Decision engine updating method and device based on large language model and electronic equipment

Abstract

The invention discloses a decision engine updating method and device based on a large language model and electronic equipment, and relates to the field of financial science and technology, and the method comprises the steps that a target instruction is analyzed through a target model, and an operation type and operation parameters corresponding to the target instruction are obtained, the target instruction is an instruction which is described through a natural language and is used for interacting with a decision engine of a financial institution, generating a target code based on an operation type and operation parameters corresponding to the target instruction, the target code being a script code which can be identified by the decision engine, and testing the target code under the condition that the target code passes the test. And updating the N sub-modules of the decision engine based on the target code. The technical problem of low interaction efficiency caused by the fact that operation and maintenance personnel need to carry out manual interaction on the decision engine based on expert experience in the prior art is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of financial technology, in particular, to a decision engine updating method and device based on a large language model and an electronic device. BACKGROUND

[0002] In the field of financial technology, the decision engine, as a core intelligent decision tool, carries out key tasks from risk assessment to credit approval. However, the current configuration and updating method of the decision engine mainly relies on the professional knowledge and manual operation of the operation and maintenance personnel. The rule setting and parameter adjustment of the traditional decision engine, especially in scenarios involving complex business logic, such as credit score adjustment and risk policy update, require the operation and maintenance personnel not only to have deep business understanding ability, but also to be proficient in specific decision script languages, thereby increasing the training cost and operation difficulty of the operation and maintenance personnel. The existing technical solutions have the following defects:

[0003] (1) High learning cost: The operation and maintenance personnel need to invest a lot of time to learn and understand the special language and complex model structure of the decision engine before they can effectively configure it.

[0004] (2) Low configuration efficiency: Manually writing and modifying rule condition expressions or decision parameters is a slow and error-prone process that is difficult to adapt to the rapidly changing market environment and business needs.

[0005] (3) Strong dependence on professionals: The maintenance and update of the decision engine rely on a small number of operation and maintenance personnel with professional knowledge, which limits the flexibility and scalability of the decision-making process.

[0006] From the above content, it can be seen that with the increasing complexity of financial business and the dynamic changes in decision-making needs, the traditional manual configuration method cannot meet the efficient and accurate decision-making requirements. The existing technical solutions have the technical problem of low interaction efficiency caused by the need for the operation and maintenance personnel to manually interact with the decision engine based on expert experience.

[0007] At present, there is no effective solution to the above problems. SUMMARY

[0008] The present application provides a decision engine updating method and device based on a large language model and an electronic device to at least solve the technical problem of low interaction efficiency caused by the need for the operation and maintenance personnel to manually interact with the decision engine based on expert experience in the prior art.

[0009] According to an aspect of the present application, a large language model-based decision engine updating method is provided, comprising: analyzing a target instruction through a target model to obtain an operation type and operation parameters corresponding to the target instruction, wherein the target instruction is an instruction described in natural language for interacting with a decision engine of a financial institution, and the target model is a large language model trained based on a historical instruction set and a historical interception rule set; generating a target code based on the operation type and operation parameters corresponding to the target instruction, wherein the target code is a script code recognizable by the decision engine; and updating N sub-modules of the decision engine based on the target code in the case that the target code passes the test, wherein N is a positive integer, and the N sub-modules at least include a rule module, a decision module, a scoring module, and a log module.

[0010] Optionally, the step of analyzing the target instruction through the target model to obtain the operation type and operation parameters corresponding to the target instruction comprises: performing data cleaning on the target instruction through the target model to obtain a first instruction, wherein the target instruction is used to remove stop words in the target instruction; performing word segmentation on the first instruction to obtain P words, wherein P is a positive integer; determining the operation parameters based on a first keyword in the P words, wherein the first keyword is a numerical type word; and determining the operation type based on a second keyword in the P words and context information corresponding to the target instruction, wherein the second keyword is used to represent action information required to be executed by the target instruction, and the operation type is at least one of:

[0011] an added type;

[0012] a deleted type;

[0013] a modified type.

[0014] Optionally, the step of generating the target code based on the operation type and operation parameters corresponding to the target instruction comprises: determining Q target sub-templates based on the operation type corresponding to the target instruction, wherein Q is a positive integer, and the Q target sub-templates are templates stored in a preset database and capable of being presented in a visual form; arranging the Q target sub-templates on a visual operation interface based on a preset process to obtain an arrangement result as a target template, wherein the preset process is used to represent a code execution logic of the decision engine; and filling the operation parameters corresponding to the target instruction into the target template to obtain the target code.

[0015] Optionally, after generating the target code based on the operation type and the operation parameter corresponding to the target instruction, the method for updating the decision engine based on the large language model further includes: performing a first test operation on the target code, wherein the first test operation is used to detect whether the target code meets the coding specification corresponding to the decision engine; in a case where the first test operation fails, generating first information, wherein the first information is used to prompt the user that the target code has a syntax error; in a case where the first test operation passes, transmitting the target code to a backup engine through a preset interface, wherein the backup engine is a backup of the decision engine of the financial institution.

[0016] Optionally, after transmitting the target code to the backup engine through the preset interface, the method for updating the decision engine based on the large language model further includes: updating N sub-modules corresponding to the backup engine based on the target code; testing the updated backup engine based on preset test data to obtain a test probability, wherein the preset test data is user transaction data of the financial institution in a historical time period, and the test probability is the accuracy rate of the updated backup engine in intercepting user transactions with transaction risks; in a case where the test probability is greater than a historical probability, determining that the target code passes the test, wherein the historical probability is the accuracy rate of the backup engine before the update in intercepting user transactions with transaction risks.

[0017] Optionally, in a case where the target code passes the test, updating the N sub-modules of the decision engine based on the target code includes: in a case where the target code passes the test, parsing the target code to obtain a parsing result, wherein the parsing result is used to represent the dependency relationship and the update order of the N sub-modules; updating the N sub-modules based on the parsing result.

[0018] Optionally, the training step of the target model includes: performing feature extraction on all historical instructions included in the historical instruction set to obtain X first features, wherein each first feature is used to represent the instruction information corresponding to a historical instruction; performing feature extraction on the configuration code of each historical interception rule in the historical interception rule set to obtain X second features, wherein each second feature is used to represent the configuration process corresponding to the configuration code of a configuration type; based on the correspondence between the X first features and the X second features, dividing the X first features and the X second features to obtain a training set and a test set; based on the feature data in the training set, iteratively training the initial large language model to obtain a first model; based on the feature data in the test set, iteratively updating the first model to obtain the target model.

[0019] According to another aspect of the present application, a large language model-based decision engine updating apparatus is also provided, comprising: an analysis unit configured to analyze a target instruction by a target model to obtain an operation type and operation parameters corresponding to the target instruction, wherein the target instruction is an instruction described in natural language for interacting with a decision engine of a financial institution, and the target model is a large language model trained based on a historical instruction set and a historical interception rule set; a first generation unit configured to generate a target code based on the operation type and operation parameters corresponding to the target instruction, wherein the target code is a script code recognizable by the decision engine; and a first updating unit configured to update N sub-modules of the decision engine based on the target code in a case where the target code passes a test, wherein N is a positive integer, and the N sub-modules at least include a rule module, a decision module, a scoring module, and a log module.

[0020] According to another aspect of the present application, a computer program product having a computer program stored therein is also provided, wherein the computer program controls the computer program product to perform the large language model-based decision engine updating method of any one of the above aspects when the computer program is executed.

[0021] According to another aspect of the present application, an electronic device is also provided, wherein the electronic device comprises one or more processors and a memory, and the memory is configured to store one or more programs, wherein the one or more programs, when executed by the one or more processors, cause the one or more processors to implement the large language model-based decision engine updating method of any one of the above aspects.

[0022] In the present application, first, a target instruction is analyzed by a target model to obtain an operation type and operation parameters corresponding to the target instruction, wherein the target instruction is an instruction described in natural language for interacting with a decision engine of a financial institution, and the target model is a large language model trained based on a historical instruction set and a historical interception rule set, and then, a target code is generated based on the operation type and operation parameters corresponding to the target instruction, wherein the target code is a script code recognizable by the decision engine, and in a case where the target code passes a test, N sub-modules of the decision engine are updated based on the target code, wherein N is a positive integer, and the N sub-modules at least include a rule module, a decision module, a scoring module, and a log module.

[0023] From the above, it can be seen that the present application adopts a manner of analyzing a target instruction described in natural language by a pre-trained large language model (i.e., a target model) to automatically generate a script code (i.e., a target code) corresponding to a decision engine, thereby achieving the purpose of quickly updating and configuring sub-modules in the decision engine.

[0024] In other words, the target instruction is deeply analyzed based on the target model, the operation type and operation parameter corresponding to the target instruction can be automatically identified, and the operation type and operation parameter are quickly converted into DSL code (i.e., target code) executable by the decision engine. On the basis that the DSL code passes the test, the sub-modules in the decision engine are dynamically updated based on the DSL code. Since the target instruction is an instruction written in natural language that can be understood by the operation and maintenance personnel, the technical solution of the present application reduces the operation and maintenance personnel of the financial institution The operation complexity of updating the decision engine, thereby realizing the technical effect of improving the efficiency of the operation and maintenance personnel interacting with the decision engine, and further solving the technical problem of low interaction efficiency caused by the operation and maintenance personnel manually interacting with the decision engine based on expert experience in the prior art. BRIEF DESCRIPTION OF DRAWINGS

[0025] The drawings described herein are used to provide further understanding of the present application, and form a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application, and do not constitute an improper limitation on the present application. In the drawings:

[0026] Figure 1 is a flowchart of an optional decision engine updating method based on a large language model according to an embodiment of the present application;

[0027] Figure 2 is a flowchart of another optional decision engine updating method based on a large language model according to an embodiment of the present application;

[0028] Figure 3 is an architecture diagram of an optional decision engine updating system based on a large language model according to an embodiment of the present application;

[0029] Figure 4 is a schematic diagram of an optional decision engine updating device based on a large language model according to an embodiment of the present application;

[0030] Figure 5 is a structural block diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0031] In order to enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should be within the scope of protection of the present application.

[0032] It should be noted that the terms "first", "second", and the like in the description and claims of the application and the above drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0033] It should also be noted that the relevant information (including but not limited to the information for display and the analyzed information) and data (including but not limited to the target instruction data, the data in the historical instruction set, and the data in the historical interception rule set) involved in the present application are all information and data authorized by the user or fully authorized by all parties. For example, an interface is provided between the system and the relevant users or institutions. Before obtaining the relevant information, the interface needs to send a request to the aforementioned user or institution, and after receiving the consent information feedback from the aforementioned user or institution, the relevant information is obtained.

[0034] In addition, the collection, storage, use, processing, transmission, provision, disclosure and application of the relevant information and data involved in the present application comply with the relevant laws, regulations and standards of the relevant regions, and necessary security measures are taken, which do not violate public order and good customs. In addition, the present application provides a corresponding operation portal for users to choose to authorize or refuse authorization. If the user chooses to refuse authorization, the corresponding expert decision-making process is entered.

[0035] In an optional embodiment, a method for interacting with a natural language processing system and a decision engine is provided. The method analyzes a target instruction input by a user through the natural language processing system, and then determines a target expression according to the analysis result, wherein the target expression is a condition expression corresponding to the target instruction. Then, the rule set, decision tree and score card in the decision engine are updated according to the target expression.

[0036] However, the conventional natural language processing system cannot recognize domain-specific terms, so the above method cannot generate configuration instructions that meet the engine specifications. Therefore, it is necessary to update the rule set, decision tree and score card in the decision engine according to the target expression in an artificial manner, thereby causing the technical problem of low interaction efficiency.

[0037] According to an embodiment of the present application, an embodiment of a large language model-based decision engine updating method is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in an order different from that shown.

[0038] The present application provides a large language model-based decision engine updating system (referred to as updating system) for executing the large language model-based decision engine updating method in the present application, Figure 1 is a flowchart of an optional large language model-based decision engine updating method according to an embodiment of the present application, as Figure 1 shown, the method comprises the following steps:

[0039] Step S101, the target instruction is parsed by the target model to obtain the operation type and operation parameter corresponding to the target instruction, wherein the target instruction is an instruction described in natural language for interacting with the decision engine of the financial institution, and the target model is a large language model trained based on the historical instruction set and the historical interception rule set.

[0040] Here, the target instruction is an instruction described in natural language, which is issued by a business personnel or an operation and maintenance personnel of a financial institution, and is used to interact with the decision engine. The target instruction can be used to adjust the transaction score threshold or update the interception rule set in the decision engine, etc.

[0041] Here, the operation type is a classification in the parsing result, indicating the decision engine operation that the target instruction aims to perform, such as rule updating, score adjustment, etc.

[0042] Here, the operation parameter is a specific numerical value or condition related to the operation type, used to refine the execution details of the instruction.

[0043] In step S101, the updating system parses the natural language instruction through the target model, extracts the operation type and parameter, and lays the foundation for subsequent DSL (Domain Specific Language, domain specific language) code generation.

[0044] Step S102, generating target code based on the operation type and operation parameter corresponding to the target instruction, wherein the target code is a script code that the decision engine can recognize.

[0045] Here, the target code is a DSL code that the decision engine can recognize and execute, which is generated based on the operation type and parameter parsed from the target instruction, and is used to perform specific decision engine logic update operations.

[0046] In step S102, the update system converts the natural language described operation instruction into a domain specific script language that can be understood by the decision engine, so as to realize the interactive intention of the operation and maintenance personnel, without manual coding, thereby improving the efficiency of the operation and maintenance personnel interacting with the decision engine.

[0047] Step S103, in the case where the target code passes the test, updating the N sub-modules of the decision engine based on the target code, wherein N is a positive integer, and the N sub-modules at least include a rule module, a decision module, a scoring module, and a log module.

[0048] Here, the N sub-modules refer to a plurality of functional modules inside the decision engine, including a rule module (responsible for defining and executing rule conditions), a decision module (making decisions based on rules and scoring results), a scoring module (quantitative assessment of user risk or credit), and a log module (recording decision process and results), etc.

[0049] Here, the test refers to executing the target code in a safe sandbox environment (for example, a test environment corresponding to the backup engine) to verify its logical correctness and functional integrity, to ensure that no errors are introduced for the decision engine, thereby causing abnormal operation.

[0050] From the above, it can be seen that the present application adopts a pre-trained large language model (i.e. target model) to analyze the target instruction described in natural language, and automatically generates the script code (i.e. target code) corresponding to the decision engine, thereby achieving the purpose of quickly updating and configuring the sub-modules in the decision engine.

[0051] In other words, the present application deeply analyzes the target instruction based on the target model, can automatically identify the operation type and operation parameters corresponding to the target instruction, and quickly converts the operation type and operation parameters into DSL code (i.e. target code) that can be executed by the decision engine. Based on the test of the DSL code, the present application dynamically updates the sub-modules in the decision engine based on the DSL code. Since the target instruction is an instruction written in natural language that can be understood by the operation and maintenance personnel, the technical solution of the present application reduces the operation complexity of the operation and maintenance personnel of the financial institution in updating the decision engine, thereby realizing the technical effect of improving the efficiency of the operation and maintenance personnel interacting with the decision engine, and further solving the technical problem of low interaction efficiency caused by the operation and maintenance personnel manually interacting with the decision engine based on expert experience in the prior art.

[0052] In an optional embodiment, the updating system first performs data cleaning on the target instruction through the target model to obtain a first instruction, wherein the target instruction is used to remove stop words in the target instruction, then the updating system performs word segmentation on the first instruction to obtain P words, wherein P is a positive integer, then the updating system determines the operation parameter based on a first keyword in the P words, wherein the first keyword is a numerical type word, and the updating system determines the operation type based on a second keyword in the P words and context information corresponding to the target instruction, wherein the second keyword is used to represent action information required by the target instruction, and the operation type is at least one of the following:

[0053] Add type;

[0054] Delete type;

[0055] Modify type.

[0056] Here, the stop word refers to a word in natural language that does not carry actual meaning, such as "of", "and", etc., thereby improving the accuracy and efficiency of target instruction analysis.

[0057] Here, the first keyword refers to a numerical type keyword identified from the P words, which is used to determine the specific numerical value of the operation, such as "1.5", "2.0".

[0058] Here, the second keyword refers to a keyword identified from the P words, which is used to represent the action of the target instruction, such as "increase", "adjust", "cancel".

[0059] Optionally, the updating system achieves the purpose of providing specific numerical basis for subsequent decision engine rule module update, scoring module update and other operations by identifying and extracting operation parameters.

[0060] Optionally, the updating system determines the specific action required by the target instruction by identifying the second keyword and its context, thereby providing type guidance for subsequent generation of execution code.

[0061] Optionally, the present application converts the originally ambiguous natural language instruction into clear operation type and operation parameter, and then generates script code recognizable by the decision engine, which greatly reduces the complexity of manual intervention, thereby improving the accuracy and efficiency of decision engine configuration.

[0062] In an optional embodiment, the updating system first determines Q target sub-templates based on the operation type corresponding to the target instruction, wherein Q is a positive integer, and the Q target sub-templates are templates stored in the preset database and capable of being presented in a visual form, then arranges the Q target sub-templates on the visual operation interface based on a preset process, takes the arrangement result as the target template, wherein the preset process is used to represent the code execution logic of the decision engine, and then fills the operation parameters corresponding to the target instruction into the target template to obtain the target code.

[0063] Here, the Q target sub-templates refer to a template set selected from the preset database, and these templates can present different decision engine operation logics in a visual form, for example, for the "modification" operation type, the preset database contains visual templates corresponding to multiple modification rule sets, scorecards or decision flows.

[0064] Optionally, the application selects appropriate visual sub-templates according to the operation type, provides an intuitive basic framework, and thus can automatically arrange the Q target sub-templates in a process based on the preset process in a visual manner, thereby reducing the generation difficulty of the target template.

[0065] Here, the preset process refers to a code execution logic sequence preset in the decision engine, which is used to guide how to construct a complete decision flow based on the operation type and the sub-template.

[0066] Optionally, the updating system automatically arranges the sub-templates on the visual interface based on the preset process to automatically generate the logic framework of the decision engine code, thereby simplifying the code generation process and operation complexity.

[0067] Optionally, through the above steps, the application not only improves the interaction efficiency between the operation and maintenance personnel and the decision engine, but also reduces the code generation error rate. The preset database and the template library, combined with the visual operation interface, enable non-technical business personnel to complete the adjustment and optimization of the decision flow, thereby promoting the modernization process of the financial decision engine management.

[0068] In an optional embodiment, the updating system first performs a first test operation on the target code, wherein the first test operation is used to detect whether the target code meets the coding specification corresponding to the decision engine, and then generates first information in the case that the first test operation fails, wherein the first information is used to prompt the user that there is a syntax error in the target code.

[0069] In addition, in the case that the first test operation passes, the updating system transmits the target code to a backup engine through a preset interface, wherein the backup engine is a backup of the decision engine of the financial institution.

[0070] Optionally, the updating system first inputs the target code into a syntax checker, and then performs the first test procedure through the syntax checker.

[0071] Here, the first test operation refers to static and dynamic testing of the target code performed on a local environment or a simulated server to ensure that it complies with the coding specifications of the decision engine, covering aspects such as syntax correctness, logical coherence, and compatibility.

[0072] Here, the first information refers to information automatically generated by the system when the first test operation fails, explicitly indicating syntax errors in the target code, such as unclosed parentheses, misspelled keywords, etc.

[0073] Optionally, through the first information, the updating system can timely feedback to the operation and maintenance personnel the problems existing in the DSL code, facilitating rapid correction and ensuring the quality of the code for subsequent deployment.

[0074] Here, the backup engine refers to a replicated version of the decision engine of a financial institution, usually used for testing, debugging, and emergency recovery purposes to ensure the safety of the production environment.

[0075] Here, the preset interface refers to a standardized communication channel established between the decision engine and external systems for transmitting target code and other operation instructions.

[0076] Optionally, through the above steps, the technical solution of the present application not only ensures the quality of the target code, but also enhances the security and reliability of the system through the use of the backup engine, providing a more robust decision-making process adjustment mechanism for financial institutions and effectively preventing business interruption risks that may be caused by code errors.

[0077] In an alternative embodiment, the updating system first updates the N sub-modules corresponding to the backup engine based on the target code, then tests the updated backup engine based on preset test data to obtain a test probability, wherein the preset test data is user transaction data of the financial institution in a historical time period, the test probability is the accuracy rate of the updated backup engine in intercepting user transactions with transaction risks, and then the updating system determines that the target code passes the test when the test probability is greater than a historical probability, wherein the historical probability is the accuracy rate of the backup engine before updating in intercepting user transactions with transaction risks.

[0078] Here, the updating operation refers to modifying the N sub-modules of the backup engine based on the operation type and parameters extracted from the target code to reflect the latest business rules and decision logic.

[0079] Optionally, in the above-mentioned testing process of the target code, the updating system applies the target code converted from the natural language instruction to the backup engine to realize dynamic updating of the decision rule without affecting the real-time operation of the main decision engine.

[0080] Optionally, the updating system verifies whether the target code improves the performance of the decision engine by comparing the test probability with the historical probability, to ensure that the new rule not only meets the coding specification, but also can more accurately identify and handle transaction risks.

[0081] In an optional embodiment, in the case that the target code passes the test, the updating system parses the target code to obtain a parsing result, wherein the parsing result is used to represent the dependency relationship and the updating order corresponding to the N sub-modules, and then the updating system updates the N sub-modules based on the parsing result.

[0082] Here, the parsing result refers to the structured analysis output of the target code, mainly representing the dependency relationship and the logical order of updating between the sub-modules, thereby providing guidance for subsequent updating of the sub-modules.

[0083] Optionally, the updating system identifies the influence degree and the mutual dependency relationship of the target code on each sub-module in the decision engine by in-depth parsing of the target code, determines the most reasonable updating order to prevent system conflicts or instability caused by improper updating order, and then updates the N sub-modules of the decision engine one by one or synchronously based on the parsing result obtained in the first step and the determined updating order and dependency relationship. The updating process ensures that it matches the parsing result, considering not only the influence of the target code on a single sub-module, but also the dependency relationship between different sub-modules, thereby avoiding engine failures caused by local updating.

[0084] In an optional embodiment, the training step of the target model includes: performing feature extraction on all historical instructions included in the historical instruction set to obtain X first features, wherein each first feature is used to represent the instruction information corresponding to a historical instruction; then the updating system performs feature extraction on the configuration code of each historical interception rule in the historical interception rule set to obtain X second features, wherein each second feature is used to represent the configuration process corresponding to a configuration type of configuration code; then the updating system divides the X first features and the X second features based on the correspondence between the X first features and the X second features to obtain a training set and a test set; and the updating system iteratively trains the initial large language model based on the feature data in the training set to obtain a first model; finally, the updating system iteratively updates the first model based on the feature data in the test set to obtain the target model.

[0085] Here, the historical instruction set records all natural language instructions of the financial institution decision engine in the past, and is an important data source for model training.

[0086] Here, the historical interception rule set contains configuration information of all interception rules in the history of the financial institution decision engine, and provides a basis for the model to understand the decision process.

[0087] Here, the correspondence refers to the association between the first feature and the second feature, reflecting the corresponding logical association between the natural language instruction and the corresponding configuration code in the decision engine.

[0088] Optionally, the updating system can ensure the performance and reliability of the test model while ensuring the amount of model training data by dividing the feature data into a training set and a test set, avoiding overfitting, and thus ensuring the effectiveness of the model in actual application. The updating system iteratively trains the initial large language model on the training set to obtain a specialized model that preliminarily adapts to the configuration requirements of the financial institution decision engine. Then, the updating system iteratively trains the model on the training set, enabling the model to learn the rules of converting natural language instructions to decision engine configuration codes in specific business scenarios of the financial institution, thereby improving the accuracy and practicality of the target model obtained by training.

[0089] From the above, it can be seen that the present application adopts a pre-trained large language model (i.e., a target model) to analyze the target instruction described in natural language, and automatically generates the corresponding script code of the decision engine (i.e., the target code), thereby achieving the purpose of quickly updating and configuring the sub-modules in the decision engine.

[0090] In other words, the present application performs in-depth analysis of the target instruction based on the target model, can automatically identify the operation type and operation parameters corresponding to the target instruction, and quickly converts the operation type and operation parameters into DSL code (i.e., target code) that the decision engine can execute. Based on the tested DSL code, the present application dynamically updates the sub-modules in the decision engine based on the DSL code. Since the target instruction is a natural language instruction written by the operation and maintenance personnel, the technical solution of the present application reduces the operation complexity of the operation and maintenance personnel of the financial institution in updating the decision engine, thereby achieving the technical effect of improving the efficiency of the operation and maintenance personnel interacting with the decision engine, and thus solving the technical problem of low interaction efficiency caused by the operation and maintenance personnel manually interacting with the decision engine based on expert experience in the prior art.

[0091] In an alternative embodiment, Figure 2 is a flowchart of another alternative large language model-based decision engine updating method according to an embodiment of the present application, as Figure 2 shown, the method comprises:

[0092] Firstly, a user natural language instruction (i.e., a target instruction) is input into an LLM (Large Language Model) model (i.e., a target model) for intent analysis, and the LLM model is used to determine the operation type of the user natural language instruction, where the operation type is an addition type, a modification type, or a deletion type. Then, a DSL template (i.e., a target template) is generated based on the operation type analyzed by the LLM model, and operation parameters analyzed by the LLM model are filled into the DSL template to obtain a DSL code (i.e., a target code).

[0093] Then, the DSL code is subjected to syntax verification by a syntax verifier. If the syntax verification fails, a prompt message is returned to prompt the user to check the input natural language instruction and the generated DSL code. If the syntax verification passes, the DSL code is deployed to a backup decision engine (i.e., a backup engine) of a financial institution, and a transaction risk interception test is performed on a decision flow corresponding to the DSL code based on the backup decision engine to obtain a test result.

[0094] If the test result shows that the transaction risk interception test is successful, the system configuration of the decision engine of the financial institution is updated based on the DSL code. If the test result shows that the transaction risk interception test fails, a prompt message is returned.

[0095] In an optional embodiment, Figure 3 is an architecture diagram of an optional decision engine updating system based on a large language model according to an embodiment of the present application, as Figure 3 shown, the system architecture includes a user interface layer, an API gateway layer, an LLM_Agent (Large Language Model_Agent) layer, a decision engine layer, and a backend system layer.

[0096] Optionally, in the user interface layer, after a user logs in to a terminal of the decision engine updating system based on a large language model, a target request (i.e., a user request carrying a target instruction) is input into the updating system. Then, the target request is transmitted to an authentication module of the API gateway layer through the API gateway, the user information carried by the target request is subjected to authority detection through the authentication module, and after it is detected that the user has the authority to interact with the decision engine, it is prompted that the verification is passed, and the target request is transmitted to a semantic understanding module of the LLM_Agent layer through a routing distribution manner for analysis to obtain an analysis result, where the analysis result at least includes an operation type and an operation parameter.

[0097] Optionally, after it is detected that the user does not have the authority to interact with the decision engine, a user prompt message is returned to prompt the user that the verification fails.

[0098] Optionally, after the authentication module prompts that the verification is passed, the update system transmits the resolution result to the DSL generator, and the DSL generator generates the DSL code (i.e., the target code) based on the resolution result, and then the update system performs syntax checking on the DSL code by the syntax checker, and in the case that the syntax checking is passed, the DSL code is transmitted to the decision engine.

[0099] Optionally, after the decision engine receives the DSL code, the DSL code is parsed by the DSL parser to obtain the decision flow corresponding to the DSL code, and then the rule execution engine updates the settings of the rule database, the model warehouse and the log system in the backend database based on the decision flow corresponding to the DSL code, wherein the backend database is a database connected with the decision engine, the rule database is used to store the rule set in the decision engine, and the model warehouse is used to store the decision module and the scoring module corresponding to the decision engine, for example as follows:

[0100] Example 1: the target instruction carried in the target request input by the user is "add a rule to suspend the loan to the user who has 2 overdue behaviors in the past 3 months, and notify the relevant operation and maintenance personnel", and the processing flow of the update system is as follows:

[0101] The target instruction carried in the target request is parsed, the operation type parsed is "add type", the corresponding operation action is "rule set update" and "notification action", and the operation parameters are "3 months" and "2 times", then the DSL code is generated based on the resolution result, and the DSL code is as follows:

[0102] {

[0103] "rule_name":"loan_denial_rule",

[0104] "condition":"user.loan_overdue_count_3m>=2",

[0105] "action":"reject_application",

[0106] "post_hook":"send_alert_to_risk_team()"

[0107] }

[0108] Then, the DSL code is syntax checked by the syntax checker, and in the case that the syntax checking is passed, the DSL code is transmitted to the decision engine and the backend database is updated based on the decision flow.

[0109] Example 2: the target instruction carried in the target request input by the user is "increase the user income weight in the scoring module from 0.3 to 0.4, and reduce the user's age weight to 0.2", the processing flow of the updating system is as follows:

[0110] The target instruction carried in the target request is analyzed, the operation type analyzed is "modification type", the operation action corresponding is "scoring card update", and the operation parameters are "0.4" and "0.2". Then, based on the analysis result, the DSL code is generated, and based on the DSL code, the scoring module is located, and the weight parameter corresponding to the scoring model is modified.

[0111] After the weight parameter corresponding to the scoring model is updated, the test flow is triggered, and by comparing the rejection rate and bad account rate of the user transaction processed by the decision engine before and after adjustment, the test result is determined.

[0112] From the above content, it can be known that the natural language driven Agent technology is used to convert the interactive instruction input by the user in the natural language description into the field special configuration code, the automatic arrangement of the rule set, the scoring card and the decision flow in the decision engine is realized, the operation complexity of the business personnel is reduced, thereby realizing the technical effect of improving the strategy iteration efficiency, and the application scenarios such as financial risk control and credit approval can be applied.

[0113] According to another aspect of the embodiment of the present application, a decision engine updating device based on a large language model is also provided, Figure 4 is a schematic diagram of an optional decision engine updating device based on a large language model according to an embodiment of the present application, as Figure 4 shown, a decision engine updating device based on a large language model includes an analysis unit 401, a first generation unit 402 and a first updating unit 403.

[0114] Optionally, the analysis unit is configured to analyze the target instruction through a target model to obtain an operation type and operation parameters corresponding to the target instruction, wherein the target instruction is an instruction described in a natural language for interacting with a decision engine of a financial institution, and the target model is a large language model trained based on a historical instruction set and a historical interception rule set; the first generation unit is configured to generate a target code based on the operation type and operation parameters corresponding to the target instruction, wherein the target code is a script code recognizable by the decision engine; and the first updating unit is configured to update N sub-modules of the decision engine based on the target code in the case that the target code passes the test, wherein N is a positive integer, and the N sub-modules at least include a rule module, a decision module, a scoring module and a log module.

[0115] In an optional embodiment, the parsing unit comprises a cleaning subunit, a word segmentation subunit, a first determining subunit, and a second determining subunit.

[0116] Optionally, the cleaning subunit is configured to perform data cleaning on the target instruction by using a target model to obtain a first instruction, wherein the target instruction is used to remove stop words in the target instruction; the word segmentation subunit is configured to perform word segmentation on the first instruction to obtain P words, wherein P is a positive integer; the first determining subunit is configured to determine the operation parameter based on a first keyword in the P words, wherein the first keyword is a numerical type word; and the second determining subunit is configured to determine the operation type based on a second keyword in the P words and context information corresponding to the target instruction, wherein the second keyword is used to represent action information required to be executed by the target instruction, and the operation type is at least one of the following:

[0117] an addition type;

[0118] a deletion type;

[0119] a modification type.

[0120] In an optional embodiment, the first generating unit comprises a third determining subunit, an arrangement subunit, and a filling subunit.

[0121] Optionally, the third determining subunit is configured to determine Q target sub-templates based on the operation type corresponding to the target instruction, wherein Q is a positive integer, and the Q target sub-templates are templates capable of being presented in a visual form and stored in a preset database; the arrangement subunit is configured to arrange the Q target sub-templates on a visual operation interface based on a preset process, and take an arrangement result as the target template, wherein the preset process is used to represent a code execution logic of the decision engine; and the filling subunit is configured to fill the operation parameter corresponding to the target instruction into the target template to obtain the target code.

[0122] In an optional embodiment, the decision engine updating apparatus based on the large language model further comprises a first testing unit, a second generating unit, and a first transmitting unit.

[0123] Optionally, the first testing unit is configured to perform a first testing operation on the target code, wherein the first testing operation is used to detect whether the target code conforms to a coding specification corresponding to the decision engine; the second generating unit is configured to generate first information in a case where the first testing operation fails, wherein the first information is used to prompt a user that there is a syntax error in the target code; and the first transmitting unit is configured to transmit the target code to a backup engine through a preset interface in a case where the first testing operation passes, wherein the backup engine is a backup of the decision engine of the financial institution.

[0124] In an optional embodiment, the big language model-based decision engine updating apparatus further comprises a second updating unit, a second testing unit, and a first determining unit.

[0125] Optionally, the second updating unit is configured to update the N sub-modules of the backup engine based on the target code; the second testing unit is configured to test the updated backup engine based on preset test data to obtain a test probability, wherein the preset test data is user transaction data of the financial institution in a historical time period, and the test probability is a correctness rate of the updated backup engine in intercepting user transactions with transaction risks; and the first determining unit is configured to determine that the target code passes the test if the test probability is greater than a historical probability, wherein the historical probability is a correctness rate of the backup engine before the update in intercepting user transactions with transaction risks.

[0126] In an optional embodiment, the big language model-based decision engine updating apparatus further comprises a first updating unit comprising an analysis subunit and an updating subunit.

[0127] Optionally, the analysis subunit is configured to analyze the target code to obtain an analysis result if the target code passes the test, wherein the analysis result is used to represent a dependency relationship and an update order of the N sub-modules, and the updating subunit is configured to update the N sub-modules based on the analysis result.

[0128] In an optional embodiment, the big language model-based decision engine updating apparatus further comprises a first extraction unit, a second extraction unit, a division unit, an iterative training unit, and an iterative first updating unit.

[0129] Optionally, the first extraction unit is configured to extract features of all historical instructions included in the historical instruction set to obtain X first features, wherein each first feature is used to represent instruction information of a historical instruction; the second extraction unit is configured to extract features of configuration code of each historical interception rule in the historical interception rule set to obtain X second features, wherein each second feature is used to represent a configuration flow of configuration code of a configuration type; the division unit is configured to divide the X first features and the X second features based on a corresponding relationship between the X first features and the X second features to obtain a training set and a test set; the iterative training unit is configured to iteratively train an initial big language model based on feature data in the training set to obtain a first model; and the iterative first updating unit is configured to iteratively update the first model based on feature data in the test set to obtain a target model.

[0130] From the above content, it can be seen that this application adopts a pre-trained large language model (i.e., target model) to parse the target instructions described in natural language, and automatically generates the script code (i.e., target code) corresponding to the decision engine, thereby achieving the purpose of rapid updating and configuration of the sub-modules in the decision engine.

[0131] In other words, the present application performs in-depth analysis of target instructions based on the target model, can automatically identify the operation type and operation parameters corresponding to the target instruction, and quickly convert the operation type and operation parameters into DSL code (i.e., target code) that can be executed by the decision engine. On the basis of the DSL code passing the test, the present application dynamically updates the sub-modules in the decision engine based on the DSL code. Since the target instruction is an instruction written in a natural language that the operation and maintenance personnel can understand, the technical solution of the present application reduces the operational complexity of the operation and maintenance personnel of the financial institution in updating the decision engine, thereby achieving the technical effect of improving the efficiency of the interaction between the operation and maintenance personnel and the decision engine, and thus solving the technical problem of low interaction efficiency caused by the need for the operation and maintenance personnel in the existing technology to manually interact with the decision engine based on expert experience.

[0132] According to another aspect of an embodiment of the present application, a computer program product is further provided, comprising a stored computer program, wherein when the computer program is running, the computer program product is controlled to execute any one of the above-mentioned decision engine update methods based on a large language model.

[0133] According to another aspect of an embodiment of the present application, an electronic device is also provided, including: a processor; and a memory for storing executable instructions of the processor; wherein the processor is configured to execute any of the above-mentioned large language model-based decision engine update methods by executing the executable instructions.

[0134] Optionally, Figure 5 This is a structural block diagram of an electronic device according to an embodiment of the present application. Figure 5 As shown, the electronic device may include: one or more ( Figure 5 Only one is shown) processor 502, memory 504, storage controller, and peripheral interface, wherein the peripheral interface is connected to the radio frequency module, audio module and display.

[0135] The processor can call information and application programs stored in the memory through the transmission device to perform the following steps: analyzing a target instruction through a target model to obtain an operation type and an operation parameter corresponding to the target instruction, wherein the target instruction is an instruction described in a natural language for interacting with a decision engine of a financial institution, and the target model is a large language model trained based on a historical instruction set and a historical interception rule set; generating a target code based on the operation type and the operation parameter corresponding to the target instruction, wherein the target code is a script code recognizable by the decision engine; and updating N sub-modules of the decision engine based on the target code in a case where the target code passes a test, wherein N is a positive integer, and the N sub-modules at least include a rule module, a decision module, a scoring module, and a log module.

[0136] The processor can call information and application programs stored in the memory through the transmission device to perform the following steps: data cleaning of a target instruction through a target model to obtain a first instruction, wherein the target instruction is used to remove stop words in the target instruction; word segmentation of the first instruction to obtain P words, wherein P is a positive integer; determination of an operation parameter based on a first keyword in the P words, wherein the first keyword is a numerical type word; determination of an operation type based on a second keyword in the P words and context information corresponding to the target instruction, wherein the second keyword is used to represent action information required to be executed by the target instruction, and the operation type is at least one of the following:

[0137] an added type;

[0138] a deleted type;

[0139] a modified type.

[0140] The processor can call information and application programs stored in the memory through the transmission device to perform the following steps: determining Q target sub-templates based on the operation type corresponding to the target instruction, wherein Q is a positive integer, and the Q target sub-templates are templates stored in a preset database and capable of being presented in a visual form; arranging the Q target sub-templates on a visual operation interface based on a preset process to obtain a target template, wherein the preset process is used to represent a code execution logic of the decision engine; and filling the operation parameter corresponding to the target instruction into the target template to obtain a target code.

[0141] The processor can call information and application programs stored in the memory through the transmission device to perform the following steps: performing a first test operation on the target code, wherein the first test operation is used to detect whether the target code meets the coding specification corresponding to the decision engine; in the case that the first test operation fails, generating first information, wherein the first information is used to prompt the user that the target code has syntax errors; in the case that the first test operation passes, transmitting the target code to the backup engine through a preset interface, wherein the backup engine is a backup of the decision engine of the financial institution.

[0142] The processor can call information and application programs stored in the memory through the transmission device to perform the following steps: updating the N sub-modules corresponding to the backup engine based on the target code; testing the updated backup engine based on preset test data to obtain a test probability, wherein the preset test data is user transaction data of the financial institution in a historical time period, and the test probability is the accuracy rate of the updated backup engine in intercepting user transactions with transaction risks; in the case that the test probability is greater than a historical probability, determining that the target code test passes, wherein the historical probability is the accuracy rate of the backup engine before updating in intercepting user transactions with transaction risks.

[0143] The processor can call information and application programs stored in the memory through the transmission device to perform the following steps: in the case that the target code test passes, parsing the target code to obtain a parsing result, wherein the parsing result is used to represent the dependency relationship and the update order of the N sub-modules; updating the N sub-modules based on the parsing result.

[0144] The processor can call information and application programs stored in the memory through the transmission device to perform the following steps: performing feature extraction on all historical instructions included in the historical instruction set to obtain X first features, wherein each first feature is used to represent the instruction information corresponding to a historical instruction; performing feature extraction on the configuration code of each historical interception rule in the historical interception rule set to obtain X second features, wherein each second feature is used to represent the configuration process corresponding to the configuration code of a configuration type; based on the correspondence between the X first features and the X second features, dividing the X first features and the X second features to obtain a training set and a test set; based on the feature data in the training set, iteratively training an initial large language model to obtain a first model; based on the feature data in the test set, iteratively updating the first model to obtain a target model.

[0145] By adopting the embodiment of the present application, a technical scheme of a user security verification method is provided, the present application adopts a pre-trained large language model (i.e., a target model) to analyze a target instruction described in natural language, and generates a corresponding script code (i.e., target code) of a decision engine through an automatic generation decision engine, thereby achieving the purpose of quickly updating and configuring a sub-module in the decision engine.

[0146] In other words, the present application performs deep analysis on the target instruction based on the target model, can automatically identify the operation type and operation parameters corresponding to the target instruction, and quickly converts the operation type and operation parameters into DSL code (i.e., target code) executable by the decision engine. On the basis that the DSL code passes the test, the present application dynamically updates the sub-module in the decision engine based on the DSL code. Since the target instruction is an instruction written in natural language that can be understood by an operation and maintenance personnel, the technical scheme of the present application reduces the operation complexity of the operation and maintenance personnel of the financial institution in updating the decision engine, thereby achieving the technical effect of improving the efficiency of the interaction between the operation and maintenance personnel and the decision engine, and further solving the technical problem of low interaction efficiency caused by the manual interaction of the operation and maintenance personnel on the decision engine based on expert experience in the prior art.

[0147] Those skilled in the art can understand that, Figure 5 The structure shown is only schematic, and the electronic device can also be a smart phone (such as an Android phone, an iOS phone, etc.), a tablet computer, a palm computer, a Mobile Internet Device (MID), a PAD, or the like. Figure 5 It does not limit the structure of the above-mentioned electronic device. For example, the electronic device can further include more or fewer components (such as a network interface, a display device, etc.) than Figure 5 or have a different configuration than Figure 5 shown.

[0148] Those skilled in the art can understand that all or part of the steps in the various methods of the above-mentioned embodiments can be completed by programs instructing the related hardware of the terminal device, and the programs can be stored in a computer readable storage medium, which can include a flash disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a magnetic disk or an optical disk, etc.

[0149] The above-mentioned serial numbers of the embodiments of the present application are only for description, and do not represent the advantages or disadvantages of the embodiments.

[0150] In the above-mentioned embodiments of the present application, the description of each embodiment has its own emphasis, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0151] In several embodiments provided in the present application, it should be understood that the disclosed technology can be implemented in other manners. For example, the described unit embodiments are merely schematic, and the division of units is merely logical function division, and there can be other division manners in actual implementation. For example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections can be indirect couplings or communication connections through some interfaces, units or modules, and can be in electrical, mechanical or other forms.

[0152] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e., can be located in one place, or can be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiments.

[0153] In addition, each functional unit in the various embodiments of the present application can be integrated into a processing unit, or each unit can exist physically, or two or more units can be integrated into one unit. The integrated unit can be implemented in the form of hardware or in the form of a software functional unit.

[0154] When the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the present application essentially or substantially or all or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The foregoing storage medium includes: a U disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a mobile hard disk, a magnetic disk or an optical disk, and various media that can store program codes.

[0155] The above description is merely the preferred embodiments of the present application, and it should be pointed out that, for those skilled in the art, without departing from the principles of the present application, some improvements and refinements can be made, and these improvements and refinements should be considered as the protection scope of the present application.

Claims

1. A decision engine update method based on a large language model, characterized in that: include: Parsing the target instruction using the target model to obtain the operation type and operation parameters corresponding to the target instruction, wherein the target instruction is an instruction described in natural language for interacting with the decision engine of the financial institution, and the target model is a large language model trained based on a historical instruction set and a historical interception rule set; Generate a target code based on the operation type and operation parameters corresponding to the target instruction, wherein the target code is a script code that can be recognized by the decision engine; If the target code test passes, N sub-modules of the decision engine are updated based on the target code, where N is a positive integer and the N sub-modules include at least a rule module, a decision module, a scoring module, and a log module.

2. The decision engine update method based on a large language model according to claim 1, characterized in that: The target instruction is parsed through the target model to obtain the operation type and operation parameters corresponding to the target instruction, including: Performing data cleaning on the target instruction using a target model to obtain a first instruction, wherein the target instruction is used to remove stop words in the target instruction; Segment the first instruction to obtain P words, where P is a positive integer; Determining the operation parameter based on a first keyword among the P words, wherein the first keyword is a word of a numerical type; The operation type is determined based on a second keyword in the P words and context information corresponding to the target instruction, wherein the second keyword is used to represent action information to be performed by the target instruction, and the operation type is at least one of the following: New types added; Delete type; Modify the type.

3. The decision engine update method based on a large language model according to claim 1, characterized in that: Generating target code based on the operation type and operation parameters corresponding to the target instruction, including: Determine Q target sub-templates based on the operation type corresponding to the target instruction, where Q is a positive integer, and the Q target sub-templates are templates that can be presented in a visual form and stored in a preset database; Arranging the Q target sub-templates in a visual operation interface based on a preset process, and using the arrangement result as the target template, wherein the preset process is used to represent the code execution logic of the decision engine; Fill the target template with the operation parameters corresponding to the target instruction to obtain the target code.

4. The decision engine update method based on a large language model according to claim 1, characterized in that: After generating the target code based on the operation type and operation parameters corresponding to the target instruction, the decision engine update method based on the large language model further includes: Performing a first test operation on the target code, wherein the first test operation is used to detect whether the target code complies with a coding specification corresponding to the decision engine; If the first test operation fails, generating first information, wherein the first information is used to prompt a user that the target code has a syntax error; In a case where the first test operation passes, the target code is transmitted to a backup engine through a preset interface, wherein the backup engine is a backup of the decision engine of the financial institution.

5. The decision engine update method based on a large language model according to claim 4, characterized in that: After the target code is transmitted to the backup engine through a preset interface, the decision engine update method based on the large language model further includes: Updating N submodules corresponding to the backup engine based on the target code; Testing the updated backup engine based on preset test data to obtain a test probability, wherein the preset test data is user transaction data of the financial institution within a historical time period, and the test probability is the accuracy rate of the updated backup engine in intercepting user transactions with transaction risks; In the case where the test probability is greater than the historical probability, it is determined that the target code test has passed, wherein the historical probability is the accuracy rate of the backup engine before the update in intercepting user transactions with transaction risks.

6. The decision engine update method based on a large language model according to claim 1, characterized in that: If the target code passes the test, updating the N submodules of the decision engine based on the target code includes: If the target code passes the test, the target code is parsed to obtain a parsing result, wherein the parsing result is used to characterize the dependency relationship and update order corresponding to the N sub-modules; The N submodules are updated based on the parsing result.

7. The decision engine update method based on a large language model according to claim 1, characterized in that: The training steps of the target model include: Performing feature extraction on all historical instructions included in the historical instruction set to obtain X first features, wherein each first feature is used to characterize instruction information corresponding to a historical instruction; Performing feature extraction on the configuration code of each historical interception rule in the historical interception rule set to obtain X second features, wherein each second feature is used to characterize a configuration process corresponding to a configuration code of a configuration type; Based on the correspondence between the X first features and the X second features, the X first features and the X second features are divided to obtain a training set and a test set; Iteratively training the initial large language model based on the feature data in the training set to obtain a first model; The first model is iteratively updated based on the feature data in the test set to obtain the target model.

8. A decision engine update method based on a large language model, characterized in that: include: a parsing unit, configured to parse a target instruction using a target model to obtain an operation type and operation parameters corresponding to the target instruction, wherein the target instruction is an instruction described in natural language for interacting with a decision engine of a financial institution, and the target model is a large language model trained based on a historical instruction set and a historical interception rule set; a first generating unit, configured to generate a target code based on an operation type and operation parameters corresponding to the target instruction, wherein the target code is a script code recognizable by the decision engine; The first update unit is used to update N sub-modules of the decision engine based on the target code when the target code test passes, wherein N is a positive integer, and the N sub-modules include at least a rule module, a decision module, a scoring module, and a log module.

9. A computer program product, characterized in that The computer program product includes a computer program, wherein when the computer program is run, the computer program product is controlled to execute the decision engine updating method based on a large language model according to any one of claims 1 to 7.

10. An electronic device, characterized in that: The invention comprises one or more processors and a memory, wherein the memory is used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the decision engine update method based on a large language model as described in any one of claims 1 to 7.

Images