A method, device, equipment and medium for generating risk control variables

Abstract

The present invention discloses a method, device, equipment and medium for generating risk control variables, and relates to the field of data processing technology. The method comprises: obtaining user description information and primary associated field information, assembling first prompt information; inputting the first prompt information into a field association derivation model, obtaining related sub-nodes and the association relationship between nodes, and assembling second prompt information; inputting the second prompt information into a calculation process description model, obtaining an indicator calculation process description and a flow chart; assembling third prompt information; inputting the third prompt information into a name specification model, obtaining standardized description information; obtaining a code sample and assembling fourth prompt information; inputting the fourth prompt information into a code generation model, obtaining a code implementation of the risk control variable. The present invention realizes an integrated process from user demand to code generation, improves processing efficiency, and reduces information transmission errors and understanding deviations caused by multi-person collaboration.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method, device, equipment and medium for generating risk control variables. Background Art

[0002] Risk control variables are key factors in assessing user credit risk. The traditional process of generating risk control variables is as follows: business demand personnel propose risk control variable requirements → business designers find relevant fields from interface specifications and deduce logical relationships → software designers write design plans and pseudocodes → developers write code based on the design plans to implement functions. However, this process is a complex and multi-step process involving collaboration among multiple departments and roles, and is highly dependent on manual operations and personnel experience. It is not only inefficient, but also prone to misunderstandings and errors, which affect the final risk control effect, make it difficult to quickly respond to changes in business needs, and increase the operating costs of the enterprise.

[0003] Therefore, it is necessary to provide a new method for generating risk control variables that can solve the above problems. Summary of the invention

[0004] In view of this, an embodiment of the present invention provides a method, device, equipment and medium for generating risk control variables, so as to solve the problem that the traditional risk control variable generation process is highly dependent on manual operation and personnel experience, resulting in low generation efficiency and prone to misunderstanding and errors.

[0005] According to a first aspect, an embodiment of the present invention provides a method for generating a risk control variable, the method comprising:

[0006] Obtaining description information of the user and first-level associated field information based on a preset node structure, and assembling the description information and the first-level associated field information into first prompt information;

[0007] Input the first prompt information as input data into the trained field association deduction model, obtain the relevant sub-nodes and the association relationship between the nodes output by the field association deduction model, and assemble the description information, the first-level association field information, the sub-nodes and the association relationship between the nodes into the second prompt information;

[0008] Input the second prompt information as input data into the trained calculation process description model to obtain the indicator calculation process description and flow chart output by the calculation process description model;

[0009] Assemble the third prompt information according to the preset code format and description information;

[0010] Inputting the third prompt information as input data into the trained name standardization model to obtain the standardized description information output by the name standardization model;

[0011] Obtain a code sample, and assemble the fourth prompt information according to the code sample, the normalized description information, the description of the associated fields, the child nodes, the association relationship between the nodes, the description of the indicator calculation process, and the flow chart;

[0012] The fourth prompt information is input as input data into the trained code generation model, and the code generation model generates code according to the code sample to obtain the code implementation of the risk control variables output by the code generation model, and generates the risk control variables based on the code implementation.

[0013] In combination with the first aspect, in a first implementation of the first aspect, obtaining the description information and the primary associated field information of the user, and assembling the description information and the primary associated field information into the first prompt information specifically includes:

[0014] Get the user's description information and first-level associated field information;

[0015] According to the preset node structure, the first prompt information is assembled with the description information as the user demand and the first-level associated field information as the top-level node in the preset node structure.

[0016] In combination with the first aspect, in the second implementation of the first aspect, the first prompt information is input as input data into the trained field association deduction model to obtain the relevant sub-nodes and the association relationship between the nodes output by the field association deduction model, and the description information, the first-level associated field information, the sub-nodes and the association relationship between the nodes are assembled into the second prompt information, specifically including:

[0017] The first prompt information is input as input data into the field association deduction model, and the field association deduction model performs association between the description information in the first prompt information and the primary association field information.

[0018] Obtain the child nodes output by the field association deduction model and the association relationship between the nodes;

[0019] According to the preset node structure, the nodes and node relationships associated with the description information as an indicator and the primary associated field information, the sub-nodes and the association relationship between the nodes as indicators are assembled to obtain the second prompt information.

[0020] In combination with the first aspect, in a third implementation manner of the first aspect, assembling the third prompt information according to the preset code format and the description information specifically includes:

[0021] Obtaining a preset code format, and establishing a preset name conversion rule according to the preset code format;

[0022] According to the preset name conversion rule and the description information, the third prompt information is assembled.

[0023] In combination with the first aspect, in a fourth implementation of the first aspect, the obtaining of the code sample, assembling the code sample, the normalized description information, the primary associated field information, the child node, the association relationship between the nodes, the indicator calculation process description, and the flowchart into the fourth prompt information, specifically includes:

[0024] Get code samples;

[0025] The fourth prompt information is assembled using the standardized description information as variable information, the first-level associated field information, sub-nodes, and the association relationship between nodes as variable associated nodes, the indicator calculation process description and the flow chart as the variable calculation process, and the code sample as the code style required by the user.

[0026] In combination with the first aspect, in a fifth implementation of the first aspect, before the step of obtaining the user's description information and the primary associated field information based on the preset node structure and assembling the description information and the primary associated field information into the first prompt information, the method further includes:

[0027] Get the preset node structure, and convert the preset node structure into a key-value style according to the hierarchical structure and field description in the preset node structure.

[0028] In combination with the first aspect, in a sixth implementation of the first aspect, the field association deduction model, the calculation process description model, the name specification model and the code generation model are all large language models.

[0029] According to a second aspect, an embodiment of the present invention further provides a device for generating risk control variables, the device comprising:

[0030] A first assembling module is used to obtain the user's description information and the first-level associated field information based on a preset node structure, and assemble the description information and the first-level associated field information into the first prompt information;

[0031] a second assembling module, for inputting the first prompt information as input data into the trained field association deduction model, obtaining the subnodes and the association relationship between the nodes output by the field association deduction model, and assembling the description information, the first-level association field information, the subnodes and the association relationship between the nodes into the second prompt information;

[0032] A process description module, used for inputting the second prompt information as input data into the trained calculation process description model, and obtaining the indicator calculation process description and flow chart output by the calculation process description model;

[0033] A third assembling module, used for assembling the third prompt information according to the preset code format and description information;

[0034] A name standardization module, used for inputting the third prompt information as input data into the trained name standardization model to obtain the standardized description information output by the name standardization model;

[0035] A fourth assembly module is used to obtain a code sample, and assemble the code sample, primary associated field information, associated field description, child nodes, association relationships between nodes, indicator calculation process description, and a flow chart into fourth prompt information;

[0036] The risk control generation module is used to input the fourth prompt information as input data into the trained code generation model, and the code generation model generates code according to the code sample to obtain the code implementation of the risk control variables output by the code generation model, and generate risk control variables based on the code implementation.

[0037] According to a third aspect, an embodiment of the present invention further provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of the method for generating risk control variables as described in any one of the above are implemented.

[0038] According to a fourth aspect, an embodiment of the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the above-described methods for generating risk control variables.

[0039] According to a fifth aspect, an embodiment of the present invention further provides a computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the steps of the method for generating risk control variables as described in any one of the above items are implemented.

[0040] The method, device, equipment and medium for generating risk control variables of the present invention construct the first prompt information through the description information and the first-level associated field information provided by the user, guide the field association deduction model to automatically derive the association relationship and logic between the risk control variables and the relevant fields in the interface data specification, and simplify the field association and logic deduction process by using LLM; by assembling the description information, the first-level associated field information, the sub-nodes and the association relationship between the nodes into the second prompt information, the powerful natural language processing and code generation capabilities of the calculation process description model are used to automatically generate business descriptions, processing procedures and codes, improve the processing efficiency of the risk control variable generation process and reduce the error rate, use LLM to assist in the generation of risk control variables, and for LLM It provides a complete prompt chain, realizing an integrated automated process from user requirements to code generation. This integrated automated process reduces information transmission errors and misunderstandings caused by multi-person collaboration, ensuring that the final implementation is consistent with the initial requirements, so that even people who lack rich experience can efficiently complete the design and implementation of risk control variables with the assistance of the large language model, greatly improving the efficiency of risk control variable generation. Through the unified processing of the name specification model, the generated variable names, logical descriptions, and code implementations all follow the preset specifications and standards, and each model uses its own unified prompts and templates to ensure the consistency and standardization of the output results, avoiding standardization problems caused by human differences. This not only helps the readability and maintainability of the code, but also facilitates team collaboration and project management. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The features and advantages of the present invention will be more clearly understood by referring to the accompanying drawings, which are schematic and should not be construed as limiting the present invention in any way. In the accompanying drawings:

[0042] Figure 1 A schematic diagram showing a flow chart of a method for generating risk control variables provided by the present invention;

[0043] Figure 2 A schematic diagram showing the structure of a device for generating risk control variables provided by the present invention is shown;

[0044] Figure 3 It is a schematic diagram of the hardware structure of the electronic device provided in the embodiment of the present application. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0046] Risk control variables are key factors in assessing user credit risk. The traditional process of generating risk control variables is as follows:

[0047] 1. Business requirements proposal: Business requirements personnel propose specific risk control variable requirements based on business requirements (which can also be risk control strategies).

[0048] For example, if it is necessary to assess a user's credit risk, the business requirements personnel may propose a risk control variable requirement of "the sum of the overdue amounts of credit card accounts in the past 12 months."

[0049] 2. Data field search and association: Business designers need to search for all fields or XML nodes related to risk control variables based on existing interface data specifications (such as the XML structure of credit information). This process specifically includes:

[0050] 2.1. Find relevant fields: Find the fields that may be related to the risk control variables from the huge XML data structure (which may contain hundreds of nodes);

[0051] The XML data structure of some samples is defined as follows:

[0052]

[0053] 2.2. Understand the meaning of fields: Read the interface documentation to understand the meaning, data type, and value range of each field;

[0054] 2.3. Establish logical associations: Derive the logical relationships between these fields and determine how to combine and calculate to obtain the required risk control variables.

[0055] 2.4. Processing multiple records: For list data (repeated nodes in XML), you need to consider the traversal and accumulation logic.

[0056] 3. Write a business process description: After finding the associated fields, the business designer writes a detailed business process description. The business process description specifically includes: data extraction logic: which nodes to extract data from, how to handle missing values ​​or abnormal values; calculation logic: how to summarize, count or calculate data, which conditions or filtering rules to use; boundary case processing: how to handle special cases, such as missing data, format inconsistency, etc.

[0057] 4. Software design plan formulation: Software system designers write software design plans and pseudocodes based on business descriptions. This process specifically includes:

[0058] 4.1. Write pseudocode: convert business logic into executable algorithm steps;

[0059] 4.2. Design system architecture: determine the module division, interface design and data flow of the program;

[0060] 4.3. Write design documents: describe in detail the functions, input and output, and exception handling of each module.

[0061] 5. Code implementation: Software developers use programming languages ​​to implement the calculation logic of risk control variables based on the design plan. The process specifically includes:

[0062] 5.1. Write code: According to the design plan, write code to extract data, calculate and output results.

[0063] 5.2. Debugging and testing: Perform unit testing to ensure the correctness of the code and handle possible exceptions.

[0064] 5.3. Optimize performance: When necessary, the code needs to be optimized to improve operating efficiency.

[0065] 6. System integration and testing: Developers integrate newly developed functions into the existing system. This process specifically includes:

[0066] 6.1. Integration testing: Verify that the new function works properly with other parts of the system.

[0067] 6.2. Performance testing: Ensure that the system performance meets the requirements under the actual data volume.

[0068] 6.3. Problem fixing: Fix the problems and bugs found according to the test results.

[0069] The above process requires the collaboration of personnel from multiple departments, such as business requirements personnel, business designers, system designers and developers. The communication process ensures that personnel from various departments have a consistent understanding of the risk control variable requirements, timely communicates problems encountered during the design and implementation process, and confirms whether the functions finally implemented meet business requirements.

[0070] The traditional risk control variable generation process is a complex and multi-step process involving the collaboration of multiple departments and multiple roles. The process is highly dependent on manual operation and personnel experience, which is not only inefficient, but also prone to misunderstandings and errors, affecting the final risk control effect. This method is difficult to respond quickly to changes in business needs and also increases the operating costs of enterprises.

[0071] Based on the above overall process, the following technical problems exist in the traditional risk control variable generation process:

[0072] 1. Field association and logic deduction are complex and time-consuming: Business designers need to find all fields related to the specified risk control variables from the huge and complex interface data specifications (such as a credit report with 470 XML nodes). This includes processing XML nodes with multiple levels and multiple records and deducing their logical relationships. The process is similar to "looking up information, highlighting key points, and marking relationships", which is very time-consuming and laborious.

[0073] 2. Manual processing is inefficient and prone to errors: The entire process relies heavily on manual operation, which is inefficient. In addition, omissions, misunderstandings or errors are prone to occur during manual search and logical deduction, affecting the accuracy of subsequent work.

[0074] 3. Multi-link collaboration leads to information bias: The generation of risk control variables requires multiple links such as business design, system design, and software development. In the process of information transmission, there may be understanding bias between the links, resulting in the final implementation logic being inconsistent with the initial design, which in turn causes rework and reduces development efficiency.

[0075] 4. High reliance on experience and high training costs: Effective completion of the above tasks requires participants to have rich business and technical experience. For new employees, the training cost is high and it is difficult for them to be competent for related work in the short term.

[0076] 5. Low standardization and high maintenance difficulty: Due to the lack of unified standards and specifications, the definition, naming, and logical implementation of risk control variables may vary from person to person, which will lead to problems such as inconsistent code style, non-standard variable naming, and differentiated logical implementation, increasing the complexity of the system and the difficulty of maintenance. At the same time, in the subsequent functional expansion and optimization, the lack of standardized code and documentation makes collaboration and handover difficult.

[0077] In summary, it is necessary to provide a new method for generating risk control variables that can solve the above problems.

[0078] The following is an explanation of the professional terms mentioned above:

[0079] Risk control variables: They are the input parameters for risk control decisions in the risk control system and are also the basic data for risk control system decision-making. For example, in the financial system, "the sum of overdue amounts of personal consumer loans in the past 24 months", "the sum of overdue amounts of historical mortgage loans for non-borrowers", and "the sum of overdue amounts of credit card accounts in the past 12 months" are all risk control variables. The calculation of these risk control variables requires analyzing, extracting, and calculating the specified variable values ​​from the data source, evaluating the risk coefficient of the specified user, and then deciding whether to grant credit to the user and the credit limit, etc.

[0080] Credit information message: refers to the important reference data retrieved from the credit information system in the risk control process. Credit information data is usually returned in XML format. Each XML node has its strict definition and is provided as a standardized interface document for user reference. Specifically, the credit information message has 24 first-level XML nodes, a 5-layer hierarchical structure, and a total of about 470 XML nodes. Since the data structure of the credit information message is complex and contains rich personal credit information of users, accurate parsing and utilization of this data is crucial for risk assessment and calculation of risk control variables.

[0081] In order to solve the above problems, a method for generating risk control variables is provided in this specification, which aims to use the artificial intelligence capabilities of the large language model (LLM) to automatically complete the field association, logical deduction and code implementation in the process of generating risk control variables, reduce manual participation and understanding deviations in each link, and improve overall efficiency and accuracy. The method for generating risk control variables provided in this specification can be applied to electronic devices with data processing capabilities. The electronic device may include a notebook, a desktop computer, a smart phone, a smart wearable device (virtual reality glasses, smart watches, etc.), a tablet computer, etc. Of course, the method for generating risk control variables provided in this specification can also be applied to applications running in the above-mentioned electronic devices. For example, the method for generating risk control variables can be applied to a browser with data processing capabilities, or it can be applied to a browser with data processing capabilities. Figure 1 FIG. 1 is a flow chart of a method for generating risk control variables according to an embodiment of the present invention. Figure 1 As shown, the method may include the following steps:

[0082] S10, obtaining the user's description information and the first-level associated field information based on the preset node structure, and assembling the description information and the first-level associated field information into a first prompt information (Prompt).

[0083] In this embodiment, the user's description information is the user's description of the risk control variables, and the first-level associated field information is the first-level XML node in the credit report, that is, the top-level XML node. For example, the user's description information is "the sum of the overdue amounts of credit card accounts in the past 12 months", and the user's first-level associated field information is "PDA-loan account information unit".

[0084] Preferably, at least one human-computer interaction method can be provided to the user to allow the user to provide corresponding descriptive information and first-level associated field information, for example, an input text box can be provided to the user to allow the user to input text-type descriptive information and first-level associated field information, and sufficient selection content can be provided to the user to allow the user to select descriptive information and first-level associated field information. It can also be a variety of human-computer interaction methods such as voice input, gesture input, etc.

[0085] It should be noted that the user provides at least one first-level associated field information, and this or these first-level associated field information will be associated using the field inference model to find the child nodes of the first-level associated field information, the child nodes of the child nodes, and the association relationship between these nodes.

[0086] Take the user's description information as "the sum of overdue amounts of credit card accounts in the past 12 months" and the user's first-level associated field information as "PDA-loan account information unit" as an example for explanation. The first prompt information (Prompt) assembled is:

[0087] The following `json` is the structure description of the nodes and sub-nodes of the XML tag {key} {key_2}

[0088] The following `question` is the `indicator calculation` problem that needs to be performed based on the current XML structure

[0089] Please find the node list that needs to participate in the current `question` calculation according to the XML structure description, and return a markdown table containing the fields: xmlPath, xmlTag, data type, element name, and description of the correlation with `indicator calculation`

[0090] ```json

[0091] {xml_content}

[0092] {xml_content_2}

[0093] ```

[0094] ```question

[0095] {input_text}

[0096] ```

[0097] That is, step S10 specifically includes:

[0098] S11. Obtain the user's description information and the first-level associated field information.

[0099] S12. According to the preset node structure, using the description information as the user's requirement and the first-level associated field information as the top-level node in the preset node structure, assemble to obtain the first prompt message. Among them, the first prompt message includes the sub-nodes required to participate in the current user's requirement and the association relationships between the nodes.

[0100] S20. Input the first prompt message as input data into the trained field association derivation model, use the field association derivation model to associate the description information and the first-level associated field information, obtain the relevant sub-nodes and the association relationships between the nodes output by the field association derivation model, and assemble the description information, the first-level associated field information, the sub-nodes and the association relationships between the nodes into the second prompt message.

[0101] In this embodiment, the field association derivation model is a trained large language model. The large language model (LLM) is a large-scale pre-trained model trained using deep learning technology, which can generate coherent and meaningful natural language text. It is based on a neural network structure and learns the latent patterns and rules of language by training on a large amount of text data. The LLM uses a large amount of Internet text data during the training process and can generate high-quality natural language text. It has achieved significant performance improvements in multiple natural language processing tasks and provides strong support for the application of artificial intelligence in the language field.

[0102] The function of this field association derivation model is to derive the relevant sub-nodes and the association relationships between the nodes according to the description information of the risk control variable and the selected first-level associated field information (the first-level XML node). The association relationships include the logical relationships between the sub-nodes and the upper-level nodes, the logical relationships between the sub-nodes and the sub-nodes, etc. That is, the second prompt message includes all the sub-nodes of the first-level associated field information derived according to the description information and the first-level associated field information and the association relationships between the nodes. Among them, the associated sub-nodes can be presented to the user in the form of a node list, and the association relationships between the nodes are more detailed description information of the associated nodes.

[0103] Take the user description information as "the sum of overdue amounts of credit card accounts in the past 12 months" and the user's first-level associated field information as "PDA-loan account information unit" as an example for explanation. The sub-nodes output by the field association deduction model and the association relationship between nodes are as follows:

[0104] | xmlPath | xmlTag | Data type | Element name | Description of correlation with `Indicator calculation` |

[0105] |---------|--------|----------|----------|-----------------------|

[0106] | / PDA / PD01 / PD01C / PD01CJ06 | PD01CJ06 | Long | Current overdue amount | Used to calculate the sum of overdue amounts of credit card accounts in the past 12 months. Especially for R2 account type, this field indicates the unpaid portion of the minimum payment amount of the previous bill. |

[0107] | / PDA / PD01 / PD01D / PD01DH / PD01DD01 | PD01DD01 | Enum | Repayment Status | Used to determine whether there is any overdue situation in each month. For R2 accounts, this field provides the monthly repayment status to help identify the specific month of overdue. |

[0108] | / PDA / PD01 / PD01A / PD01AD01 | PD01AD01 | Enum | Account Type| Used to filter out credit card accounts (R2 or R3) to ensure that calculations are performed only on credit card accounts. |

[0109] These child nodes provide the necessary information to calculate the sum of overdue amounts of credit card accounts in the past 12 months, including confirmation of overdue status, acquisition of overdue amounts, and screening of account types.

[0110] Take the user's description information as "the sum of overdue amounts of credit card accounts in the past 12 months" and the user's primary association field information as "PDA-debit and credit account information unit" as an example, and the assembled second prompt information is:

[0111] The following `index items` are the contents that need to be parsed from the XML data

[0112] The following `Indicator Item and XML Node Relationship` is a list of XML nodes and their relationships associated with `Indicator Item`

[0113] The following `json` is the structure description of the associated XML node

[0114] Based on the above, output the description of the `indicator calculation process`, mark the `xmlPath` of the data during the processing, and output the flow chart in the `plantuml` format. Note: If there is 'occurrence' at any level of the node: '[0..*]' means multiple records, and the value must be traversed

[0115] #### Indicator items

[0116] {input}

[0117] #### Relationship between indicator items and XML nodes

[0118] {tables}

[0119] ```json

[0120] {nodes_info}

[0121] ```

[0122] ```plantuml

[0123] @startuml

[0124] start

[0125] xxx

[0126] @enduml

[0127] ```

[0128] #### Indicator calculation process

[0129] - **Indicator explanation**: xx

[0130] - **Process**:

[0131] - xx

[0132] - xx

[0133] That is, step S20 specifically includes:

[0134] S21. Input the first prompt information as input data into the field association deduction model, and the field association deduction model performs association between the description information in the first prompt information and the primary association field information to obtain the sub-nodes output by the field association deduction model and the association relationship between the nodes.

[0135] S22. According to the preset node structure, the nodes and node relationships associated with the description information as an indicator and the primary associated field information, the sub-nodes and the association relationship between the nodes as indicators are assembled to obtain the second prompt information.

[0136] S30. Input the second prompt information as input data into the trained calculation process description model. The calculation process description model parses the second prompt information and generates a business processing process description to obtain the indicator calculation process description and flowchart output by the calculation process description model.

[0137] In this embodiment, the calculation process description model is also a trained large language model. Since the trained field association derivation model performs the association between nodes during the derivation process, the irrelevant sub-nodes have been filtered out in the second prompt information generated according to the description information and the first-level associated field information. The calculation process description model is used to generate a detailed description and flow chart of the indicator calculation process according to the structural description information of the associated sub-nodes.

[0138] Taking the user description information as "the sum of overdue amounts of credit card accounts in the past 12 months" and the user's first-level associated field information as "PDA-debit and credit account information unit" as an example, the indicator calculation process description and flow chart output by the calculation process description model are as follows:

[0139] #### Indicator calculation process

[0140] - **Indicator explanation**: The sum of overdue amounts of credit card accounts in the past 12 months, used to assess the credit card repayment ability of individuals or companies.

[0141] - **Process**:

[0142] -Parse all debit and credit account information units (` / PDA / PD01`) from the XML data.

[0143] -For each debit and credit account information unit, check if the account type (` / PDA / PD01 / PD01A / PD01AD01`) is a credit card account (R2 or R3).

[0144] - If it is a credit card account, further check the current overdue amount (` / PDA / PD01 / PD01C / PD01CJ06`) in the latest monthly performance information segment (` / PDA / PD01 / PD01C`).

[0145] - At the same time, check the repayment status information (` / PDA / PD01 / PD01D / PD01DH`) in the repayment record information segment (` / PDA / PD01 / PD01D`) of the last 24 months, especially the repayment status (` / PDA / PD01 / PD01D / PD01DH / PD01DD01`) to determine whether there is any overdue situation.

[0146] -Traverse the repayment status of the last 24 months and filter out the data of the last 12 months.

[0147] - For each month with an overdue record, add up the current overdue total to get the sum of the overdue amounts of the credit card accounts in the past 12 months.

[0148] ```plantuml

[0149] @startuml

[0150] start

[0151] : Parse loan account information unit / PDA / PD01;

[0152] while (Does the debit and credit account information unit exist?)

[0153] :Check account type / PDA / PD01 / PD01A / PD01AD01;

[0154] if (account type is R2 or R3?)

[0155] :Get the current overdue total amount / PDA / PD01 / PD01C / PD01CJ06;

[0156] : Parse the repayment record information segment / PDA / PD01 / PD01D for the last 24 months;

[0157] while (Does repayment status exist?)

[0158] :Check repayment status / PDA / PD01 / PD01D / PD01DH / PD01DD01;

[0159] if (repayment status indicates overdue?)

[0160] : Add up the current overdue amount;

[0161] endif

[0162] :Move to the next repayment status information; 3.Generated code

[0163] endwhile

[0164] endif

[0165] : Move to the next debit and credit account information unit;

[0166] endwhile

[0167] : Calculate the sum of overdue amounts of credit card accounts in the past 12 months;

[0168] stop

[0169] @enduml

[0170] ```

[0171] S40, assembling the third prompt information according to the preset code format and the description information. In this embodiment, the purpose of assembling the third prompt information is to enable the subsequent code generation model to generate a standardized English name for the current risk control variable, that is, to generate a corresponding standardized English name from the standardized Chinese name in the description information, so as to facilitate code implementation.

[0172] The third prompt information obtained by assembling may be:

[0173] ### Java function naming convention

[0174] {naming_example}

[0175] ### question

[0176] {user_input}

[0177] The above is a series of Java function standard Chinese names and their corresponding standard English names.

[0178] The canonical Chinese name of a function generates its corresponding canonical English name.

[0179] ### Output format

[0180] ```Standard English name```

[0181] That is, step S40 specifically includes:

[0182] S41, obtaining a preset code format, and establishing a preset name conversion rule according to the preset code format. The preset name conversion rule is the above-mentioned conversion of the standard Chinese name into the corresponding standard English name.

[0183] S42: Assemble and obtain third prompt information according to the preset name conversion rule and description information.

[0184] S50: Input the third prompt information as input data into the trained name standardization model, and convert the name in the description information by the name standardization model to obtain the standardized description information output by the name standardization model. It can be understood that the standardized description information is text information that complies with the preset code format.

[0185] In this embodiment, the name standardization model is also a trained large language model.

[0186] S60, obtaining a code sample, and assembling fourth prompt information according to the code sample, the normalized description information, the primary associated field information, the child nodes, the association relationship between the nodes, the indicator calculation process description, and the flow chart.

[0187] Taking the user description information as "the sum of overdue amounts of credit card accounts in the past 12 months" and the user's primary association field information as "PDA-debit and credit account information unit" as an example, the assembled fourth prompt information is:

[0188] Refer to the code sample to complete the complete code for the following functions:

[0189] Note: Need to `import util.XmlUtil;`, the current environment contains

[0190] #### Code Example

[0191] {code_sample}

[0192] #### Variable ID

[0193] {variableId}

[0194] #### Variable name

[0195] {variableName}

[0196] #### Associate XML nodes

[0197] {xmlTag}

[0198] #### Associated XML node structure description

[0199] {nodes_info}

[0200] #### Calculation process

[0201] {input_text}

[0202] That is, step S60 specifically includes:

[0203] S61. Obtain code samples.

[0204] S62, using the standardized description information as variable information, the associated field description, sub-nodes, and the relationship between nodes as variable associated nodes, the indicator calculation process description and the flowchart as the variable calculation process, and the code sample as the code style required by the user, assemble to obtain the fourth prompt information.

[0205] S70. Input the fourth prompt information as input data into the trained code generation model, and the code generation model generates code according to the code sample to obtain the code implementation of the risk control variables output by the code generation model. Once the code implementation of the risk control variables is obtained, the required risk control variables can be generated based on the code implementation.

[0206] In this embodiment, the code generation model is also a trained large language model. The code generation model is used to generate complete code based on the provided variable information (including standardized description information, associated field description, sub-nodes, association relationships between nodes, indicator calculation process description and flow chart) and code samples. The complete code is the generated risk control variable, which can be integrated into the existing risk control system later.

[0207] It should be noted that the above-mentioned field association inference model, calculation process description model, name specification model and code generation module are all large language models, which can be four independent LLMs. Considering the reasoning quality, multiple models can also jointly select one LLM. For example, the above-mentioned four models select an LLM with sufficiently high reasoning quality.

[0208] In particular, the credit data involved in this application, such as descriptive information, first-level associated field information, preset node structure, preset name conversion rules, code samples, etc., have been obtained with full consent and authorization, and the collection, use and processing of relevant information must comply with relevant laws, regulations and standards of relevant countries and regions.

[0209] The method for generating risk control variables of the present invention constructs the first prompt information through the description information and the first-level associated field information provided by the user, guides the field association deduction model to automatically derive the association relationship and logic between the risk control variables and the relevant fields in the interface data specification, and uses LLM to simplify the field association and logic deduction process; by assembling the description information, the first-level associated field information, the sub-nodes and the association relationship between the nodes into the second prompt information, the powerful natural language processing and code generation capabilities of the calculation process description model are used to automatically generate business descriptions, processing procedures and codes, improve the processing efficiency of the risk control variable generation process and reduce the error rate, use LLM to assist in the generation of risk control variables, and provide a complete The prompt chain realizes an integrated automated process from user needs to code generation. This integrated automated process reduces information transmission errors and misunderstandings caused by multi-person collaboration, ensures that the final implementation is consistent with the initial requirements, and enables even those who lack rich experience to efficiently complete the design and implementation of risk control variables with the assistance of the large language model, greatly improving the efficiency of risk control variable generation. Through the unified processing of the name specification model, the generated variable names, logical descriptions, and code implementations all follow the preset specifications and standards, and each model uses its own unified prompts and templates to ensure the consistency and standardization of the output results, avoiding standardization problems caused by human differences. This not only helps the readability and maintainability of the code, but also facilitates team collaboration and project management.

[0210] Considering that the first large language model, that is, the field association deduction model, is convenient for node association, this method also needs to be prepared. Specifically, the preset node structure will be obtained from the external interface, and according to the hierarchical structure (XML node structure) and field description in the preset node structure, the preset node structure will be converted into a key-value structure and stored in a file or database. This ensures that the external interface required for risk control variable generation is relatively clear, the meaning of each field is clear, and the data structure is clear.

[0211] The following describes a device for generating risk control variables provided by an embodiment of the present invention. The device for generating risk control variables described below and the method for generating risk control variables described above can refer to each other.

[0212] In order to solve the above problems, a device for generating risk control variables is provided in this specification, aiming to provide a database performance optimization solution that is efficient, low-cost, highly automated and has good scalability. Figure 2 is a schematic diagram of the structure of a device for generating risk control variables according to an embodiment of the present invention, such as Figure 2 As shown, the device may include:

[0213] The first assembling module 10 is used to obtain the user's description information and the first-level associated field information based on the preset node structure, and assemble the description information and the first-level associated field information into the first prompt information (Prompt).

[0214] In this embodiment, the user's description information is the user's description of the risk control variables, and the first-level associated field information is the first-level XML node in the credit report, that is, the top-level XML node. For example, the user's description information is "the sum of the overdue amounts of credit card accounts in the past 12 months", and the user's first-level associated field information is "PDA-loan account information unit".

[0215] Preferably, at least one human-computer interaction method can be provided to the user to allow the user to provide corresponding descriptive information and first-level associated field information, for example, an input text box can be provided to the user to allow the user to input text-type descriptive information and first-level associated field information, and sufficient selection content can be provided to the user to allow the user to select descriptive information and first-level associated field information. It can also be a variety of human-computer interaction methods such as voice input, gesture input, etc.

[0216] It should be noted that the user provides at least one first-level associated field information, and this or these first-level associated field information will be associated using the field inference model to find the child nodes of the first-level associated field information, the child nodes of the child nodes, and the association relationship between these nodes.

[0217] The second assembly module 20 is used to input the first prompt information as input data into the trained field association inference model, use the field association inference model to associate the description information and the first-level associated field information, obtain the relevant sub-nodes and the association relationship between nodes output by the field association inference model, and assemble the description information, the first-level associated field information, the sub-nodes and the association relationship between nodes into the second prompt information.

[0218] In this embodiment, the field association derivation model is a trained large language model. LLM is a large-scale pre-trained model trained using deep learning technology, which can generate coherent and meaningful natural language text. It is based on a neural network structure and learns the potential patterns and laws of language by training on a large amount of text data. LLM uses a large amount of Internet text data during training and can generate high-quality natural language text. It has achieved significant performance improvements in multiple natural language processing tasks, providing strong support for the application of artificial intelligence in the field of language.

[0219] The function of the field association derivation model is to derive the related child nodes and the association relationship between nodes based on the description information of the risk control variable and the selected first-level association field information (first-level XML node). The association relationship includes the logical relationship between the child node and the upper-level node, the logical relationship between the child node and the child node, etc., that is, the second prompt information includes all the child nodes of the first-level association field information derived from the description information and the first-level association field information and the association relationship between nodes. Among them, the associated child nodes can be presented to the user in the form of a node list, and the association relationship between nodes is a more detailed description information of the associated nodes.

[0220] The process description module 30 is used to input the second prompt information as input data into the trained calculation process description model. The calculation process description model parses the second prompt information and generates a business processing process description to obtain the indicator calculation process description and flowchart output by the calculation process description model.

[0221] In this embodiment, the calculation process description model is also a trained large language model. Since the trained field association derivation model performs the association between nodes during the derivation process, the irrelevant sub-nodes have been filtered out in the second prompt information generated according to the description information and the first-level associated field information. The calculation process description model is used to generate a detailed description and flow chart of the indicator calculation process according to the structural description information of the associated sub-nodes.

[0222] The third assembly module 40 is used to assemble the third prompt information according to the preset code format and description information. In this embodiment, the purpose of assembling the third prompt information is to enable the subsequent code generation model to generate a standardized English name for the current risk control variable, that is, to generate the corresponding standardized English name from the standardized Chinese name in the description information, so as to facilitate code implementation.

[0223] The name standardization module 50 is used to input the third prompt information as input data into the trained name standardization model, and the name standardization model converts the name in the description information to obtain the standardized description information output by the name standardization model. It can be understood that the standardized description information is text information that complies with the preset code format.

[0224] In this embodiment, the name standardization model is also a trained large language model.

[0225] The fourth assembly module 60 is used to obtain code samples, and assemble the fourth prompt information according to the code samples, normalized description information, primary associated field information, sub-nodes, association relationships between nodes, indicator calculation process description and flow chart.

[0226] The risk control generation module 70 is used to input the fourth prompt information as input data into the trained code generation model, and the code generation model generates code according to the code sample to obtain the code implementation of the risk control variables output by the code generation model. Once the code implementation of the risk control variables is obtained, the required risk control variables can be generated based on the code implementation.

[0227] In this embodiment, the code generation model is also a trained large language model. The code generation model is used to generate complete code based on the provided variable information (including standardized description information, associated field description, sub-nodes, association relationships between nodes, indicator calculation process description and flow chart) and code samples. The complete code is the generated risk control variable, which can be integrated into the existing risk control system later.

[0228] In particular, the credit data involved in this application, such as descriptive information, first-level associated field information, preset node structure, preset name conversion rules, code samples, etc., have been obtained with full consent and authorization, and the collection, use and processing of relevant information must comply with relevant laws, regulations and standards of relevant countries and regions.

[0229] The risk control variable generation device of the present invention constructs the first prompt information through the description information and the first-level associated field information provided by the user, guides the field association deduction model to automatically derive the association relationship and logic between the risk control variable and the relevant fields in the interface data specification, and uses LLM to simplify the field association and logic deduction process; by assembling the description information, the first-level associated field information, the sub-nodes and the association relationship between the nodes into the second prompt information, the powerful natural language processing and code generation capabilities of the calculation process description model are used to automatically generate business descriptions, processing procedures and codes, thereby improving the processing efficiency of the risk control variable generation process and reducing the error rate, using LLM to assist in the generation of risk control variables, and providing a complete The prompt chain realizes an integrated automated process from user needs to code generation. This integrated automated process reduces information transmission errors and misunderstandings caused by multi-person collaboration, ensures that the final implementation is consistent with the initial requirements, and enables even those who lack rich experience to efficiently complete the design and implementation of risk control variables with the assistance of the large language model, greatly improving the efficiency of risk control variable generation. Through the unified processing of the name specification model, the generated variable names, logical descriptions, and code implementations all follow the preset specifications and standards, and each model uses its own unified prompts and templates to ensure the consistency and standardization of the output results, avoiding standardization problems caused by human differences. This not only helps the readability and maintainability of the code, but also facilitates team collaboration and project management.

[0230] Figure 3An example of a physical structure diagram of an electronic device is shown in FIG. Figure 3 As shown, the electronic device may include: a processor 310 (processor), a communication interface 320 (Communications Interface), a memory 330 (memory) and a communication bus 340, wherein the processor 310, the communication interface 320, and the memory 330 communicate with each other through the communication bus 340. The processor 310 may call the logic command in the memory 330 to execute the method for generating risk control variables, which method includes:

[0231] Obtaining description information of the user and first-level associated field information based on a preset node structure, and assembling the description information and the first-level associated field information into first prompt information;

[0232] Input the first prompt information as input data into the trained field association deduction model, obtain the relevant sub-nodes and the association relationship between the nodes output by the field association deduction model, and assemble the description information, the first-level association field information, the sub-nodes and the association relationship between the nodes into the second prompt information;

[0233] Input the second prompt information as input data into the trained calculation process description model to obtain the indicator calculation process description and flow chart output by the calculation process description model;

[0234] Assemble the third prompt information according to the preset code format and description information;

[0235] Inputting the third prompt information as input data into the trained name standardization model to obtain the standardized description information output by the name standardization model;

[0236] Obtain a code sample, and assemble the code sample, normalized description information, primary associated field information, child nodes, association relationships between nodes, indicator calculation process description, and a flow chart into fourth prompt information;

[0237] The fourth prompt information is input as input data into the trained code generation model to obtain the code implementation of the risk control variables output by the code generation model, and the risk control variables are generated based on the code implementation.

[0238] In addition, the logic instructions in the above-mentioned memory 330 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memor), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0239] On the other hand, the present invention further provides a computer program product, the computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions, when the program instructions are executed by a computer, the computer can execute the method for generating risk control variables provided by the above methods, the method comprising:

[0240] Obtaining description information of the user and first-level associated field information based on a preset node structure, and assembling the description information and the first-level associated field information into first prompt information;

[0241] Input the first prompt information as input data into the trained field association deduction model, obtain the relevant sub-nodes and the association relationship between the nodes output by the field association deduction model, and assemble the description information, the first-level association field information, the sub-nodes and the association relationship between the nodes into the second prompt information;

[0242] Input the second prompt information as input data into the trained calculation process description model to obtain the indicator calculation process description and flow chart output by the calculation process description model;

[0243] Assemble the third prompt information according to the preset code format and description information;

[0244] Inputting the third prompt information as input data into the trained name standardization model to obtain the standardized description information output by the name standardization model;

[0245] Obtain a code sample, and assemble the code sample, normalized description information, primary associated field information, child nodes, association relationships between nodes, indicator calculation process description, and a flow chart into fourth prompt information;

[0246] The fourth prompt information is input as input data into the trained code generation model to obtain the code implementation of the risk control variables output by the code generation model, and the risk control variables are generated based on the code implementation.

[0247] In another aspect, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, which is implemented when the computer program is executed by a processor to perform the above-mentioned methods for generating risk control variables, the method comprising:

[0248] Obtaining description information of the user and first-level associated field information based on a preset node structure, and assembling the description information and the first-level associated field information into first prompt information;

[0249] Input the first prompt information as input data into the trained field association deduction model, obtain the relevant sub-nodes and the association relationship between the nodes output by the field association deduction model, and assemble the description information, the first-level association field information, the sub-nodes and the association relationship between the nodes into the second prompt information;

[0250] Input the second prompt information as input data into the trained calculation process description model to obtain the indicator calculation process description and flow chart output by the calculation process description model;

[0251] Assemble the third prompt information according to the preset code format and description information;

[0252] Inputting the third prompt information as input data into the trained name standardization model to obtain the standardized description information output by the name standardization model;

[0253] Obtain a code sample, and assemble the code sample, normalized description information, primary associated field information, child nodes, association relationships between nodes, indicator calculation process description, and a flow chart into fourth prompt information;

[0254] The fourth prompt information is input as input data into the trained code generation model to obtain the code implementation of the risk control variables output by the code generation model, and the risk control variables are generated based on the code implementation.

[0255] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0256] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0257] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for generating risk control variables, characterized in that: The method comprises: Obtaining description information of the user and first-level associated field information based on a preset node structure, and assembling the description information and the first-level associated field information into first prompt information; Input the first prompt information as input data into the trained field association deduction model, obtain the relevant sub-nodes and the association relationship between the nodes output by the field association deduction model, and assemble the description information, the first-level association field information, the sub-nodes and the association relationship between the nodes into the second prompt information; Input the second prompt information as input data into the trained calculation process description model to obtain the indicator calculation process description and flow chart output by the calculation process description model; Assemble the third prompt information according to the preset code format and description information; Inputting the third prompt information as input data into the trained name standardization model to obtain the standardized description information output by the name standardization model; Obtain a code sample, and assemble the code sample, normalized description information, primary associated field information, child nodes, association relationships between nodes, indicator calculation process description, and a flow chart into fourth prompt information; The fourth prompt information is input as input data into the trained code generation model to obtain the code implementation of the risk control variables output by the code generation model, and the risk control variables are generated based on the code implementation.

2. The method for generating risk control variables according to claim 1, characterized in that: The obtaining of the user's description information and primary associated field information, and assembling the description information and the primary associated field information into the first prompt information specifically includes: Get the user's description information and first-level associated field information; According to the preset node structure, the first prompt information is assembled with the description information as the user demand and the first-level associated field information as the top-level node in the preset node structure.

3. The method for generating risk control variables according to claim 1, characterized in that: The first prompt information is input as input data into the trained field association deduction model to obtain the relevant sub-nodes and the association relationship between the nodes output by the field association deduction model, and the description information, the first-level association field information, the sub-nodes and the association relationship between the nodes are assembled into the second prompt information, specifically including: The first prompt information is input as input data into the field association deduction model, and the field association deduction model performs association between the description information in the first prompt information and the primary association field information. Obtain the child nodes output by the field association deduction model and the association relationship between the nodes; According to the preset node structure, the nodes and node relationships associated with the description information as an indicator and the primary associated field information, the sub-nodes and the association relationship between the nodes as indicators are assembled to obtain the second prompt information.

4. The method for generating risk control variables according to claim 1, characterized in that: The third prompt information is assembled according to the preset code format and the description information, specifically including: Obtaining a preset code format, and establishing a preset name conversion rule according to the preset code format; According to the preset name conversion rule and the description information, the third prompt information is assembled.

5. The method for generating risk control variables according to claim 1, characterized in that: The obtaining of the code sample, assembling the code sample, the normalized description information, the primary associated field information, the child nodes, the association relationship between the nodes, the indicator calculation process description and the flow chart into the fourth prompt information, specifically includes: Get code samples; The fourth prompt information is assembled using the standardized description information as variable information, the first-level associated field information, sub-nodes, and the association relationship between nodes as variable associated nodes, the indicator calculation process description and the flow chart as the variable calculation process, and the code sample as the code style required by the user.

6. The method for generating risk control variables according to claim 1, characterized in that: Before the step of obtaining the user's description information and the primary associated field information based on the preset node structure and assembling the description information and the primary associated field information into the first prompt information, the method further includes: Get the preset node structure, and convert the preset node structure into a key-value style according to the hierarchical structure and field description in the preset node structure.

7. The method for generating risk control variables according to claim 1, characterized in that: The field association deduction model, the calculation process description model, the name specification model and the code generation model are all large language models.

8. A device for generating risk control variables, characterized in that: The device comprises: A first assembling module is used to obtain the user's description information and the first-level associated field information based on a preset node structure, and assemble the description information and the first-level associated field information into the first prompt information; a second assembling module, for inputting the first prompt information as input data into the trained field association deduction model, obtaining the subnodes and the association relationship between the nodes output by the field association deduction model, and assembling the description information, the primary association field information, the subnodes and the association relationship between the nodes into the second prompt information; A process description module, used for inputting the second prompt information as input data into the trained calculation process description model, and obtaining the indicator calculation process description and flow chart output by the calculation process description model; A third assembling module, used for assembling the third prompt information according to the preset code format and description information; A name standardization module, used for inputting the third prompt information as input data into the trained name standardization model to obtain the standardized description information output by the name standardization model; A fourth assembly module is used to obtain a code sample, and assemble the code sample, the normalized description information, the primary associated field information, the child nodes, the association relationship between the nodes, the indicator calculation process description, and the flow chart into fourth prompt information; The risk control generation module is used to input the fourth prompt information as input data into the trained code generation model, obtain the code implementation of the risk control variables output by the code generation model, and generate the risk control variables based on the code implementation.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the method for generating risk control variables according to any one of claims 1 to 7 are implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for generating risk control variables according to any one of claims 1 to 7 are implemented.

Images