Strategy generation model training method and device, strategy generation method and device, and terminal

Abstract

The embodiment of the invention discloses a strategy generation model training method and device, a strategy generation method and device and a terminal, and the method comprises the steps: adding a risk knowledge label for a sample risk supervision corpus according to professional experience, and obtaining a standard training corpus; constructing at least one strategy generation model based on the at least one basic large model, inputting a standard training corpus into each strategy generation model, and training each strategy generation model; and in the training process of each strategy generation model, each strategy generation model is controlled to output a plurality of risk control strategies for the standard training corpus, and parameters of each strategy generation model are adjusted according to the plurality of risk control strategies and the risk knowledge labels until each strategy generation model converges. Through the training corpus marked based on professional experience, the strategy generation model constructed based on the basic large model is trained, so that the strategy generation model can learn professional knowledge and experience and is used for generating the strategy, and the accuracy and robustness of strategy generation are improved through the self-learning ability of the large model.

Description

Technical Field

[0001] The embodiments of this specification relate to the field of computer technology, and in particular to a strategy generation model training and strategy generation method, device, and terminal. Background Art

[0002] Rule- or template-based risk control strategy formulation has long played a crucial role in various risk control scenarios, such as finance and information security. This approach relies heavily on a carefully pre-designed and defined set of rules and standardized templates. These rules and templates typically cover a wide range of risk types and compliance requirements. However, in a rapidly evolving regulatory environment and complex risk scenarios, these approaches are increasingly facing challenges such as high maintenance costs, poor adaptability, insufficient flexibility, and limited identification and response capabilities. Summary of the Invention

[0003] The embodiments of this specification provide a strategy generation model training and strategy generation method, device, and terminal, which can solve the technical problem in related technologies that it is difficult to quickly and flexibly generate risk control strategies.

[0004] In a first aspect, an embodiment of this specification provides a strategy generation model training method, the method comprising: Based on professional experience, risk knowledge annotations are added to the sample risk supervision corpus to obtain standard training corpus; Construct at least one strategy generation model based on at least one basic large model, input the above-mentioned standard training corpus into each strategy generation model, and train each strategy generation model; During the training process of each strategy generation model, each strategy generation model is controlled to output multiple risk control strategies for the above-mentioned standard training corpus, and the parameters of each strategy generation model are adjusted according to the above-mentioned multiple risk control strategies and the above-mentioned risk knowledge annotations until each strategy generation model converges.

[0005] In one possible implementation, the above-mentioned control of each strategy generation model to output multiple risk control strategies for the above-mentioned standard training corpus includes: using at least one prompt to instruct each strategy generation model to output multiple risk control strategies for the above-mentioned standard training corpus; the above-mentioned prompt is used to specify the recognition purpose and output rules of each strategy generation model.

[0006] In a possible implementation, the above-mentioned adjusting the parameters of each strategy generation model according to the above-mentioned multiple risk control strategies and the above-mentioned risk knowledge annotations until each strategy generation model converges includes: selecting at least one candidate risk control strategy that meets the preset validity conditions from the above-mentioned multiple risk control strategies; adjusting the parameters of each strategy generation model according to each candidate risk control strategy and the above-mentioned risk knowledge annotations until each strategy generation model converges.

[0007] In a possible implementation, the above-mentioned selecting at least one candidate risk control strategy that meets the preset validity conditions from the above-mentioned multiple risk control strategies includes: based on a self-service voting mechanism, mixing the above-mentioned multiple risk control strategies and selecting at least one candidate risk control strategy that meets the preset validity conditions from the above-mentioned multiple risk control strategies according to preset validity screening rules.

[0008] In a possible implementation, the above-mentioned adjustment of the parameters of each strategy generation model according to each candidate risk control strategy and the above-mentioned risk knowledge annotation until each strategy generation model converges includes: calculating the semantic similarity between each candidate risk control strategy and the above-mentioned risk knowledge annotation, feeding back the semantic similarity corresponding to each candidate risk control strategy to the strategy generation model corresponding to each risk control strategy, and controlling each strategy generation model to adjust the parameters based on the received semantic similarity until each strategy generation model converges.

[0009] In a possible implementation, the above-mentioned adding risk knowledge annotations to the sample risk supervision corpus based on professional experience to obtain standard training corpus includes: adding risk knowledge annotations of at least one risk dimension to the sample risk supervision corpus based on professional experience to obtain standard training corpus; the above-mentioned risk dimensions include at least risk type, risk subcategory, risk control strategy, relevant data indicators, discrimination rules and corpus reference templates.

[0010] In a possible implementation, the method further includes: collecting a plurality of risk supervision corpora in a network environment based on a preset crawler program, and selecting at least one risk supervision corpus as a sample risk supervision corpus.

[0011] In a possible implementation, after the above-mentioned risk supervision corpus is collected in the network environment based on a preset crawler program, it also includes: identifying at least one title in each risk supervision corpus, generating directory information of each risk supervision corpus based on the title of each risk supervision corpus; and storing the directory information of each risk supervision corpus in a structured database.

[0012] In a second aspect, an embodiment of this specification provides a policy generation method, the method comprising: Obtain target risk supervision corpus, input the target risk supervision corpus into at least one converged strategy generation model, and obtain multiple risk control strategies output by each strategy generation model; Perform supervision in risk scenarios based on the above-mentioned multiple risk control strategies; The converged strategy generation model is obtained by training based on the strategy generation model training method in the above embodiment.

[0013] In one possible implementation, the above-mentioned supervision in the risk scenario based on the above-mentioned multiple risk control strategies includes: judging the applicability of each candidate risk control strategy, determining at least one target risk control strategy that meets the preset applicability conditions; and performing supervision in the risk scenario based on each target risk control strategy.

[0014] In one possible implementation, the above-mentioned applicability judgment of each candidate risk control strategy to determine at least one target risk control strategy that meets the preset applicability conditions includes: inputting each candidate risk control strategy into a target deep neural network, determining the applicability score output by the above-mentioned target deep neural network for each candidate risk control strategy based on a text classification algorithm; and determining at least one target risk control strategy that meets the preset applicability conditions based on the applicability score corresponding to each candidate risk control strategy.

[0015] In one possible implementation, the applicability score is used to quantitatively represent at least one of the following information: the content quality of each candidate risk control strategy, its relevance to the target risk regulatory corpus, and its actual operational value.

[0016] In one possible implementation, the above-mentioned supervision in the risk scenario based on the above-mentioned multiple risk control strategies includes: converting the policy text of each risk control strategy into an executable monitoring indicator; connecting the above-mentioned monitoring indicators to the risk scenario, and supervising the risks in the above-mentioned risk scenario based on the above-mentioned monitoring indicators.

[0017] In a possible implementation, the obtaining of the target risk supervision corpus includes: obtaining the target risk supervision corpus from a structured database, where the structured database is used to store directory information of at least one risk supervision corpus.

[0018] In a third aspect, an embodiment of this specification provides a strategy generation model training device, which includes: The corpus annotation module is used to add risk knowledge annotations to the sample risk supervision corpus based on professional experience to obtain standard training corpus; A sample input module is used to construct at least one strategy generation model based on at least one basic large model, input the above-mentioned standard training corpus into each strategy generation model, and train each strategy generation model; The model training module is used to control each strategy generation model to output multiple risk control strategies for the above-mentioned standard training corpus during the training process of each strategy generation model, and adjust the parameters of each strategy generation model according to the above-mentioned multiple risk control strategies and the above-mentioned risk knowledge annotations until each strategy generation model converges.

[0019] In a fourth aspect, an embodiment of this specification provides a policy generation device, the device comprising: A corpus acquisition module is used to obtain target risk supervision corpus, input the target risk supervision corpus into at least one converged strategy generation model, and obtain multiple candidate risk control strategies output by each strategy generation model; Strategy control module, used to perform supervision in risk scenarios based on the above multiple risk control strategies; The converged strategy generation model is obtained by training based on the strategy generation model training method in the above embodiment.

[0020] In a fifth aspect, an embodiment of this specification provides a computer program product comprising instructions, which, when executed on a computer or a processor, enables the computer or the processor to execute the method of the first or second aspect.

[0021] In a sixth aspect, an embodiment of this specification provides a computer storage medium, wherein the computer storage medium stores a plurality of instructions, and the instructions are suitable for being loaded by a processor and executing the method of the first aspect or the second aspect.

[0022] In a seventh aspect, an embodiment of this specification provides a terminal, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is suitable for being loaded by the processor and executing the method of the first or second aspect.

[0023] The beneficial effects of the technical solutions provided by some embodiments of this specification include at least: The embodiments of this specification provide a method for training a strategy generation model, which adds risk knowledge annotations to the sample risk supervision corpus based on professional experience to obtain standard training corpus; constructs at least one strategy generation model based on at least one basic large model, inputs the standard training corpus into each strategy generation model, and trains each strategy generation model; during the training process of each strategy generation model, controls each strategy generation model to output multiple risk control strategies for the standard training corpus, and adjusts the parameters of each strategy generation model based on multiple risk control strategies and risk knowledge annotations until each strategy generation model converges. By training the strategy generation model constructed based on the basic large model with training corpus annotated based on professional experience, the strategy generation model can learn professional knowledge and experience and use it to generate strategies. The self-learning ability of the large model improves the accuracy and robustness of subsequent strategy generation, and also significantly reduces the maintenance cost of the strategy generation process. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the embodiments of this specification or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of this specification. For those skilled in the art, other drawings can be obtained based on these drawings without paying any creative work.

[0025] Figure 1 An exemplary system architecture diagram of a strategy generation model training method provided in an embodiment of this specification; Figure 2 A flow chart of a strategy generation model training method provided in an embodiment of this specification; Figure 3 A flow chart of a strategy generation model training method provided in an embodiment of this specification; Figure 4 A flow chart of a strategy generation method provided in an embodiment of this specification; Figure 5 A flow chart of a strategy generation method provided in an embodiment of this specification; Figure 6 A structural block diagram of a strategy generation model training device provided in an embodiment of this specification; Figure 7 A structural block diagram of a strategy generation device provided in an embodiment of this specification; Figure 8 A schematic diagram of the structure of a terminal provided in an embodiment of this specification. DETAILED DESCRIPTION

[0026] To make the features and advantages of the embodiments of this specification more obvious and easy to understand, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the embodiments described are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without making creative efforts shall fall within the scope of protection of the embodiments of this specification.

[0027] When the following description refers to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the embodiments of this specification. Instead, they are merely examples of devices and methods consistent with some aspects of the embodiments of this specification, as detailed in the appended claims. And in the description of the embodiments of this specification, unless otherwise indicated, " / " means or, for example, A / B can mean A or B: "and / or" in the text is only a way to describe the association relationship of associated objects, indicating that there can be three relationships, for example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, in the description of the embodiments of this specification, "multiple" refers to two or more than two.

[0028] In the following, the terms "first" and "second" are used for descriptive purposes only and should not be understood to imply or suggest relative importance or implicitly indicate the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features.

[0029] Currently, risk monitoring and control in risk management scenarios such as finance, information security, and privacy protection typically relies on rules or templates to generate risk control strategies. The core of this strategy development approach lies in its heavy reliance on a set of carefully pre-designed and defined rules and standardized templates. These rules and templates typically cover a wide range of risk types and compliance requirements, including but not limited to customer identity verification processes and indicators of suspicious activity monitoring. In practice, professionals adapt and apply pre-defined rules and templates designed to ensure the normal and correct conduct of transactions based on specific transaction scenarios, thereby automatically generating or adjusting targeted risk control strategies.

[0030] From an operational perspective, the advantages of this approach lie in its intuitiveness and predictability. With clear rules and standardized templates, the risk control strategy development process is relatively straightforward, easy to understand, and easy to implement. It also provides a relatively stable framework to ensure that transactions adhere to established compliance requirements and risk control standards.

[0031] However, with the rapid evolution of current risk control requirements, the limitations of this traditional approach are becoming increasingly apparent. First, high maintenance costs are a significant issue. With the constant adjustment of regulatory rules and the increasing diversification of risk types, the development of risk control strategies requires constant updating and optimization of risk control rules and templates, which undoubtedly increases the maintenance workload. Second, rule-based or template-based approaches often struggle to adapt to complex and changing risk scenarios. For example, in information security protection scenarios, new risk types and malicious attack methods are constantly emerging, and traditional rules and templates often fail to respond quickly to these emerging threats.

[0032] Furthermore, lack of flexibility is a major drawback of this approach. Different transaction scenarios and customer groups often require more targeted risk control strategies. For example, in information security protection scenarios, flexible risk control strategies are required for different types of user operations and users with varying permissions. However, rule-based or template-based approaches often lack the flexibility to meet these individual needs. Finally, while this approach can identify and address certain risks to a certain extent, its identification capabilities and effectiveness are relatively limited. It may not provide sufficient insight and accuracy when faced with complex or hidden risk behaviors.

[0033] Therefore, the embodiments of this specification provide a strategy generation model training and strategy generation method to solve the above-mentioned technical problem of difficulty in quickly and flexibly generating risk control strategies.

[0034] See also Figure 1 , Figure 1 This is an exemplary system architecture diagram of a strategy generation model training method provided in an embodiment of this specification.

[0035] like Figure 1 As shown, the system architecture may include a terminal 101, a network 102, and a server 103. The network 102 is used to provide a medium for a communication link between the terminal 101 and the server 103. The network 102 may include various types of wired communication links or wireless communication links. For example, the wired communication link may include an optical fiber, a twisted pair, or a coaxial cable, and the wireless communication link may include a Bluetooth communication link, a Wireless-Fidelity (Wi-Fi) communication link, or a microwave communication link.

[0036] The terminal 101 can interact with the server 103 via the network 102 to receive messages from the server 103 or send messages to the server 103, or the terminal 101 can interact with the server 103 via the network 102 to receive messages or data sent to the server 103 by other users. The terminal 101 can be hardware or software. When the terminal 101 is hardware, it can be various electronic devices, including but not limited to smart watches, smart phones, tablet computers, laptop computers, and desktop computers. When the terminal 101 is software, it can be installed in the electronic devices listed above. It can be implemented as multiple software or software modules (for example, to provide distributed services), or it can be implemented as a single software or software module, which is not specifically limited here.

[0037] In an embodiment of the present specification, in the corpus preparation stage, the terminal 101 first adds risk knowledge annotations to the sample risk supervision corpus based on professional experience to obtain standard training corpus; after obtaining the corpus, the terminal 101 constructs at least one strategy generation model based on at least one basic large model, inputs the standard training corpus into each strategy generation model, and trains each strategy generation model; during the training process of each strategy generation model, the terminal 101 controls each strategy generation model to output multiple risk control strategies for the standard training corpus, and adjusts the parameters of each strategy generation model according to the multiple risk control strategies and risk knowledge annotations until each strategy generation model converges.

[0038] Server 103 may be a business server that provides various services. It should be noted that server 103 may be either hardware or software. If server 103 is hardware, it may be implemented as a distributed server cluster consisting of multiple servers, or as a single server. If server 103 is software, it may be implemented as multiple software programs or software modules (e.g., to provide distributed services), or as a single software program or software module, without further limitation.

[0039] Alternatively, the system architecture may also not include the server 103. In other words, the server 103 may be an optional device in the embodiments of this specification, that is, the method provided in the embodiments of this specification may be applied to a system structure that only includes the terminal 101, and the embodiments of this specification do not limit this.

[0040] It should be understood that Figure 1 The number of terminals, networks, and servers in the example is only illustrative and can be any number of terminals, networks, and servers depending on the implementation requirements. It should be noted that the exemplary system architecture diagram of the policy generation method is similar to the exemplary system architecture diagram of the policy generation model training method, and therefore will not be repeated here.

[0041] See also Figure 2 , Figure 2 This is a flowchart of a policy generation model training method provided in an embodiment of this specification. The execution subject of an embodiment of this specification can be a terminal executing policy generation model training, a processor in a terminal executing the policy generation model training method, or a policy generation model training service in a terminal executing the policy generation model training method. For ease of description, the specific execution process of the policy generation model training method is described below using the example of a processor in a terminal as the execution subject.

[0042] like Figure 2 As shown, the strategy generation model training method may at least include: S202. Add risk knowledge annotations to the sample risk supervision corpus based on professional experience to obtain standard training corpus.

[0043] Alternatively, when developing risk control strategies, the first step is to collect relevant regulatory documents from various sources (such as regulatory websites and regulatory releases) as risk management data. These documents are then processed briefly to extract key information and store it in a database. Legal experts, compliance personnel, and others then define a series of rules and templates based on existing regulatory rules and internal procedures. These rules and templates are typically stored in a structured manner, such as database tables or configuration files. Based on the predefined rules and templates, key information extracted from the risk management data is populated into the corresponding templates, resulting in a specific risk control strategy. These risk control strategies are then applied to real-world scenarios, and appropriate monitoring mechanisms are established to ensure effective implementation. Strategies are regularly adjusted based on implementation effectiveness and rule updates to maintain the overall system's compliance and internal control levels. However, with rapidly evolving risk control requirements, this manual process of interpreting rules, developing templates, and formulating strategies clearly faces numerous challenges in terms of timeliness and accuracy.

[0044] Optionally, considering that the big model has the ability to self-learn rules and knowledge, in order to cope with the rapidly changing regulatory environment and complex risk scenarios, a strategy generation model can be built based on the basic big model to achieve automatic learning of regulatory rules and automatic generation of risk control strategies, thereby improving the problems of strategy accuracy and risk control efficiency faced when formulating risk control strategies based on established rules and templates.

[0045] Specifically, the strategy generation model can self-learn knowledge and information from a relatively small amount of training data. Once converged, it can intelligently interpret various raw risk management data and adaptively formulate risk control strategies. Therefore, in order for the strategy generation model to learn the professional experience and knowledge required for strategy formulation, it first requires a standard training corpus that is precisely annotated based on professional experience.

[0046] Furthermore, in terms of training costs, the strategy generation model benefits from the powerful data processing capabilities and optimization algorithms of the underlying large model. This allows it to significantly reduce its reliance on large amounts of labeled data while ensuring efficient learning, thus reducing the cost of tedious steps such as data collection, cleaning, and labeling. Therefore, regulatory texts collected in advance from various channels (such as regulatory websites and documents released by regulatory agencies) are used as risk management corpus. Only a small amount of sample risk management corpus needs to be annotated with risk knowledge based on the professional knowledge and industry experience of experts, compliance personnel, and others. This allows the expert knowledge and experience to be converted into a machine-understandable format, providing high-quality training data for subsequent strategy generation model training.

[0047] Optionally, after completing the risk knowledge annotation, high-quality standard training corpus can be obtained. This standard training corpus not only contains rich risk and regulatory information, but also, through meticulous annotation, provides clear learning objectives and directions for the model. It should be noted that the number of sample risk and regulatory corpus is at least one. Even with a very small amount of training corpus, the policy generation model based on a large model base can achieve excellent learning and convergence.

[0048] S204: Construct at least one strategy generation model based on at least one basic large model, input standard training corpus into each strategy generation model, and train each strategy generation model.

[0049] Optionally, during the model training phase, at least one strategy generation model can be constructed based on at least one basic large model. A basic large model refers to a large pre-trained model with extensive knowledge representation and powerful reasoning capabilities, which can adapt to a variety of application scenarios and task requirements. These basic large models have generally performed well in fields such as natural language processing and text classification, such as Tongyi Qianwen, ChatGPT, and LLM3. When constructing a strategy generation model, multiple strategy generation models can be constructed using multiple basic large models. The basic large models used between the strategy generation models can be the same or different, and this specification does not limit this.

[0050] Furthermore, the standard training corpus is input into each strategy generation model so that the strategy generation model can automatically generate appropriate risk control strategies for the risk supervision corpus by learning the expert knowledge and experience in the standard training corpus.

[0051] S206. During the training process of each strategy generation model, control each strategy generation model to output multiple risk control strategies for the standard training corpus, and adjust the parameters of each strategy generation model according to the multiple risk control strategies and risk knowledge annotations until each strategy generation model converges.

[0052] Optionally, during the training process, we input standard training corpus into each strategy generation model, control each strategy generation model to output multiple risk control strategies based on the standard training corpus, and continuously adjust the parameters and structure of each strategy generation model based on multiple risk control strategies and risk knowledge annotations, so that it can more accurately learn the relationship between risk strategies and corpus until each strategy generation model converges.

[0053] Specifically, strategy generation model convergence means the model reaches preset performance and becomes stable. At this point, the model can be considered to have fully learned the information in the standard training corpus, and the risk control strategy it outputs has achieved a relatively ideal balance between risk identification and control. At this point, each strategy generation model has completed training and can be put into practical risk control applications to provide strong support for risk management decisions.

[0054] It should be noted that in the embodiments of this specification, when there are multiple policy generation models, when there are a sufficient number of converged policy generation models to cover the usage requirements in the actual application scenario, the training of each policy generation model can be stopped, and the converged policy generation model can be used in the actual scenario. In one possible implementation, five policy generation models can be constructed for training. When 60% of the models, that is, three policy generation models, reach convergence, the current fine-tuning training can be terminated, and the three converged policy generation models can be used to generate risk control strategies in the actual scenario.

[0055] In an embodiment of the present specification, a strategy generation model training method is provided, which adds risk knowledge annotations to the sample risk supervision corpus based on professional experience to obtain standard training corpus; constructs at least one strategy generation model based on at least one basic large model, inputs the standard training corpus into each strategy generation model, and trains each strategy generation model; during the training process of each strategy generation model, controls each strategy generation model to output multiple risk control strategies for the standard training corpus, and adjusts the parameters of each strategy generation model based on multiple risk control strategies and risk knowledge annotations until each strategy generation model converges. By training the strategy generation model constructed based on the basic large model with the training corpus annotated based on professional experience, the strategy generation model can learn professional knowledge and experience and use it to generate strategies. The self-learning ability of the large model improves the accuracy and robustness of subsequent strategy generation, and also significantly reduces the maintenance cost of the strategy generation process.

[0056] See also Figure 3 , Figure 3 A flow chart of a strategy generation model training method provided in an embodiment of this specification.

[0057] like Figure 3 As shown, the strategy generation model training method may at least include: S302: Collect multiple risk supervision corpora in the network environment based on a preset crawler program, and select at least one risk supervision corpus as a sample risk supervision corpus.

[0058] Optionally, predefined rules and templates, as the basic framework for compliance management and risk control for many organizations, are often static and relatively fixed in their essential characteristics. This static nature limits their adaptability to the ever-changing and rapidly evolving regulatory environment to a certain extent. In a complex and ever-changing regulatory context, new regulatory rules may be introduced frequently, and existing rules may also undergo major adjustments or revisions. Faced with such dynamic changes, existing rules and templates may not be adjusted in a timely manner, and this lag increases compliance risks. In order to effectively meet this challenge, in the embodiments of this specification, an automated collection technology based on a public data interface is used to accurately locate and efficiently collect various types of rule files, reserve them as risk supervision corpus, and then select at least one risk supervision corpus as a sample risk supervision corpus.

[0059] Specifically, it is first necessary to clearly define the scope of rule file types that need to be collected. This scope not only covers traditional laws, regulations, regulatory guidelines and other formal documents, but also includes more specific and detailed regulations such as management methods and internal operating procedures. A comprehensive source of corpus ensures the accuracy and effectiveness of subsequent risk control rules. In order to realize the automated collection task, the public data interface can be monitored based on a preset automated crawler program. When various rule files are located, they are immediately collected and stored. This automated access process greatly improves the efficiency and accuracy of corpus collection, while reducing the burden and error rate of manual operations. It should be noted that the implementation method of the preset automated crawler program in the embodiment of this specification does not involve high-frequency, violent and other non-compliant processing methods. The collection, use and processing of relevant data in its implementation process are fully authorized by all parties and need to comply with the relevant laws, regulations and standards of relevant countries and regions. For example, the regulatory rule documents, risk supervision corpus, etc. involved in the embodiments of this specification are all obtained under public or fully authorized circumstances.

[0060] Furthermore, the preset crawler program can be set to collect and update the corpus according to a preset period (such as one day, one week, etc.). This enables the system to respond quickly to regulatory changes, ensuring that the system can incorporate relevant information into the analysis framework as soon as new rules are introduced or existing rules are changed, and adjust compliance strategies in a timely manner, overcoming the limitations of traditional static rules and templates, thereby effectively reducing compliance risks.

[0061] S304: Identify at least one title in each risk management corpus, generate directory information of each risk management corpus based on the title of each risk management corpus; and store the directory information of each risk management corpus in a structured database.

[0062] Alternatively, collected regulatory documents often exist in unstructured text, which affects their readability and management difficulty. Therefore, it is necessary to pre-process the collected risk regulatory corpus to lay the data foundation for subsequent analysis and strategy generation.

[0063] Optionally, during the preprocessing phase, the original text can be meticulously and accurately annotated to extract structural information from the text, transforming unstructured data into structured data. Furthermore, named entity recognition (NER) technology can be introduced. NER intelligently identifies various meaningful entity names within the text, which is crucial for understanding and analyzing regulatory documents.

[0064] Specifically, NER technology can identify core elements in texts such as file names, paragraphs, and clauses. In the embodiments of this specification, this technology can be used to identify the title directory structure in the file, including key information such as file names, chapters, and subsections. After identifying this information, it is stored in a database in a structured manner. The structured data format not only improves the readability and searchability of the data, but also facilitates more efficient subsequent use of the data, thereby greatly improving work efficiency. Storing the information of the regulatory rules file in a structured manner also provides convenient data support for subsequent policy generation.

[0065] S306. Add risk knowledge annotations of at least one risk dimension to the sample risk supervision corpus based on professional experience to obtain standard training corpus; the risk dimension at least includes risk type, risk subcategory, risk control strategy, relevant data indicators, discrimination rules and corpus reference template.

[0066] Optionally, during the expert experience input phase, risk knowledge annotations can be added to the sample risk management corpus based on professional experience, in accordance with at least one risk dimension, with the goal of constructing a comprehensive, multi-dimensional standard training corpus. Risk knowledge annotations may include, but are not limited to: identifying and marking risk-related keywords, phrases, or sentences; categorizing and annotating different types of risks, such as credit risk, market risk, and operational risk, which are common in financial scenarios; or information leakage risk, virus implantation risk, and other common risk types in information security protection scenarios; and adding quantitative indicators such as risk level and impact range to the corpus, so that subsequent models can more accurately learn and identify risk characteristics.

[0067] Specifically, the risk dimensions covered during risk knowledge annotation include but are not limited to: Risk type: clearly annotate the types of risks involved in the corpus, such as information leakage risk and virus implantation risk in information security scenario protection, credit risk, market risk, operational risk, etc. in financial scenarios, to lay the foundation for subsequent risk classification and assessment; Risk subclass: further subdivide based on the risk type, such as subdividing the virus implantation risk in the information security protection scenario into file bundling implantation risk, link implantation risk, email implantation risk, etc., and subdividing the market risk in the financial scenario into interest rate risk, exchange rate risk, etc., to provide more specific risk identification information; Risk control strategy: propose corresponding risk management strategies for the identified risks. Risk control measures or strategic recommendations, such as risk avoidance, risk transfer, and risk reduction, provide guidance for risk management practices; Related data indicators: Extract key data indicators related to risk, such as information data sources and user behavior flow data in information security protection scenarios, and non-performing loan rates and capital adequacy ratios in financial scenarios, as important bases for risk assessment and monitoring; Discrimination rules: Develop clear discrimination rules to distinguish different risk levels or risk states, and provide support for the establishment of risk warning and response mechanisms; Corpus citation templates: Design standardized corpus citation formats and templates to facilitate accurate citation of annotated corpora in subsequent analysis, report writing, and other links, ensuring the traceability and consistency of information. The above-mentioned multi-dimensional risk knowledge annotation provides a solid data foundation for the training and optimization of risk identification models.

[0068] S308. Construct at least one strategy generation model based on at least one basic large model, input standard training corpus into each strategy generation model, and train each strategy generation model.

[0069] Regarding step S308 , please refer to the detailed description in step S204 , which will not be repeated here.

[0070] S310. During the training process of each strategy generation model, at least one prompt is used to instruct each strategy generation model to output multiple risk control strategies for a standard training corpus; the prompt is used to specify the recognition purpose and output rules of each strategy generation model.

[0071] Optionally, prompts play a crucial role in large models, significantly impacting the quality and effectiveness of the model's output. To enable the policy generation model built on the large model to more accurately execute a series of audit operations, pre-configured prompts can be used. These pre-configured prompts specify the policy generation model's identification purpose and output rules. These prompts help the policy generation model identify the task type and required operations, making it easier for the policy generation model to understand the required operations, their purpose, and the rules to follow when returning. By providing more specific prompts, the policy generation model can more easily generate desired outputs.

[0072] Furthermore, to achieve more accurate risk control strategy results, the strategy generation model can be instructed to progressively parse the standard training corpus through multiple stages of prompts, ultimately outputting multiple risk control strategies. For example, in the first stage, prompts are provided to the strategy generation model, informing it that its role as a risk control expert is to interpret documents, with the goal of outputting relevant data, indicators, strategies, and methods. Once the strategy generation model completes its current output, the prompts in the second stage inform the model that its role is to structure the previous round of data, with the goal of processing strategies and indicators in a structured form (e.g., item 1, item 2, etc.). This process continues, ultimately resulting in the strategy generation model outputting standardized, actionable strategies and regulations. By varying the prompts, the strategy generation model can be guided to make appropriate decisions as it parses the corpus, improving the stability of the strategy generation model in generating risk control strategies.

[0073] S312: Select at least one candidate risk control strategy that meets a preset validity condition from multiple risk control strategies.

[0074] Optionally, in order to improve the accuracy and robustness of strategy generation, multiple risk control strategies can be screened and filtered first, and at least one candidate risk control strategy that meets the preset validity conditions can be selected from the multiple risk control strategies. The relatively reliable risk control strategy can be used to train the model, so that the model can achieve better training results.

[0075] Specifically, a self-service voting mechanism (bagging) is employed to combine multiple risk control strategies and select at least one candidate risk control strategy that meets pre-set effectiveness criteria based on pre-set effectiveness screening rules. This mechanism aims to improve overall risk control effectiveness through ensemble learning. When implementing the bagging voting mechanism, each strategy generation model is treated as an "expert" and independently outputs a corresponding risk control strategy. Then, based on the bagging voting mechanism, the system comprehensively analyzes the output of these risk control strategies based on pre-set effectiveness screening rules. These screening rules may include, but are not limited to, performance indicators such as the strategy's source, recall rate, and reliability, as well as practical considerations such as stability, responsiveness, and resource consumption in actual applications. Using these evaluation criteria, the system objectively measures the performance of each strategy, ensuring that only truly effective strategies that meet transaction requirements are selected as candidate risk control strategies.

[0076] S314. Adjust the parameters of each strategy generation model according to each candidate risk control strategy and risk knowledge annotation until each strategy generation model converges.

[0077] Optionally, for each selected risk control strategy, the semantic similarity between each selected risk control strategy and the risk knowledge annotation can be calculated, and the semantic similarity corresponding to each selected risk control strategy can be fed back to the strategy generation model corresponding to each risk control strategy, and each strategy generation model can be controlled to adjust the parameters based on the received semantic similarity until each strategy generation model converges.

[0078] Specifically, the generated multiple risk control strategies are compared and adjusted with the existing risk knowledge annotations. Through the introduction of the above embodiments, it can be seen that the risk knowledge annotations are carefully formulated by field experts based on historical cases, industry standards and the latest risk trends, and represent the industry's consensus and practice on risk identification. By carefully comparing these professional knowledge with the strategies output by the model, the shortcomings in the strategies can be identified. Based on the results of the above comparative analysis, the system can further adjust the parameters of each strategy generation model. This adjustment process involves modifying the algorithm weights, increasing or decreasing the feature dimensions, adjusting the model complexity and other aspects, aiming to make the strategies output by the model closer to the actual scenario requirements and improve the effectiveness and accuracy of the strategy. This iterative optimization process often needs to be repeated. After each adjustment, the model will be retrained and its performance will be evaluated until the predetermined convergence standard is reached.

[0079] It should be noted that the convergence standard can be formulated according to actual needs. For example, if the policy generation model is required to output a risk control strategy with a high semantic similarity to the expert strategy, the semantic similarity between the risk control strategy output by the policy generation model and the risk knowledge annotation can be set to be above 90%, and the model is considered to have reached convergence. At the same time, the policy generation model can be standardized in the prompt words to output the results according to the expert's language style. Then, on the basis of the standardized output based on the prompt words, the policy generation model can output the risk control strategy in the expert's language style. After the policy generation model outputs the risk control strategy, its text content is directly compared with the text content of the risk knowledge annotation provided by the expert for semantic similarity calculation, so as to judge whether the policy generation model has reached convergence based on the preset semantic similarity threshold. The embodiments of this specification do not specifically limit the convergence judgment method and related thresholds involved in the convergence standard.

[0080] In an embodiment of this specification, a policy generation model training method is provided. First, an automated collection technology is implemented through a crawler program to periodically and accurately locate and efficiently collect various rule files, which are then stored as risk management corpus. Further, through named entity recognition technology, the titles and directories in the text are intelligently identified, and the text is structured and stored based on the title and directory information of the text. In the expert experience input stage, risk knowledge annotations of at least one risk dimension are added to the sample risk management corpus based on professional experience to construct a comprehensive, multi-dimensional standard training corpus, which provides a solid data foundation for the training and optimization of the risk identification model. In order to enable the policy generation model built based on the large model to perform a series of audit operations more accurately, multi-stage prompts are also pre-configured. In the configured prompts, the recognition purpose and output rules of the policy generation model are specified to help the policy generation model clarify the task type and the operations to be performed, and guide the policy generation model to make corresponding analysis and judgment operations when parsing the corpus, which is conducive to improving the stability of the policy generation model when generating risk control strategies. For the multiple risk control strategies output by each model, a self-service voting mechanism is also used to mix multiple risk control strategies and select at least one candidate risk control strategy that meets the preset validity criteria from the multiple risk control strategies according to preset validity screening rules. This ensures that only those strategies that are truly effective and meet the business requirements are selected as candidate risk control strategies. Finally, by calculating the semantic similarity between each candidate risk control strategy and the risk knowledge annotation, the control strategy generation model adjusts the parameters based on the semantic similarity, making the model output strategy more closely aligned with the actual scenario requirements, improving the strategy's effectiveness and accuracy until the model converges.

[0081] See also Figure 4 , Figure 4 A flow chart of a strategy generation method provided in an embodiment of this specification.

[0082] like Figure 4 As shown, the strategy generation method may at least include: S402: Obtain target risk supervision corpus, input the target risk supervision corpus into at least one converged strategy generation model, and obtain multiple risk control strategies output by each strategy generation model.

[0083] Optionally, after fine-tuning the strategy generation model constructed based on the large model base using the strategy generation model training method in the above embodiment, the converged strategy generation model obtained can generate risk control strategies for each risk supervision corpus.

[0084] Specifically, the target risk management corpus for parsing and generating risk control strategies is first determined. This corpus is then fed into at least one converged strategy generation model. Each strategy generation model, drawing on the expertise and experience acquired during training, outputs multiple risk control strategies for the target risk management corpus.

[0085] S404. Perform supervision in risk scenarios based on multiple risk control strategies.

[0086] Optionally, after obtaining these risk control strategies, supervision can be performed in risk scenarios based on multiple risk control strategies. That is, these risk control strategies can be deployed to actual risk control scenarios and implemented to ensure that the system can respond to changes in regulatory rules in a timely manner in actual scenarios and improve the overall level of compliance and internal control.

[0087] In an embodiment of this specification, a strategy generation method is provided, which obtains target risk supervision corpus, inputs the target risk supervision corpus into at least one converged strategy generation model, and obtains multiple risk control strategies output by each strategy generation model; supervision is performed in a risk scenario based on the multiple risk control strategies; wherein the converged strategy generation model is trained based on the strategy generation model training method of any of the above embodiments. The automated strategy generation process implemented based on each strategy generation model reduces human intervention, improves the consistency and reliability of the strategy, and provides a more intelligent risk control strategy solution, thereby improving the system's compliance and internal control level.

[0088] See also Figure 5 , Figure 5 A flow chart of a strategy generation method provided in an embodiment of this specification.

[0089] like Figure 5 As shown, the strategy generation method may at least include: S502. Obtain target risk supervision corpus from a structured database; input the target risk supervision corpus into at least one converged strategy generation model to obtain multiple risk control strategies output by each strategy generation model.

[0090] Alternatively, as can be seen from the above embodiments, various risk management rules files can be collected from public file interfaces using automated crawler programs. Information such as titles and directories from these files is extracted and stored in a structured database for easy query and analysis. Therefore, when outputting risk control strategies based on a converged model, the target risk management corpus can be obtained from the structured database.

[0091] Furthermore, the target risk management corpus is fed into at least one converged strategy generation model. After the target risk management corpus is fed, the converged strategy generation model analyzes, learns, and infers the corpus based on pre-set algorithms and logic, and generates corresponding risk control strategies accordingly.

[0092] S504: Perform applicability assessment on each candidate risk control strategy to determine at least one target risk control strategy that meets preset applicability conditions.

[0093] Optionally, after obtaining multiple risk control strategies output by each strategy generation model, these strategies may reflect different focuses due to the potential use of multiple strategy generation models with different parameters. For example, different prevention and control measures (such as user behavior monitoring, high-risk behavior monitoring, and the establishment of targeted firewalls) may be implemented for different types of risks (such as information leakage risk and virus implantation risk). However, not all strategies may be feasible in actual scenarios. Therefore, in order to determine applicable strategies, it is necessary to conduct a suitability assessment on these output strategies to determine at least one target risk control strategy that meets the preset suitability conditions.

[0094] Optionally, in the applicability determination stage, the embodiment of this specification uses a deep neural network text classification algorithm to score the applicability of each risk control strategy that passes the preliminary generation process, that is, each candidate risk control strategy is input into the target deep neural network to determine the applicability score output by the target deep neural network for each candidate risk control strategy based on the text classification algorithm.

[0095] Specifically, the deep neural network text classification algorithm possesses powerful text understanding and analysis capabilities. When applied to assessing the applicability of risk control strategies, it not only focuses on the literal meaning of the strategy text, but also deeply explores the underlying logic, intent, and relevance to the specific context. Specifically, the target deep neural network algorithm comprehensively considers multiple dimensions, including the level of detail in the strategy description, the rationality of the logical structure, the clarity of the targeted risk types, and the potential positive and negative impacts of strategy implementation. Furthermore, the algorithm more accurately assesses the strategy's applicability and feasibility based on the specific transactional and regulatory context within which the strategy text exists.

[0096] Furthermore, the target deep neural network in the embodiments of this specification can also dynamically adjust according to the regulatory requirements related to the policy. As the regulatory environment continues to change, the algorithm can learn and absorb new regulatory requirements in real time, ensuring that the policy evaluation is always synchronized with the feasibility standards in the actual scenario.

[0097] Optionally, each risk control strategy is evaluated using a deep neural network text classification algorithm, and a quantitative suitability score is assigned. This score quantifies at least one of the following aspects of each candidate risk control strategy: content quality, relevance to the target risk regulatory corpus, and practical operational value. This score intuitively reflects the strategy's applicability in real-world applications. A higher score indicates greater compliance with regulatory requirements and greater practical operational value. Based on these scores, strategies can be sorted and screened to identify target risk control strategies that meet pre-set suitability criteria, ensuring that the final target risk control strategy deployed meets both regulatory requirements and practical operational value.

[0098] S506: Convert the policy text of each risk control policy into executable monitoring indicators; connect the monitoring indicators to the risk scenarios, and supervise the risks in the risk scenarios based on the monitoring indicators.

[0099] Optionally, during the strategy deployment and launch phase, a series of indicator conversion operations are required to smoothly transform the risk control strategy selected after strict applicability assessment into actual monitoring capabilities and quickly integrate it into the system's risk management system.

[0100] First, the risk control strategies selected after suitability assessment are fed back to the corresponding risk category managers in the system. To facilitate their quick understanding and application, text2SQL technology is used to automatically convert the policy text into executable monitoring indicators. This technology intelligently parses key information from the policy text, including risk type, monitoring conditions, and warning thresholds, and automatically converts it into structured SQL queries. These SQL queries can be directly integrated into existing risk control systems, enabling automated policy deployment and monitoring. This allows risk category managers to convert risk control strategies into executable monitoring indicators without having to manually write complex SQL code. These monitoring indicators not only cover the various risk types faced in the current scenario but can also be flexibly adjusted to suit different transaction scenarios and regulatory requirements. Once the system detects that a risk indicator reaches or exceeds the preset warning threshold, it immediately triggers an alert mechanism, alerting relevant personnel to take timely intervention and resolution measures. This approach greatly improves the efficiency of risk control strategy implementation, helps the system promptly identify and respond to potential risks, and significantly enhances the proactive and accurate nature of risk prevention and control.

[0101] In an embodiment of the present specification, a strategy generation method is provided, which obtains a target risk supervision corpus from a structured database; the target risk supervision corpus is input into at least one converged strategy generation model to obtain multiple risk control strategies output by each strategy generation model. A deep neural network text classification algorithm is used to score the applicability of each risk control strategy that passes the preliminary generation process, thereby quantitatively representing at least one aspect of the content quality, relevance to the target risk supervision corpus, and actual operational value of each candidate risk control strategy. Based on these scores, the strategies are sorted and screened to ensure that the target risk control strategy that is finally launched meets regulatory requirements and has actual operational value. The strategy text of each risk control strategy is converted into an executable monitoring indicator through text2sql technology, and the monitoring indicator is connected to the risk scenario, and the risks in the risk scenario are supervised based on the monitoring indicator, so that the system can respond to changes in regulatory rules more flexibly, thereby effectively responding to changes in regulatory rules and improving the overall compliance and internal control level.

[0102] See also Figure 6 , Figure 6 This is a structural block diagram of a strategy generation model training device provided in the embodiment of this specification. Figure 6 As shown, the strategy generation model training device 600 includes: Corpus annotation module 610, used to add risk knowledge annotations to the sample risk supervision corpus based on professional experience to obtain standard training corpus; A sample input module 620 is configured to construct at least one strategy generation model based on at least one basic large model, input standard training corpus into each strategy generation model, and train each strategy generation model; The model training module 630 is used to control each strategy generation model to output multiple risk control strategies for the standard training corpus during the training process of each strategy generation model, and adjust the parameters of each strategy generation model according to the multiple risk control strategies and risk knowledge annotations until each strategy generation model converges.

[0103] Optionally, the model training module 630 is further used to use at least one prompt to instruct each strategy generation model to output multiple risk control strategies for the standard training corpus; the prompt is used to specify the recognition purpose and output rules of each strategy generation model.

[0104] Optionally, the model training module 630 is also used to select at least one candidate risk control strategy that meets preset validity conditions from multiple risk control strategies; adjust the parameters of each strategy generation model according to each candidate risk control strategy and risk knowledge annotation until each strategy generation model converges.

[0105] Optionally, the model training module 630 is further used to mix multiple risk control strategies based on a self-service voting mechanism and select at least one candidate risk control strategy that meets preset validity conditions from the multiple risk control strategies according to preset validity screening rules.

[0106] Optionally, the model training module 630 is also used to calculate the semantic similarity between each candidate risk control strategy and the risk knowledge annotation, and feed back the semantic similarity corresponding to each candidate risk control strategy to the strategy generation model corresponding to each risk control strategy, and control each strategy generation model to adjust the parameters based on the received semantic similarity until each strategy generation model converges.

[0107] Optionally, the corpus annotation module 610 is also used to add risk knowledge annotations of at least one risk dimension to the sample risk supervision corpus based on professional experience to obtain standard training corpus; the risk dimension includes at least risk type, risk subcategory, risk control strategy, relevant data indicators, judgment rules and corpus reference templates.

[0108] Optionally, the policy generation model training device 600 further includes: a corpus collection module, configured to collect a plurality of risk supervision corpora in a network environment based on a preset crawler program, and select at least one risk supervision corpus as a sample risk supervision corpus.

[0109] Optionally, the strategy generation model training device 600 also includes: a corpus preprocessing module, used to identify at least one title in each risk management corpus, generate directory information of each risk management corpus based on the title of each risk management corpus; and store the directory information of each risk management corpus in a structured database.

[0110] In an embodiment of the present specification, a policy generation model training device is provided, wherein a corpus annotation module is used to add risk knowledge annotations to the sample risk supervision corpus based on professional experience to obtain a standard training corpus; a sample input module is used to construct at least one policy generation model based on at least one basic large model, input the standard training corpus into each policy generation model, and train each policy generation model; a model training module is used to control each policy generation model to output multiple risk control strategies for the standard training corpus during the training of each policy generation model, and adjust the parameters of each policy generation model according to multiple risk control strategies and risk knowledge annotations until each policy generation model converges. The policy generation model constructed based on the basic large model is trained by training the training corpus based on professional experience annotations, so that the policy generation model can learn professional knowledge and experience and use it to generate strategies. The self-learning ability of the large model improves the accuracy and robustness of subsequent strategy generation, and also significantly reduces the maintenance cost of the policy generation process.

[0111] See also Figure 7 , Figure 7This is a structural block diagram of a strategy generation device provided in the embodiment of this specification. Figure 7 As shown, the strategy generating device 700 includes: Corpus acquisition module 710 is used to acquire target risk management corpus, input the target risk management corpus into at least one converged strategy generation model, and obtain multiple candidate risk control strategies output by each strategy generation model; Strategy deployment module 720, for performing supervision in risk scenarios based on multiple risk control strategies; The converged strategy generation model is obtained by training based on the strategy generation model training method in the above embodiment.

[0112] Optionally, the strategy deployment module 720 is further configured to perform applicability assessment on each candidate risk control strategy, determine at least one target risk control strategy that meets preset applicability conditions, and perform supervision in risk scenarios based on each target risk control strategy.

[0113] Optionally, the strategy deployment module 720 is further used to input each candidate risk control strategy into the target deep neural network, determine the applicability score output by the target deep neural network for each candidate risk control strategy based on the text classification algorithm; and determine at least one target risk control strategy that meets the preset applicability conditions based on the applicability score corresponding to each candidate risk control strategy.

[0114] Optionally, the applicability score is used to quantitatively represent at least one of the following information: the content quality, relevance to the target risk regulatory corpus, and actual operational value of each candidate risk control strategy.

[0115] Optionally, the strategy deployment module 720 is further configured to convert the strategy text of each risk control strategy into executable monitoring indicators; connect the monitoring indicators to the risk scenarios, and supervise the risks in the risk scenarios based on the monitoring indicators.

[0116] Optionally, the corpus acquisition module 710 is further configured to acquire target risk management corpus from a structured database, where the structured database is configured to store directory information of at least one risk management corpus.

[0117] In an embodiment of the present specification, a policy generation device is provided, wherein a corpus acquisition module is used to obtain target risk supervision corpus, input the target risk supervision corpus into at least one converged policy generation model, and obtain multiple candidate risk control strategies output by each policy generation model; a policy deployment module is used to perform supervision in risk scenarios based on multiple risk control strategies; wherein the converged policy generation model is trained based on the policy generation model training method of the above embodiment. The automated policy generation process implemented based on each policy generation model reduces human intervention, improves the consistency and reliability of the policy, and provides a more intelligent risk control strategy solution, thereby improving the compliance and internal control level of the system.

[0118] The embodiments of this specification provide a computer program product including instructions. When the computer program product is run on a computer or a processor, the computer or the processor is caused to perform the steps of any one of the methods in the above embodiments.

[0119] The embodiments of this specification also provide a computer storage medium, which can store multiple instructions, and the instructions are suitable for being loaded by a processor and executing the steps of any method in the above embodiments.

[0120] See Figure 8 , Figure 8 This is a schematic diagram of the structure of a terminal provided in the embodiment of this specification. Figure 8 As shown, the terminal 800 may include: at least one processor 801 , at least one network interface 804 , a user interface 803 , a memory 805 , and at least one communication bus 802 .

[0121] The communication bus 802 is used to implement the connection and communication between these components.

[0122] The user interface 803 may include a display screen (Display) and a camera (Camera). Optionally, the user interface 803 may also include a standard wired interface and a wireless interface.

[0123] The network interface 804 may optionally include a standard wired interface or a wireless interface (such as a WI-FI interface).

[0124] The processor 801 may include one or more processing cores. Using various interfaces and circuits, the processor 801 connects various components within the terminal 800. It executes instructions, programs, code sets, or instruction sets stored in the memory 805 and accesses data stored in the memory 805 to perform various functions and process data for the terminal 800. Optionally, the processor 801 may be implemented using at least one of the following hardware forms: a digital signal processing (DSP), a field-programmable gate array (FPGA), or a programmable logic array (PLA). The processor 801 may integrate one or a combination of a central processing unit (CPU), a graphics processing unit (GPU), and a modem. The CPU primarily processes the operating system, user interface, and application programs; the GPU is responsible for rendering and drawing content displayed on the display; and the modem handles wireless communications. It is understood that the modem may also be implemented independently of the processor 801 and implemented on a separate chip.

[0125] Among them, the memory 805 may include a random access memory (RAM) or a read-only memory (ROM). Optionally, the memory 805 includes a non-transitory computer-readable storage medium. The memory 805 can be used to store instructions, programs, codes, code sets or instruction sets. The memory 805 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as a touch function, a sound playback function, an image playback function, etc.), instructions for implementing the above-mentioned various method embodiments, etc.; the data storage area may store data involved in the above-mentioned various method embodiments, etc. The memory 805 may also be optionally at least one storage device located away from the aforementioned processor 801. As Figure 8 As shown, the memory 805 as a computer storage medium may include an operating system, a network communication module, a user interface module, and a policy generation model training program or a policy generation program.

[0126] exist Figure 8In the terminal 800 shown, the user interface 803 is primarily used to provide an input interface for the user and obtain user input data; and the processor 801 can be used to call the policy generation model training program or policy generation program stored in the memory 805. When executing the policy generation model training program or policy generation program, the processor 801 specifically implements the steps of the method in any of the above embodiments.

[0127] In the several embodiments provided in this specification, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of modules is merely a logical function division. In actual implementation, there may be other division methods, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed can be through some interface, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.

[0128] Modules described as separate components may or may not be physically separate, and components shown as modules may or may not be physical modules, that is, they may be located in one place or distributed across multiple network modules. Some or all of these modules may be selected to achieve the purpose of this embodiment based on actual needs.

[0129] In the above embodiments, all or part of the embodiments may be implemented using software, hardware, firmware, or any combination thereof. When implemented using software, all or part of the embodiments may be implemented in the form of a computer program product. The computer program product comprises one or more computer instructions. When loaded and executed on a computer, the computer program instructions fully or partially perform the processes or functions described above in accordance with the embodiments of this specification. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted via the computer-readable storage medium. The computer instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium accessible by a computer or a data storage device such as a server or data center that integrates one or more available media. The above-mentioned available media can be magnetic media (for example, floppy disks, hard disks, tapes), optical media (for example, digital versatile discs (DVDs)), or semiconductor media (for example, solid state drives (SSDs)).

[0130] It should be noted that for the aforementioned method embodiments, for the sake of simplicity of description, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the embodiments of this specification are not limited by the order of the actions described, because according to the embodiments of this specification, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the embodiments of this specification.

[0131] In addition, it should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, storage, display, etc.), and signals involved in the embodiments of this specification are all authorized by the user or fully authorized by all parties, and the collection, use, and processing of relevant data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. For example, the regulatory rules and documents, risk management corpus, etc. involved in this specification are all obtained publicly or with full authorization.

[0132] The foregoing description of this specification describes specific embodiments. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be performed in an order different from that described in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require the specific order shown or the sequential order to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0133] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0134] The above is a description of a strategy generation model training and strategy generation method, device and terminal provided in the embodiments of this specification. For technical personnel in this field, based on the ideas of the embodiments of this specification, there may be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the embodiments of this specification.

Claims

1. A strategy generation model training method, the method comprising: Based on professional experience, risk knowledge annotations are added to the sample risk supervision corpus to obtain standard training corpus; Constructing at least one strategy generation model based on at least one basic large model, inputting the standard training corpus into each strategy generation model, and training each strategy generation model; the basic large model refers to a large pre-trained model with knowledge representation and reasoning capabilities; During the training process of each strategy generation model, each strategy generation model is controlled to output multiple risk control strategies for the standard training corpus, and the parameters of each strategy generation model are adjusted according to the multiple risk control strategies and the risk knowledge annotations until each strategy generation model converges; The adjusting the parameters of each strategy generation model according to the multiple risk control strategies and the risk knowledge annotations until each strategy generation model converges includes: selecting at least one candidate risk control strategy that meets a preset validity condition from the multiple risk control strategies; adjusting the parameters of each strategy generation model according to each candidate risk control strategy and the risk knowledge annotations until each strategy generation model converges; The method of adjusting the parameters of each strategy generation model according to each candidate risk control strategy and the risk knowledge annotation until each strategy generation model converges includes: calculating the semantic similarity between each candidate risk control strategy and the risk knowledge annotation, feeding back the semantic similarity corresponding to each candidate risk control strategy to the strategy generation model corresponding to each risk control strategy, and controlling each strategy generation model to adjust the parameters based on the received semantic similarity until each strategy generation model converges.

2. According to the method of claim 1, the control strategy generation model outputs multiple risk control strategies for the standard training corpus, including: Using at least one prompt to instruct each strategy generation model to output multiple risk control strategies based on the standard training corpus; The prompts are used to specify the identification purpose and output rules of each strategy generation model.

3. The method according to claim 1, wherein the step of selecting at least one candidate risk control strategy that meets a preset validity condition from the plurality of risk control strategies comprises: Based on the self-service voting mechanism, the multiple risk control strategies are mixed and at least one candidate risk control strategy that meets the preset validity conditions is selected from the multiple risk control strategies according to the preset validity screening rules.

4. The method according to claim 1, wherein the step of adding risk knowledge annotations to the sample risk supervision corpus based on professional experience to obtain standard training corpus includes: Based on professional experience, risk knowledge annotations of at least one risk dimension are added to the sample risk supervision corpus to obtain standard training corpus; The risk dimensions include at least risk type, risk subcategory, risk control strategy, relevant data indicators, judgment rules and corpus citation template.

5. The method according to claim 1, further comprising: Based on a preset crawler program, multiple risk supervision corpora in the network environment are collected, and at least one risk supervision corpus is selected as a sample risk supervision corpus.

6. The method according to claim 5, after collecting risk supervision data in the network environment based on the preset crawler program, further comprising: Identifying at least one title in each risk management corpus, and generating directory information of each risk management corpus based on the title of each risk management corpus; Store the directory information of each risk supervision corpus in a structured database.

7. A strategy generation method, the method comprising: Obtaining target risk supervision corpus, inputting the target risk supervision corpus into at least one converged strategy generation model, and obtaining multiple risk control strategies output by each strategy generation model; Perform supervision in risk scenarios based on the multiple risk control strategies; The converged strategy generation model is obtained by training based on the strategy generation model training method according to any one of claims 1 to 8.

8. The method according to claim 7, wherein performing supervision in risk scenarios based on the multiple risk control strategies comprises: Conduct applicability assessment on each candidate risk control strategy and determine at least one target risk control strategy that meets the preset applicability conditions; Perform supervision in risk scenarios based on each target risk control strategy.

9. The method according to claim 8, wherein the applicability assessment of each candidate risk control strategy and determination of at least one target risk control strategy that meets a preset applicability condition comprises: Inputting each candidate risk control strategy into a target deep neural network, and determining a suitability score output by the target deep neural network for each candidate risk control strategy based on a text classification algorithm; Based on the applicability scores corresponding to the candidate risk control strategies, at least one target risk control strategy that meets the preset applicability conditions is determined.

10. The method according to claim 9, wherein the applicability score is used to quantitatively represent at least one of the following information: content quality, relevance to the target risk regulatory corpus, and actual operational value of each candidate risk control strategy.

11. The method according to claim 7, wherein the step of performing supervision in a risk scenario based on the multiple risk control strategies comprises: Convert the strategy texts of each risk control strategy into executable monitoring indicators; The monitoring indicators are connected to the risk scenarios, and the risks in the risk scenarios are supervised based on the monitoring indicators.

12. The method according to claim 9, wherein obtaining target risk supervision corpus comprises: A target risk management corpus is obtained from a structured database, where the structured database is used to store directory information of at least one risk management corpus.

13. A strategy generation model training device, comprising: The corpus annotation module is used to add risk knowledge annotations to the sample risk supervision corpus based on professional experience to obtain standard training corpus; A sample input module is used to construct at least one strategy generation model based on at least one basic large model, input the standard training corpus into each strategy generation model, and train each strategy generation model; the basic large model refers to a large pre-trained model with knowledge representation and reasoning capabilities; A model training module is used to control each strategy generation model to output multiple risk control strategies for the standard training corpus during the training process of each strategy generation model, and adjust the parameters of each strategy generation model according to the multiple risk control strategies and the risk knowledge annotations until each strategy generation model converges; The model training module is further configured to select at least one candidate risk control strategy that meets a preset validity condition from the multiple risk control strategies; and adjust the parameters of each strategy generation model according to each candidate risk control strategy and the risk knowledge annotation until each strategy generation model converges; The model training module is also used to calculate the semantic similarity between each candidate risk control strategy and the risk knowledge annotation, and feed back the semantic similarity corresponding to each candidate risk control strategy to the strategy generation model corresponding to each risk control strategy, and control each strategy generation model to adjust the parameters based on the received semantic similarity until each strategy generation model converges.

14. A strategy generation device, comprising: A corpus acquisition module is used to acquire target risk supervision corpus, input the target risk supervision corpus into at least one converged strategy generation model, and obtain multiple candidate risk control strategies output by each strategy generation model; A strategy control module, configured to perform supervision in risk scenarios based on the multiple risk control strategies; The converged strategy generation model is obtained by training based on the strategy generation model training method according to any one of claims 1 to 8.

15. A computer program product comprising instructions, which, when run on a computer or a processor, causes the computer or the processor to perform the steps of the method according to any one of claims 1 to 6 or 7 to 12.

16. A computer storage medium storing a plurality of instructions, wherein the instructions are suitable for being loaded by a processor and executing the steps of the method according to any one of claims 1 to 6 or 7 to 12.

17. A terminal comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 or 7 to 12 are implemented.

Images