Dynamic rule risk judgment method and system based on large language model

Abstract

The invention relates to the technical field of risk analysis, discloses a dynamic rule risk judgment method and system based on a large language model, and is used for solving the technical problems that in the prior art, a risk judgment rule cannot be automatically and dynamically constructed, so that the automation of risk identification and management and control is poor, and the real-time performance is poor. The method comprises the following steps: calling a large language model to extract basic rule information based on obtained real-time supervision information and historical experience information; performing structured expansion on the basic rule information to obtain expanded rule information; taking the basic rule information and the extended rule information as risk rules, and constructing a risk rule knowledge base; when a risk judgment request is received, analyzing and acquiring target information contained in the risk judgment request; calling a large language model to extract key elements contained in the target information, and judging whether the key elements hit risk rules contained in a risk rule knowledge base or not in combination with the risk rule knowledge base; and if yes, determining that the target information has a violation risk, and outputting risk assessment information.

Description

Technical Field

[0001] This invention relates to the field of risk analysis technology, and in particular to a dynamic rule-based risk assessment method, system, electronic device, computer storage medium, and computer program product based on a large language model. Background Technology

[0002] In today's context of increasingly stringent and refined regulatory requirements across industries, achieving efficient and precise risk control amidst a vast and complex system of frequently updated rules has become a key challenge for many sectors. For example, in areas such as import / export trade, financial compliance, data security, healthcare, and production safety, regulatory agencies or competent authorities typically issue a series of highly binding regulations based on multiple objectives, including industry development, public interest, and international standards. These rules can be extensive in content and frequently updated, placing extremely high demands on the compliance responsiveness of relevant institutions and enterprises.

[0003] However, in practice, due to the sheer number of rules, their complex semantic structure, frequent revisions, and high professional barriers, existing methods commonly employ a hybrid approach combining expert rules with artificial intelligence. This involves domain experts extracting judgment rules from policy texts, which are then combined with AI models for information extraction and preliminary identification. For example, named entity recognition technology is used to extract key elements from unstructured text and match them with pre-defined rules to identify risks.

[0004] However, these methods heavily rely on the accumulated experience of experts and the completeness of rule modeling. When facing new fields and new clauses, they often require high manual customization costs and development cycles. Moreover, in cases of complex rules, overlaps or conflicts may occur, leading to a significant increase in the complexity of system maintenance and interpretation. Therefore, there is an urgent need for an intelligent method that can automatically parse rule text, dynamically construct judgment logic, and achieve accurate risk identification and control. This method would be suitable for diverse scenarios with high-frequency rule iteration, numerous entries, and stringent accuracy requirements, providing more agile and reliable technical support for various risk control tasks. Summary of the Invention

[0005] The main objective of this invention is to solve the technical problem that the existing technology cannot automatically and dynamically construct risk assessment rules, resulting in poor automation and real-time performance of risk identification and control.

[0006] The first aspect of this invention provides a dynamic rule-based risk assessment method based on a large language model, comprising: Obtain real-time regulatory information and historical experience information, and call a large language model to extract basic rule information based on the real-time regulatory information and the historical experience information; The basic rule information is structurally expanded to obtain extended rule information; The basic rule information and the extended rule information are used as risk rules to construct a risk rule knowledge base; When a risk assessment request is received, the target information contained in the risk assessment request is parsed and obtained; The key elements contained in the target information are extracted by calling a large language model, and combined with the risk rule knowledge base, the key elements are judged by retrieval enhancement generation technology to determine whether they match the risk rules contained in the risk rule knowledge base. If so, the target information is considered to have a violation risk, and risk assessment information is output based on the hit risk rules.

[0007] Optionally, in a first implementation of the first aspect of the present invention, the real-time regulatory information includes policy announcements, compliance systems, and risk control notices; the historical experience information includes expert experience information and historical case information. The basic rule information extracted by calling the large language model based on the real-time monitoring information and the historical experience information includes: The large language model is invoked to perform structured decomposition, semantic mapping, and standard classification of the policy announcements, compliance systems, and risk control notices, and to identify key information blocks; The large language model is invoked to extract knowledge, model rules, and match context from prior knowledge in expert experience information and historical case information to obtain prior knowledge information. Extract entity objects and logical relationships between them from the key information block and the prior knowledge information, and construct basic rule information based on the entity objects and the logical relationships.

[0008] Optionally, in a second implementation of the first aspect of the present invention, the step of extracting the entity object and the logical relationship between the entity object from the key information block and the prior knowledge information includes: Extract initial entity objects from the key information block and the prior knowledge information; The obtained initial entity objects are identified, semantic disambiguation is performed on initial entity objects with aliases based on context information, and intelligent completion is performed on initial entity objects with unclear referents to obtain optimized entity objects. Based on the optimized entity object, logical relationships are extracted from the key information block and the prior knowledge information.

[0009] Optionally, in a third implementation of the first aspect of the present invention, the step of structurally expanding the basic rule information to obtain expanded rule information includes: The large language model is invoked to mine the derived dimension information of the basic rule information, wherein the derived dimension information contains data of multiple modalities; Based on a large language model combined with a dynamic attention mechanism, the derived dimension information is fused and aligned across modal heterogeneous data to obtain the expanded new rules; The newly added rules obtained from the expansion are subject to compliance and security verification, and the newly added rules that pass the compliance and security verification are used as the expansion rule information.

[0010] Alternatively, in a fourth implementation of the first aspect of the invention, The process of calling a large language model to extract key elements from the target information, and combining this with the risk rule knowledge base, uses retrieval-enhanced generation techniques to determine whether the key elements match risk rules contained in the risk rule knowledge base. Based on pre-built prompt word generation rules, construct recognition prompt words; Based on the identified prompt words, a large language model is invoked to extract the key elements contained in the target information; The key elements are converted into semantic vectors. Based on the retrieval enhancement generation technology, the semantic vectors are used to perform retrieval and matching in the risk rule knowledge base to find whether the risk rule knowledge base contains candidate rules whose similarity meets the preset hit threshold. If so, then determine whether the target information matches the candidate rule based on the candidate rule.

[0011] Optionally, in a fifth implementation of the first aspect of the present invention, after the risk assessment information is output based on the hit-based risk rule, the method further includes: Obtain expert verification feedback on the risk assessment information; Based on the expert verification feedback, the risk rules in the risk rule knowledge base are adjusted in weight, added, or deleted to update the risk rules in the risk rule knowledge base.

[0012] A second aspect of this invention provides a dynamic rule-based risk assessment system based on a large language model, comprising: The rule building module is used to obtain real-time regulatory information and historical experience information, and call the large language model to extract basic rule information based on the real-time regulatory information and the historical experience information; The rule extension module is used to structurally extend the basic rule information to obtain extended rule information; The knowledge base construction module is used to construct a risk rule knowledge base by taking the basic rule information and the extended rule information as risk rules. The risk assessment module is used to parse and obtain the target information contained in the risk assessment request when a risk assessment request is received; to call the large language model to extract the key elements contained in the target information; and to combine the risk rule knowledge base to determine whether the key elements match the risk rules contained in the risk rule knowledge base through retrieval enhancement generation technology; if so, the target information has a violation risk, and risk assessment information is output based on the matched risk rules.

[0013] A third aspect of the present invention provides a dynamic rule risk assessment device based on a large language model, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the dynamic rule risk assessment device based on a large language model to perform the steps of the above-described dynamic rule risk assessment method based on a large language model.

[0014] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the steps of the above-described dynamic rule risk assessment method based on a large language model.

[0015] A fifth aspect of the present invention provides a computer program product, including a computer program / instruction that, when executed by a processor, implements the steps of the dynamic rule risk assessment method based on a large language model as described above.

[0016] The technical solution provided by this invention involves acquiring real-time regulatory information and historical experience information, and then using a large language model to extract basic rule information based on these information. This basic rule information is then structurally expanded to obtain extended rule information. The basic and extended rule information are used as risk rules to construct a risk rule knowledge base. When a risk assessment request is received, the system parses and obtains the target information contained in the request. The large language model is then used to extract key elements from the target information, and combined with the risk rule knowledge base, a retrieval-enhanced generation technique is used to determine whether the key elements match any risk rules contained in the knowledge base. If so, the target information is considered to have a violation risk, and risk assessment information is output based on the matched risk rules. This method can automatically parse rule text and dynamically construct judgment logic, achieving accurate risk identification and intelligent risk management. It is also adaptable to scenarios with high-frequency rule iteration, improving the automation and real-time performance of risk identification and management.

[0017] Furthermore, the system, electronic device, computer-readable storage medium, and computer program product provided by this invention can also solve the corresponding technical problems. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating the first embodiment of the dynamic rule-based risk assessment method based on a large language model in this invention. Figure 2 This is a flowchart illustrating the second embodiment of the dynamic rule-based risk assessment method based on a large language model in this invention. Figure 3 This is a schematic diagram of the model architecture of the second embodiment of the dynamic rule risk assessment method based on a large language model in this invention. Figure 4 This is a schematic diagram of an embodiment of the dynamic rule risk assessment system based on a large language model according to the present invention; Figure 5 This is a schematic diagram of an embodiment of the dynamic rule risk assessment device based on a large language model according to the present invention; Figure 6 This is a schematic diagram illustrating the principle of a computer-readable medium according to an embodiment of the present invention. Detailed Implementation

[0019] Exemplary embodiments of the invention will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limiting the invention to the embodiments set forth herein. Rather, these exemplary embodiments are provided to make the invention more comprehensive and complete, and to facilitate a full communication of the inventive concept to those skilled in the art. The same reference numerals in the drawings denote the same or similar elements, components, or parts, and therefore repeated descriptions of them will be omitted.

[0020] Subject to the technical concept of this invention, the features, structures, characteristics or other details described in a particular embodiment may be combined in one or more other embodiments in a suitable manner.

[0021] In the description of specific embodiments, the features, structures, characteristics, or other details described in this invention are intended to enable those skilled in the art to fully understand the embodiments. However, it is not excluded that those skilled in the art can practice the technical solutions of this invention without one or more of the specific features, structures, characteristics, or other details.

[0022] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0023] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0024] The terms “and / or” or “and / or” include all combinations of any one or more of the listed items.

[0025] Please see Figure 1 The first embodiment of the dynamic rule risk assessment method based on a large language model in this invention includes: S101. Obtain real-time regulatory information and historical experience information, and call the large language model to extract basic rule information based on real-time regulatory information and historical experience information; It is understood that the executing entity of this invention can be a dynamic rule-based risk assessment system based on a large language model, or it can be a terminal or a server; no specific limitation is made here. This embodiment of the invention will be described using a server as the executing entity as an example.

[0026] The real-time regulatory information includes policy announcements, compliance regulations, and risk control notices; the historical experience information includes expert experience information and historical case information; a large language model is used to perform structured decomposition, semantic mapping, and standard classification of the policy announcements, compliance regulations, and risk control notices to identify key information blocks; the large language model is used to extract knowledge, model rules, and match context from the prior knowledge in the expert experience information and historical case information to obtain prior knowledge information; entity objects and logical relationships between entity objects are extracted from the key information blocks and the prior knowledge information, and basic rule information is constructed based on the entity objects and the logical relationships.

[0027] S102. The basic rule information is structurally expanded to obtain extended rule information; The large language model is invoked to mine the derived dimension information of the basic rule information, wherein the derived dimension information contains data from multiple modalities; based on the large language model and combined with a dynamic attention mechanism, the derived dimension information is fused and aligned across modalities to obtain the expanded new rules; the expanded new rules are subjected to compliance and security verification, and the new rules that pass the compliance and security verification are used as the expanded rule information.

[0028] S103. Use basic rule information and extended rule information as risk rules to construct a risk rule knowledge base; The basic rule information and the extended rule information that has passed compliance and security verification are used as the final risk rules to build a risk rule knowledge base. This knowledge base is then combined with a large language model to implement retrieval-enhanced generation technology, which is used to determine whether there is a risk of violation.

[0029] S104. When a risk assessment request is received, parse and obtain the target information contained in the risk assessment request; Once the server is ready to perform specific risk assessments, it can receive and respond to risk assessment requests sent by clients. Upon receiving a risk assessment request, it parses and retrieves the target information contained in the request. This target information refers to the source information for which risk assessment is required, such as product declaration information.

[0030] S105. Call the large language model to extract the key elements contained in the target information, and combine it with the risk rule knowledge base to determine whether the key elements hit the risk rules contained in the risk rule knowledge base through retrieval enhancement generation technology. Based on pre-built prompt word generation rules, identification prompt words are constructed; based on the identification prompt words, a large language model is invoked to extract key elements contained in the target information; the key elements are converted into semantic vectors, and through retrieval enhancement generation technology, a retrieval and matching is performed in the risk rule knowledge base based on the semantic vectors to find whether the risk rule knowledge base contains candidate rules with similarity meeting a preset hit threshold; if so, it is determined whether the target information hits the candidate rule based on the candidate rule.

[0031] S106. If so, the target information has a risk of violation, and risk assessment information is output based on the hit risk rules.

[0032] The large language model uses its logical reasoning ability to determine whether the target information matches the candidate rules. If it matches a risk rule, it outputs the conclusion that there is a risk of violation. At the same time, the large language model evaluates the risk rule information and outputs the risk level and risk avoidance suggestions.

[0033] The method provided in this embodiment of the invention can automatically parse rule text and dynamically construct judgment logic to achieve accurate risk identification and intelligent risk management. It can also adapt to scenarios with high-frequency rule iteration, improving the automation and real-time performance of risk identification and management.

[0034] Please refer to Figure 2-3 The second embodiment of the dynamic rule risk assessment method based on a large language model in this invention includes: It is understood that the executing entity of this invention can be a dynamic rule-based risk assessment system based on a large language model, or it can be a terminal or a server; no specific limitation is made here. This embodiment of the invention will be illustrated using a server as the executing entity in the scenario of commodity information declaration during import and export.

[0035] The specific approach in this embodiment can be implemented based on the ES-Redis-NeuralDB three-dimensional hybrid architecture. It introduces rule graph relationship storage by storing rule text combined with basic inverted index; updates the weight pipeline in real time through real-time weight calculation and rule lineage tracking; and achieves fused graph neural network analysis through vector retrieval and policy network centrality calculation. The specific processing steps are explained below.

[0036] S201. The large language model is used to perform structured decomposition, semantic mapping, and standard classification of policy announcements, compliance systems, and risk control notices to identify key information blocks; In this embodiment, before determining the specific risk situation based on the information of the goods to be declared, the server will pre-build a risk rule knowledge base. For example, it can summarize and categorize relevant experience and knowledge, such as policy announcements issued by the Ministry of Commerce and the General Administration of Customs, as well as non-compliance notification information, to form rule texts that can be used to determine the risk of goods.

[0037] This embodiment utilizes a large language model to perform end-to-end semantic segmentation and tagging of unstructured risk control corpora. The unstructured risk control corpora can include policy announcements, compliance regulations, risk control notices, and due diligence reports, among other things. By integrating rule-guided templates, entity recognition, and text structure analysis with the large language model, key parts are automatically identified, thereby structurally decomposing these unstructured risk control corpora. The content within the key parts obtained from the structural decomposition is then semantically mapped and classified according to standards to obtain key information blocks.

[0038] S202. Call the large language model to extract knowledge, model rules, and match context from the prior knowledge in the expert experience information and historical case information to obtain prior knowledge information. In this embodiment, in addition to obtaining policy announcements, compliance systems, and risk control notifications from the aforementioned steps, prior knowledge information can be further extracted based on expert experience and historical case information. Specifically, through steps such as knowledge extraction, rule modeling, context matching, and continuous optimization, knowledge contained in expert experience and historical case information can be extracted from the expert-verified rules and sample sets, enabling the reusability, interpretability, and evolvability of risk control knowledge.

[0039] S203. Extract entity objects and logical relationships between entity objects from key information blocks and prior knowledge information, and construct basic rule information based on entity objects and logical relationships; After obtaining the key information blocks and prior knowledge information, entity objects and logical relationships are extracted from these contents. During the extraction process, this embodiment also uses entity preprocessing techniques to optimize the entity information contained in the key information blocks and prior knowledge information. Specifically, this includes: extracting initial entity objects from the key information blocks and prior knowledge information; identifying the obtained initial entity objects; performing entity normalization correction and alias aggregation on initial entity objects with aliases based on context information; performing semantic disambiguation; intelligently completing initial entity objects with unclear referents to obtain optimized entity objects; and preprocessing the entity objects using timeline binding and associated context extraction to optimize issues such as multiple spellings of the same entity, semantic ambiguity, unclear referents, and cross-segment distribution in the corpus. After obtaining the preprocessed entity objects, logical relationships can be extracted from the key information blocks and prior knowledge information.

[0040] In one specific implementation, this step employs a semantic role labeling framework for semantic recognition, thereby accurately identifying the "subject-condition-behavior-punishment" quadruple information in key information blocks and prior knowledge information. Simultaneously, this embodiment compares the extracted quadruple information with the existing rule base in real time, using techniques such as edit distance and semantic similarity dual thresholds (edit distance ≤ 3 and cosine similarity ≥ 0.92) to identify conflicting clauses, directly retaining clauses without conflicts. If the basic rule information extracted based on the quadruple information conflicts with existing rules, a large language model is invoked to determine whether it is a rule update. If it is a rule update, the basic rule information is retained and conflicting old clauses are deleted when constructing the rule knowledge base subsequently.

[0041] S204. Call the large language model to mine the derived dimension information of the basic rule information. Based on the large language model and the dynamic attention mechanism, perform cross-modal heterogeneous data fusion and alignment of the derived dimension information to obtain the expanded new rules. After completing rule extraction based on a large language model, this embodiment needs to leverage the model's cross-modal knowledge fusion capabilities to deeply structure and expand the rules in order to build a highly complete policy knowledge base. Through a pre-trained extended model, this embodiment can expand the content based on basic rule information to obtain relevant and reasonable new rules.

[0042] Specifically, knowledge tracing can be performed based on the multilingual corpora used during the training of large language models. This activates the implicit associations learned by the models from trillions of multilingual corpora, and mines derived dimensional information from basic rule information. This derived dimensional information includes data from multiple modalities. For example, based on the names of chemicals, combined with the IUPAC naming system for chemicals, reaction pathway topology, and process technology parameter matrix information, the content of policy rules and other information can be deeply expanded, such as mining alternative names for chemicals, supplementing corresponding chemical equations, and analyzing related process technologies.

[0043] Subsequently, by combining a business knowledge engine built through a dynamic attention mechanism, the fusion and alignment of heterogeneous data, including text, molecular formulas, and process flow diagrams, are achieved. Following a preprocessing scheme for entity objects similar to that in step S203, this step also performs alias disambiguation and intelligent completion operations on each entity object contained in the mined derived dimension information; and maps the technical standard information and related descriptions existing in the derived dimension information to the aforementioned basic rule information, resulting in expanded new rules.

[0044] S205. Perform compliance and security verification on the newly added rules obtained from the extension, and use the newly added rules that pass the compliance and security verification as the extended rule information; To reduce the impact of output illusions in large language models on new rules and prevent the generation of content that does not conform to the facts, compliance and security verification is required after obtaining the expanded new rules. Error information is removed through review to ensure the formation of an accurate, reliable, and usable rule set, providing solid data support for subsequent risk assessment and other work.

[0045] S206. Use basic rule information and extended rule information as risk rules to construct a risk rule knowledge base; The basic rule information and the extended rule information that has passed compliance and security verification are used as the final risk rules to construct a risk rule knowledge base. The risk rule knowledge base can then be combined with a large language model to determine whether there is a risk of violation through retrieval-enhanced generation technology.

[0046] In one specific implementation, please refer to Figure 3The risk rule knowledge base can be the rule example library shown in the figure. After obtaining the rule example library, it can also be converted into vector form based on the text vector conversion method of the Embedding model to obtain the rule example vector library.

[0047] In a preferred embodiment, the method further includes writing the acquired basic rule information and extended rule information into the Redis transaction log in real time, triggering a cross-departmental collaborative verification process, and using the verification differences to fine-tune the dataset.

[0048] In a preferred embodiment, when constructing the risk knowledge base, the obtained basic and extended rule information is further optimized based on a reranking model (such as bge-reranker-base). This reranking model can assess the semantic importance of the included rules by combining context and business objectives. Specifically, this includes: strengthening the weight of rules strongly related to the current risk, identifying and suppressing redundant or contradictory clauses through a heterogeneous conflict detection mechanism, and dynamically placing high-confidence rules at the top. Based on this, the quality of the input large model rule set can be significantly improved, noise interference can be reduced, and high-quality knowledge input can be provided for subsequent evaluation.

[0049] S207. When a risk assessment request is received, the target information contained in the risk assessment request is parsed and obtained. Once the server is ready to perform specific risk assessments, it can receive and respond to risk assessment requests sent by clients. Upon receiving a risk assessment request, it parses and retrieves the target information contained in the risk assessment request.

[0050] To illustrate with a specific example, the target information refers to the source information that requires risk assessment, such as commodity declaration information. This commodity declaration information could be as follows: "1. (Tungsten rod, 0 / 0 / rod shape / tungsten / forged / 99.95% tungsten, 0.05% impurities / 15.2G / no brand)...K. (Tungsten wire, 0 / 2 / wire shape / pure tungsten / forged / 99.95% tungsten, 0.05% impurities / 15.2G / no brand / CCW-91 / 1.00MM)."

[0051] S208. Call the large language model to extract the key elements contained in the target information, and combine it with the risk rule knowledge base to determine whether the key elements match the risk rules contained in the risk rule knowledge base through retrieval enhancement generation technology. Based on pre-built prompt word generation rules, identification prompt words are constructed; based on the identification prompt words, a large language model is invoked to extract key elements contained in the target information; wherein, when the target information is commodity declaration information, the key elements may include information such as commodity name, specifications, and trade flow.

[0052] The key elements are converted into semantic vectors. Based on the retrieval enhancement generation technology, the semantic vectors are used to perform retrieval and matching in the risk rule knowledge base to find whether the risk rule knowledge base contains candidate rules with similarity that meet a preset hit threshold. If so, it is determined whether the target information hits the candidate rule based on the candidate rule.

[0053] To illustrate with a concrete example, the prompt word generation rule can be based on obtaining a pre-configured prompt word template and definable information, whereby the definable information can be task information and rule information. A practical example of the prompt word template is: "You are a risk control assessment assistant. You can achieve the 'task objective' based on the 'task information' and 'rule information'." Here, the "task information" can be: "Based on existing policy announcement rules, determine which rule the user's declared product information matches, and provide the match number and the reason for the match." Furthermore, the output format can also be defined within the "task information."

[0054] In a preferred embodiment, the prompt word engineering can be implemented based on PAC (Prompt As Code) technology, which automatically generates domain-adaptive prompt words through template syntax tree parsing; Redis Graph is used to store the relationship between prompt words and output mapping, which can support version backtracking and impact analysis through methods such as timestamp-based node version chains; during the model testing phase, illegal outputs can be automatically blocked based on dual filtering using regular expressions and semantic rules.

[0055] Please continue reading. Figure 3 To illustrate this with a concrete example, using the specific product declaration information from S207, this product declaration information is converted into a text vector using the Embedding model. Then, rules and examples are matched based on the content of the rule example vector library. The matching results are input into a rerank model for reranking, yielding the rule example result that best matches the product declaration information. Furthermore, in actual matching, an ES-Redis-NeuralDB three-dimensional hybrid architecture can be used to implement retrieval enhancement, further optimizing the retrieval enhancement generation method used in this solution.

[0056] In this step, the result determined by the enhanced generation technology is: "The first commodity declaration information hit the risk rule, the reason for which is 'the commodity is solid tungsten and the tungsten content is greater than 97%, hitting the solid tungsten import and export control rule'; the kth commodity declaration information did not hit the risk rule, the reason for which is 'the commodity is tungsten wire, and tungsten wire is a non-controlled item'." S209. If so, the target information has a risk of violation, and risk assessment information is output based on the hit risk rules.

[0057] The large language model uses its logical reasoning ability to determine whether the target information matches the candidate rules. If it matches a risk rule, it outputs a conclusion that there is a risk of violation. Simultaneously, the large language model evaluates the risk rule information and outputs the risk level and risk avoidance suggestions. For example, after obtaining the risk rules contained in the risk rule knowledge base that match the key elements using retrieval enhancement generation technology in the aforementioned steps, it can generate relevant prompts containing product declaration information and rule example matching results. The large language model then outputs the final evaluation result and structured risk information based on these prompts.

[0058] In a preferred embodiment, the method further includes obtaining expert verification feedback on the risk assessment information; based on the expert verification feedback, adjusting the weights, adding or deleting risk rules in the risk rule knowledge base to update the risk rules in the risk rule knowledge base, thereby realizing automatic feedback of the difference between the model output and the expert secondary verification to the sample library, forming a closed-loop optimization, and continuously improving the accuracy of risk identification and standard adaptability.

[0059] The method provided in this embodiment of the invention can automatically parse rule text and dynamically construct judgment logic. Through the high-performance similarity retrieval capability of the vector library, it provides accurate rule basis for subsequent risk assessment and provides underlying support for the multi-path recall mechanism in the risk control vertical domain, improving the comprehensiveness and timeliness of risk identification. The construction method based on the risk control rule base can further improve the quality of the rule items included in the risk control rule base and reduce noise interference. Simultaneously, the similarity calculation method and judgment method described in this embodiment can accurately quantify the correlation strength between the declared information and the rule clauses, providing a quantitative basis for the risk assessment conclusion based on the similarity score of the matched rules, achieving traceability. The method in this embodiment achieves accurate risk identification and intelligent risk management, adapts to scenarios with high-frequency rule iteration, and improves the automation and real-time performance of risk identification and management.

[0060] The above describes the dynamic rule risk assessment method based on a large language model in the embodiments of the present invention. The following describes the dynamic rule risk assessment system based on a large language model in the embodiments of the present invention. Please refer to [link / reference]. Figure 4 One embodiment of the dynamic rule-based risk assessment system based on a large language model in the proposed embodiments includes: The rule construction module 401 is used to obtain real-time regulatory information and historical experience information, and call the large language model to extract basic rule information based on the real-time regulatory information and the historical experience information; The rule extension module 4O2 is used to structurally extend the basic rule information to obtain extended rule information; Knowledge base construction module 403 is used to construct a risk rule knowledge base by taking the basic rule information and the extended rule information as risk rules. The risk assessment module 404 is used to parse and obtain the target information contained in the risk assessment request when a risk assessment request is received; call the large language model to extract the key elements contained in the target information, and combine the risk rule knowledge base to determine whether the key elements match the risk rules contained in the risk rule knowledge base through retrieval enhancement generation technology; if so, the target information has a violation risk, and output risk assessment information based on the matched risk rules.

[0061] The system provided in this embodiment of the invention can automatically parse rule text and dynamically construct judgment logic to achieve accurate risk identification and intelligent risk management. It can also adapt to scenarios with high-frequency rule iteration, improving the automation and real-time performance of risk identification and management.

[0062] In another embodiment of this application, the real-time regulatory information includes policy announcements, compliance systems, and risk control notices; the historical experience information includes expert experience information and historical case information. The rule construction module 401 is further specifically used for: The large language model is invoked to perform structured decomposition, semantic mapping, and standard classification of the policy announcements, compliance systems, and risk control notices, and to identify key information blocks; The large language model is invoked to extract knowledge, model rules, and match context from prior knowledge in expert experience information and historical case information to obtain prior knowledge information. Extract entity objects and logical relationships between them from the key information block and the prior knowledge information, and construct basic rule information based on the entity objects and the logical relationships.

[0063] In another embodiment of this application, the step of extracting the entity object and the logical relationship between the entity object from the key information block and the prior knowledge information includes: Extract initial entity objects from the key information block and the prior knowledge information; The obtained initial entity objects are identified, semantic disambiguation is performed on initial entity objects with aliases based on context information, and intelligent completion is performed on initial entity objects with unclear referents to obtain optimized entity objects. Based on the optimized entity object, logical relationships are extracted from the key information block and the prior knowledge information.

[0064] In another embodiment of this application, the rule extension module 4O2 is further used for: The large language model is invoked to mine the derived dimension information of the basic rule information, wherein the derived dimension information contains data of multiple modalities; Based on a large language model combined with a dynamic attention mechanism, the derived dimension information is fused and aligned across modal heterogeneous data to obtain the expanded new rules; The newly added rules obtained from the expansion are subject to compliance and security verification, and the newly added rules that pass the compliance and security verification are used as the expansion rule information.

[0065] In another embodiment of this application, the step of calling a large language model to extract key elements contained in the target information, and combining this with the risk rule knowledge base, using retrieval-enhanced generation technology to determine whether the key elements match risk rules contained in the risk rule knowledge base includes: Based on pre-built prompt word generation rules, construct recognition prompt words; Based on the identified prompt words, a large language model is invoked to extract the key elements contained in the target information; The key elements are converted into semantic vectors. Based on the retrieval enhancement generation technology, the semantic vectors are used to perform retrieval and matching in the risk rule knowledge base to find whether the risk rule knowledge base contains candidate rules whose similarity meets the preset hit threshold. If so, then determine whether the target information matches the candidate rule based on the candidate rule.

[0066] In another embodiment of this application, the dynamic rule-based risk assessment system based on a large language model further includes a feedback learning module, which is specifically used for: Obtain expert verification feedback on the risk assessment information; Based on the expert verification feedback, the risk rules in the risk rule knowledge base are adjusted in weight, added, or deleted to update the risk rules in the risk rule knowledge base.

[0067] The system provided in this embodiment of the invention can automatically parse rule text and dynamically construct judgment logic to achieve accurate risk identification and intelligent risk management. It can also adapt to scenarios with high-frequency rule iteration, improving the automation and real-time performance of risk identification and management.

[0068] Based on the same inventive concept, this specification also provides an electronic device for dynamic rule risk assessment based on a large language model. The electronic device for dynamic rule risk assessment based on a large language model in this embodiment of the invention will be described in detail below from the perspective of hardware processing.

[0069] Figure 5 This is a schematic diagram of an electronic device provided as an embodiment of this specification. Refer to the following... Figure 5To describe the electronic device 500 according to this embodiment of the invention. Figure 5 The electronic device 500 shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of the present invention.

[0070] like Figure 5 As shown, the electronic device 500 is presented in the form of a general-purpose computing device. The components of the electronic device 500 may include, but are not limited to: at least one processing unit 510, at least one storage unit 520, a bus 530 connecting different system components (including storage unit 520 and processing unit 510), a display unit 540, etc.

[0071] The storage unit stores program code that can be executed by the processing unit 510, causing the processing unit 510 to perform the steps described in the processing method section of this specification according to various exemplary embodiments of the present invention. For example, the processing unit 510 can perform, for example... Figure 1 and Figure 2 The steps are shown.

[0072] The storage unit 520 may include a readable medium in the form of a volatile storage unit, such as a random access memory unit (RAM) 5201 and / or a cache storage unit 5202, and may further include a read-only memory unit (ROM) 5203.

[0073] The storage unit 520 may also include a program / utility 5204 having a set (at least one) program module 5205, such program module 5205 including but not limited to: an operating system, one or more application programs, other program modules and program data, each or some combination of these examples may include an implementation of a network environment.

[0074] Bus 530 can represent one or more of several types of bus structures, including a memory cell bus or memory cell controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of the various bus structures.

[0075] Electronic device 500 can also communicate with one or more external devices 100 (e.g., keyboard, pointing device, Bluetooth device, etc.), and with one or more devices that enable a user to interact with electronic device 500, and / or with any device that enables electronic device 500 to communicate with one or more other computing devices (e.g., router, modem, etc.). This communication can be performed via input / output (I / O) interface 550. Furthermore, electronic device 500 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 560. Network adapter 560 can communicate with other modules of electronic device 500 via bus 530. It should be understood that, although... Figure 5 As not shown, other hardware and / or software modules may be used in conjunction with the electronic device 500, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0076] Through the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described in this invention can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this invention can be embodied in the form of a software product, which can be stored in a computer-readable storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, or network device, etc.) to execute the method described above according to this invention. When the computer program is executed by a data processing device, it enables the computer-readable medium to implement the method described above, i.e.: as... Figure 1 or Figure 2 The method shown.

[0077] Figure 6 This is a schematic diagram of a computer-readable medium provided for embodiments of this specification.

[0078] accomplish Figure 1 or Figure 2The computer program of the method shown can be stored on one or more computer-readable media. A computer-readable medium can be a readable signal medium or a readable storage medium. A readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0079] The computer-readable storage medium may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The readable storage medium may also be any readable medium other than a readable storage medium, capable of transmitting, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0080] In addition, the present invention also provides a computer program product, including a computer program / instruction that, when executed by a processor, implements the dynamic rule risk assessment method based on a large language model as described in any of the above embodiments.

[0081] Program code for performing the operations of this invention can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can execute entirely on the user's computing device, partially on the user's device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0082] In summary, the present invention can be implemented in hardware, or as software modules running on one or more processors, or a combination thereof. Those skilled in the art will understand that in practice, general-purpose data processing devices such as microprocessors or digital signal processors (DSPs) can be used to implement some or all of the functions of some or all of the components according to the embodiments of the present invention. The present invention can also be implemented as a device or apparatus program (e.g., a computer program and computer program product) for performing part or all of the methods described herein. Such programs implementing the present invention can be stored on a computer-readable medium or can be in the form of one or more signals. Such signals can be downloaded from an Internet website, provided on a carrier signal, or provided in any other form.

[0083] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the present invention is not inherently related to any specific computer, virtual device, or electronic device, and various general-purpose devices can also implement the present invention. The above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0084] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0085] If the technical solution of this application involves personal information, the product using this technical solution has clearly informed the user of the personal information processing rules and obtained the user's voluntary consent before processing the personal information. If the technical solution of this application involves sensitive personal information, the product using this technical solution has obtained the user's separate consent before processing the sensitive personal information, and also meets the requirement of "express consent". For example, at personal information collection devices such as cameras, clear and prominent signs are set up to inform users that they have entered the scope of personal information collection and that personal information will be collected. If an individual voluntarily enters the collection scope, it is deemed that they have agreed to the collection of their personal information; or on the personal information processing device, with clear signs / information informing users of the personal information processing rules, authorization is obtained from the individual through pop-up information or by asking the individual to upload their personal information; wherein, the personal information processing rules may include information such as the personal information processor, the purpose of personal information processing, the processing method, and the types of personal information processed.

[0086] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A dynamic rule-based risk assessment method based on a large language model, characterized in that, include: Obtain real-time regulatory information and historical experience information, and call a large language model to extract basic rule information based on the real-time regulatory information and the historical experience information; The basic rule information is structurally expanded to obtain extended rule information; The basic rule information and the extended rule information are used as risk rules to construct a risk rule knowledge base; When a risk assessment request is received, the target information contained in the risk assessment request is parsed and obtained; The key elements contained in the target information are extracted by calling a large language model, and combined with the risk rule knowledge base, the key elements are judged by retrieval enhancement generation technology to determine whether they match the risk rules contained in the risk rule knowledge base. If so, the target information is considered to have a violation risk, and risk assessment information is output based on the hit risk rules.

2. The dynamic rule-based risk assessment method based on a large language model according to claim 1, characterized in that, The real-time regulatory information includes policy announcements, compliance regulations, and risk control notices; the historical experience information includes expert experience information and historical case information. The basic rule information extracted by calling the large language model based on the real-time monitoring information and the historical experience information includes: The large language model is invoked to perform structured decomposition, semantic mapping, and standard classification of the policy announcements, compliance systems, and risk control notices, and to identify key information blocks; The large language model is invoked to extract knowledge, model rules, and match context from prior knowledge in expert experience information and historical case information to obtain prior knowledge information. Extract entity objects and logical relationships between them from the key information block and the prior knowledge information, and construct basic rule information based on the entity objects and the logical relationships.

3. The dynamic rule-based risk assessment method based on a large language model according to claim 2, characterized in that, The extraction of entity objects and the logical relationships between entity objects from the key information block and the prior knowledge information includes: Extract initial entity objects from the key information block and the prior knowledge information; The obtained initial entity objects are identified, semantic disambiguation is performed on initial entity objects with aliases based on context information, and intelligent completion is performed on initial entity objects with unclear referents to obtain optimized entity objects. Based on the optimized entity object, logical relationships are extracted from the key information block and the prior knowledge information.

4. The dynamic rule-based risk assessment method based on a large language model according to claim 1, characterized in that, The structured expansion of the basic rule information to obtain expanded rule information includes: The large language model is invoked to mine the derived dimension information of the basic rule information, wherein the derived dimension information contains data of multiple modalities; Based on a large language model combined with a dynamic attention mechanism, the derived dimension information is fused and aligned across modal heterogeneous data to obtain the expanded new rules; The newly added rules obtained from the expansion are subject to compliance and security verification, and the newly added rules that pass the compliance and security verification are used as the expansion rule information.

5. The dynamic rule-based risk assessment method based on a large language model according to claim 1, characterized in that, The process of calling a large language model to extract key elements from the target information, and combining this with the risk rule knowledge base, uses retrieval-enhanced generation techniques to determine whether the key elements match risk rules contained in the risk rule knowledge base. Based on pre-built prompt word generation rules, construct recognition prompt words; Based on the identified prompt words, a large language model is invoked to extract the key elements contained in the target information; The key elements are converted into semantic vectors. Based on the retrieval enhancement generation technology, the semantic vectors are used to perform retrieval and matching in the risk rule knowledge base to find whether the risk rule knowledge base contains candidate rules whose similarity meets the preset hit threshold. If so, then determine whether the target information matches the candidate rule based on the candidate rule.

6. The dynamic rule-based risk assessment method based on a large language model according to any one of claims 1-5, characterized in that, After the risk assessment information is output based on the hit-based risk rules, the following is also included: Obtain expert verification feedback on the risk assessment information; Based on the expert verification feedback, the risk rules in the risk rule knowledge base are adjusted in weight, added, or deleted to update the risk rules in the risk rule knowledge base.

7. A dynamic rule-based risk assessment system based on a large language model, characterized in that, include: The rule building module is used to obtain real-time regulatory information and historical experience information, and call the large language model to extract basic rule information based on the real-time regulatory information and the historical experience information; The rule extension module is used to structurally extend the basic rule information to obtain extended rule information; The knowledge base construction module is used to construct a risk rule knowledge base by taking the basic rule information and the extended rule information as risk rules. The risk assessment module is used to parse and obtain the target information contained in the risk assessment request when a risk assessment request is received. The key elements contained in the target information are extracted by calling a large language model, and combined with the risk rule knowledge base, the retrieval enhancement generation technology is used to determine whether the key elements match the risk rules contained in the risk rule knowledge base; if so, the target information has a violation risk, and risk assessment information is output based on the matched risk rules.

8. A dynamic rule-based risk assessment device based on a large language model, characterized in that, The dynamic rule-based risk assessment device based on a large language model includes: a memory and at least one processor, wherein the memory stores instructions; The at least one processor invokes the instructions in the memory to cause the dynamic rule risk assessment device based on a large language model to perform the steps of the dynamic rule risk assessment method based on a large language model as described in any one of claims 1-6.

9. A computer-readable storage medium storing a computer program / instructions thereon, characterized in that, When the program / instruction is executed by the processor, it implements the steps of the dynamic rule risk assessment method based on a large language model as described in any one of claims 1-6.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps of the dynamic rule risk assessment method based on a large language model as described in any one of claims 1-6.

Images