Financial supervision case rule extraction method and system based on large-scale language model

Abstract

The invention discloses a financial supervision case rule extraction method and system based on a large-scale language model, and the method comprises the steps: carrying out the cleaning and tagging of a collected supervision case text, carrying out the deep semantic analysis of a preprocessed text through a large-scale language model, recognizing key elements, and forming a structured semantic representation; clustering the cases subjected to the structured semantic representation so as to classify similar cases into the same theme group, and performing automatic abstraction and induction on the common logic of the cases in the group by combining the context learning capability of a large-scale language model so as to generate executable candidate rules; verifying an execution result of the candidate rule in the case set, selecting a minimum rule set which can completely cover the target case and is logically consistent, and performing confidence evaluation on the minimum rule set to screen a final rule; and carrying out storage and index management, traceability and evolution maintenance on the final rule. In this way, the accuracy, timeliness and interpretability of rule generation can be remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of text extraction technology, specifically relating to a method and system for extracting rules from financial regulatory cases based on a large-scale language model. Background Technology

[0002] There is a technical bottleneck in financial regulation and compliance review scenarios where case knowledge cannot be effectively transformed into enforceable regulatory rules. Currently, regulatory agencies typically guide the identification of suspicious behavior and risk prevention by publishing typical cases. However, these cases only describe specific situations and outcomes, and often do not provide clear rule statements or threshold standards, making it difficult for existing systems to automate the transformation from "case experience" to "enforceable rules".

[0003] Current technologies for transforming cases into rules primarily rely on human experts for analysis and summarization, which has significant limitations. First, they lack a systematic semantic extraction and structuring mechanism. Traditional keyword matching or template recognition methods struggle to capture the implicit logical relationships and judgment elements within case texts, such as implicit patterns like "excessively high transaction frequency" or "frequent fund transfers between accounts without reasonable business explanations," resulting in one-sided and incomplete rule extraction results. Second, existing methods struggle to automatically identify common features from diverse cases with significant variations in expression, lacking the ability to summarize universal rules from a large number of cases and thus failing to formulate directly enforceable regulatory conditions.

[0004] Meanwhile, due to the lack of a unified mechanism for verifying logical consistency and assessing confidence, rules extracted by different analysts often suffer from logical conflicts, incomplete coverage, or redundant definitions, making it difficult to guarantee the reliability and consistency of the rule system. Furthermore, the existing rule extraction process relies heavily on manual labor and has a long rule update cycle, which cannot meet the real-world needs of rapidly changing regulatory requirements and continuously growing case data.

[0005] Existing research on case knowledge utilization mainly focuses on technical approaches based on rule templates, knowledge graphs, or information extraction. For example, some systems construct rule fragments through keyword matching and domain dictionaries, or manually set logical conditions with the help of expert experience to achieve partial automated detection. These methods are effective in structured text or explicitly rule-based environments, but they often struggle to maintain integrity and consistency when faced with long natural language cases, complex semantic logic, and cross-case knowledge fusion.

[0006] In recent years, large-scale language models (LLMs) have made significant progress in semantic understanding and knowledge reasoning, providing new technical means for the automatic abstraction and logical induction of case knowledge. By leveraging their contextual understanding and generation capabilities, core elements and implicit logic in cases can be identified to a certain extent, providing a semantic foundation for further rule extraction.

[0007] Therefore, large-scale language models with semantic understanding and knowledge induction capabilities can be used to perform semantic decomposition and logical extraction on typical regulatory cases, extracting the suspicious behavioral characteristics and judgment criteria contained therein, laying the foundation for subsequent automated rule generation and executable expression. Summary of the Invention

[0008] To address the technical challenges in case rule extraction, such as insufficient semantic understanding of cases, difficulty in summarizing common features, and difficulty in ensuring rule consistency and credibility, this invention aims to provide a method and system for extracting financial regulatory case rules based on a large-scale language model. This method can automatically extract suspicious features from regulatory cases and transform them into executable rules, achieving the structured, parameterized, and automated transformation of case knowledge, and significantly improving the accuracy, timeliness, and interpretability of rule generation.

[0009] To achieve the above-mentioned objectives, an embodiment provides a method for extracting financial regulatory case rules based on a large-scale language model, comprising the following steps: Data preprocessing and semantic understanding: After cleaning and labeling the collected regulatory case texts, a large-scale language model is used to perform deep semantic analysis on the preprocessed texts to identify key elements and logical relationships in the cases and form structured semantic representations. Rule generation: Clustering of cases with structured semantic representation to group similar cases into the same topic group, and combining the context learning capabilities of large-scale language models to automatically abstract and summarize the common logic of cases within the group to generate executable candidate rules; Rule conflict resolution: Verify the execution results of candidate rules in the case set, select the smallest set of rules that can fully cover the target cases and are logically consistent, and evaluate the confidence of the smallest set of rules to filter the final rules; Rule management and output: Store and index the final rules, trace their origins, and maintain their evolution.

[0010] Preferably, the collected regulatory case texts are cleaned and tagged, including: The text content of policies and cases is extracted by OCR recognition and document parsing algorithms, and irrelevant symbols and redundant fields are cleaned up to ensure data consistency and usability. Then, based on the extracted content, the plot descriptions and key conclusions in the cases are identified, and semantic tags indicating whether each case violates the rules are automatically assigned.

[0011] Preferably, a large-scale language model is used to perform deep semantic analysis on the preprocessed text to identify key elements and implicit logical relationships in the case, and to form a structured semantic representation, including: Based on assigning semantic labels to each case to indicate whether it violates regulations, a large-scale language model is invoked to extract key elements from the case text. These key elements include: subject, behavior type, object, time, conditional relationship, and regulatory conclusion. Then, the identified key elements are formed into a structured semantic representation.

[0012] Preferably, clustering cases of structured semantic representation to group similar cases into the same topic group includes: The similarity of key elements is calculated based on the structured semantic representation of the case, and the cases are clustered according to the similarity. Cases with similar characteristics are divided into the same topic group, where the topic includes plot, behavior and violation pattern.

[0013] Preferably, by combining the context learning capabilities of large-scale language models, the common logic of cases within a group is automatically abstracted and summarized to generate executable candidate rules, including: First, design prompts for each topic group; Then, several cases are selected from each topic group as the basis for generating rules for the large-scale language model. Under the guidance of prompt words, logical abstraction and induction are performed on several cases within the group to automatically generate candidate rules for each topic group. These candidate rules describe the violation conditions and violation conclusions.

[0014] Preferably, the execution results of the candidate rules in the case set are verified, and the smallest set of rules that can completely cover the target cases and are logically consistent is selected, including: Each generated candidate rule is applied to all cases within the group, including those not involved in rule generation. The judgment result for each case is then recorded to verify whether the rule accurately identifies violations and whether the generated rule set can cover all violations. Based on the verification results, the candidate rules are combined and merged to form a minimal rule set that can accurately identify all violation cases.

[0015] Preferably, the confidence level of the minimum rule set is evaluated to filter the final rules, including: Confidence scores are calculated from four dimensions: ① Generate confidence scores, i.e., the model outputs probabilities or confidence scores; ② Consistency of source, i.e., the number and similarity of supporting cases; ③ Logical stability, i.e., the hit rate and false alarm rate when executed on sample data; ④ Manual feedback results: i.e., the weighted score confirmed after review; Rules with a comprehensive confidence score higher than the threshold are automatically entered into the database to form the final rules; otherwise, they enter the manual review stage.

[0016] Preferably, the final rules are stored and indexed, including: The generated rules are standardized and uniformly stored, recording key fields, including logical conditions, judgment conclusions, source information, and confidence levels. All final rules are organized through an indexing mechanism, supporting rapid retrieval and clustering based on key fields, feature fields, time intervals, and semantic similarity, thereby achieving centralized management and efficient querying of rule knowledge.

[0017] Preferably, the final rules are traced back to their origins and their evolution is maintained, including: The entire process of generating, updating, merging, and deleting the final rules is recorded to form a version chain and operation log, in order to support the historical tracking and change auditing of rules; When the model is updated or new cases are added, the rules can be automatically regenerated and replaced, and consistency verification can be performed based on the differences between the old and new rules to ensure the stability and continuous evolution of the rule system.

[0018] To achieve the above-mentioned objectives, embodiments of the present invention also provide a financial regulatory case rule extraction system based on a large-scale language model, comprising: The data preprocessing and semantic understanding module is used to clean and label the collected regulatory case texts, and then use a large-scale language model to perform deep semantic analysis on the preprocessed texts to identify key elements and implicit logical relationships in the cases and form a structured semantic representation. The rule generation module is used to cluster cases with structured semantic representations to group similar cases into the same topic group. It also combines the context learning capabilities of large-scale language models to automatically abstract and summarize the common logic of cases within the group and generate executable candidate rules. The rule conflict resolution module is used to verify the execution results of candidate rules in the case set, select the minimum set of rules that can fully cover the target cases and are logically consistent, and perform confidence evaluation on the minimum set of rules to filter the final rules. The rule management and output module is used for storing and indexing the final rules, as well as tracing and maintaining their evolution.

[0019] Compared with the prior art, the beneficial effects of the present invention include at least the following: This invention employs a large-scale language model to automate semantic understanding and element extraction of case texts, and generates common rules based on semantic clustering. Simultaneously, conflict resolution is performed based on the rule execution results, forming a complete and accurate rule set. The entire process reduces manual intervention and reliance on labeled data, improving the efficiency and accuracy of rule generation. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of the financial regulatory case rule extraction method based on a large-scale language model provided in the embodiment; Figure 2 This is a schematic diagram of the structure of the financial regulatory case rule extraction system based on a large-scale language model provided in the embodiment. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of this invention.

[0023] While existing technologies have attempted to analyze cases and support decisions through rule-based reasoning, case matching, or logic networks, they generally suffer from the following shortcomings: 1. The construction of rule and case libraries relies heavily on manual annotation and rule definition, making it difficult to cope with complex and ever-changing real-world scenarios; 2. Existing methods largely depend on manual induction or template matching during rule extraction, making it difficult to automatically abstract common logic from diverse cases with significant differences in expression; 3. Existing technologies lack efficient rule verification and conflict resolution mechanisms. Rules extracted by different analysts may have logical conflicts, repetitions, or incomplete coverage, leading to an unstable rule system that is difficult to apply. To address these technical problems, this invention provides a financial regulatory case rule extraction scheme based on a large-scale language model, which can automatically identify suspicious behavioral characteristics, extract judgment logic, and generate executable and verifiable rules from typical cases.

[0024] like Figure 1 As shown in the embodiment, a method for extracting financial regulatory case rules based on a large-scale language model is provided, including the following steps: S1, Data Preprocessing and Semantic Understanding: After cleaning and labeling the collected regulatory case texts, a large-scale language model is used to perform deep semantic analysis on the preprocessed texts, identify key elements and logical relationships in the cases, and form a structured semantic representation.

[0025] Specifically, step S1 is responsible for organizing, annotating, and semantically analyzing the collected monitoring case texts, transforming the natural language case texts into high-quality structured input data suitable for large-scale model analysis. Utilizing the semantic understanding and logical reasoning capabilities of large-scale language models, the core elements and implicit logical relationships in the cases are identified, laying the semantic foundation for subsequent rule generation. This includes: Case data processing and tagging: First, the system processes regulatory case files in different formats (such as PDF, Word, HTML, scanned documents, etc.), extracting the main text from the case documents through OCR recognition and document parsing algorithms, and cleaning up irrelevant symbols and redundant fields to ensure data consistency and usability. Then, based on the extracted content, the system identifies the plot descriptions and key conclusions in the cases, automatically assigning semantic tags to each case to indicate whether it violates regulations, generating a labeled structured dataset as input for subsequent analysis.

[0026] Semantic element extraction and structured modeling: Based on semantic labeling, the system calls a language model to extract key elements from the case text, automatically identifying key elements such as the subject, type of behavior, object, time, conditional relationships, and conclusion. Finally, the system represents the identified semantic elements in a structured semantic representation, making them easier for computers to understand and store, and providing data support for subsequent feature clustering, cross-case analysis, and rule generation.

[0027] When extracting key elements, an example of a prompt is provided: "You are an expert in extracting key elements from case information. Please extract the following key information from the following case text and output it in a structured format:" Main body: Extract the main people or entities involved in the case (such as companies, individuals, organizations, etc.).

[0028] Behavior type: Extract words that describe the behavior or activities in the case (e.g., fund transfer, account manipulation, illegal transactions, etc.).

[0029] Object: The specific object involved in the extraction action (such as: account, funds, transaction, etc.).

[0030] Time: Extract time information from the case (such as date, time period, etc.).

[0031] Conditional relationships: Extract the logical relationship between behavior and other elements (e.g., if condition X is true, then behavior Y will occur).

[0032] Regulatory conclusions: Extract the regulatory conclusions or judgments of the case (e.g., whether it is a violation, the type of violation, the penalty measures, etc.).

[0033] Please ensure that the extracted information is accurate and complete, and output it in the following format: {"Subject": "Subject Information","Behavior Type": "Behavior Type","Object": "Object","Time": "Time","Conditional Relationship": "Conditional Relationship","Regulatory Conclusion": "Regulatory Conclusion"} Case text: [Case text content].

[0034] In step S1, the proposed case element extraction method based on semantic hierarchical decomposition automatically identifies and extracts key information from cases through a large-scale language model, transforming case features into structured data, reducing manual intervention, and improving the efficiency and accuracy of rule generation.

[0035] S2, Rule Generation: Cluster the cases represented by structured semantics to group similar cases into the same topic group, and combine the context learning capabilities of large-scale language models to automatically abstract and summarize the common logic of cases within the group to generate executable candidate rules.

[0036] Step S2 is primarily responsible for automatically extracting rules from regulatory cases. Through case clustering, cases with similar themes are grouped into different groups, with each group containing cases sharing similar plots and characteristics. Next, using cue engineering and in-context learning (ICL) techniques, the system generates rules for each theme, thus achieving automated rule generation.

[0037] Case Clustering: The system first calculates the similarity of semantic elements in the organized and labeled case data (structured semantic representation), and then clusters the cases based on the similarity, grouping cases with similar characteristics into the same topic group. Cases in each group are highly similar in terms of plot, behavior, and violation patterns. For example, in a financial scenario, cases of "abnormal fund transfers" and cases of "failure to report as required" may be divided into different groups. The clustered topic groups provide a targeted data foundation for subsequent rule generation.

[0038] Hint Engineering and ICL Applications: Within each topic group, the system guides the large-scale language model to generate corresponding rules through design hints. The system selects several cases from each group as the basis for generating rules for the large-scale language model, ensuring that the generated rules accurately reflect the core features of the group, while effectively reducing computational overhead and improving system processing efficiency.

[0039] Specifically, taking abnormal fund transfers as an example, the following warning words can be designed: You are a case rule summarization expert. Please summarize the common characteristics of violations related to fund transfers from the following cases, and derive the following rules: Behavior type: Describes the specific method of fund transfer (e.g., cross-border transfer, large-amount fund transfer, etc.).

[0040] Violation criteria include: frequency of fund transfers, amount thresholds, and relationships between accounts.

[0041] Conclusion of violation: Based on whether the behavior meets the violation criteria, determine whether it constitutes an abnormal transfer of funds and provide a conclusion.

[0042] Regulatory conclusion: Regulatory measures for this behavior.

[0043] Case elements: [Case 1 element] [Case 2 element] [Case 3 element].

[0044] Intra-group rule generation: The system calls a large-scale language model, and prompts generated based on ICL technology are logically abstracted and summarized within the group to automatically generate candidate rules for each topic group. These candidate rules describe specific violation conditions and conclusions based on the characteristics of the group. The generated rules are expressed in natural language. For example, for the topic group "abnormal fund transfers," the generated rules might include: "If the number of fund transfers exceeds 3 times, and the single amount exceeds 500,000 yuan, it is judged as abnormal fund transfer behavior." Here, "the number of fund transfers exceeds 3 times" and "the single amount exceeds 500,000 yuan" are common conditions summarized from cases within the group. These rules can be directly applied to monitoring and decision support systems.

[0045] In step S2, the rule generation method based on case clustering and context learning involves automatically grouping cases with similar semantic features into the same topic group through semantic similarity calculation and clustering algorithms. Combined with the context learning capabilities of a large-scale language model, the common logic of cases within the group is automatically abstracted and summarized to generate corresponding rule expressions, thus achieving the intelligent transformation of case knowledge into a rule system. S3, Rule Conflict Resolution: Verify the execution results of candidate rules in the case set, select the smallest set of rules that can fully cover the target cases and are logically consistent, and evaluate the confidence of the smallest set of rules to filter the final rules.

[0046] Conflict detection and rule merging: After generating initial rules, each candidate rule is applied to all cases within the group, including those not involved in rule generation. Then, the judgment result for each case is recorded to verify whether the rules accurately identify violations and whether the generated rule set covers all violations. Based on the execution results, the system combines and merges the rules to form a minimal rule set capable of accurately identifying all violation cases. In this way, the system ensures that the generated rules are neither redundant nor ineffective in covering all violation scenarios, thereby improving the accuracy and applicability of the rules.

[0047] Rule confidence assessment: To measure the reliability of rules, the system calculates confidence scores from four dimensions: ① Generate confidence scores : Output probability or confidence score directly output by the model; ② Consistency of source :in This represents the number of cases that meet the criteria. This represents the total number of cases within the group. This represents the average similarity between the abstract semantic vectors in the supporting cases and the rules. Indicates weight; ③ Logical stability :in , TP represents the number of samples that violated the rule but were matched, FP represents the number of samples that did not violate the rule but were matched, and FN represents the number of samples that violated the rule but were not matched. Indicates weight; ④ Manual feedback results :in The number of times a rule is deemed "reasonable / approved" by human reviewers. This represents the number of times that a rule is deemed "needs modification / not approved" by human reviewers.

[0048] Ultimately, the overall confidence score was: The weight .

[0049] Rules with a comprehensive confidence level higher than the threshold are automatically added to the final rule set; otherwise, they undergo manual review. This mechanism allows the system to select the final rule set that possesses high credibility and enforceability.

[0050] In step S3, a conflict resolution method based on execution feedback is used, which involves verifying the execution results of the rules in the case set to evaluate their coverage and identification accuracy. Based on the verification results, the smallest set of rules that can fully cover the target cases and have no logical duplication or decision conflicts is selected, ensuring the logical consistency and optimal coverage of the generated rules, and improving the interpretability and practicality of the rule system.

[0051] S4, Rule Management and Output: Store and index the final rules, trace their origins, and maintain their evolution.

[0052] In this embodiment, step S4 is mainly used to perform structured storage, unified management, and traceable output of the verified rules, thereby achieving standardized expression and continuous maintenance of rule knowledge. This ensures that the rules have consistency, parsability, and long-term availability, providing reliable support for subsequent system calls and analysis.

[0053] Structured rule storage and index management: The system standardizes and uniformly stores the generated rules, recording key fields such as logical conditions, judgment conclusions, source information, and confidence levels. All rules are organized through an indexing mechanism, supporting rapid retrieval and clustering based on key fields, feature fields, time intervals, and semantic similarity, achieving centralized management and efficient querying of rule knowledge.

[0054] Rule tracing and evolution maintenance: The system records the entire process of rule generation, updating, merging, and deletion, forming a version chain and operation log to support historical rule tracking and change auditing. When the model is updated or new cases are added, the system can automatically trigger rule regeneration and version replacement, and perform consistency verification based on the differences between the old and new rules to ensure the stability and continuous evolution of the rule system.

[0055] like Figure 2 As shown, the embodiment also provides a financial regulatory case rule extraction system 20 based on a large-scale language model, including: a data preprocessing and semantic understanding module 21, a rule generation module 22, a rule conflict resolution module 23, and a rule management and output module 24, wherein, The data preprocessing and semantic understanding module 21 is used to clean and label the collected regulatory case texts, and then use a large-scale language model to perform deep semantic analysis on the preprocessed texts to identify key elements and implicit logical relationships in the cases and form a structured semantic representation. The rule generation module 22 is used to cluster cases of structured semantic representation to group similar cases into the same topic group, and combined with the context learning ability of large-scale language models, automatically abstracts and summarizes the common logic of cases within the group to generate executable candidate rules. The rule conflict resolution module 23 is used to verify the execution results of candidate rules in the case set, select the smallest set of rules that can completely cover the target cases and are logically consistent, and perform confidence evaluation on the smallest set of rules to filter the final rules. The rule management and output module 24 is used for storing and indexing the final rules, as well as tracing and maintaining their evolution.

[0056] It should be noted that the financial regulatory case rule extraction system based on a large-scale language model provided in the above embodiments should be illustrated using the above-described functional module division as an example when extracting financial regulatory case rules. The functions can be assigned to different functional modules as needed, i.e., the internal structure of the terminal or server can be divided into different functional modules to complete all or part of the functions described above. Furthermore, the financial regulatory case rule extraction system based on a large-scale language model provided in the above embodiments and the financial regulatory case rule extraction method embodiments based on a large-scale language model belong to the same concept. For details of its implementation process, please refer to the financial regulatory case rule extraction method embodiments based on a large-scale language model, which will not be repeated here.

[0057] The specific embodiments described above illustrate the technical solution and beneficial effects of the present invention in detail. It should be understood that the above description is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for extracting financial regulatory case rules based on a large-scale language model, characterized by, Comprising the following steps: Data preprocessing and semantic understanding: After cleaning and labeling the collected supervision case texts, use large-scale language models to perform deep semantic analysis on the preprocessed texts, identify the key elements and logical relationships in the cases, and form structured semantic representations; Rule generation: Cluster the structured semantic representations of the cases to group similar cases into the same topic group, and use the context learning ability of large-scale language models to automatically abstract and induce the common logic of the cases in the group, generating executable candidate rules; Rule conflict resolution: Verify the execution results of the candidate rules in the case set, select the smallest rule set that can completely cover the target cases and is logically consistent, and perform confidence evaluation on the smallest rule set to select the final rules; Rule management and output: Store and index manage, trace and evolution maintain the final rules. 2.The method of claim 1, wherein, Clean and label the collected supervision case texts, including: Extract the main content of the policy and case text through OCR recognition and document parsing algorithm, and clean up irrelevant symbols and redundant fields to ensure data consistency and availability, then identify the plot description and conclusion points in the case according to the extracted content, and automatically assign semantic labels to each case whether it is a violation. 3.The method of claim 1, wherein, Use large-scale language models to perform deep semantic analysis on the preprocessed texts, identify the key elements and implicit logical relationships in the cases, and form structured semantic representations, including: Based on the semantic labels assigned to each case whether it is a violation, call large-scale language models to extract key elements from the case text, including: subject, behavior type, object, time, condition relationship and supervision conclusion, then form structured semantic representations from the identified key elements. 4.The method of claim 1, wherein, Cluster the structured semantic representations of the cases to group similar cases into the same topic group, including: Calculate the similarity of key elements based on the structured semantic representations of the cases, and cluster the cases according to the similarity, divide the cases with similar characteristics into the same topic group, where the topics include plot, behavior and violation mode. 5.The method of claim 1, wherein, Combine the context learning ability of large-scale language models to automatically abstract and induce the common logic of the cases in the group, generate executable candidate rules, including: First, design prompt words for each topic group; Then select several cases from each topic group as the basis for large-scale language models to generate rules, perform logical abstraction and induction on several cases within the group under the guidance of prompt words, automatically generate candidate rules for each topic group, which describe the violation conditions and violation conclusions. 6.The method of claim 1, wherein, Verify the execution results of the candidate rules in the case set, select the smallest rule set that can completely cover the target cases and is logically consistent, including: The generated each candidate rule is applied to all cases in the group, including cases that did not participate in rule generation, and then the determination result of each case is recorded to verify whether the rule accurately identifies the violation behavior and whether the generated rule set can cover all violation behaviors. According to the test execution result, the candidate rules are combined and merged to form a minimum rule set that can accurately identify all violation cases. 7.The method of claim 1, wherein, The minimum rule set is subjected to confidence evaluation to screen the final rules, including: The confidence score is calculated from four dimensions: ① Generation confidence, that is, model output probability or self-confidence score; ② Source consistency, that is, the number of supporting cases and similarity; ③ Logical stability, that is, hit rate, false alarm rate and the like in sample data execution; ④ Artificial feedback result: that is, the weighted score of review and confirmation; Rules with a comprehensive confidence score higher than the threshold value are automatically stored to form the final rules, otherwise they enter the artificial review link. 8.The method of claim 1, wherein, The final rules are stored and indexed, including: The generated rules are standardized and uniformly stored, and key fields are recorded, including logical conditions, determination conclusions, source information and confidence; all final rules are organized through an indexing mechanism to support fast retrieval and clustering based on key fields, element fields, time intervals and semantic similarity, realizing centralized management and efficient query of rule knowledge. 9.The financial regulatory case rule extraction method based on a large-scale language model according to claim 1, wherein, The final rules are traced and evolutionally maintained, including: The generation, update, merging and deletion of the final rules are recorded to form version chains and operation logs to support historical tracking and change audit of the rules; When the model is updated or new cases are input, rule regeneration and version replacement can be automatically triggered, and consistency verification is performed according to the differences between the old and the new, ensuring the stability and continuous evolution of the rule system. 10.A system for extracting financial regulatory case rules based on a large-scale language model, characterized by, including: The data preprocessing and semantic understanding module is used to clean and label the collected supervision case text, and then use a large-scale language model to perform deep semantic analysis on the preprocessed text, identify the key elements and implicit logical relationships in the case, and form a structured semantic representation; The rule generation module is used to cluster the structured semantic representation cases to group similar cases into the same topic group, and use the context learning ability of the large-scale language model to automatically abstract and induce the common logic of the cases in the group to generate executable candidate rules; The rule conflict resolution module is used to verify the execution result of the candidate rules in the case set, select the minimum rule set that can completely cover the target cases and has consistent logic, and perform confidence evaluation on the minimum rule set to screen the final rules; The rule management and output module is used to store and index the final rules, trace and evolutionally maintain them.

Images