Financial regulatory rule dynamic distillation method, apparatus, medium, and device

Abstract

The application relates to the technical field of financial informatization, and discloses a financial supervision rule dynamic distillation method, device, medium and equipment, which comprises the following steps: preprocessing unstructured rule text, and converting the unstructured rule text into structured training samples based on a logic tree annotation system; generating teacher output results based on the structured training samples, and performing distillation training on a student model by using the structured training samples and the teacher output results to obtain a distillation model; performing constraint decoding on the output process of the distillation model to generate candidate structured executable rules, and performing multidimensional effectiveness verification to determine target structured executable rules, which are deployed into a rule engine of a target business system. The above method forms a complete automatic closed loop from rule text acquisition, structured conversion, model distillation to rule generation and deployment, realizes instant analysis of unstructured rule text, structured extraction and automatic deployment, shortens a rule updating cycle, and reduces maintenance cost.

Description

Technical Field

[0001] This invention relates to the field of financial information technology, and in particular to a method, apparatus, storage medium, and computer equipment for the dynamic distillation of financial regulatory rules. Background Technology

[0002] In the field of financial regulation and compliance, existing technical solutions typically rely on rule engines based on fixed thresholds or traditional machine learning models to process and parse regulatory texts. Such systems achieve the mechanized execution of regulatory requirements through preset judgment conditions or statistical features.

[0003] However, such methods have significant drawbacks in practical applications. On the one hand, rule updates are severely delayed, taking a considerable amount of time from the release of regulatory texts to the deployment of system rules, making it difficult to adapt to rapidly changing regulatory environments. On the other hand, the regulatory semantics expressed in natural language are difficult to quantify into executable logic through fixed patterns, resulting in significant semantic understanding biases. Furthermore, system maintenance costs are high, requiring continuous investment of human resources for rule verification and iteration. In addition, existing technologies also consider applying large language models to process regulatory texts. Although large language models have demonstrated powerful capabilities in the field of natural language processing, no effective adaptation solution has yet been developed for financial regulatory scenarios.

[0004] Therefore, a method that can automatically parse the semantics of unstructured regulatory text and transform it into structured executable logic is urgently needed. Summary of the Invention

[0005] In view of this, this application provides a method, apparatus, storage medium and computer equipment for dynamic distillation of financial regulatory rules. The main purpose is to solve the problems of lagging rule updates, semantic understanding deviations and high maintenance costs in the prior art, and the fact that large language models have not yet formed an effective adaptation solution for financial regulatory scenarios.

[0006] According to a first aspect of the present invention, a method for dynamic distillation of financial regulatory rules is provided, comprising: Obtain the unstructured rule text to be processed, and preprocess the unstructured rule text; Based on the logic tree annotation system, the preprocessed unstructured rule text is transformed into structured training samples; Using a pre-trained large language model as the teacher model, a large language model with parameters to be trained is constructed as the student model. The teacher output is generated based on the structured training samples, and the student model is trained by distillation using the structured training samples and the teacher output to obtain a distillation model. The distillation model is used to map the unstructured rule text into structured executable rules. Based on preset output format constraint rules, the output process of the distillation model is constrained and decoded to generate candidate structured executable rules. The candidate structured executable rules are then subjected to multi-dimensional validity verification. After the verification is passed, the target structured executable rule is determined. Deploy the target structured executable rules into the rule engine of the target business system.

[0007] Optionally, the step of obtaining the unstructured rule text to be processed and preprocessing the unstructured rule text includes: obtaining the original unstructured rule text from a preset rule publishing source; performing format cleaning on the unstructured rule text and converting the format-cleaned unstructured rule text into a unified encoding format, wherein the format cleaning includes removing preset format tags and redundant information; and classifying and storing the unstructured rule text in the unified encoding format according to preset rule categories.

[0008] Optionally, the step of transforming preprocessed unstructured rule text into structured training samples based on a logic tree annotation system includes: constructing a logic tree annotation system, wherein the logic tree annotation system uses entity types as nodes, relation types as edges, and conditional expressions as node attributes or edge attributes; semantically annotating the preprocessed unstructured rule text based on the logic tree annotation system to extract rule logic; and converting the rule logic into a structured data format to form structured training samples, wherein the structured training samples include a logic structure field and a basis reference field, the logic structure field being used to characterize the conditional branch structure in the rule logic, and the basis reference field being used to characterize the source information of the unstructured rule text corresponding to the rule logic.

[0009] Optionally, the step of using the structured training samples and the teacher output to distill the student model to obtain a distillation model includes: inputting the structured training samples into the teacher model to obtain teacher output, wherein the teacher output includes structured rule logic, rule field probability distribution, and confidence weights; inputting the structured training samples into the student model to obtain student output, adjusting the parameters of the student model with the objectives of minimizing the supervised loss between the student output and the ground truth labels corresponding to the structured training samples, and minimizing the distillation loss between the student output and the teacher output, to obtain target model parameters; and determining the distillation model based on the target model parameters.

[0010] Optionally, the target structured executable rule includes: a relationship type for characterizing the logical association between entities, a conditional expression for characterizing the rule trigger judgment condition, and a comprehensive confidence level for characterizing the overall reliability of the rule, wherein the comprehensive confidence level is calculated based on the output probability of the distillation model and the multi-dimensional validity verification.

[0011] Optionally, the multi-dimensional validity verification includes: field integrity verification, syntax validity verification, rule conflict detection, and sandbox execution verification; wherein, the rule conflict detection includes: detecting threshold conflicts, action conflicts, and conflicts during the effective period between the candidate structured executable rule and the deployed rule based on the rule's applicable subject, triggering event, data source field, effective device, and execution action; the sandbox execution verification includes: loading the candidate structured executable rule in an isolated rule engine environment, and verifying the rule triggering result using normal samples, violation samples, and boundary samples.

[0012] Optionally, after deploying the target structured executable rule to the rule engine of the target business system, the method further includes: performing an accuracy evaluation on the deployed target structured executable rule to obtain an evaluation result; verifying the target structured executable rule based on the evaluation result, identifying erroneous rules in the target structured executable rule, and correcting the erroneous rules; and using the corrected target structured executable rule as an incremental training sample, feeding it back to the structured training sample for iterative optimization of the distillation model.

[0013] According to a second aspect of the present invention, a dynamic distillation apparatus for financial regulatory rules is provided, comprising: The text preprocessing module is used to obtain the unstructured rule text to be processed and to preprocess the unstructured rule text. The sample construction module is used to transform preprocessed unstructured rule text into structured training samples based on the logic tree annotation system. The distillation training module is used to construct a large language model with parameters to be trained as a student model, using a pre-trained large language model as a teacher model, generating teacher output results based on the structured training samples, and using the structured training samples and the teacher output results to perform distillation training on the student model to obtain a distillation model. The distillation model is used to map the unstructured rule text into structured executable rules. The rule generation module is used to perform constraint decoding on the output process of the distillation model based on preset output format constraint rules, generate candidate structured executable rules, and perform multi-dimensional validity verification on the candidate structured executable rules. After the verification is passed, the target structured executable rule is determined. The rule deployment module is used to deploy the target structured executable rules to the rule engine of the target business system.

[0014] According to a third aspect of the present invention, a storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described dynamic distillation method for financial regulatory rules.

[0015] According to a fourth aspect of the present invention, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described dynamic distillation method for financial regulatory rules.

[0016] This invention provides a method, apparatus, storage medium, and computer equipment for dynamic distillation of financial regulatory rules. By automating the access and standardization of unstructured rule text, data preparation can be completed without manual intervention, reducing the time delay caused by manual collection and processing. Through a pre-defined logic tree structure, semantic information in the rule text expressed in natural language is extracted and expressed in a structured manner, transforming ambiguous natural language descriptions into clear and unified logical forms, ultimately quantifying them precisely into executable logic, avoiding semantic comprehension bias. Through parameter adjustment, a large language model learns and masters the mapping rules from unstructured text to structured rules, enabling the distillation model to output corresponding structured executable rules based on the received unstructured rule text, eliminating the need for manual parsing and encoding of each rule, shortening the rule update cycle, and reducing system maintenance costs. The target structured executable rules are deployed to the rule engine of the target business system, realizing an end-to-end automated process from new rule text input to executable rule deployment, shortening rule update time, and ensuring that the generated rules can be directly recognized and executed by the business system's rule engine without modifying the existing system architecture, further reducing deployment and maintenance complexity. The above method forms a complete automated closed loop from rule text acquisition, structured transformation, model distillation to rule generation and deployment, realizing real-time parsing of unstructured rule text, as well as structured extraction and automated deployment, shortening the rule update cycle and reducing maintenance costs.

[0017] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, specific embodiments of this application are given below. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings: Figure 1 A flowchart illustrating a dynamic distillation method for financial regulatory rules provided by an embodiment of the present invention is shown. Figure 2 A flowchart illustrating another method for dynamic distillation of financial regulatory rules provided by an embodiment of the present invention is shown. Figure 3 This diagram illustrates the structure of a dynamic distillation device for financial regulatory rules provided in an embodiment of the present invention. Figure 4 This invention provides a schematic diagram of another dynamic distillation device for financial regulatory rules, according to an embodiment of the present invention. Figure 5 A schematic diagram of the device structure of a computer device provided in an embodiment of the present invention is shown. Detailed Implementation

[0019] Exemplary embodiments of the present application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present application are shown in the drawings, it should be understood that the present application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this application will be thorough and complete, and will fully convey the scope of the present application to those skilled in the art.

[0020] This application provides a method for the dynamic distillation of financial regulatory rules, such as... Figure 1 As shown, the method includes the following steps: 101. Obtain the unstructured rule text to be processed and preprocess the unstructured rule text.

[0021] Unstructured rule text refers to rule-based descriptive text expressed in natural language that lacks a fixed data model or format specification. Typical characteristics include flexible sentence structure and no predefined field structure, such as policy documents, industry standards, and legal provisions issued by regulatory agencies. Preprocessing refers to a series of automated cleaning and format standardization operations on the original unstructured rule text. The purpose is to remove redundant information in the text that is not related to the semantics of the rules and convert the text into a unified encoding and storage format to provide standardized input for subsequent structured transformation.

[0022] Specifically, the original unstructured rule text is obtained from a pre-defined rule publication source. Taking the financial regulatory scenario as an example, policy documents issued by regulatory agencies are usually presented in PDF or web page format. The documents contain elements such as titles, body text, clause numbers, format marks, and possible charts and annotations. After obtaining the documents through automated means, preprocessing operations are performed, including removing format marks from the documents, deleting redundant content that is not related to the rule logic, and converting the processed text into a standard encoding format. Finally, the text is classified and stored according to the pre-defined rule categories to form standardized text data that can be directly used later.

[0023] In this embodiment, the time delay and operational errors caused by manual collection and organization of rule text are avoided, laying a data foundation for subsequent automated processing. By removing format tags and redundant information, preprocessing operations reduce noise interference in the text and improve the accuracy and efficiency of subsequent semantic parsing. Through unified encoding and classified storage, the consistency and manageability of data are ensured, facilitating batch processing and retrieval of large-scale rule text. This step provides high-quality, standardized input data for the entire dynamic rule transformation process, enabling the transformation from unstructured to structured data.

[0024] 102. Based on the logic tree annotation system, the preprocessed unstructured rule text is transformed into structured training samples.

[0025] Among them, the logic tree annotation system refers to a framework specification for the structured decomposition and annotation of rule texts expressed in natural language. Specifically, it organizes the core elements in the rules in a tree structure, with entity types as nodes, relation types as edges, and conditional expressions as node attributes or edge attributes, in order to fully characterize the logical composition of the rules. Structured training samples refer to sample data with a unified data format formed by extracting and organizing the semantic information in the original unstructured rule text according to the specification of the logic tree annotation system, which can be directly used for parameter adjustment and training of large language models.

[0026] Specifically, a logic tree annotation system is first constructed, defining how to decompose rule text described in natural language into computable logical units. Taking a financial regulatory scenario as an example, suppose the original policy text contains the following rule: "If the issuer's credit rating is lower than AA, then the bonds issued by this issuer shall not be sold to individual investors." Based on this method, the original policy text is semantically annotated according to the logic tree annotation system. First, the entity types are identified, including issuers, bonds, and individual investors. Then, the relationship types between entities are extracted, including issuance and sales. Finally, the conditional expression, namely, the issuer's credit rating is lower than AA, is extracted. The above annotation results are organized into a structured data format, where the conditional expression, credit rating lower than AA, is used as an attribute of the issuer in the node or as an attribute of issuance in the edge. Together with the entities and relationships, they constitute the complete rule logic. The structured representation is finally stored in key-value pair format, forming structured training samples for model training.

[0027] In this embodiment, the fuzzy and scattered natural language rule text is transformed into a clear and unified structured representation, fundamentally solving the technical obstacle that natural language semantics are difficult to accurately quantify into executable logic and eliminating semantic understanding bias. The logic tree annotation system provides standardized operating procedures for rule decomposition and annotation, ensuring consistency between different rules and different annotators, and improving the quality and usability of training samples. The generated structured training samples retain the complete logical information of the original rules, providing high-quality supervision signals for subsequent parameter adjustment of the large language model, thereby improving the rule generation accuracy of the final distillation model.

[0028] 103. Using a pre-trained large language model as the teacher model, construct a large language model with parameters to be trained as the student model. Generate teacher output results based on structured training samples, and use the structured training samples and teacher output results to perform distillation training on the student model to obtain a distillation model. The distillation model is used to map unstructured rule text into structured executable rules.

[0029] Among them, the teacher model refers to a large-scale language model that has been pre-trained on a large-scale general corpus, used to output structured rule logic, rule field probability distribution, and confidence weights based on structured training samples; the student model refers to a language model with a smaller parameter size than the teacher model, used to learn the output patterns of the teacher model in the task of parsing financial regulatory rules; distillation training refers to introducing the rule field probability distribution, logical structure, and confidence weights output by the teacher model as soft labels on the basis of supervised training objectives, so that the student model can learn the rule parsing ability of the teacher model; the distillation model refers to the target model obtained after distillation training, which inherits the semantic understanding ability of the teacher model, and is endowed with the specific function of mapping unstructured rule text into structured executable rules, and the output content can be directly recognized and executed by the rule engine of the business system.

[0030] Specifically, a pre-trained large language model is selected as the teacher model, and a language model with a smaller parameter size is selected as the student model. The structured training samples generated in the previous steps are input into the teacher model to obtain the teacher output, which includes structured rule logic, citation information, and confidence weights. Then, the same unstructured rule text is input into the student model to obtain the student output. After that, the student model parameters are adjusted with the student output being close to the target output in the structured training samples as the first training objective and the student output being close to the teacher output as the second training objective. During the parameter adjustment process, regularization constraints can also be applied to the student model parameters to prevent the model from overfitting to the training samples. After the distillation training process, the resulting distillation model has the ability to automatically convert unstructured rule text of a preset type into structured executable rules.

[0031] In this embodiment, the ability of a large language model to transfer from general semantic understanding to domain-specific rule transformation tasks is realized, enabling the model to automatically complete the rule parsing and logic extraction work that originally required human experts. Joint supervision based on soft labels output by the teacher model and real annotations ensures high accuracy and consistency of the structured executable rules output by the model. The output rules can be directly called by the rule engine of the business system without secondary conversion. The distillation model obtained through distillation training has generalization ability; for new rule texts not appearing in the training samples, the model can still effectively transform them based on the learned mapping rules, thus realizing real-time rule parsing and dynamic updates. This step realizes the automated mapping from rule text input to executable rule output, providing a model foundation for rapid rule deployment and dynamic updates.

[0032] 104. Based on the preset output format constraint rules, the output process of the distillation model is constrained and decoded to generate candidate structured executable rules. The candidate structured executable rules are then validated in multiple dimensions. After the validation is passed, the target structured executable rule is determined.

[0033] Output format constraint rules refer to a pre-defined set of constraints used to standardize the output data format of the distillation model. The purpose is to ensure that the structured rules generated by the model meet the parsing requirements and execution specifications of the target business system. Specifically, output format constraint rules include a preset logic tree schema and a rule engine DSL syntax template. The preset logic tree schema defines the field structure, field types, and logical hierarchy relationships between fields that the structured rules should include, while the rule engine DSL syntax template defines the specific grammatical expression of the rules in the target rule engine. Constraint decoding refers to applying the above format constraint rules during the model's output generation process, restricting the model's output to conforming to the format specifications, rather than freely generating arbitrary text formats. Multi-dimensional validity verification refers to a systematic check of the correctness and reliability of candidate structured executable rules from multiple different dimensions, including verification of structural integrity, grammatical correctness, logical consistency, and execution feasibility.

[0034] Specifically, taking financial regulatory scenarios as an example, the pre-defined logic tree schema requires that the output structured rules must include three core fields: entity type, relation type, and condition expression. Furthermore, the condition expression must be in the form of a triple of comparison object, operator, and threshold. The rule engine DSL syntax template requires rules to be organized using the syntax format "IF condition THEN action." When the distillation model decodes the input unstructured rule text, the constraint decoding mechanism forces the model's output to adhere to the constraints of the aforementioned schema and DSL template. If the model's output lacks a condition expression field, or if the operator in the condition expression uses a symbol not defined in the DSL template, the constraint decoding will automatically correct or restrict such output. The generation process ensures that the output conforms to the preset format specifications. After constraint decoding, candidate structured executable rules are generated. Then, the candidate rules undergo multi-dimensional validity verification. Field integrity verification checks whether the rule contains all the required fields defined in the schema. Syntax validity verification verifies whether the rule conforms to the syntax specifications of the DSL template. Rule conflict detection determines whether there are logical contradictions between the newly generated rule and the rules in the existing rule base. Sandbox execution verification simulates the execution of the rule in an isolated test environment to verify whether it will produce exceptions or errors in actual operation. After all verifications pass, the candidate rule is determined as the target structured executable rule and can be used for subsequent deployment.

[0035] In this embodiment, constraint decoding and multi-dimensional verification steps ensure that the rules generated by the distillation model conform to the format requirements of the target business system, avoiding parsing errors or deployment failures caused by format incompatibility. The multi-dimensional verification mechanism comprehensively verifies candidate rules from four levels: structure, syntax, logic, and execution, significantly reducing the risk of erroneous rules entering the deployment stage and improving the reliability of rule generation. Embedding constraint decoding into the model generation process rather than correcting it afterward reduces the cost of repeated generation and correction due to format errors, improves overall processing efficiency, and ensures the stable operation of the business system.

[0036] 105. Deploy the target structured executable rules to the rule engine of the target business system.

[0037] The target business system refers to the downstream system that needs to apply rules to make business judgments, such as risk control system, transaction monitoring system, compliance review system, etc.; the rule engine refers to the software component in the business system used to store, manage and execute structured rules, which can parse the input rule logic and trigger corresponding judgments or operations based on real-time data.

[0038] Specifically, during implementation, structured executable rules are automatically deployed to the rule engine of the target business system. The rule engine can then call these rules in real time during subsequent business processing to monitor and determine the concentration of holdings for each bond. The entire process, from rule text input to rule deployment completion, requires no manual coding or system configuration adjustments.

[0039] In this embodiment, the structured executable rules output by the distillation model include confidence weights, providing a quantitative basis for subsequent manual verification and risk classification. Rules with low confidence can be prioritized for manual review, balancing automation efficiency and output accuracy. The generated target structured executable rules can be directly deployed to the rule engine of the existing business system without any modification to the business system architecture, reducing the complexity of technical integration and implementation costs. Finally, whenever a new rule text is published, it can be instantly converted and deployed through the same distillation model, forming an automated closed loop from rule acquisition to rule effectiveness.

[0040] This invention provides a dynamic distillation method for financial regulatory rules. It automates the access and standardization of unstructured rule text, completing data preparation without manual intervention and reducing time delays caused by manual collection and processing. Through a pre-defined logic tree structure, it extracts and expresses semantic information from the natural language rule text in a structured manner, transforming ambiguous natural language descriptions into clear and unified logical forms, ultimately quantifying them into executable logic and avoiding semantic comprehension bias. By adjusting parameters, a large language model learns and masters the mapping rules between unstructured text and structured rules, enabling the distillation model to output corresponding structured executable rules based on the received unstructured rule text, eliminating the need for manual parsing and encoding of each rule, shortening the rule update cycle, and reducing system maintenance costs. Deploying the target structured executable rules to the rule engine of the target business system achieves an end-to-end automated process from new rule text input to executable rule deployment, shortening rule update time and ensuring that the generated rules can be directly recognized and executed by the business system's rule engine without modifying the existing system architecture, further reducing deployment and maintenance complexity. The above method forms a complete automated closed loop from rule text acquisition, structured transformation, model distillation to rule generation and deployment, realizing real-time parsing of unstructured rule text, as well as structured extraction and automated deployment, shortening the rule update cycle and reducing maintenance costs.

[0041] This application provides another method for the dynamic distillation of financial regulatory rules, such as... Figure 2 As shown, it specifically includes: Step 201: Obtain the unstructured rule text to be processed and perform preprocessing.

[0042] Specifically, the original unstructured rule text is obtained from the preset rule publishing source; the unstructured rule text is cleaned and converted into a unified encoding format, wherein the format cleaning includes removing preset format tags and redundant information; and the unstructured rule text in the unified encoding format is classified and stored based on the preset rule categories.

[0043] In this implementation, the rule release sources first include the official websites of financial regulatory agencies, such as the policy document release pages publicly disclosed by various regulatory agencies. The system monitors the policy document updates on the website periodically or in real time through automated programs. When a new bond market-related policy document is detected, it automatically downloads and obtains the original content of the document. The obtained policy document is usually presented in PDF, WORD, or HTML web page format, and the content includes various elements such as title, body clauses, format marks, tables, charts, issuing agency signature, issuance date, and copy units, which are typical unstructured rule texts.

[0044] Then, the system removes formatting marks and redundant information from the text that are irrelevant to the semantics of the rules. Formatting marks include typesetting information such as font style, font size, line spacing, paragraph indentation, and page breaks. Redundant information includes the issuing authority's name, issuance date, copy recipients, layout design elements, headers and footers, watermarks, and non-standard annotations that are not directly related to the rule logic. After cleaning, the system converts the remaining unstructured rule text into a unified encoding format. In this embodiment, the UTF-8 encoding standard is used to ensure the compatibility and consistency of the text under different operating systems and software environments.

[0045] Finally, after the encoding conversion is completed, the system classifies and stores the unstructured rule text in a unified encoding format based on preset rule categories. The rule categories are divided according to the financial business area to which the policy belongs, such as bond issuance management, bond trading supervision, information disclosure requirements, investor suitability management, etc. Taking the bond market as an example, if a policy document mainly stipulates the access conditions and approval process for bond issuance, it will be stored in the bond issuance management category. If the policy document involves the monitoring and handling of abnormal transactions in bond trading, it will be stored in the bond trading supervision category. The classified and stored text data is organized into a structured file directory or database record. Each record contains metadata information such as text content, category, acquisition time, and original source link, providing standardized and indexable input data for subsequent structured conversion steps.

[0046] Step 202: Based on the logic tree annotation system, the preprocessed unstructured rule text is transformed into structured training samples.

[0047] This involves constructing a logic tree annotation system, where entity types are nodes, relationship types are edges, and conditional expressions are node or edge attributes. Based on this system, semantic annotation is performed on the preprocessed unstructured rule text to extract rule logic. The rule logic is then converted into a structured data format to form structured training samples.

[0048] Furthermore, the structured training samples include: logical structure fields and basis reference fields; wherein, the logical structure fields are used to represent the conditional branch structure in the rule logic, and the basis reference fields are used to represent the source information of the unstructured rule text corresponding to the rule logic.

[0049] In this implementation, the logic tree annotation system uses entity types as nodes, relationship types as edges, and condition expressions as node attributes or edge attributes. Taking the bond market regulatory scenario as an example, entity types include the subjects or objects involved in the rules, such as issuers, bonds, investors, underwriters, and rating agencies. Relationship types are used to represent the logical connections between entities, including issuance (the connection between issuers and bonds), holding (the connection between investors and bonds), transaction (the connection between buyers and sellers and bonds), and rating (the connection between rating agencies and bonds). Condition expressions are used to represent the judgment conditions on which the rule is triggered, such as a credit rating lower than A, a holding ratio exceeding 3%, or an issuance time after 2025. Condition expressions can be attached to entity nodes as node attributes, such as imposing a rating of not lower than X on issuers, or they can be attached to relationship edges as edge attributes, such as imposing a holding period of more than 30 days on holding relationships. The above three elements together constitute a complete rule logic unit, organized in a tree or graph structure.

[0050] Based on the logic tree annotation system, a team of financial experts performs semantic annotation on the preprocessed unstructured rule text, extracting the underlying rule logic. The annotation team then deconstructs and annotates the clauses according to the logic tree annotation system. First, they identify the entity types in the text, then extract the relationship types between entities, and finally extract the conditional expressions, attaching them as node attributes. Through the annotation process, the original scattered and ambiguous natural language expression is transformed into a structured logical representation with clear nodes, edges, and attributes. After completing the semantic annotation, the extracted rule logic is converted into a structured data format to form structured training samples. In this embodiment, the annotation results are stored in JSON format. Each structured training sample contains two core parts: a logical structure field and a basis reference field. The logical structure field is used to represent the conditional branch structure in the rule logic, specifically expressed in the form of IF-THEN rules. The basis reference field is used to represent the source information of the original unstructured rule text corresponding to the rule logic, including the name of the original policy document, the issuing agency, the clause number, the publication date, and the original text excerpt. All annotated structured training samples are aggregated into a training sample set for subsequent parameter adjustment steps of the large language model.

[0051] By using logical tree annotation and structured sample construction, fuzzy and scattered natural language rules are transformed into clear and unified logical representations, fundamentally overcoming the problem of semantic understanding bias. The combination of logical structure fields in IF-THEN form and reference fields ensures both the computability of the rule logic and the traceability of the original source. The standardized storage in JSON format facilitates efficient parsing and processing by computer programs, providing high-quality input data for subsequent model training.

[0052] Step 203: Using a knowledge distillation architecture, a distillation model is obtained to map unstructured rule text into structured executable rules.

[0053] The process involves inputting structured training samples into a teacher model to obtain teacher output, which includes structured rule logic, rule field probability distribution, and confidence weights. Then, structured training samples are input into a student model to obtain student output. The parameters of the student model are adjusted to minimize the supervision loss between the student output and the corresponding ground truth labels of the structured training samples, and to minimize the distillation loss between the student output and the teacher output, thus obtaining target model parameters. Finally, a distillation model is determined based on these target model parameters.

[0054] In this embodiment, structured training samples are input into the teacher model to obtain teacher output results, which include structured rule logic, rule field probability distribution, and confidence weights. At the same time, structured training samples are input into the student model to obtain student output results. Each training sample contains an input part and a labeled part, namely unstructured rule text and the corresponding structured executable rule.

[0055] Subsequently, the parameters of the student model are adjusted with the goal of minimizing the supervised loss between the student's output and the real labels corresponding to the structured training samples, and minimizing the distillation loss between the student's output and the teacher's output. To achieve efficient and stable parameter adjustment, this embodiment uses LoRA technology for efficient parameter fine-tuning. LoRA technology injects low-rank decomposition matrices into some layers of the model while keeping the original parameters of the pre-trained model unchanged. The model adapts to specific tasks by training the injected matrix parameters, thereby significantly reducing the number of parameters that need to be trained. The optimization objective can be summarized as follows: to make the structured rules output by the student model as close as possible to the standard rules labeled in the training samples, while making the student model imitate the behavior output of the teacher model as much as possible, and to control the complexity of the model through regularization constraints to prevent overfitting. Specifically, the training process aims to minimize the difference between the student model's prediction and the real labels, and to minimize the difference between the student model's output and the teacher model's output, while applying regularization constraints to the model parameters to balance the model's fitting ability and generalization ability.

[0056] Regarding hyperparameter settings, in this embodiment, the rank parameter for low-rank adaptive fine-tuning is set to r=8, and the learning rate is set to 10. 4The regularization coefficient is set to λ=0.01, where the rank parameter r controls the dimension of the injection matrix, affecting the model's expressive power and the number of training parameters; the learning rate controls the step size of parameter updates, affecting the convergence speed and stability of training; the regularization coefficient controls the strength of the regularization constraint, affecting the model's generalization performance; the optimized target model parameters are obtained through parameter adjustment, and the distillation model is determined based on the target model parameters. The distillation model inherits the semantic understanding ability of the large language model and is endowed with the specific function of mapping unstructured rule text into structured executable rules.

[0057] Step 204: Based on the preset output format constraint rules, generate candidate structured executable rules and perform multi-dimensional validity checks to determine the target structured executable rule.

[0058] The multi-dimensional validity verification includes: field integrity verification, syntax validity verification, rule conflict detection, and sandbox execution verification. Rule conflict detection includes: detecting threshold conflicts, action conflicts, and conflicts during the effective period between candidate structured executable rules and deployed rules based on the rule's applicable subject, triggering event, data source field, effective device, and execution action. Sandbox execution verification includes: loading candidate structured executable rules in an isolated rule engine environment and verifying the rule triggering results using normal samples, violation samples, and boundary samples.

[0059] In this implementation, rule conflict detection is introduced to prevent newly generated rules from logically contradicting existing rules deployed in the rule engine. Specifically, rule conflict detection compares and analyzes multiple core dimensions of rules, including the applicable subject, triggering event, data source field, effective device, and execution action. Based on these dimensions, conflict detection focuses on identifying the following three types of conflicts: First, threshold conflict, where candidate rules and deployed rules set different thresholds for the same judgment condition. For example, a candidate rule stipulates that an alarm is triggered when the concentration exceeds 10%, while a deployed rule stipulates that an alarm is triggered when the concentration exceeds 8%, resulting in inconsistent triggering results under the same business scenario. Second, action conflict, where the execution actions triggered by candidate rules and deployed rules under the same conditions contradict each other. For example, one rule requires allowing a transaction, while another requires rejecting a transaction. Third, effective period conflict, where the effective time range of candidate rules and deployed rules overlaps but the logical requirements are inconsistent. For example, two rules propose mutually exclusive compliance requirements for the same business scenario within the same time period. Through the above conflict detection, abnormal system behavior or incorrect business judgments caused by rule conflicts can be effectively avoided.

[0060] In its implementation, after completing field integrity verification, syntax validity verification, and rule conflict detection, this method further performs sandbox execution verification on candidate structured executable rules. Sandbox execution verification involves loading candidate rules in an isolated rule engine test environment independent of the production environment and verifying whether the actual execution behavior of the rules meets expectations by simulating input data. Specifically, this method constructs three types of test samples to verify candidate rules: first, normal samples, i.e., compliant business data that does not trigger rule violations, used to verify that the rules will not generate false alarms; second, violation samples, i.e., abnormal business data that is certain to trigger rule violations, used to verify that the rules can correctly identify violations and trigger preset actions; and third, boundary samples, i.e., edge business data that is exactly at the rule threshold, used to verify whether the rule's judgment logic under critical conditions is accurate. Through sandbox execution verification, potential logical defects or execution anomalies in candidate rules during actual operation can be discovered in advance in an isolated environment, avoiding the business risks caused by directly deploying problematic rules to the production environment.

[0061] Step 205: Deploy the target structured executable rules to the rule engine of the target business system.

[0062] The target structured executable rules include: relationship types used to characterize the logical association between entities, conditional expressions used to characterize the rule triggering judgment conditions, and comprehensive confidence scores used to characterize the overall reliability of the rules. The comprehensive confidence scores are calculated by fusing the output probability of the distillation model with multi-dimensional validity verification.

[0063] In this implementation, the relation type refers to the semantic label used in the structured executable rule to characterize the logical relationship between entities, defining the action or state relationship between the subject and object involved in the rule; the condition expression refers to the judgment condition used in the structured executable rule to characterize the triggering of the rule, usually including three parts: comparison object, comparison operator, and threshold, such as credit rating below AA, concentration greater than 10%, holding period exceeding 30 days, etc.; the confidence weight refers to the quantitative score of the certainty of the rule parsing result when the distillation model generates structured executable rules, usually represented by a value between 0 and 1, with higher values ​​indicating higher certainty. The more confident the model is in the correctness of the rule, the better. Similarly, taking the financial regulatory scenario as an example, suppose the newly released regulatory policy stipulates that the investment ratio of a fund in a single bond shall not exceed 10% of the fund's net assets. After processing the text, the distillation model outputs a relationship type of "holding", which is used to characterize the holding association between the fund and the bond and the comparison relationship that the association must meet. The condition expression is "investment ratio exceeds 10%", which is used to characterize the specific judgment condition for triggering the rule. Here, the investment ratio is the comparison object, "exceeds" is the comparison operator, 10% is the threshold, and the confidence weight is 0.96, indicating that the distillation model has a high degree of certainty in the rule parsing result.

[0064] The above three elements together constitute a complete structured and executable rule. When the rule is deployed to the rule engine of the target business system, the system continuously calculates the fund's actual investment ratio in each bond while processing transaction data in real time, and determines whether the ratio exceeds 10%. If the actual ratio exceeds 10%, preset compliance actions are triggered, such as rejecting the transaction, generating an alarm, and restricting subscriptions. In addition, the confidence weight plays an important auxiliary role in the actual application of the rule. For rules with high confidence, the system can directly use them without manual intervention. For rules with medium or low confidence, the system can include them in the priority manual review queue, and experts will conduct secondary confirmation before deciding whether to deploy them. For rules with extremely low confidence, the system can automatically mark them as questionable rules, triggering manual re-labeling and model iteration processes. The hierarchical processing mechanism based on confidence weight in this application takes into account both automation efficiency and rule accuracy.

[0065] This application designs three key elements for structured executable rules. Firstly, the combination of relation types and conditional expressions fully characterizes the entity relationships and triggering conditions within the rule, enabling accurate parsing and execution by the business system's rule engine. Secondly, the introduction of comprehensive confidence provides a deterministic measure of the distillation model's output, integrating the model's output probability with multi-dimensional verification results. This not only considers the distillation model's original confidence in rule parsing but also integrates the pass / fail status and anomaly feedback from field integrity verification, syntax validity verification, rule conflict detection, and sandbox execution verification. This makes the confidence assessment more comprehensive and reliable, allowing the system to adopt differentiated processing strategies based on the comprehensive confidence level, effectively controlling the risks associated with model output errors while maintaining high efficiency. The standardized output format based on these three elements has good scalability; regulatory rules from different sources and of different types can all be uniformly represented in the same format, facilitating unified management and retrieval within a large-scale rule base. The comprehensive confidence level provides a more reliable quantitative basis for subsequent feedback iteration mechanisms, continuously improving the rule generation accuracy of the distillation model.

[0066] Step 206: Evaluate the accuracy of the target structured executable rules to iteratively optimize the distillation model.

[0067] The process involves: evaluating the accuracy of deployed target structured executable rules to obtain evaluation results; verifying the target structured executable rules based on the evaluation results, identifying erroneous rules in the target structured executable rules, and correcting the erroneous rules; and using the corrected target structured executable rules as incremental training samples, feeding them back into the structured training samples for iterative optimization of the distillation model.

[0068] In this embodiment, after deploying the target structured executable rules to the rule engine of the target business system, this application further introduces a system performance monitoring and feedback iteration mechanism to ensure the continuous accuracy of the rules and the dynamic evolution capability of the model.

[0069] Specifically, this application first conducts an accuracy assessment of the deployed structured executable rules to obtain the assessment results. The assessment process can be implemented in various ways. For example, the system can periodically sample a certain proportion of the deployed rules and compare their execution results in actual business scenarios with the expected results. Alternatively, when a business system triggers a judgment action of a rule, the system can automatically record the context information and judgment result of this trigger for subsequent analysis and evaluation. Or, a rule accuracy assessment index system can be established to quantitatively evaluate the rule quality from multiple dimensions such as rule hit rate, false positive rate, and false negative rate.

[0070] Based on the evaluation results, this application validates the target structured executable rules, identifies erroneous rules, and corrects them. For example, suppose the distillation model generates and deploys a rule: if the concentration of a single bond exceeds 10% of the fund's net assets, an alarm is triggered. However, in actual operation, accuracy assessments revealed that this rule produced false alarms in certain situations. For instance, when the fund's net assets temporarily decreased due to market fluctuations, previously within-limit holdings were mistakenly judged as exceeding limits. Through manual verification, experts discovered that the conditional expression lacked a constraint on the net asset calculation benchmark. Therefore, the rule was corrected, and the conditional expression was revised. The formula was adjusted to have a concentration exceeding 10% of the fund's net assets on the previous trading day, thus eliminating the false alarm problem. After the correction was completed, the corrected target structured executable rules were used as incremental training samples and fed back into the structured training samples for iterative optimization of the distillation model. The corrected rules were added to the original training sample set to form an updated training sample set. Subsequently, the updated training sample set was used to perform a new round of parameter adjustments on the distillation model, enabling the model to learn the correct rule mapping rules from this correction, so that it can output more accurate structured executable rules when encountering similar rule texts in the future.

[0071] Based on the above mechanism, this application constructs a complete closed-loop feedback system, namely, the deployed rules are evaluated in actual operation, the problematic rules found in the evaluation are verified and corrected, the corrected rules are fed back to the training set, the update of the training set drives the iterative optimization of the model, the optimized model generates more accurate rules, thereby reducing the error rate in subsequent rule deployments, and as the number of iterations increases, the rule generation accuracy of the distillation model continues to improve, and the error rate gradually converges.

[0072] Furthermore, as Figure 1In terms of specific implementation, this application provides a dynamic distillation device for financial regulatory rules, such as... Figure 3 As shown, the device includes: a text preprocessing module 301, a sample construction module 302, a distillation training module 303, a rule generation module 304, and a rule deployment module 305.

[0073] The text preprocessing module 301 is used to obtain the unstructured rule text to be processed and to preprocess the unstructured rule text. The sample construction module 302 is used to convert preprocessed unstructured rule text into structured training samples based on the logic tree annotation system. The distillation training module 303 is used to construct a large language model with parameters to be trained as a student model, using a pre-trained large language model as a teacher model, generating teacher output results based on structured training samples, and using the structured training samples and teacher output results to perform distillation training on the student model to obtain a distillation model. The distillation model is used to map unstructured rule text into structured executable rules. The rule generation module 304 is used to perform constraint decoding on the output process of the distillation model based on preset output format constraint rules, generate candidate structured executable rules, and perform multi-dimensional validity verification on the candidate structured executable rules. After the verification is passed, the target structured executable rule is determined. The rule deployment module 305 is used to deploy the target structured executable rules to the rule engine of the target business system.

[0074] In specific application scenarios, the text preprocessing module 301 is specifically used to obtain the original unstructured rule text from the preset rule publishing source; perform format cleaning on the unstructured rule text, and convert the format-cleaned unstructured rule text into a unified encoding format, wherein the format cleaning includes removing preset format tags and redundant information; and classify and store the unstructured rule text in the unified encoding format according to the preset rule categories.

[0075] In specific application scenarios, the sample construction module 302 is specifically used to construct a logic tree annotation system, in which entity types are nodes, relation types are edges, and conditional expressions are node attributes or edge attributes. Based on the logic tree annotation system, semantic annotation is performed on the preprocessed unstructured rule text to extract the rule logic. The rule logic is converted into a structured data format to form structured training samples, which include: logical structure fields and basis reference fields. The logical structure fields are used to represent the conditional branch structure in the rule logic, and the basis reference fields are used to represent the source information of the unstructured rule text corresponding to the rule logic.

[0076] In specific application scenarios, the distillation training module 303 is used to input structured training samples into the teacher model to obtain teacher output results, which include structured rule logic, rule field probability distribution, and confidence weights; input structured training samples into the student model to obtain student output results; and adjust the parameters of the student model to obtain target model parameters with the goal of minimizing the supervision loss between the student output results and the corresponding real labels of the structured training samples, and minimizing the distillation loss between the student output results and the teacher output results; and determine the distillation model based on the target model parameters.

[0077] In specific application scenarios, the target structured executable rules in the rule generation module 304 include: relationship types used to characterize the logical association between entities, conditional expressions used to characterize the rule trigger judgment conditions, and comprehensive confidence scores used to characterize the overall reliability of the rules. The comprehensive confidence scores are calculated by fusing the output probability of the distillation model and the multi-dimensional validity verification.

[0078] In specific application scenarios, the multi-dimensional validity verification in the rule generation module 304 includes: field integrity verification, syntax validity verification, rule conflict detection, and sandbox execution verification. Among them, rule conflict detection includes: based on the rule's applicable subject, triggering event, data source field, effective device, and execution action, detecting threshold conflicts, action conflicts, and conflicts during the effective period between candidate structured executable rules and deployed rules. Sandbox execution verification includes: loading candidate structured executable rules in an isolated rule engine environment, and verifying the rule triggering results using normal samples, violation samples, and boundary samples.

[0079] In specific application scenarios, such as Figure 4 As shown, the above-mentioned device also includes a model iteration module 306, which is specifically used to evaluate the accuracy of the deployed target structured executable rules and obtain the evaluation results; to verify the target structured executable rules based on the evaluation results, identify erroneous rules in the target structured executable rules, and correct the erroneous rules; and to use the corrected target structured executable rules as incremental training samples and feed them back to the structured training samples for iterative optimization of the distillation model.

[0080] It should be noted that other corresponding descriptions of the functional units involved in the dynamic distillation device for financial regulatory rules provided in this embodiment can be found in [reference needed]. Figure 1 and Figure 2 The corresponding descriptions in [the document] will not be repeated here.

[0081] Based on the above, Figure 1 and Figure 2Accordingly, this embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the above-described dynamic distillation method for financial regulatory rules.

[0082] Based on this understanding, the technical solution of this application can be embodied in the form of a software product. This software product can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, or portable hard drive), and includes several instructions to enable a computer device (such as a personal computer, server, or network device) to execute the dynamic distillation method for financial regulatory rules in various implementation scenarios of this application.

[0083] Based on the above, Figure 1 and Figure 2 The method shown, and Figure 3 and Figure 4 The illustrated embodiment of the dynamic distillation apparatus for financial regulatory rules is designed to achieve the aforementioned objectives, such as... Figure 5 As shown, this embodiment also provides a physical device for the dynamic distillation of financial regulatory rules. This device includes a communication bus, a processor, a memory, and a communication interface. It may also include input / output interfaces and a display device. The various functional units can communicate with each other via the bus. The memory stores a computer program, and the processor executes the program stored in the memory to perform the dynamic distillation method for financial regulatory rules described in the above embodiment.

[0084] Optionally, the physical device may also include a user interface, a network interface, a camera, radio frequency (RF) circuitry, sensors, audio circuitry, a Wi-Fi module, etc. The user interface may include a display screen, input units such as a keyboard, etc., and optional user interfaces may also include USB interfaces, card reader interfaces, etc. The network interface may optionally include standard wired interfaces, wireless interfaces (such as Wi-Fi interfaces), etc.

[0085] Those skilled in the art will understand that the structure of the physical device for dynamic distillation of financial regulatory rules provided in this embodiment does not constitute a limitation on the physical device, and may include more or fewer components, or combine certain components, or have different component arrangements.

[0086] The storage medium may also include an operating system and a network communication module. The operating system is a program that manages the hardware and software resources of the aforementioned physical device, supporting the operation of information processing programs and other software and / or programs. The network communication module is used to enable communication between the various components within the storage medium, as well as communication with other hardware and software in the information processing physical device.

[0087] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware platform, or it can be implemented by hardware. By applying the technical solution of this application, data preparation can be completed without manual intervention through automated access and standardization of unstructured rule text, reducing the time delay caused by manual collection and organization. Through a preset logic tree structure, the semantic information in the rule text expressed in natural language is extracted and expressed in a structured way, transforming the ambiguous natural language description into a clear and unified logical form, and finally accurately quantifying it into executable logic, avoiding semantic understanding bias. Through parameter adjustment, the large language model learns and masters the mapping rules between unstructured text and structured rules, thereby enabling the distillation model to output corresponding structured executable rules based on the received unstructured rule text, without the need for manual parsing and encoding of each rule, shortening the rule update cycle and reducing system maintenance costs. The target structured executable rules are deployed to the rule engine of the target business system, realizing an end-to-end automated process from new rule text input to executable rule deployment, shortening the rule update time, and the generated rules can be directly recognized and executed by the rule engine of the business system without modifying the existing system architecture, further reducing the complexity of deployment and maintenance. The above method forms a complete automated closed loop from rule text acquisition, structured transformation, model distillation to rule generation and deployment, realizing real-time parsing of unstructured rule text, as well as structured extraction and automated deployment, shortening the rule update cycle and reducing maintenance costs.

[0088] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing this application. Those skilled in the art will understand that the modules in the apparatus of the embodiment can be distributed within the apparatus of the embodiment as described, or can be modified to be located in one or more apparatuses different from this embodiment. The modules of the above-described embodiment can be combined into one module, or further divided into multiple sub-modules.

[0089] The serial numbers in this application are for descriptive purposes only and do not represent the superiority or inferiority of any particular implementation scenario. The above disclosures are merely a few specific implementation scenarios of this application; however, this application is not limited thereto, and any variations conceived by those skilled in the art should fall within the protection scope of this application.

Claims

1. A method for dynamic distillation of financial regulatory rules, characterized in that, include: Obtain the unstructured rule text to be processed, and preprocess the unstructured rule text; Based on the logic tree annotation system, the preprocessed unstructured rule text is transformed into structured training samples; Using a pre-trained large language model as the teacher model, a large language model with parameters to be trained is constructed as the student model. The teacher output is generated based on the structured training samples, and the student model is trained by distillation using the structured training samples and the teacher output to obtain a distillation model. The distillation model is used to map the unstructured rule text into structured executable rules. Based on preset output format constraint rules, the output process of the distillation model is constrained and decoded to generate candidate structured executable rules. The candidate structured executable rules are then subjected to multi-dimensional validity verification. After the verification is passed, the target structured executable rule is determined. Deploy the target structured executable rules into the rule engine of the target business system.

2. The method according to claim 1, characterized in that, The step of obtaining the unstructured rule text to be processed and preprocessing the unstructured rule text includes: Obtain the original unstructured rule text from the preset rule publishing source; The unstructured rule text is cleaned to remove formatting, and the cleaned unstructured rule text is converted into a unified encoding format. The format cleaning includes removing preset formatting tags and redundant information. Unstructured rule text in a unified encoding format is classified and stored based on preset rule categories.

3. The method according to claim 1, characterized in that, The logic tree-based annotation system transforms preprocessed unstructured rule text into structured training samples, including: Construct a logic tree annotation system, wherein the logic tree annotation system uses entity type as nodes, relation type as edges, and condition expression as node attributes or edge attributes; Based on the aforementioned logic tree annotation system, semantic annotation is performed on the preprocessed unstructured rule text to extract rule logic; The rule logic is converted into a structured data format to form a structured training sample. The structured training sample includes a logical structure field and a basis reference field. The logical structure field is used to characterize the conditional branch structure in the rule logic, and the basis reference field is used to characterize the source information of the unstructured rule text corresponding to the rule logic.

4. The method according to claim 1, characterized in that, The process of using the structured training samples and the teacher output to perform distillation training on the student model to obtain a distillation model includes: The structured training samples are input into the teacher model to obtain the teacher output results, wherein the teacher output results include structured rule logic, rule field probability distribution and confidence weight; The structured training samples are input into the student model to obtain student output results. The parameters of the student model are adjusted to minimize the supervision loss between the student output results and the real labels corresponding to the structured training samples, and to minimize the distillation loss between the student output results and the teacher output results, so as to obtain the target model parameters. The distillation model is determined based on the target model parameters.

5. The method according to claim 1, characterized in that, The target structured executable rule includes: a relationship type for characterizing the logical association between entities, a conditional expression for characterizing the rule trigger judgment condition, and a comprehensive confidence level for characterizing the overall reliability of the rule, wherein the comprehensive confidence level is calculated based on the output probability of the distillation model and the multi-dimensional validity verification.

6. The method according to claim 1, characterized in that, The multi-dimensional validity verification includes: field integrity verification, syntax validity verification, rule conflict detection, and sandbox execution verification; The rule conflict detection includes: detecting threshold conflicts, action conflicts, and conflicts during the effective period between the candidate structured executable rule and the deployed rule based on the rule's applicable subject, triggering event, data source field, effective device, and execution action; The sandbox execution verification includes: loading the candidate structured executable rules in an isolated rule engine environment, and using normal samples, violation samples, and boundary samples to verify the rule trigger results.

7. The method according to claim 1, characterized in that, After deploying the target structured executable rule to the rule engine of the target business system, the method further includes: The accuracy of the deployed target structured executable rules is evaluated to obtain the evaluation results; Based on the evaluation results, the target structured executable rules are validated, erroneous rules in the target structured executable rules are identified, and the erroneous rules are corrected. The modified target structured executable rules are used as incremental training samples and fed back to the structured training samples for iterative optimization of the distillation model.

8. A dynamic distillation device for financial regulatory rules, characterized in that, include: The text preprocessing module is used to obtain the unstructured rule text to be processed and to preprocess the unstructured rule text. The sample construction module is used to transform preprocessed unstructured rule text into structured training samples based on the logic tree annotation system. The distillation training module is used to construct a large language model with parameters to be trained as a student model, using a pre-trained large language model as a teacher model, generating teacher output results based on the structured training samples, and using the structured training samples and the teacher output results to perform distillation training on the student model to obtain a distillation model. The distillation model is used to map the unstructured rule text into structured executable rules. The rule generation module is used to perform constraint decoding on the output process of the distillation model based on preset output format constraint rules, generate candidate structured executable rules, and perform multi-dimensional validity verification on the candidate structured executable rules. After the verification is passed, the target structured executable rule is determined. The rule deployment module is used to deploy the target structured executable rules to the rule engine of the target business system.

9. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images