive matching score and a preset threshold, it is determined whether any node to be matched is a target node associated with the target regulatory semantic element. For example, if the comprehensive matching score is greater than the preset threshold, the node to be matched is determined to be a target node associated with the target regulatory semantic element; if the comprehensive matching score is less than or equal to the preset threshold, the node to be matched is determined not to be a target node associated with the target regulatory semantic element.

[0101] Method 2: The three matching methods are sequentially filtered according to a preset matching priority order, where string matching has a higher priority than vector similarity matching, and vector similarity matching has a higher priority than context matching. If a match is successful at any matching stage, that node is directly determined to be a target node associated with the target regulatory semantic element. If a match fails at any matching stage, the next matching stage is executed. If a match fails in all three matching stages, that node is determined not to be a target node associated with the target regulatory semantic element.

[0102] In this embodiment, the server uses a collaborative matching mechanism of string matching, vector similarity matching, and context matching to not only improve the coverage and semantic depth of target node identification, but also ensure its high relevance and adaptability in specific business scenarios. This provides a high-precision and high-reliability starting point for subsequent inference path generation, thereby improving the speed and accuracy of risk identification and realizing reliable risk identification to effectively cope with massive and rapidly changing regulatory rules.

[0103] S410: Perform a traversal search on the associated subgraph to obtain at least one reasoning path that starts with the target node and ends with the regulatory rule node.

[0104] Optionally, in the associated subgraph, the server uses each target node as a starting node and employs a heuristic graph search algorithm (such as the A* algorithm) to perform path search, obtaining at least one reasoning path that starts with a target node and ends with a regulatory rule node. Specifically, during the search process, the heuristic graph search algorithm calculates the path from the starting node to the current node based on the edge weights of each edge in the path. The cumulative path cost is used as the actual path cost. Based on the current node Construct a heuristic function to predict the relevance to any regulatory rule node. The above-mentioned estimated correlation is derived from the current node. The comprehensive weight of reachable regulatory rule nodes is determined; subsequently, based on the aforementioned actual path costs... Compared with the above heuristic functions Calculate the total assessment value The system dynamically selects nodes to expand first. When any regulatory rule node is reached for the first time during the search process, the complete search path corresponding to that search process is determined as a reasoning path.

[0105] Furthermore, the actual path cost mentioned above It can be from the starting node to the current node. The sum of the weights of all edges on the path; the above heuristic function It can be from the current node The maximum overall weight value among all regulatory rule nodes reachable within the pre-path length (e.g., within the preset number of hops); the above overall evaluation function can be: ,in These are configurable balance parameters.

[0106] S412: Based on each reasoning path, determine at least one target regulatory rule.

[0107] Optionally, the server determines at least one target regulatory rule by identifying the regulatory rules corresponding to the regulatory rule nodes in each inference path as the target regulatory rule.

[0108] In this embodiment, the server determines the associated subgraph corresponding to the target regulatory semantic elements and the event context information of the target business event from the dynamic knowledge graph. This effectively reduces the interference of irrelevant nodes and edges in the dynamic knowledge graph, improves the overall reasoning efficiency of the applicable rule set for the target business event, and thus increases the speed of risk identification. Subsequently, the server uses a collaborative matching mechanism of string matching, vector similarity matching, and context matching to select target nodes in the associated subgraph that are highly correlated and adaptable to the target regulatory semantic elements. These nodes serve as the starting points for the reasoning path, providing a high-precision and high-reliability starting point for subsequent reasoning path generation, thereby improving the accuracy of risk identification. Next, the server performs a traversal search in the associated subgraph starting from the target node. It calculates the actual path cost using the semantic relationship strength represented by edge weights and constructs a heuristic function based on the comprehensive weights of regulatory rule nodes. This dynamically guides the search to find at least one complete and reasonable reasoning path. Based on the regulatory rule node at the end of the obtained reasoning path, the corresponding regulatory rule is directly extracted as the target regulatory rule. This effectively ensures that the final output target regulatory rule not only carries a complete reasoning path chain but also has high authority and high adaptability, improving the reliability and interpretability of risk identification.

[0109] It is worth noting that this embodiment only describes a preferred implementation. In S410, the server can also perform a directed graph traversal search in the associated subgraph to find all feasible paths with the target node as the starting node and the regulatory rule node as the ending node, and take all feasible paths as at least one inference path. In S412, the server can also select a target inference path from at least one inference path based on the edge weights of the edges in each inference path and the comprehensive weights of the regulatory rule nodes in each inference path, and determine at least one target regulatory rule based on at least one target inference path. This method will be described in detail later and will not be repeated here.

[0110] S414: Perform multiple different modal verifications on the target business event according to at least one target regulatory rule to obtain multimodal verification results.

[0111] Optionally, the server performs multiple different modal verifications on the target business event according to at least one target regulatory rule, such as performing at least two different modal verifications among text semantic verification, numerical logic verification, graph structure verification, and temporal behavior verification, to obtain a multimodal verification result.

[0112] Specifically, text semantic verification is used to determine whether the event description text of the target business event matches the target regulatory rules at the semantic level. In the embodiments of this specification, the server can process the text files of the target business event through a natural language processing model (such as BERT) and output the text semantic verification results. Numerical logic verification is used to verify whether the numerical attributes in the target business event meet the quantitative constraints of the target regulatory rules. In the embodiments of this specification, the server can process the numerical files of the target business event through a DSL (Domain-Specific Language) interpreter and an SQL engine and output the numerical logic verification results. Graph structure verification is used to detect hidden risk paths in the target business event that are not explicitly declared but violate regulatory rules based on the network relationship graph of the target business event. In the embodiments of this specification, the server can process the relationship network and / or entity relationship graph of the target business event through a graph neural network model or a link prediction model and output the graph structure verification results. Temporal behavior verification is used to analyze whether there are regulatory evasion patterns in the time series data of target business events. In the embodiments of this specification, the server can process the time series data of target business events through an LSTM (Long Short-Term Memory) anomaly detection model or a GAN (Generative Adversarial Network) reconstruction error model and output the temporal behavior verification results.

[0113] S416: Perform fusion processing on the multimodal verification results to obtain the risk identification results of the target business event.

[0114] Understandably, all the verification results mentioned above can be verification scores. For example, text semantic verification results can be text semantic verification scores, numerical logic verification results can be numerical logic verification scores, graph structure verification results can be graph structure verification scores, and temporal behavior verification results can be temporal behavior verification scores, etc. The server can ultimately perform weighted fusion processing on verification scores from multiple different modalities to obtain a comprehensive risk score for the target business event. Based on the comprehensive risk score of the target business event and a preset risk threshold, the server obtains the risk identification result for the target business event. For example, if the comprehensive risk score of the target business event is greater than the first preset risk threshold, the server determines the target business event as a high-risk business event and generates a risk warning message. If the comprehensive risk score of the target business event is greater than the second preset risk threshold but less than or equal to the first preset risk threshold, the server determines the target business event as a medium-risk business event and generates compliance recommendation information. If the comprehensive risk score of the target business event is less than or equal to the second preset risk threshold, the server determines the target business event as a low-risk business event and generates compliance recommendation information.

[0115] It is worth noting that the above is merely a preferred embodiment of the embodiments described in this specification. In practical applications, the server can directly determine the target business event as high-risk when the verification score of any modality exceeds the corresponding preset risk threshold; alternatively, weighted voting or other integrated strategies can be used to synthesize the verification results of each modality to obtain the risk identification result of the target business event.

[0116] In this embodiment, the server employs a multimodal verification and result fusion strategy to enable comprehensive risk identification by integrating heterogeneous information from multiple sources, including semantics, numerical values, relationships, and time series, to fully assess the compliance status of target business events. On one hand, this enhances the ability to identify complex and concealed violations, improving the comprehensiveness of risk identification. On the other hand, through a structured scoring and tiered response mechanism, it improves the accuracy, interpretability, and operability of risk identification results, thereby achieving reliable risk identification to effectively address the massive and rapidly changing regulatory rules.

[0117] In one embodiment, such as Figure 5 As shown, another risk identification method is provided. Taking the application of this method to a server as an example, the method includes the following steps: S502: Obtain the event description text of the target business event.

[0118] Specifically, S502 is the same as S202, and will not be repeated here.

[0119] S504: Semantically align the event semantic vector corresponding to the event description text with the regulatory semantic vector corresponding to the regulatory rule text to obtain the target regulatory semantic elements associated with the target business event.

[0120] Specifically, S504 is the same as S204, and will not be repeated here.

[0121] S506: Based on the target regulatory semantic elements and the event context information of the target business events, determine the associated subgraph of the target business events from the dynamic knowledge graph.

[0122] The dynamic knowledge graph comprises multiple nodes and edges connecting them. The nodes include regulatory entity nodes, regulatory condition nodes, and regulatory rule nodes. Each edge represents the semantic relationship between the nodes. Each edge has an edge weight determined based on the strength of the semantic relationship; each regulatory rule node has a comprehensive weight determined based on multiple rule weight factors.

[0123] Specifically, S506 is the same as S406, which will not be repeated here.

[0124] S508: Identify at least one target node in the association subgraph that is associated with the target regulatory semantic element.

[0125] Specifically, S508 is identical to S408, which will not be repeated here.

[0126] S510: Perform a traversal search on the associated subgraph to obtain at least one reasoning path that starts with the target node and ends with the regulatory rule node.

[0127] Optionally, the server performs a directed graph traversal search in the associated subgraph to find all feasible paths starting from the target node and ending at a regulatory rule node, and treats all feasible paths as at least one inference path. That is, each inference path is a directed path that starts from a target node, passes through several intermediate nodes, and finally reaches a regulatory rule node, where the target node and intermediate nodes are not regulatory rule nodes (i.e., regulatory entity nodes or regulatory condition nodes). For example, an inference path can be represented as: [Imported cardiac stent] - [Unit price greater than 6000 yuan] - [Article X of the "Medical Insurance Payment Management Measures"].

[0128] S512: Scoring each inference path is based on the edge weights of the edges in each inference path and the combined weights of the regulatory rule nodes in each inference path.

[0129] Optionally, for each inference path: the server obtains the edge weights of all edges and the combined weights of the regulatory rule nodes, then performs an arithmetic average of the edge weights of all edges to obtain the first path score, and uses the combined weights of the regulatory rule nodes as the second path score. Then, the first path score and the second path score are weighted and summed to obtain the comprehensive score result of the inference path.

[0130] S514: Based on the scoring results of each reasoning path, select at least one target reasoning path from at least one reasoning path.

[0131] Optionally, the server selects K paths with higher scores from at least one inference path as target inference paths; where K is a positive integer greater than 1.

[0132] S516: Based on at least one target reasoning path, determine at least one target regulatory rule.

[0133] Optionally, the server determines the regulatory rules corresponding to the regulatory rule nodes in each target inference path as candidate regulatory rules, and then selects at least one target regulatory rule from at least one candidate regulatory rule. Specifically, the process of the server selecting at least one target regulatory rule from at least one candidate regulatory rule is as follows: The server first obtains the applicable conditions and behavioral constraints of each candidate regulatory rule based on the node description information and node context information of the regulatory rule nodes in each target inference path, and determines whether there are multiple candidate regulatory rules with consistent applicable conditions and behavioral constraints based on the applicable conditions and behavioral constraints of each candidate regulatory rule. If there are multiple candidate regulatory rules with consistent applicable conditions and behavioral constraints, the server uses a large language model to detect whether there are rule conflicts among the multiple candidate regulatory rules with consistent applicable conditions and behavioral constraints. If there are rule conflicts, the server selects the candidate regulatory rule with the highest comprehensive weight among the multiple candidate regulatory rules as the target regulatory rule to achieve conflict resolution; if there are no rule conflicts, the server selects all of the multiple candidate regulatory rules with consistent applicable conditions and behavioral constraints as the target regulatory rule.

[0134] In this embodiment, the server obtains a first path score reflecting the overall semantic coherence and relational reliability of the path by arithmetically averaging the edge weights. The comprehensive weight of the regulatory rule nodes is then used as a second path score. A weighted summation is then used to generate a comprehensive score. Based on the comprehensive scores of each inference path, the inference path with the highest score is selected as the target inference path. This ensures that the path score considers both the quality of the inference link from event features to regulatory rules and the business authority of the regulatory rules themselves. Furthermore, for multiple candidate regulatory rules with consistent applicable conditions and behavioral constraints, the server introduces a large language model for semantic-level conflict detection. If a rule conflict is detected, the rule with higher confidence is automatically selected based on the comprehensive weight of the regulatory rules to resolve the conflict. If there is no conflict, all relevant rules are retained, ensuring that the final output target regulatory rule is logically reachable, semantically reasonable, and free from contradictions. The entire process effectively guarantees the reliability of the target regulatory rule inference, thereby improving the reliability of risk identification.

[0135] S518: Perform multimodal verification of the target business event based on at least one target regulatory rule to obtain the risk identification result of the target business event.

[0136] Specifically, S518 is the same as S208, and will not be repeated here.

[0137] In the aforementioned risk identification method, during the event understanding phase, the server acquires the event description text of the target business event and automatically aligns the event semantic vector corresponding to the event description text with the regulatory semantic vector corresponding to the regulatory rule text. This accurately extracts the target regulatory semantic elements associated with the target business event within the regulatory context, achieving automated dynamic mapping between event descriptions and regulatory rules. This ensures that the extracted regulatory semantic elements simultaneously reflect the specific characteristics of the business scenario and the applicable logic of the regulatory rules, thus providing accurate and reliable semantic input for subsequent reasoning and verification. During the reasoning and verification phase, the server automatically reasones about the acquired target regulatory semantic elements based on a dynamic knowledge graph updated with regulatory dynamics (such as the dynamic updates of regulatory rule texts). In the reasoning process, through quantitative scoring, multi-path optimization, and conflict resolution, it accurately, efficiently, and reasonably identifies the target regulatory rules applicable to the target business event. Based on the target regulatory rules, multimodal verification of the target business event is performed to obtain the risk identification results. This not only effectively suppresses risk misidentification caused by single-modal verification but also ensures the executability and high reliability of the target regulatory rules. By adopting the above-mentioned risk identification methods, the speed and accuracy of risk identification are improved, thereby effectively addressing the massive and rapidly changing regulatory rules.

[0138] To explain in detail the technical solution of the risk control method in the embodiments of this specification, the following will be combined with Figure 6The entire process is described, and it specifically includes the following steps: Phase 1: System Construction Phase 1. The server obtains the regulatory rule text.

[0139] 2. The server performs semantic parsing on the regulatory rule text to obtain multiple regulatory semantic elements; these multiple regulatory semantic elements include regulatory entities, regulatory conditions, and regulatory rules.

[0140] 3. The server generates a regulatory semantic vector for each regulatory semantic element and stores it in the regulatory rule feature library.

[0141] 4. The server constructs an initial knowledge graph based on multiple regulatory semantic elements. This initial knowledge graph includes multiple nodes and edges connecting them; the nodes include regulatory entity nodes, regulatory condition nodes, and regulatory rule nodes; each edge represents the semantic relationship between the nodes.

[0142] 5. The server determines the comprehensive weight of each regulatory rule node based on multiple rule weight factors corresponding to each regulatory rule node; the multiple rule weight factors include rule effectiveness level weight factor, rule application frequency weight factor, rule false alarm rate weight factor, and rule importance weight factor.

[0143] 6. The server associates the comprehensive weights of each regulatory rule node with the initial knowledge graph to form a dynamic knowledge graph, which includes multiple independent and callable subgraphs.

[0144] Phase Two: System Reasoning Phase 1. The server obtains the event description text of the target business event.

[0145] 2. The server semantically aligns the event semantic vector corresponding to the event description text with the regulatory semantic vector corresponding to the regulatory rule text to obtain the target regulatory semantic elements associated with the target business event, thereby mapping internal event terms to regulatory terms. Specifically, the processing is as follows: 2.1 After obtaining the event description text of the target business event, the server performs vectorization encoding on the event description text to obtain the event semantic vector.

[0146] 2.2 Calculate the similarity between the event semantic vector and each regulatory semantic vector pre-stored in the regulatory rule feature library.

[0147] 2.3. Use regulatory semantic vectors with similarity higher than a preset threshold as target regulatory semantic vectors.

[0148] 2.4. The structured regulatory semantic elements corresponding to the target regulatory semantic vector are taken as the target regulatory semantic elements.

[0149] 3. The server loads the dynamic knowledge graph and performs context-aware reasoning. Specifically, it performs the following processing: 3.1 Obtain the event context information of the target business event. Based on the target regulatory semantic elements and the event context information of the target business event, determine the association subgraph of the target business event from the dynamic knowledge graph. Wherein: 3.1.1 The dynamic knowledge graph is updated along with the regulatory dynamics, which include the updates to the regulatory rules text and the regulatory concerns of external institutions. The updates to the regulatory rules text include, but are not limited to, the addition, revision or repeal of the regulatory rules text.

[0150] 3.1.2 The dynamic knowledge graph includes multiple independent and callable subgraphs, each containing multiple nodes and edges connecting them. These nodes include regulatory entity nodes, regulatory condition nodes, and regulatory rule nodes; the edges connecting them represent the semantic relationships between the nodes.

[0151] 3.2 The server determines at least one target node in the associated subgraph that is associated with the target regulatory semantic element. Specifically, it performs the following processing: 3.2.1 The server performs string matching between the node description information of the node to be matched in the associated subgraph and the event description information of the target business event.

[0152] 3.2.2 Perform similarity matching between the node semantic vector of the node to be matched in the associated subgraph and the event semantic vector of the target business event.

[0153] 3.2.3 Perform context matching between the node attribute information of the node to be matched in the associated subgraph and the context information of the target business event.

[0154] 3.2.4 Based on the matching results of string matching, similarity matching, and context matching, at least one target node associated with the target regulatory semantic element is selected from the association subgraph.

[0155] 3.3 The server performs a traversal search on the associated subgraph to obtain at least one reasoning path from the target node to the regulatory rule node.

[0156] 3.4. Based on each inference path, the server determines at least one target regulatory rule. Specifically, it performs the following processing: 3.4.1. Based on the edge weights of the edges in each inference path and the combined weights of the regulatory rule nodes in each inference path, each inference path is scored.

[0157] 3.4.2 Based on the scoring results of each reasoning path, select at least one target reasoning path from at least one reasoning path.

[0158] 3.4.3. Based on at least one target reasoning path, determine at least one target regulatory rule.

[0159] 4. The server performs multimodal verification on the target business event according to at least one target regulatory rule to obtain the risk identification result of the target business event. Specifically, the processing is as follows: 4.1 The server performs multiple different modal verifications on the target business event according to at least one target supervision rule to obtain multimodal verification results; the multiple different modal verifications include at least two of the following: text semantic verification, numerical logic verification, graph structure verification, and temporal behavior verification.

[0160] 4.2 The server performs fusion processing on the multimodal verification results to obtain the risk identification results of the target business event.

[0161] 5. The server determines whether the risk identification result of the target business event indicates that the target business event is a high-risk business event. If yes, proceed to step 6 below; if no, proceed to step 7 below.

[0162] 6. The server generates risk warning information for the target business event.

[0163] 7. The server generates compliance recommendation information for the target business event.

[0164] Phase 3: System Update Phase 1. Based on the verification feedback of the risk identification results, the server adjusts the rule weight factors corresponding to the relevant regulatory rule nodes in the dynamic knowledge graph.

[0165] 2. The server updates the comprehensive weight of the corresponding regulatory rule nodes based on the adjusted rule weight factors, thereby updating the dynamic knowledge graph based on the risk identification results.

[0166] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0167] Based on the above-mentioned risk identification method, such as Figure 7As shown in the embodiments of this specification, a risk identification device 700 for implementing the risk identification method described above is also provided. The risk identification device 700 includes: The acquisition module 701 is used to acquire the event description text of the target business event; The mapping module 702 is used to semantically align the event semantic vector corresponding to the event description text with the regulatory semantic vector corresponding to the regulatory rule text to obtain the target regulatory semantic elements associated with the target business event. The reasoning module 703 is used to reason about the target regulatory semantic elements based on the dynamic knowledge graph to determine at least one target regulatory rule; the dynamic knowledge graph is updated with the updates of the regulatory dynamics, including the updates of the regulatory rule text. The verification module 704 is used to perform multimodal verification of the target business event according to at least one target regulatory rule, and obtain the risk identification result of the target business event.

[0168] In one possible implementation, the dynamic knowledge graph includes multiple nodes and edges connecting the nodes; the multiple nodes include regulatory entity nodes, regulatory condition nodes, and regulatory rule nodes; each edge is used to represent the semantic relationship between the nodes; the reasoning module 703 is specifically used to: determine the association subgraph of the target business event from the dynamic knowledge graph based on the event context information of the target regulatory semantic elements and the target business event; determine at least one target node in the association subgraph that is associated with the target regulatory semantic elements; perform a traversal search on the association subgraph to obtain at least one reasoning path that starts with the target node and ends with the regulatory rule node; and determine at least one target regulatory rule based on each reasoning path.

[0169] In one possible implementation, the reasoning module 703 is specifically used to: perform string matching between the node description information of the node to be matched in the association subgraph and the event description information of the target business event; perform vector similarity matching between the node semantic vector of the node to be matched in the association subgraph and the event semantic vector of the target business event; perform context matching between the node attribute information of the node to be matched in the association subgraph and the event context information of the target business event; and, based on the matching results of string matching, vector similarity matching, and context matching, filter out at least one target node associated with the target regulatory semantic element from the association subgraph.

[0170] In one possible implementation, each edge has an edge weight determined based on the strength of semantic relationships; each regulatory rule node has a comprehensive weight determined based on multiple rule weight factors; the inference module 703 is specifically used to: score each inference path based on the edge weights of the edges in each inference path and the comprehensive weights of the regulatory rule nodes in each inference path; select at least one target inference path from at least one inference path based on the scoring results of each inference path; and determine at least one target regulatory rule based on at least one target inference path.

[0171] In one possible implementation, the verification module 704 is specifically used to: perform multiple different modal verifications on the target business event according to at least one target regulatory rule to obtain multimodal verification results; the multiple different modal verifications include at least two of the following: text semantic verification, numerical logic verification, graph structure verification, and temporal behavior verification; and perform fusion processing on the multimodal verification results to obtain the risk identification results of the target business event.

[0172] In one possible implementation, the risk identification device 700 further includes a construction module for: acquiring regulatory rule text; performing semantic parsing on the regulatory rule text to obtain multiple regulatory semantic elements; the multiple regulatory semantic elements include regulatory entities, regulatory conditions, and regulatory rules; constructing an initial knowledge graph based on the multiple regulatory semantic elements; the initial knowledge graph includes multiple nodes and edges connecting each node; the multiple nodes include regulatory entity nodes, regulatory condition nodes, and regulatory rule nodes; each edge is used to represent the semantic relationship between nodes; determining the comprehensive weight of each regulatory rule node based on multiple rule weight factors corresponding to each regulatory rule node; the multiple rule weight factors include rule effectiveness level weight factor, rule application frequency weight factor, rule false alarm rate weight factor, and rule importance weight factor; and associating the comprehensive weight of each regulatory rule node with the initial knowledge graph to form a dynamic knowledge graph.

[0173] In one possible implementation, the risk identification device 700 further includes an update module, used to: adjust the rule weight factors corresponding to the corresponding regulatory rule nodes in the dynamic knowledge graph based on the risk identification results; and update the comprehensive weight of the corresponding regulatory rule nodes based on the adjusted rule weight factors to form an updated dynamic knowledge graph.

[0174] Each module in the aforementioned risk identification device 700 can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0175] This specification also provides an electronic device, which may be a server, and its internal structure diagram may be as follows: Figure 8 As shown, this electronic device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The application database stores dynamic knowledge graphs, etc. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network. The processor executes computer programs to implement a risk identification method.

[0176] Those skilled in the art will understand that Figure 8 The structures shown are merely block diagrams of a portion of the structure related to the scheme described in this specification, and do not constitute a limitation on the electronic devices to which the scheme described in this specification is applied. Specific electronic devices may include more or fewer components than those shown in the figures, or may combine certain components, or may have different component arrangements.

[0177] In one possible implementation, an electronic device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0178] This specification also provides a computer storage medium storing instructions that, when executed on a computer or processor, cause the computer or processor to perform one or more steps in the above embodiments. If the constituent modules of the above-described electronic device are implemented as software functional units and sold or used as independent products, they can be stored in the aforementioned computer storage medium.

[0179] In one possible implementation, a computer storage medium is provided that stores a computer program, which, when executed by a processor, implements the steps in the above method embodiments.

[0180] This specification also provides a computer program product, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0181] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this specification are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in or transmitted through a computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, Digital Subscriber Line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., Digital Versatile Discs (DVDs)), or semiconductor media (e.g., Solid State Disks (SSDs)).

[0182] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. The aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks. Unless otherwise specified, the technical features of this embodiment and its implementation can be combined arbitrarily.

[0183] The embodiments described above are merely preferred embodiments of this specification and are not intended to limit the scope of this specification. Any modifications and improvements made by those skilled in the art to the technical solutions of this specification without departing from the spirit of this specification should fall within the protection scope defined by the claims.

[0184] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired results. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

Claims

1. A risk identification method, characterized in that, The method includes: Obtain the event description text of the target business event; Semantically align the event semantic vector corresponding to the event description text with the regulatory semantic vector corresponding to the regulatory rule text to obtain the target regulatory semantic elements associated with the target business event. The target regulatory semantic elements are inferred based on a dynamic knowledge graph to determine at least one target regulatory rule; the dynamic knowledge graph is updated as the regulatory dynamics are updated, including the update dynamics of the regulatory rule text; The target business event is validated using multiple modalities based on the at least one target regulatory rule to obtain a multimodal validation result; the multiple modal validations include at least two of the following: text semantic validation, numerical logic validation, graph structure validation, and temporal behavior validation; The multimodal verification results are fused to obtain the risk identification results of the target business event; The method further includes: Obtain the text of the regulatory rules; The regulatory rule text is semantically parsed to obtain multiple regulatory semantic elements; these multiple regulatory semantic elements include regulatory entities, regulatory conditions, and regulatory rules. An initial knowledge graph is constructed based on the aforementioned multiple regulatory semantic elements; the initial knowledge graph includes multiple nodes and edges connecting each node; the multiple nodes include regulatory entity nodes, regulatory condition nodes, and regulatory rule nodes; each edge is used to represent the semantic relationship between nodes; Based on multiple rule weight factors corresponding to each regulatory rule node, the comprehensive weight of each regulatory rule node is determined; the multiple rule weight factors include rule effectiveness level weight factor, rule application frequency weight factor, rule false alarm rate weight factor, and rule importance weight factor. The combined weights of each regulatory rule node are associated with the initial knowledge graph to form the dynamic knowledge graph.

2. The method as described in claim 1, characterized in that, The dynamic knowledge graph includes multiple nodes and edges connecting each node; the multiple nodes include regulatory entity nodes, regulatory condition nodes, and regulatory rule nodes; each edge is used to represent the semantic relationship between nodes; The step of reasoning about the target regulatory semantic elements based on a dynamic knowledge graph to determine at least one target regulatory rule includes: Based on the target regulatory semantic elements and the event context information of the target business event, the associated subgraph of the target business event is determined from the dynamic knowledge graph; Identify at least one target node in the associated subgraph that is associated with the target regulatory semantic element; Perform a traversal search on the associated subgraph to obtain at least one reasoning path that starts with the target node and ends with the regulatory rule node; Based on each of the aforementioned reasoning paths, at least one target regulatory rule is determined.

3. The method as described in claim 2, characterized in that, Determining at least one target node in the associated subgraph that is associated with the target regulatory semantic element includes: The node description information of the node to be matched in the associated subgraph is matched with the event description information of the target business event. Perform vector similarity matching between the node semantic vector of the node to be matched in the associated subgraph and the event semantic vector of the target business event; The node attribute information of the node to be matched in the associated subgraph is matched with the event context information of the target business event. Based on the matching results of the string matching, the matching results of the vector similarity matching, and the matching results of the context matching, at least one target node associated with the target regulatory semantic element is selected from the association subgraph.

4. The method as described in claim 2, characterized in that, Each edge has an edge weight determined based on the strength of its semantic relationship; Each of the aforementioned regulatory rule nodes has a comprehensive weight determined based on multiple rule weight factors; The determination of at least one target regulatory rule based on each of the aforementioned inference paths includes: Each inference path is scored based on the edge weights of the edges in each inference path and the combined weights of the regulatory rule nodes in each inference path. Based on the scoring results of each of the aforementioned reasoning paths, select at least one target reasoning path from at least one reasoning path; Based on the at least one target reasoning path, at least one target regulatory rule is determined.

5. The method as described in claim 1, characterized in that, After performing multimodal verification on the target business event according to the at least one target regulatory rule to obtain the risk identification result of the target business event, the method further includes: If the risk identification result of the target business event indicates that the target business event is a high-risk business event, a risk warning information for the target business event is generated.

6. The method as described in claim 1, characterized in that, After performing multimodal verification on the target business event according to the at least one target regulatory rule to obtain the risk identification result of the target business event, the method further includes: Based on the risk identification results, the rule weight factors corresponding to the relevant regulatory rule nodes in the dynamic knowledge graph are adjusted. The overall weight of the corresponding regulatory rule node is updated based on the adjusted rule weight factor to form the updated dynamic knowledge graph.

7. A risk identification device, characterized in that, The device includes: The acquisition module is used to acquire the event description text of the target business event; The mapping module is used to semantically align the event semantic vector corresponding to the event description text with the regulatory semantic vector corresponding to the regulatory rule text to obtain the target regulatory semantic elements associated with the target business event. The reasoning module is used to reason about the target regulatory semantic elements based on a dynamic knowledge graph to determine at least one target regulatory rule; the dynamic knowledge graph is updated as the regulatory dynamics are updated, and the regulatory dynamics include the update dynamics of the regulatory rule text; The verification module is used to perform multiple different modal verifications on the target business event according to the at least one target regulatory rule to obtain multimodal verification results; the multiple different modal verifications include at least two of the following: text semantic verification, numerical logic verification, graph structure verification, and temporal behavior verification; the multimodal verification results are fused to obtain the risk identification results of the target business event; A construction module is used to acquire regulatory rule text; perform semantic parsing on the regulatory rule text to obtain multiple regulatory semantic elements; the multiple regulatory semantic elements include regulatory entities, regulatory conditions, and regulatory rules; construct an initial knowledge graph based on the multiple regulatory semantic elements; the initial knowledge graph includes multiple nodes and edges connecting each node; the multiple nodes include regulatory entity nodes, regulatory condition nodes, and regulatory rule nodes; each edge is used to represent the semantic relationship between nodes; determine the comprehensive weight of each regulatory rule node based on multiple rule weight factors corresponding to each regulatory rule node; the multiple rule weight factors include rule effectiveness level weight factor, rule application frequency weight factor, rule false alarm rate weight factor, and rule importance weight factor; associate the comprehensive weight of each regulatory rule node with the initial knowledge graph to form the dynamic knowledge graph.

8. An electronic device, characterized in that, include: Processor and memory; The memory stores a computer program, and when the processor executes the computer program, it implements the method steps of any one of claims 1-6.

9. A computer storage medium, characterized in that, The computer storage medium stores a plurality of instructions adapted for loading by a processor and executing the method steps as claimed in any one of claims 1-6.

10. A computer program product, characterized in that, It includes a computer program that, when executed by a processor, implements the steps of the method according to any one of claims 1-6.

Images