Risk control analysis method and device, electronic equipment and storage medium

Abstract

The invention relates to the field of natural language large model analysis, in particular to a risk control analysis method and device, electronic equipment and a storage medium, and the analysis method comprises the steps: constructing an incremental pre-training corpus according to risk control business knowledge; performing fine adjustment on the pre-training language model according to the incremental pre-training corpus to obtain a pre-training language model after fine adjustment; historical risk user behavior data are collected, and supervised fine tuning data corpora are constructed according to the collected historical risk user behavior data; performing fine tuning on the pre-training language model after fine tuning according to the supervised fine tuning data corpus to obtain a trained risk control business large model; and obtaining real-time risk user behavior data, inputting the data into the trained risk control business large model, and obtaining a risk control analysis condition. Compared with the prior art, through two times of fine adjustment, the finally obtained risk control business large model can effectively analyze the risk control condition of the user based on the risk control business knowledge, and the risk control analysis efficiency is effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of natural language large model analysis, and more specifically, to a risk control analysis method, device, electronic device and storage medium. Background Art

[0002] With the continuous development of the live broadcast industry, live broadcast platforms have attracted a large number of users and streamers through their content creation. At the same time, they have also attracted some black production, which conducts illegal activities on the live broadcast platforms, bringing security risks to the live broadcast platforms and also affecting the normal use of users. Therefore, live broadcast platforms need to conduct risk assessments on users or streamers with abnormal behaviors in the live broadcast rooms, timely discover and handle risk behaviors, and ensure the safety of the live broadcast rooms. In the prior art, usually through manual means, a targeted rule system is set for the operations of users, and the rule system is used to determine whether there are risks for users and their operations. However, the coverage of the rule system is limited and it is difficult to cover all risk behaviors, and manual intervention is required, resulting in low efficiency. Summary of the Invention

[0003] The present invention aims to overcome at least one defect of the above prior art, and provides a risk control analysis method, device, electronic device and storage medium for improving the coverage and analysis efficiency of risk analysis.

[0004] According to a first aspect of the present application, a risk analysis method is provided, and the analysis method includes:

[0005] Construct an incremental pre-training corpus according to risk control business knowledge;

[0006] Fine-tune a pre-trained language model according to the incremental pre-training corpus to obtain a fine-tuned pre-trained language model;

[0007] Collect historical risk user behavior data, and construct a supervised fine-tuning data corpus according to the historical risk user behavior data;

[0008] Fine-tune the fine-tuned pre-trained language model again according to the supervised fine-tuning data corpus to obtain the trained pre-trained language model as a risk control business large model;

[0009] Obtain real-time risk user behavior data, input the real-time risk user behavior data into the risk control business large model for processing, and obtain risk control analysis situations.

[0010] Optionally, the constructing a supervised fine-tuning data corpus according to the historical risk user behavior data specifically includes:

[0011] Obtain a number of historical risk user groups according to the preset risk characteristics and the historical risk user behavior data; several historical risk users with the same risk characteristics are included in the historical risk user groups;

[0012] Obtain all the historical business behaviors of each historical risk user in each historical risk user group according to the historical risk user behavior data;

[0013] Obtain the historical high-frequency business behaviors of each historical risk user group according to the number of historical risk users corresponding to each historical business behavior; in the historical risk user group, the ratio of the number of historical risk users corresponding to the historical high-frequency business behavior to the number of historical risk users in the historical risk user group exceeds the prediction threshold;

[0014] Extract the sample parameters of each historical risk user group from the historical risk user behavior data according to the preset label parameters;

[0015] Obtain the risk control analysis samples of the historical risk user group under each corresponding historical high-frequency business behavior according to the corresponding sample parameters;

[0016] Construct the supervised fine-tuning data corpus according to each historical risk user group, and the corresponding sample parameters and risk control analysis samples.

[0017] Optionally, the obtaining the risk control analysis samples of the historical risk user group under each corresponding historical high-frequency business behavior according to the corresponding sample parameters specifically includes:

[0018] Use the corresponding historical high-frequency business behavior as the business scenario, utilize the pre-trained chat generation pre-trained model, conduct risk analysis on the historical risk user group based on the business scenario and the sample parameters, and output the risk control analysis samples according to the preset template.

[0019] Optionally, the inputting the real-time risk user behavior data into the risk control business large model for processing to obtain the risk control analysis situation specifically includes:

[0020] Obtain a number of real-time risk user groups according to the preset risk characteristics and the real-time risk user behavior data; several real-time risk users with the same risk characteristics are included in the real-time risk user groups;

[0021] Obtain all the real-time business behaviors of each real-time risk user in each real-time risk user group according to the real-time user behavior data;

[0022] Obtain the real-time high-frequency business behaviors of each of the real-time risk user groups according to the number of real-time risk users corresponding to each of the real-time business behaviors; in the real-time risk user group, the ratio of the number of real-time risk users corresponding to the real-time high-frequency business behavior to the number of real-time risk users in the real-time risk user group exceeds a prediction threshold;

[0023] Extract the label data of each real-time risk user group from the real-time risk user behavior data according to the preset label parameters;

[0024] Input each of the real-time risk user groups and the corresponding label data into the trained large risk control business model according to each real-time high-frequency business behavior, and obtain the risk control analysis situation of each real-time risk user group under each real-time high-frequency business behavior.

[0025] Optionally, the construction of the incremental pre-training corpus according to risk control business knowledge specifically includes:

[0026] Construct several risk control rules, several risk control factors, and several risk control labels in the risk control business knowledge into json strings respectively according to the preset character format, and use the constructed json strings as the incremental pre-training corpus.

[0027] Optionally, the fine-tuning of the pre-trained language model according to the incremental pre-training corpus specifically includes:

[0028] Fine-tune the pre-trained language model by Lora according to the incremental pre-training corpus.

[0029] Optionally, the re-fine-tuning of the fine-tuned pre-trained language model according to the supervised fine-tuning data corpus specifically includes:

[0030] Fine-tune the fine-tuned pre-trained language model by Lora according to the supervised fine-tuning data corpus.

[0031] According to the second aspect of the present application, a risk analysis device is provided, and the analysis device includes:

[0032] An incremental corpus construction module, configured to construct an incremental pre-training corpus according to risk control business knowledge;

[0033] An incremental fine-tuning module, configured to fine-tune a pre-trained language model according to the incremental pre-training corpus to obtain a fine-tuned pre-trained language model;

[0034] A supervised corpus construction module, configured to collect historical risk user behavior data and construct a supervised fine-tuning data corpus according to the historical risk user behavior data;

[0035] A supervised fine-tuning module for further fine-tuning the fine-tuned pre-trained language model according to the supervised fine-tuning data corpus to obtain the trained pre-trained language model as a large risk control business model;

[0036] A risk control analysis module for obtaining real-time risk user behavior data, inputting the real-time risk user behavior data into the large risk control business model for processing, and obtaining a risk control analysis situation.

[0037] According to the third aspect of the present application, an electronic device is provided, including a memory and a processor. A computer-readable instruction is stored on the memory, and the processor executes the computer-readable instruction to implement the risk control analysis method described in the first aspect above.

[0038] According to the fourth aspect of the present application, a computer storage medium is provided, on which a computer-readable program is stored. When the computer-readable program is executed, the risk control analysis method described in the first aspect above is implemented.

[0039] According to any of the above aspects, the risk control analysis method, device, electronic device and storage medium provided by the present application use an incrementally pre-trained corpus constructed by risk control business knowledge to pre-train a language model, so that the pre-trained language model can learn risk control business knowledge during the fine-tuning process based on the language learning ability of the model itself. Then, by collecting historical risk user behavior data to construct a supervised fine-tuning data corpus, the fine-tuned pre-trained language model is further fine-tuned according to the supervised fine-tuning data corpus, so that it can learn for actual user behaviors, obtain users with risk behaviors, and make the finally trained large risk control business model have a larger coverage of risk analysis; at the same time, using the model to analyze the risk situation of users can reduce manual intervention, thereby effectively improving the efficiency of risk analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0041] Figure 1 It is a schematic diagram of the application scenario of the analysis method provided in this embodiment.

[0042] Figure 2 It is a flowchart of the steps of the analysis method provided in this embodiment.

[0043] Figure 3Flow chart of steps for constructing supervised fine-tuning data corpus provided in this embodiment.

[0044] Figure 4 Flow chart of steps for obtaining risk control analysis situation provided in this embodiment.

[0045] Figure 5 Device structure diagram of the analysis device provided in this embodiment.

[0046] Figure 6 Device structure diagram of the electronic device provided in this embodiment.

[0047] Appended drawing annotation: Server 100, terminal 200, incremental corpus construction module 11, incremental fine-tuning module 12, supervised corpus construction module 13, supervised fine-tuning module 14, risk control analysis module 15, memory 21, processor 22, bus 23, communication interface 24. Detailed implementation manners

[0048] The attached drawings of this application are only for illustrative purposes and cannot be construed as a limitation to this application. For better illustrating the following embodiments, some components in the drawings will be omitted, enlarged or reduced, which do not represent the dimensions of the actual products; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0049] In order to enable those skilled in the art of this technology to better understand the solution of this application, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the attached drawings in the embodiments of this application. Obviously, the described embodiments are only a part of the embodiments of this application, rather than all the embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of this application.

[0050] It should be noted that the terms "first", "second", etc. in the description and claims of this application and the above attached drawings are used to distinguish similar objects, and do not have to be used to describe a specific order or sequence. It should be understood that such used data can be interchanged under appropriate circumstances so that the embodiments of this application described here can be implemented in an order other than those illustrated or described here. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units does not have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0051] Embodiment 1

[0052] With the continuous development of the live broadcast industry, live broadcast platforms have attracted a large number of users and streamers through their content creation. At the same time, they have also attracted some black production, which conducts illegal activities on the live broadcast platforms, bringing security risks to the live broadcast platforms and also affecting the normal use of users. Therefore, live broadcast platforms need to conduct risk assessments on users or streamers with abnormal behaviors in the live broadcast rooms, timely discover and handle risk behaviors, and ensure the security of the live broadcast rooms. In the prior art, the risk control analysis is usually achieved through the following several methods:

[0053] 1. In an artificial way, a targeted rule system is set for the operations of users, and the set rule system is used to judge whether there are risks in users and user behaviors. However, the coverage of the rule system is limited and it is difficult to cover all risk behaviors. Moreover, manual intervention is required, resulting in low efficiency.

[0054] 2. In a machine learning way, a risk control analysis model is constructed, and the risk control analysis model is used to analyze users and user behaviors to judge whether there are risks. Although it can improve the efficiency of risk control analysis to a certain extent, the interpretability of machine learning is poor, making it difficult to summarize and analyze the risk behaviors of users, and thus difficult to obtain a specific risk analysis report. At the same time, the versatility of machine learning is poor and it is difficult to adapt to new risk behaviors.

[0055] This embodiment provides a technical solution that can solve the above problems. The following will describe the specific implementation manners of this embodiment in detail with reference to the accompanying drawings.

[0056] Exemplarily, it is a schematic diagram of an application scenario of a risk control analysis method provided by an embodiment of the present application. As Figure 1 shown, the application scenario at least includes a server 100 and a terminal 200 that can communicate with the server 100. The server 100 has an image processing function and can also have a data transmission function for video streams and audio streams; the terminal 200 has a streaming media playback function and can also have an image processing function.

[0057] It can be understood that the server 100 can be an independent electronic device or a cluster composed of multiple electronic devices; the terminal 200 can be a smart phone terminal, a personal computer, a tablet computer, a vehicle-mounted terminal, etc., but is not limited thereto.

[0058] In an implementable manner, the server 100 and the terminal 200 can respectively execute a risk control analysis method provided by an embodiment of the present application. Or, optionally, a part of the risk control analysis method provided by an embodiment of the present application is executed in the server 100 and a part is executed in the terminal 200.

[0059] As Figure 2 shown, this embodiment provides a risk control analysis method, and the analysis method can specifically include:

[0060] S1: Construct an incremental pre-trained corpus according to risk control business knowledge;

[0061] In this embodiment, the constructing of the incremental pre-trained corpus according to risk control business knowledge may specifically include:

[0062] Construct several risk control rules, several risk control factors, and several risk control labels in the risk control business knowledge into json strings respectively according to a preset character format, and use the constructed json strings as the incremental pre-trained corpus.

[0063] Specifically, the risk control business knowledge can be obtained by studying relevant contents such as industry reports, case analyses, and regulatory policies in the video live broadcast industry. At the same time, customer behavior data can be analyzed to identify risk patterns and abnormal behaviors therein, or summarized according to industry experience to sort out the risk control business knowledge.

[0064] After obtaining the risk control business knowledge, several risk control rules, several risk control factors, and several risk control labels can be extracted from the risk control business knowledge by using a knowledge graph, and / or data analysis, and / or machine learning, etc.

[0065] In this embodiment, after the several risk control rules, several risk control factors, and several risk control labels are respectively constructed into json strings according to a preset character format, the scattered several risk control rules, several risk control factors, and several risk control labels can have a more systematic data expression, thereby improving the learning efficiency and learning effect of learning according to the incremental pre-trained corpus.

[0066] S2: Fine-tune a pre-trained language model according to the incremental pre-trained corpus to obtain a fine-tuned pre-trained language model;

[0067] In this embodiment, the pre-trained language model can adopt an existing pre-trained language model. Preferably, the pre-trained language model can adopt a pre-trained CPT (Chinese Pre-trained Transformer) model. The pre-trained CPT model is a pre-trained language model for Chinese natural language processing tasks. It is based on the Transformer architecture and pre-trained on a large amount of Chinese text data to capture the characteristics and rules of the Chinese language.

[0068] Therefore, in this embodiment, by using the pre-trained pre-trained language model for fine-tuning, the learning efficiency of the model can be improved, and at the same time, the ability of the pre-trained language model to effectively capture language characteristics can be utilized to effectively learn the risk control business knowledge in the incremental pre-trained corpus.

[0069] In this embodiment, the incremental pre-trained language can be achieved through LoRA fine-tuning. The LoRA fine-tuning technology is a technology for fine-tuning pre-trained models in natural language processing tasks, which can improve the learning efficiency and interpretability of the pre-trained language model.

[0070] Specifically, in this step, the hyperparameter settings for LoRA fine-tuning include: the learning rate is set to 0.0001, the learning batch size is 4, the number of epochs is set to 2, and the learning rate scheduler is set to cosine.

[0071] S3: Collect historical risk user behavior data, and construct a supervised fine-tuning data corpus based on the historical risk user behavior data;

[0072] Specifically, the historical risk user behavior data mainly includes historical user information, user behavior, user device status, and user geographical location, etc., which are historical business behavior data generated when users use the services of the video live streaming platform or application at historical time points. It can be obtained by collecting historical user data recorded in the background of the video live streaming platform or application.

[0073] Specifically, in this step, constructing a supervised fine-tuning data corpus based on the historical risk user behavior data, as Figure 3 shown, specifically includes:

[0074] A1: According to the preset risk features and the historical risk user behavior data, obtain a number of historical risk user groups; each historical risk user group includes a number of historical risk users with the same risk features;

[0075] Risk behaviors usually have a certain group nature, which is manifested as multiple different people corresponding to different accounts, or the same person using different accounts to uniformly execute the same risk behavior. Therefore, in this embodiment, the user groups with risks are mainly analyzed, so it is necessary to first screen out the user groups with risks.

[0076] Specifically, the preset risk features can be used to query and compare the historical risk user behavior data, and the user groups that meet the risk features are screened out. Exemplarily, the video live streaming platform or application usually sets corresponding risk judgment rules as its risk features, and records the user events that meet the risk judgment rules and forms risk events for reporting. The user information therein can be sorted out according to the risk events to obtain users with the same risk features, that is, users who meet the same risk judgment rules, and user groups are obtained.

[0077] A2: Obtain all historical business behaviors of each historical risky user in each of the historical risky user groups according to the historical risky user behavior data; the historical business behaviors represent various operation events of the historical risky users.

[0078] A3: Obtain the historical high-frequency business behaviors of each historical risky user group according to the number of historical risky users corresponding to each historical business behavior; in the historical risky user group, the ratio of the number of historical risky users corresponding to the historical high-frequency business behavior to the number of historical risky users in the historical risky user group exceeds the prediction threshold.

[0079] Understandably, for the historical risky user group, if a certain historical business behavior is executed simultaneously or within a certain time range in the historical risky user group, such as multiple users sending black production bullet screens in the same video or live broadcast room, it indicates that the historical risky user group has a relatively high risk, and the probability that the historical risky user group is a black production group is higher.

[0080] A4: Extract the sample parameters of each historical risky user group from the historical risky user behavior data according to the preset label parameters.

[0081] In this embodiment, the label parameters mainly may include parameters such as user labels, address labels, and device labels that reflect specific user information, as a reference basis for analyzing the risk control situation of the historical risky users. Among them, the user label identifies the information of the user, the address label identifies the geographical location information of the user, and the device label represents the information of the user terminal, such as whether there is modification, rooting, or whether the device fingerprint is abnormal, etc.

[0082] A5: Obtain the risk control analysis samples of the historical risky user group under the corresponding historical high-frequency business behaviors according to the corresponding sample parameters.

[0083] Specifically, in this embodiment, the acquisition of the risk control analysis samples may specifically include:

[0084] Use the corresponding historical high-frequency business behavior as the business scenario, and use the pre-trained chat generation pre-training model to perform risk analysis on the historical risky user group based on the business scenario and the sample parameters, and output the risk control analysis samples according to the preset template.

[0085] Understandably, the analysis results of the historical risky user group under different historical high-frequency business behaviors may be different. Therefore, for each historical risky user group, it is necessary to use its corresponding respective historical high-frequency business behaviors as business scenarios to analyze and obtain the risk control analysis samples of the historical risky user group under its corresponding respective high-frequency business behaviors.

[0086] Among them, the chat generation pre-trained model can be set using existing chat generation pre-trained models, such as ChatGPT (Chat Generative Pre-trained Transformer). Through the chat generation pre-trained model, the risk control analysis sample is output according to a preset template.

[0087] Specifically, the preset template may include a risk control analysis result and a risk control analysis process. The risk control analysis result may be: whether the historical risk user group belongs to a risk group under the corresponding historical high-frequency business behavior; the risk control analysis process may be: under the corresponding historical high-frequency business behavior, according to the sample parameters of the historical risk user group, the reasoning process of judging and analyzing whether the historical risk user group belongs to / does not belong to a risk group.

[0088] Through the natural language template, the results of the risk control analysis and the risk control analysis process are output, so that the output risk control analysis sample can be more intuitive and easier to understand, and thus the reasoning and analysis results of the finally obtained model have better interpretability.

[0089] A6: Construct the supervised fine-tuning data corpus according to each of the historical risk user groups, as well as the corresponding sample parameters and risk control analysis samples.

[0090] S4: Perform secondary fine-tuning on the fine-tuned pre-trained language model according to the supervised fine-tuning data corpus to obtain the trained pre-trained language model, which is used as the risk control business large model;

[0091] In this embodiment, specifically, Lora can be used to fine-tune the fine-tuned pre-trained language model.

[0092] Specifically, in this step, the hyperparameter settings for Lora fine-tuning include: the learning rate is set to 0.0001, the learning batch size is 8, epochs is set to 3, and the learning rate scheduler is set to cosine.

[0093] It can be understood that since the data used for the fine-tuning of the pre-trained language model is not real-time data, the fine-tuning of the pre-trained language model in steps S2 and S4 can be performed offline.

[0094] S5: Obtain real-time risk user behavior data, input the real-time risk user behavior data into the risk control business large model for processing, and obtain the risk control analysis situation.

[0095] Corresponding to the historical risk user behavior data, the real-time risk user behavior data mainly includes real-time user information, user behavior, user device situation, and user geographical location, etc.

[0096] Specifically, in this step, inputting the real-time risk user behavior data into the risk control business large model for processing to obtain the risk control analysis situation, as Figure 4 shown, specifically may include:

[0097] B1: According to the preset risk characteristics and the real-time risk user behavior data, obtain a number of real-time risk user groups; the real-time risk user groups include a number of real-time risk users with the same risk characteristics;

[0098] B2: According to the real-time user behavior data, obtain all the real-time business behaviors of each real-time risk user in each real-time risk user group;

[0099] B3: According to the number of real-time risk users corresponding to each real-time business behavior, obtain the real-time high-frequency business behaviors of each real-time risk user group; in the real-time risk user group, the ratio of the number of real-time risk users corresponding to the real-time high-frequency business behavior to the number of real-time risk users in the real-time risk user group exceeds the prediction threshold;

[0100] B4: According to the preset label parameters, extract the label data of each real-time risk user group from the real-time risk user behavior data;

[0101] It can be understood that steps B1 - B4 mainly include processing the real-time risk user behavior data so that it can be better recognized by the risk control business large model, and then obtaining the specific risk control analysis situation. The specific operation content of the processing is similar to the processing of the historical risk user behavior data in steps A1 - A4 above, and the specific content of steps B1 - B4 can refer to the description of steps A1 - A4 above.

[0102] B5: Input each real-time risk user group and the corresponding label data into the trained risk control business large model according to each real-time high-frequency business behavior, and obtain the risk control analysis situation of each real-time risk user group under each real-time high-frequency business behavior.

[0103] In this embodiment, after processing the real-time risk user behavior data, a number of real-time risk user groups are obtained. Each real-time risk user group contains a number of real-time risk users. Each real-time risk user group corresponds to a number of real-time high-frequency business behaviors, and each real-time risk user group has corresponding label data.

[0104] Similar to the fine-tuning step of the historical risk user behavior data, in this step, the real-time risk user group, one of the real-time high-frequency business behaviors corresponding to the real-time risk user group, and the corresponding tag data are input into the risk control business large model. The risk control business large model uses the input real-time high-frequency business behavior as the business scenario, and based on the tag data, analyzes the risk control situation of the real-time risk user group under this real-time high-frequency business behavior, including whether the real-time risk user group is a risk group under this real-time high-frequency business behavior, and the analysis process of whether the real-time risk user group is a risk group.

[0105] By analogy, for each of the real-time user groups, the risk analysis situations corresponding to all of their corresponding real-time high-frequency business behaviors are respectively analyzed and obtained.

[0106] In a preferred implementation manner of this embodiment, the real-time high-frequency business behaviors can be screened, and the real-time high-frequency business behaviors with high-risk are screened out. Based on the screened real-time high-frequency business behaviors, the risk control analysis situations of the corresponding real-time risk user groups are analyzed and obtained through the risk control business large model.

[0107] In this embodiment, the pre-trained language model is fine-tuned twice. In the first fine-tuning, the pre-trained language model is enabled to learn risk control business knowledge. In the second fine-tuning, the pre-trained language model is enabled to output the risk control analysis results of users through natural language for specific user data. Through the two fine-tunings, the finally obtained risk control business large model can effectively analyze the risk control situations of users based on risk control business knowledge, effectively improving the efficiency of risk control analysis. At the same time, the risk control business large model can output analysis results based on natural language, making the analysis process more interpretable.

[0108] Based on the same inventive concept, this embodiment also provides a risk control analysis device, as Figure 5 shown. The analysis device may specifically include:

[0109] An incremental corpus construction module 11, configured to construct an incremental pre-trained corpus according to risk control business knowledge;

[0110] In this embodiment, the incremental corpus construction module 11 can be used to execute Figure 2 the step S1 shown. The specific description of the incremental corpus construction module 11 can refer to the description of the step S1.

[0111] An incremental fine-tuning module 12, configured to fine-tune the pre-trained language model according to the incremental pre-trained corpus to obtain a fine-tuned pre-trained language model;

[0112] In this embodiment, the incremental fine-tuning module 12 can be used to execute Figure 2 the steps S2 shown. For the specific description of the incremental fine-tuning module 12, reference can be made to the description of the steps S2.

[0113] The supervised corpus construction module 13 is used to collect historical risk user behavior data and construct a supervised fine-tuning data corpus according to the historical risk user behavior data;

[0114] In this embodiment, the supervised corpus construction module 13 can be used to execute Figure 2 the steps S3 shown, and Figure 3 the steps A1 - A6 shown. For the specific description of the supervised corpus construction module 13, reference can be made to the description of the steps S3.

[0115] The supervised fine-tuning module 14 is used to further fine-tune the fine-tuned pre-trained language model according to the supervised fine-tuning data corpus to obtain the trained pre-trained language model as the large risk control business model;

[0116] In this embodiment, the supervised fine-tuning module 14 can be used to execute Figure 2 the steps S4 shown. For the specific description of the supervised fine-tuning module 14, reference can be made to the description of the steps S4.

[0117] The risk control analysis module 15 is used to obtain real-time risk user behavior data, input the real-time risk user behavior data into the large risk control business model for processing, and obtain the risk control analysis situation.

[0118] In this embodiment, the risk control analysis module 15 can be used to execute Figure 2 the steps S5 shown, and Figure 4 the steps B1 - B5 shown. For the specific description of the risk control analysis module 15, reference can be made to the description of the steps S5.

[0119] This embodiment also provides an electronic device, Figure 6 which shows the structural diagram of the electronic device of this embodiment, including a memory 21 and a processor 22. The memory 21 stores computer-readable instructions, and the processor 22 executes the computer-readable instructions to implement the risk control analysis method of this embodiment.

[0120] Preferably, the electronic device further includes a bus 23 and a communication interface 24, and the processor 22, the communication interface 24 and the memory 21 are connected through the bus 23.

[0121] Among them, the memory 21 may include a high-speed random access memory (RAM), and may also include a non-volatile memory, such as at least one disk memory. The communication connection between this system network element and at least one other network element is realized through at least one communication interface 24 (which can be wired or wireless), and the Internet, wide area network, local area network, metropolitan area network, etc. can be used. The bus 23 can be an ISA bus, a PCI bus, an EISA bus, etc. The bus 23 can be divided into an address bus, a data bus, a control bus, etc. (not fully drawn in the figure).

[0122] The processor 22 can be an integrated circuit chip with signal processing capabilities. In a specific implementation process, the steps in the embodiments of the above method can be completed by the integrated logic circuit in the hardware of the processor 22 or instructions in software form. The above-mentioned processor 22 can be a general-purpose processor, including a central processing unit (CPU for short), a network processor (NP for short), etc.; it can also be a digital signal processor (DSP for short), an application specific integrated circuit (ASIC for short), a field programmable gate array (FPGA for short), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, and can implement or execute the various methods, steps, and logic block diagrams disclosed in the embodiments of the present invention. The general-purpose processor can be a microprocessor or the processor 22 can also be any conventional processor 22, etc. The steps of the method disclosed in combination with the embodiments of the present invention can be directly embodied as being executed and completed by a hardware decoding processor, or executed and completed by a combination of hardware and software modules in the decoding processor. The software module can be located in a mature storage medium in the art such as a random access memory, flash memory, read-only memory, programmable read-only memory, or electrically erasable programmable memory, register, etc. This storage medium is located in the memory 21, and the processor 22 reads the information in the memory 21 and combines its hardware to complete the steps of the method in the foregoing embodiments.

[0123] The embodiments of the present invention also provide a computer-readable storage medium. The computer-readable storage medium stores computer-executable instructions. When the computer-executable instructions are called and executed by the processor 22, the computer-executable instructions cause the processor 22 to implement the above-mentioned risk control analysis method. For the specific implementation, reference can be made to the embodiments, and details are not described herein again.

[0124] When the above-mentioned function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical discs.

[0125] Obviously, the above-mentioned embodiments of the present invention are merely examples for clearly illustrating the technical solutions of the present invention, rather than limitations on the specific implementation manners of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the claims of the present invention shall be included within the protection scope of the claims of the present invention.

Claims

1. A risk analysis method, characterized in that: The analysis method comprises: Construct incremental pre-training corpus based on risk control business knowledge; Fine-tune the pre-trained language model according to the incremental pre-trained corpus to obtain a fine-tuned pre-trained language model; Collect historical risk user behavior data, and construct a supervised fine-tuning data corpus based on the historical risk user behavior data; Fine-tune the fine-tuned pre-trained language model again according to the supervised fine-tuning data corpus to obtain the trained pre-trained language model as a risk control business large model; Real-time risk user behavior data is obtained, and the real-time risk user behavior data is input into the risk control business big model for processing to obtain risk control analysis results.

2. A risk analysis method according to claim 1, characterized in that: The step of constructing a supervised fine-tuning data corpus based on the historical risk user behavior data specifically includes: According to the preset risk characteristics and the historical risk user behavior data, a number of historical risk user groups are obtained; the historical risk user groups include a number of historical risk users having the same risk characteristics; Acquire all historical business behaviors of each of the historical risk users in each of the historical risk user groups according to the historical risk user behavior data; According to the number of historical risk users corresponding to each of the historical business behaviors, the historical high-frequency business behaviors of each of the historical risk user groups are obtained; in the historical risk user group, the ratio of the number of historical risk users corresponding to the historical high-frequency business behaviors to the number of historical risk users in the historical risk user group exceeds a prediction threshold; Extracting sample parameters of each historical risk user group from the historical risk user behavior data according to preset label parameters; According to the corresponding sample parameters, the risk control analysis samples of the historical risk user group under the corresponding historical high-frequency business behaviors are obtained; The supervised fine-tuning data corpus is constructed according to each of the historical risk user groups, and the corresponding sample parameters and risk control analysis samples.

3. A risk analysis method according to claim 2, characterized in that: The step of obtaining risk control analysis samples of the historical risk user group under corresponding historical high-frequency business behaviors according to corresponding sample parameters specifically includes: The corresponding historical high-frequency business behaviors are used as business scenarios, and a pre-trained model is generated using pre-trained chats. Risk analysis is performed on the historical risk user group based on the business scenarios and the sample parameters, and the risk control analysis samples are output according to a preset template.

4. A risk analysis method according to any one of claims 2 or 3, characterized in that: The inputting the real-time risk user behavior data into the risk control business big model for processing to obtain risk control analysis information specifically includes: According to the preset risk characteristics and the real-time risk user behavior data, a number of real-time risk user groups are obtained; the real-time risk user groups include a number of real-time risk users having the same risk characteristics; Acquire all real-time business behaviors of each real-time risk user in each real-time risk user group according to the real-time user behavior data; According to the number of real-time risk users corresponding to each real-time business behavior, the real-time high-frequency business behavior of each real-time risk user group is obtained; in the real-time risk user group, the ratio of the number of real-time risk users corresponding to the real-time high-frequency business behavior to the number of real-time risk users in the real-time risk user group exceeds a prediction threshold; According to preset label parameters, extracting label data of each real-time risk user group from the real-time risk user behavior data; Each of the real-time risk user groups and the corresponding label data are input into the trained risk control business model according to each real-time high-frequency business behavior to obtain the risk control analysis situation of each of the real-time risk user groups under each real-time high-frequency business behavior.

5. A risk analysis method according to any one of claims 1 to 3, characterized in that: The step of constructing incremental pre-training corpus based on risk control business knowledge specifically includes: Several risk control rules, several risk control factors and several risk control labels in the risk control business knowledge are constructed into json strings according to preset character formats, and the constructed json strings are used as the incremental pre-training corpus.

6. A risk analysis method according to any one of claims 1 to 3, characterized in that: The fine-tuning of the pre-trained language model according to the incremental pre-trained corpus specifically includes: According to the incremental pre-training corpus, the pre-training language model is fine-tuned through Lora.

7. A risk analysis method according to any one of claims 1 to 3, characterized in that: The fine-tuning of the fine-tuned pre-trained language model according to the supervised fine-tuning data corpus specifically includes: According to the supervised fine-tuning data corpus, the fine-tuned pre-trained language model is fine-tuned again through Lora.

8. A risk analysis device, characterized in that: The analysis device comprises: Incremental corpus construction module, used to construct incremental pre-training corpus based on risk control business knowledge; An incremental fine-tuning module, used to fine-tune the pre-trained language model according to the incremental pre-training corpus to obtain a fine-tuned pre-trained language model; A supervised corpus construction module is used to collect historical risk user behavior data and construct a supervised fine-tuning data corpus based on the historical risk user behavior data; A supervised fine-tuning module is used to further fine-tune the fine-tuned pre-trained language model according to the supervised fine-tuning data corpus to obtain the trained pre-trained language model as a risk control business large model; The risk control analysis module is used to obtain real-time risk user behavior data, input the real-time risk user behavior data into the risk control business big model for processing, and obtain risk control analysis results.

9. An electronic device, comprising a memory and a processor, characterized in that: The memory stores computer-readable instructions, and the processor executes the computer-readable instructions to implement a risk control analysis method as described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that: A computer-readable program is stored thereon, and when the computer-readable program is executed, a risk control analysis method as described in any one of claims 1 to 7 is implemented.

Images