AI model-based auxiliary data auditing method and system

Abstract

The invention relates to an AI model-based auxiliary data auditing method and system, and the system comprises a data auditing auxiliary module which comprises a pre-auditing agent which is configured to check whether data submitted by a merchant is complete or not and carry out the data extraction of the submitted data; a risk auditing agent configured to verify the authenticity, consistency and risk of the data based on the extracted data; a compliance auditing agent configured to verify whether the data is legal and compliant based on the extracted data; the manual auditing module is configured to be used for manually rechecking auditing results from the pre-auditing agent, the risk auditing agent and the compliance auditing agent; and the routing module is configured to be used for selecting the AI model for constructing the corresponding sub-agent from the plurality of AI models according to the auditing task to be executed.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of electronic commerce, and more particularly, to merchant information auditing of a three-party payment platform in online transactions. BACKGROUND

[0002] With the vigorous development of Internet finance, the online electronic commerce industry has risen. This in turn has driven the development of third-party payment.

[0003] In a virtual intangible market, the transaction parties do not know each other and do not know the root cause. Therefore, the payment problem has been one of the bottlenecks of the development of electronic commerce. The seller is unwilling to ship first, for fear that the goods cannot be recovered after being shipped. The buyer is unwilling to pay first, for fear of not being able to get the goods or the quality of the goods not being guaranteed after payment. The result of the game is that both parties are unwilling to take risks first, and online shopping cannot be carried out.

[0004] To meet the market demand for synchronous exchange, the third-party payment has emerged as the times require. The third party is a "middle platform" for payment of funds between the buyer and the seller in the absence of credit guarantee or legal support. The buyer pays the money to the third party other than the buyer and the seller. The third party provides a safe trading service. Its operation is essentially to set up an intermediate transition account between the payer and the payee, so that the transfer of funds is controllable. Only when the opinions of both parties are reached can the funds be determined. In the transaction through the third-party platform, after the buyer selects and purchases the goods, the buyer uses the account provided by the third-party platform to pay the money, and the other party notifies the seller that the money has arrived and the goods are shipped. After the buyer inspects the goods, the buyer can notify the seller to pay. The emergence of the third-party payment platform theoretically completely eliminates fraudulent behavior in electronic transactions.

[0005] In practical applications, in order to improve the security of third-party transactions and avoid fraudulent behavior, it is necessary to audit the merchants accessing the third-party payment platform.

[0006] The core purpose of merchant auditing is:

[0007] ‌1) Preventing financial risks‌

[0008] By auditing the qualifications of merchants (such as the authenticity of business licenses) and operating scenarios (such as real photo verification), the risk of illegal fund flow can be effectively reduced. In actual transactions, unverified merchants may become a channel for illegal fund flow, for example, a large proportion of transactions involving unverified merchants are fake transactions.

[0009] 2)‌ Safeguarding fund safety‌

[0010] Auditing ensures that merchants have legal operating qualifications, avoiding user fund loss due to merchant running away or illegal operation. A licensed payment institution needs to isolate the funds of the merchant from its own funds, and manage the funds through a reserve fund account to ensure fund safety.

[0011] 3) Maintaining payment system stability

[0012] The review can screen out high-risk merchants (such as various illegal transactions) to prevent them from causing system congestion or collapse due to large-scale abnormal transactions.

[0013] However, for merchant review, the current platform still mainly relies on a large number of manual reviews. The review content mainly includes the following categories: legal person module review (legal person information, legal person certificate attachment), license review (license information, license attachment), entry review (entry information, entry attachment), basic information review (merchant name, MCC, storefront and interior, etc.), rate information review (business opening restrictions, business rate), agreement review, terminal review, etc.

[0014] Extracting the review of these modules into review rules involves more than 80 rules, such as whether the front and back of the legal person's ID card are consistent, whether the storefront and interior are the same store, whether the business license is real, whether the seal is real and clear, whether the merchant address is valid, etc.

[0015] Although some rules have been automatically judged by software, there are still a large number of scenarios that cannot be implemented by software. For example:

[0016] 1. Determine whether the front and back of the legal person's ID card are consistent;

[0017] 2. Compare the storefront and the name of the receipt;

[0018] 3. Compare whether the storefront and interior are the same store;

[0019] 4. Compare the storefront sign and contact address;

[0020] 5. Compare the interior and MCC;

[0021] 6. Whether the attachment material is PS or screen flipping;

[0022] 7. Compare the authorized settlement letter and the system;

[0023] 8. Compare the paper agreement and the system, etc.

[0024] Therefore, the review of merchant information has always been heavily dependent on manpower, resulting in a significant increase in the cost of the platform.

[0025] On the other hand, with the strong rise of AI model technology, AI model technology has penetrated into multiple core links in the financial industry, significantly improving efficiency and intelligent level. For example, AI models have been widely used in risk assessment of financial transactions, financial credit approval, quantitative trading and investment decision-making, investment research and information processing, etc. However, in the platform merchant audit aspect, there is no special model trained and developed to assist in auditing.

[0026] Therefore, there is a demand to provide a scheme capable of assisting in auditing merchant information based on AI model to realize efficient auditing of merchant information. SUMMARY

[0027] The present application relates to a scheme capable of assisting in auditing merchant information based on AI model to realize efficient auditing of merchant information.

[0028] According to a first aspect of the present application, a system for assisting in auditing information based on AI model is provided, comprising:

[0029] an information auditing assistance module, the information auditing assistance module comprising:

[0030] a pre-audit intelligent agent configured to check whether the information submitted by the merchant is complete and to extract data from the submitted information;

[0031] a risk audit intelligent agent configured to verify the authenticity, consistency and risk of the data based on the extracted data;

[0032] a compliance audit intelligent agent configured to verify whether the data is legal and compliant based on the extracted data;

[0033] an artificial audit module configured to manually review the audit results from the pre-audit intelligent agent, the risk audit intelligent agent and the compliance audit intelligent agent; and

[0034] a routing module configured to select an AI model for constructing a corresponding sub-intelligent agent from a plurality of AI models according to an audit task to be performed.

[0035] According to a second aspect of the present application, a method for assisting in auditing information based on AI model using the system of the first aspect is provided, comprising:

[0036] the merchant submits information to the payment platform to request for audit;

[0037] the pre-audit intelligent agent checks whether the information submitted by the merchant is complete, extracts data from the submitted information, and forwards the audit and extraction results to the artificial audit module;

[0038] The risk audit intelligence verifies the authenticity, consistency and risk of the extracted data, and forwards the risk audit result to the manual audit module;

[0039] The compliance audit intelligence verifies whether the extracted data is legal and compliant, and forwards the compliance audit result to the manual audit module;

[0040] The manual audit module receives the audit results forwarded from the pre-audit intelligence agent, the risk audit intelligence agent and the compliance audit intelligence agent respectively, and manually reviews them; and

[0041] The manual audit module generates a merchant audit report according to the result of manual review, and returns it to the merchant.

[0042] This summary is provided to introduce some concepts in a simplified form that are further described below in the detailed description. This summary is not intended to identify key or essential features of the claimed subject matter, nor is it intended to be used to limit the scope of the claimed subject matter. BRIEF DESCRIPTION OF DRAWINGS

[0043] In order to describe the manner in which the above-recited and other advantages and features of the application can be obtained, a more particular description will be rendered by reference to specific embodiments thereof which are illustrated in the appended drawings. Understanding that these drawings depict only typical embodiments of the application and are not therefore to be considered to be limiting of its scope, the application will be described and explained with additional specificity and detail by the use of the accompanying drawings in which:

[0044] Figure 1 A schematic structural diagram of a system for assisting in auditing materials based on an AI model is shown according to one embodiment of the present application.

[0045] Figure 2 An example interface of API configuration of an AI model is shown according to one embodiment of the present application.

[0046] Figure 3 A schematic flow chart of a method for assisting in auditing materials based on an AI model is shown according to one embodiment of the present application. DETAILED DESCRIPTION

[0047] In general, the method and system for assisting in auditing materials based on an AI model of the present application can achieve the following three improvements:

[0048] 1) In the scheme of the present application, the current mainstream AI models are used to assist in auditing merchant information. And for each link of the merchant information audit, a corresponding routing mechanism is put forward to ensure that the most suitable AI model can be used to assist in auditing at each link, thereby improving the efficiency and accuracy of the audit.

[0049] First of all, thanks to the popularity of artificial intelligence technology, many domestic Internet companies have developed their own AI model tools and released them to users for free. For example, Alibaba has launched the Tongyi series (such as Qwen3) AI model, Tencent has launched the Hunyuan series large model, and ByteDance has launched the Doubao AI model, etc.

[0050] These models developed by these companies have their own advantages and disadvantages. Some are good at processing text; some are better at processing pictures and videos; some have low training cost and fast speed, suitable for enterprise deployment; and some are more focused on scenario services.

[0051] This leads to the fact that it is not efficient to use only one enterprise's AI model platform to complete all the auxiliary operations in the merchant information audit. Therefore, in the present application, a routing mechanism is provided to allow the platform's merchant information audit system to use the most suitable AI model for each audit task in the merchant audit to assist in auditing, thereby improving the efficiency and accuracy of the audit.

[0052] 2) Use an integrated tool such as "Agent" + "RAG" + "Fine-tuning" to replace the traditional pre-training of AI models, thereby saving a lot of training time and training cost.

[0053] As we all know, in traditional technology, in order for an AI model to successfully replace human labor to automatically complete some tasks, it is necessary to first train the AI model with training data. Only trained AI models can give accurate results. Therefore, in order to improve the work efficiency of AI models, high-quality training data is needed for extensive training. Undoubtedly, such training is costly and time-consuming.

[0054] In the present application, an integrated tool containing "Agent" (preferably plus "RAG" and "Fine-tuning") is proposed to enable AI models to understand what you say without training and complete the corresponding tasks according to your words. This undoubtedly greatly speeds up the development process and saves training costs.

[0055] 3) Accurate extraction of data for various forms of merchant information.

[0056] The information submitted by the merchant includes many forms, such as paper documents, photocopies, photos, electronic documents, digital images and videos, etc. The existing various identification technologies, such as OCR technology, have various defects in identifying corresponding data from these files, especially in scenarios such as identification and splicing of paginated text, identification of text with ink interference lines, identification of tables in documents with mixed text and graphics, etc. The accuracy of identification will decrease sharply.

[0057] In view of the defects, the present application adopts a corresponding improved data extraction tool to overcome these problems, thereby improving the accuracy of data extraction.

[0058] The following will be further described in combination with the accompanying Figure 1 to further illustrate the schematic structural diagram of a system for assisting in auditing information based on an AI model according to an embodiment of the present application.

[0059] As shown in the figure, a merchant information auditing system and the modules contained therein are disclosed.

[0060] The merchant information auditing system includes an information auditing assistance module 110, an artificial auditing module 120 and a routing module 130. These modules can exchange data with each other through various wired / wireless communication technologies.

[0061] The information auditing assistance module 110 is composed of multiple agents based on AI models, and is configured to assist in auditing the information of the merchant by using AI models.

[0062] The "agent" refers to an autonomous decision-making entity simulating the collaboration mode of human experts, and is a computing entity with autonomous perception, decision-making and execution capabilities, which can complete specific tasks independently or with limited human intervention. In the context of the merchant information auditing system of the present application, the agent can be understood as an automated auxiliary auditing entity equipped with AI algorithms.

[0063] In many AI model tools provided by Internet platforms, users are allowed to create their own agents through API configuration (as shown in Figure 2 By inputting corresponding prompt words in the text box of the API configuration, adding various custom variables and calling the model template, an agent that can automatically complete the specified task can be constructed.

[0064] As mentioned above, there are currently about 80 audit rules related to merchant information review, that is, there are about 80 tasks that need to be reviewed. Some of these tasks involve the completeness of the submitted information, such as whether the corresponding files are missing, whether the format of the files meets the requirements; some tasks involve the authenticity of the submitted information, such as whether the legal identity information is real, whether the merchant address is valid, etc.; some tasks involve whether the information is compliant, such as whether the store address is a commercial or residential address, whether the account name of the settlement personnel is the legal person, etc. Therefore, in the actual merchant information review process, according to the nature of the review task, it can be divided into three links, including: information pre-audit (also known as "primary review") review, information risk review, and information compliance review.

[0065] Correspondingly, for these links, by setting the prompt words, variables and templates of various audit tasks involved in these links in the API configuration function of the AI model on the Internet platform as described above, the corresponding sub-agents can be constructed, and then by integrating all sub-agents in the same link, the pre-audit review agent 112, the risk review agent 114 and the compliance review agent 116 can be constructed. Each agent is a collection of configured sub-agents of the tasks of the corresponding link.

[0066] Specifically: for each audit task, the corresponding sub-agent can be constructed by inputting the corresponding prompt words, variables and templates in the API configuration interface as shown in the following figure. Figure 2

[0067] Although the existing AI model has developed many template tools that can be used for the completion of some audit tasks, such as OCR recognition tools, identity card authenticity verification tools, PS trace recognition tools, etc. In the API configuration process, many audit tasks can actually complete the task by simply calling the corresponding template tools of the model, which greatly facilitates the configuration for the tasks. However, since these template tools are not specifically developed for the merchant information review of the third-party payment platform, therefore, the result of directly using these template tools for automatic review may have a high error rate, and the efficiency is not high.

[0068] In order to make the original template tools provided by the AI model more efficient and accurate, the corresponding prompt words and parameters also need to be set to let the model understand the true purpose of the audit task, and in the Figure 2 part of the prompt words for the "identity card background consistency" audit task are given. By adding these prompt words and parameters to provide more conditions and details for the model, the working efficiency and accuracy of the sub-agent can be further improved.

[0069] ​The prompt words are constructed based on various factors that a technician needs to consider when manually performing the audit task, so different audit tasks have different prompt words. Moreover, with the modification of the corresponding legal norms, the prompt words also need to be changed accordingly, and not remain unchanged.

[0070] While the parameters can be divided into input / output specification parameters and performance tuning parameters. The former is used to limit the variable type (text / number / JSON, etc.), to avoid format errors that cause process interruption, and the latter can be used to fine-tune the AI model and the called template, to make the audit result more accurate.

[0071] The pre-audit audit agent 112 is configured to check whether the information submitted by the merchant is complete (i.e. integrity audit) and to extract data from the submitted information, so it contains a sub-agent that verifies the audit rules related to information integrity and the corresponding data extraction tools.

[0072] In order to open a transaction account on a third-party payment platform, a merchant must first submit various information required by the platform, which can be divided into the following categories:

[0073] 1) Basic qualification documents, including:

[0074] ‌Business license: must be within the valid period, and the scope of business is related to payment business.

[0075] ‌ID card of legal person: clear scanned copy of front and back, valid for at least 1 month.

[0076] ‌Bank account information: for enterprises: bank account information for public; for individual businesses: bank account of legal person or public account.

[0077] …

[0078] 2) Industry-specific qualifications, including:

[0079] ‌Financial merchants: need to provide financial qualification certificates such as Payment Business License.

[0080] ‌Food and beverage merchants: need to provide Food Business License.

[0081] ‌Network service merchants: need ICP license or EDI license.

[0082] …

[0083] 3) Proof of business premises, including:

[0084] ‌Store photos: require photos that reflect the true business premises, including storefront photos, interior photos, cash register photos, etc.

[0085] ‌Proof of address: Lease contract or property deed, accurate to the house number.

[0086] …

[0087] 4) Other supplementary materials, including:

[0088] ‌Merchant management specification: Contains transaction scenario description and user protection measures.

[0089] ‌System function proof: Need to show screenshots of the platform merchant management system.

[0090] ‌Special industry proof: Such as education needs to be licensed, etc.

[0091] …

[0092] Each type of material has corresponding form requirements, for example, some materials must be original, not copies or scans. Some materials such as photos must be in color and have a resolution that meets the requirements.

[0093] If the merchant's submitted materials are not complete or do not meet the specifications, the pre-audit review agent 112 can automatically identify missing items and generate a list of materials to be submitted to the merchant. In this way, the merchant can resubmit the missing materials according to the list.

[0094] In order for the pre-audit review agent 112 to have the verification function, as previously described, the API configuration of each sub-agent of the review task (each review task can correspond to one or more related review rules) included in the agent can be added with the prompt words, parameters and templates corresponding to the review rules of the task to inform the model what materials are required and what form of materials is acceptable. After the user gives enough prompts, the pre-audit review agent 112 can use the analysis capabilities and tools of the AI model itself to automatically verify whether the materials submitted by the merchant meet the requirements for submission.

[0095] After the merchant's materials are verified, the effective data in various forms of merchant materials needs to be accurately extracted for review.

[0096] As previously described, the user's submitted materials can be in various forms, such as paper files, physical photos, electronic documents, digital pictures, etc.

[0097] For different forms of materials, there may be different problems when extracting effective data from them, and various corresponding data extraction tools need to be used to ensure the accuracy of the extracted effective data.

[0098] For example, for general paper documents such as photocopies of identity cards, business licenses, etc., the existing OCR technology can be directly used to extract the key fields in the business license and identity card, or the paper documents can be photographed first and then extracted through OCR technology. Because these certificates all use a unified format template, and the AI models provided by each platform generally have developed an OCR module specially used for identifying these certificates, therefore, by calling these modules, the accuracy of identification can reach more than 98%.

[0099] However, some paper documents may not be satisfactory when directly identified by existing OCR tools.

[0100] For example, the original paper agreement is an A3 paper with a large amount of agreement text information, among which the deduction rate information is located in the middle and lower part of the right table and is filled in by the user's handwriting. After the user takes a photo of the paper agreement and uploads the photo, if the existing OCR identification module is directly called, the following problems may be found:

[0101] 1. Some people take pictures separately on the left and right sides and then upload them, and some people take pictures as a whole and then upload them.

[0102] 2. Some uploaded photos are upside down.

[0103] 3. Handwriting recognition in the table is difficult.

[0104] To solve these problems in this scenario, the pre-audit intelligent agent 112 of the present application further provides the following data extraction tools (means):

[0105] For the first problem, the titles on the left and right sides of the agreement (bold font) can be identified to determine whether the picture is taken separately or as a whole.

[0106] For the second problem, the OCR scanning can be used to determine the rotation angle of the text in the photo, and then the text in the photo is executed in reverse and then transmitted to the large model.

[0107] For the third problem, the cause of this situation is analyzed. After analysis, it is found that when the picture text content is large, the accuracy of the large model in identifying the information in the table will decrease, especially for handwritten text, the accuracy will be even lower.

[0108] According to the analysis result, the applicant finds through repeated experiments that as long as the table is cut out and then transmitted to the large model alone, the accuracy of the table can be greatly improved.

[0109] Then, the next problem is how to accurately crop the table from the photo.

[0110] The most easily thought of solution to this problem is that a large model can easily locate the coordinates of the table, and then cut according to the coordinates. But in actual application, it is found that the accuracy of cutting the table by this means is not very high.

[0111] Later, after continuous attempts, the present application adopts a new mechanism to identify the table coordinates, as follows:

[0112] Using image processing techniques, such as the open source computer vision and machine learning software library OpenCV, the image is first preprocessed to convert it to a grayscale image, then the line is made more prominent through binarization technology, and then noise reduction is used to enhance the line and remove small noise points. After preprocessing, Hough transform is used to detect straight lines, and after straight line detection, horizontal and vertical lines are separated, and finally the leftmost, rightmost, uppermost and lowermost straight lines are taken as the coordinates of the table.

[0113] Through the above image processing technique, the accuracy of cutting the table from the photo has been significantly improved, but it is not perfect. Because in actual recognition, it is found that this image processing technique brings new problems:

[0114] Because the text in the A3 paper is printed, some of the agreement content will have ink interference lines (straight lines appear randomly), which eventually leads to the processing technology considering the ink interference lines as the outer frame lines of the table, making the recognition much larger, which leads to a decrease in the accuracy of data extraction of the table's rate information.

[0115] To solve this problem, the present application is further improved: Since the agreement used is a fixed template, the straight line coordinates of the table identified by the image processing technique can be roughly checked first, and the obvious interference lines that cannot be the table are filtered out as noise. At the same time, a rough judgment is made on the text around the line, which is also a noise filtering recognition point. For example, if the title of the table header in the left upper corner is "Business Information", it means that when this text is recognized, the nearest line to the left and above it should be the left and upper outer frame lines of the table. By combining the title text in the table to identify the frame, the accuracy of identifying the outer frame of the table is greatly improved.

[0116] By providing these improved means in the form of templates or tools to the intelligent agent, it can be found that the accuracy of data extraction of the agreement rate information has been greatly improved, from about 78% to 98%.

[0117] The above is only a data extraction tool (means) developed for one form (table). There are other data extraction means to help extract data from the materials, such as if the submitted store interior photo is blurred (insufficient light), the store interior arrangement and background wall can be distinguished by first adjusting the brightness and contrast. If the customer submits a storefront photo with a too-small storefront subject, making it difficult to see the store number, the storefront subject can be cropped and enlarged before recognizing the store number, … These data extraction means that improve recognition can be implemented by converting the operations and experiences of manual execution of the corresponding audit rules into corresponding prompt words and inputting them into the API configuration of the corresponding sub-agent, wherein the tools used can be added to the sub-agent in the form of a model calling template. In this way, the auditor only needs to adjust the appropriate parameters according to the actual situation to let the sub-agent automatically perform the task without having to repeatedly input the prompt words and template information each time. For example, when auditing a store interior photo, when the submitted store interior photo is blurred, the auditor only needs to set the brightness and contrast parameters according to the actual situation of the photo, and the sub-agent can first call the image processing tool to perform brightness and contrast adjustment on the photo, and then perform image recognition to extract valid data. This is more effective than directly using an OCR tool. Even in some AI models that are advantageous in image processing, the brightness and contrast parameters can be automatically selected by the model through preliminary analysis of the photo, without manual setting.

[0118] By introducing corresponding data extraction tools / means (such as OpenCV, image processing tools) when extracting data from submitted materials, the data extraction efficiency and accuracy of the pre-audit audit agent 112 can be greatly improved.

[0119] It should be understood that even the most powerful agent cannot achieve 100% accuracy in recognizing text images, therefore, the pre-audit audit agent 112 can only be used as an auxiliary means for pre-audit, and the audit results of the material integrity and the extracted data are submitted to the manual audit module 120 for review. However, this can reduce more than 90% of the workload of manual audit. Moreover, with the development of technology, the accuracy of data extraction will continue to improve, further reducing labor consumption.

[0120] The risk audit agent 114 is configured to verify the authenticity, consistency, and risk of the data based on the data extracted from the submitted materials by the pre-audit audit agent 112. The risk audit agent 114 includes sub-agents that verify audit rules related to the authenticity, consistency, and risk of the data, and corresponding verification tools (such as a personal identity verification module (such as a face recognition module), an enterprise identity verification module that accesses the “National Enterprise Credit Information Publicity System”).

[0121] The authenticity of the data refers to verifying whether the information provided by the merchant is real and reliable. For example, whether the data extracted from the ID card is real and valid, whether the business license provided by the merchant is real, and the like. This can be achieved by calling the relevant tools of the AI model (such as the personal identity verification module) or accessing the relevant website (such as the national enterprise credit information public system). The authenticity verification ensures that the information submitted by the merchant is real and valid.

[0122] The consistency of the data refers to verifying whether the relevant data in the information submitted by the merchant is consistent. For example, whether the scope of business in the business license is consistent with the MCC applied for by the merchant, whether the name of the ID card is consistent with the name of the legal person in the business license, whether the business address filled in is consistent with the address on the business license, whether the store name filled in is consistent with the name on the store photo, whether the personal settlement account is completely consistent with the name of the legal ID card, and the like.

[0123] The risk of the data refers to the credit investigation of individual merchants, the credit record review of enterprise merchants, whether the submitted pictures have PS traces, and the like.

[0124] The above-mentioned authenticity, consistency and risk of the data can be achieved by converting the operation and experience when manually performing the corresponding audit rules into corresponding prompt words and inputting them into the API configuration of the corresponding sub-agent. Among them, the verification tool, the network verification module and / or the PS trace identification tool can be added to the sub-agent in the form of calling the model template.

[0125] It should be understood that the risk audit agent 114 cannot achieve 100% accuracy when auditing, therefore, the risk audit agent 114 can only be used as an auxiliary means for risk audit, and the final audit result data still needs to be submitted to the manual audit module 120 for review. However, this auxiliary audit can greatly reduce the workload of manual audit and speed up the audit. Moreover, with the development of technology, the accuracy of the risk audit will continue to improve, further reducing the labor consumption.

[0126] The compliance audit agent 116 is configured to verify whether the data is legal and compliant based on the data extracted from the submitted information by the pre-audit agent 112. The compliance audit agent 116 includes sub-agents of the audit rules related to verifying whether the data is legal and whether the data is compliant, and corresponding verification tools.

[0127] The verification of whether the data is legal and compliant can include whether the merchant name contains misleading or prohibited words, whether the business license is within the valid period, whether the legal person's ID is valid for more than one month, whether there is an authorization letter when the signatory is not the legal representative, whether the special industry merchant holds the corresponding industry license, etc.

[0128] Similarly, the legal and compliant audit of the data can be achieved by converting the operations and experiences in manually performing the corresponding audit rules into corresponding prompt words and inputting them into the API configuration of the corresponding sub-agent. Among them, the verification tools can be added to the sub-agent in the form of a template calling the model.

[0129] It should be understood that since the relevant regulations and rules are amended every year, the corresponding prompt words of each sub-agent in the compliance audit agent 116 also need to be updated to adapt to the latest legal provisions.

[0130] It should be understood that the compliance audit agent 116 cannot achieve 100% accuracy when auditing, so the qualified audit agent 116 can only be used as an auxiliary means for compliance audit, and the final audit result data still needs to be submitted to the manual audit module 120 for review. But this kind of auxiliary audit can greatly reduce the workload of manual audit and speed up the audit. Moreover, with the development of technology, the accuracy of the compliance audit will continue to improve, further reducing the consumption of manpower.

[0131] The manual audit module 120 is configured to manually review the audit result data from the pre-audit agent 112, the risk audit agent 114 and the compliance audit agent 116 to review the accuracy of the audit result. And in the preferred embodiment, the manual audit module 120 is further configured to adjust the prompt words, parameters and templates of the sub-agent performing the corresponding audit task according to the review results, so as to further improve its audit ability.

[0132] For example, if the sub-agent performing the identification of the paper protocol in A3 format containing the protocol rate information table as described above has a low identification accuracy of the rate information in the table, the manual audit module can improve the identification accuracy of the sub-agent by adjusting the "text around the line" setting, adjusting the parameters in OpenCV, etc.

[0133] The manual audit module 120 generates a merchant audit report according to the results of manual review and returns it to the merchant. The merchant audit report includes various problems identified by the pre-audit agent 112, the risk audit agent 114 and the compliance audit agent 116 and reviewed, as well as suggestions for solving the problems.

[0134] The routing module 130 is configured to select a suitable AI model from a plurality of AI model tools for constructing a corresponding agent according to an audit task (i.e., an audit rule) to be performed.

[0135] As mentioned above, due to the differences in research and development capabilities and areas of focus of various enterprises, the AI model tools they launch have their respective performance advantages and disadvantages. Some have strong text processing capabilities; some are better at processing pictures and videos; some have low training costs and fast speeds, making them suitable for enterprise deployment; and some are more scene-oriented. There is currently no single AI model that can handle all scenarios.

[0136] On the other hand, in the audit of merchant information, various application scenarios such as character recognition, image recognition, video authentication, and picture comparison are involved. Therefore, if only the AI model tools of a certain Internet platform are focused on, it may lead to the embarrassing situation that the audit accuracy is very high in some scenarios, but very low in other scenarios.

[0137] To solve the above problems, the present application further provides a routing module 130. The routing module 130 adopts a microservice architecture design, separates the routing function into a middleware layer, and connects multiple AI model platforms through a unified API gateway. At the same time, a model performance monitoring system is built in to monitor the response time, accuracy, and other key indicator data of each model in real time and store them in a log database, so as to accumulate historical audit data and performance of each model. In addition, the routing module 130 also supports a disaster recovery mechanism that switches to other large models when a selected large model is abnormal.

[0138] The routing module 130 provides a routing switching mechanism. The switching mechanism mainly determines which AI model is suitable for constructing a corresponding sub-agent based on the characteristics of the audit task to be performed, the performance of each AI model, and / or its historical indicator data in performing the audit task.

[0139] Specifically, first, the routing module 130 receives a request from the information audit assistance module 110 to access the AI model of an Internet platform to perform a certain audit task therein.

[0140] Based on the request, the routing module 130 determines the characteristics of the audit task to be performed (i.e., the nature of the audit task, such as character recognition, image recognition, or identity authentication).

[0141] According to the characteristics, the routing module 130 selects one or more candidate models that are good at handling the audit task from a plurality of AI models, for example, based on the performance of the AI model in other (involving text, image, audio and video analysis and recognition) application scenarios.

[0142] Subsequently, the routing module 130 retrieves the historical indicator data of the candidate models from the log database, which includes the historical response time of the models in performing the audit task, the historical audit accuracy rate of the models in performing the audit task, the historical risk omission rate of the models in performing the audit task, and so on.

[0143] Finally, the routing module 130 selects the most suitable AI model based on the comprehensive analysis of the historical indicator data (for example, scoring each candidate model using a weighted summation algorithm), and interfaces the audit task with the selected AI model to build and use a sub-agent to automatically perform the audit.

[0144] In addition, when the selected AI model encounters an abnormal situation in performing the audit task, the routing module 130 will also switch the audit task from the abnormal AI model to another candidate model (usually the second highest scoring model) to ensure the smooth completion of the task audit.

[0145] In addition to building the agent, in order to achieve efficient material audit without the need for a large amount of model training, in the preferred embodiment, the present application further introduces the "RAG" and "fine-tuning" technologies.

[0146] RAG, which stands for Knowledge Augmented Retrieval Generation, provides a real-time updated database for large models. In this way, AI models with RAG function not only rely on outdated information, but also actively search for the latest and most relevant data to answer customer questions. Unlike traditional LLMs that generate responses based solely on pre-trained knowledge, RAG retrieves relevant information from external databases or documents and uses this information to generate more accurate and contextually appropriate responses. This approach ensures that the output is based on the latest and most authoritative data, making it particularly useful for applications that require up-to-date information. It combines the extensive knowledge base of AI models with the accuracy of real-time data retrieval. This not only provides accurate responses, but also perfectly adapts to the context.

[0147] The RAG technology mainly includes an information retrieval (IR) system and a natural language generation (NLG) model.

[0148] The IR system uses advanced search algorithms to scan large datasets and quickly find relevant information. These algorithms use semantic search techniques that go beyond simple keyword matching, allowing them to understand the context and meaning of queries. The RAG system can retrieve data from multiple sources, including internal databases, online repositories, and real-time web searches. The quality and range of these data sources are crucial to the effectiveness of the system.

[0149] NLG components rely on powerful language models, such as GPT-3, which are capable of generating human-like text. These models are trained on vast datasets, enabling them to create coherent and contextually relevant responses. NLG involves multiple techniques to ensure that the generated text is relevant and accurate. This includes fine-tuning models with domain-specific data and using advanced text generation algorithms.

[0150] With the RAG technology, the following advantages can be achieved in the review of merchant information:

[0151] 1) RAG can achieve real-time review standard updates by connecting to an external regulatory policy library in real time, without the need to retrain the model to adapt to the latest regulatory requirements. For example, when the rules for business registration change, the knowledge base of RAG can be updated to take effect.

[0152] 2) Multi-source data verification is achieved, that is, by simultaneously retrieving multi-dimensional data such as business registration information, credit records, and operating permits of the merchant, cross-verification is performed to improve the accuracy of the review.

[0153] 3) By automatically comparing the merchant's submitted information with the latest requirements of industry standards, potential compliance risk points are identified.

[0154] Fine-tuning, also commonly known as fine-tuning, is a core concept in the field of deep learning. It mainly involves adjusting the parameters of a pre-trained general-purpose large model to improve its performance on specific tasks, enabling it to more professionally handle various challenges. In this application, the fine-tuning refers to the artificial review module 120 continuously adjusting the relevant parameters in the prompt words according to the review of the agent's review of the data to improve the accuracy of the review.

[0155] For example, by fine-tuning the labeled data of merchant information (such as business license, operating category, qualification documents, etc.) in the prompt words, the large model can master industry terminology, review rules, and anomaly pattern recognition capabilities, improving its sensitivity to specific scenarios. Fine-tune the model output for specific review tasks (such as authenticity verification, compliance judgment), reduce the "illusion" or ambiguous answers of the general model, and ensure that the results meet business standards.

[0156] Compared with full training, fine-tuning only needs a small amount of data to take effect, reduces the computing cost, and is suitable for rapid iteration of audit strategies. By making the pre-audit agent 112, the risk audit agent 114, and the compliance audit agent 116 in the merchant information audit auxiliary module 110 interface with the appropriate AI model through the scheduling of the routing module 130 to assist in auditing the merchant information, and then reviewing the assisted audit results through the manual audit module, the accurate audit of the merchant information can be realized while reducing the labor consumption, greatly reducing the cost of the payment company audit and improving the audit efficiency.

[0157] Based on the above system structure, the schematic flowchart of a method for assisting in auditing information based on an AI model according to an embodiment of the present application is further illustrated below in combination with the accompanying Figure 3

[0158] As shown in the figure, first, in the initialization stage:

[0159] In step 302, the merchant information audit auxiliary module 110 analyzes the process of auditing the merchant information of the payment platform to determine the audit tasks contained in the pre-audit agent 112, the risk audit agent 114, and the compliance audit agent 116.

[0160] In step 304, for each audit task, the routing module 130 selects the AI model to be interfaced from the plurality of AI models according to the characteristics of the audit task, the performance of the AI model, and / or the historical index data.

[0161] Subsequently, in step 306, for each audit task, the corresponding sub-agent is generated by setting the prompt words, parameters, and templates of the API configuration of the selected AI model.

[0162] Finally, in step 308, the corresponding pre-audit agent 112, risk audit agent 114, and compliance audit agent 116 are constructed by integrating the sub-agents of the same link.

[0163] Subsequently, in the audit stage:

[0164] In step 310, the merchant submits information to the payment platform to request for audit.

[0165] Subsequently, in step 312, the pre-audit agent 112 checks whether the information submitted by the merchant is complete (i.e., integrity audit).

[0166] If the information is complete, in step 314, the pre-audit agent 112 performs data extraction on the submitted information, and forwards the pre-audit result and the extracted data to the manual audit module 120, and the process proceeds to step 316.

[0167] ​If the information is missing, the process proceeds to step 318, in which the pre-audit review intelligent agent 112 generates a list of missing information based on the missing information and returns the list to the merchant.

[0168] In step 316, the risk review intelligent agent 114 verifies the authenticity, consistency and risk of the data based on the data extracted from the submitted information by the pre-audit review intelligent agent 112.

[0169] Subsequently, in step 320, the risk review intelligent agent 114 forwards the risk review result to the manual review module 120, and the process proceeds to step 322.

[0170] Then, in step 322, the compliance review intelligent agent 116 verifies whether the data is legal and compliant based on the data extracted from the submitted information by the pre-audit review intelligent agent 112.

[0171] Then, in step 324, the compliance review intelligent agent 116 forwards the compliance review result to the manual review module 120.

[0172] In the review stage:

[0173] In step 326, the manual review module 120 receives the review results forwarded from the pre-audit review intelligent agent 112, the risk review intelligent agent 114 and the compliance review intelligent agent 116 respectively, and manually reviews them.

[0174] In step 328, the manual review module 120 generates a merchant review report based on the results of the manual review, and returns it to the merchant. The merchant review report includes various problems identified by the pre-audit review intelligent agent 112, the risk review intelligent agent 114 and the compliance review intelligent agent 116 after review and review, and suggestions for solving the problems.

[0175] In step 330, the manual review module 120 adjusts the prompt words, parameters and templates of the sub-intelligent agents performing the corresponding review tasks according to the results of the manual review, so as to further improve their review capabilities.

[0176] In a preferred embodiment, while the merchant information is being reviewed in the review stage, the routing module 130 also monitors the response time, accuracy and other key indicators of each AI model when performing information review in real time, so as to accumulate historical indicator data and performance of each model. And when the routing module 130 monitors that the selected AI model is abnormal, switch the pending review task to other candidate AI model to realize disaster recovery.

[0177] In another preferred embodiment, in addition to using the agent in the review stage, the "RAG" and "tuning" techniques are further introduced in the review to achieve efficient material review without the need for a large amount of model training.

[0178] Thus far, the method of assisting in reviewing materials based on the AI model ends.

[0179] Although the technology has been described in language specific to structural features and / or methodological acts, it is to be understood that the appended claims are not necessarily limited to the features or acts described. Rather, the features and acts are described as example forms of implementing the claims.

[0180] The operations of the example processes are illustrated in individual blocks and summarized with reference to those blocks. The processes are illustrated as logical flow graphs, each operation of which can represent one or more operations that can be implemented in hardware, software, or a combination thereof. In the context of software, the operations represent computer-executable instructions stored on one or more computer-readable media that, when executed by one or more processors, cause one or more processors to perform the recited operations. Generally, computer-executable instructions include routines, programs, objects, modules, components, data structures, and the like that perform particular functions or implement particular abstract data types. The order in which the operations are described is not intended to be construed as a limitation, and any number of the described operations can be executed in any order, combined in any order, subdivided into multiple sub-operations, and / or executed in parallel to implement the described processes. The described processes can be performed by resources associated with one or more computing devices, such as one or more internal or external CPUs or GPUs, and / or one or more segments of hardware logic, such as FPGAs, DSPs, or other types of accelerators.

[0181] All of the methods and processes described above can be embodied in, and fully automated via, software code modules executed by one or more general purpose computers or processors. Such code modules can be stored in any type of computer-readable storage medium or other computer storage device. The code modules can also be encoded in a computer program product embodied on a computer readable medium. Some or all of the methods can alternatively be embodied in specialized computer hardware.

[0182] Any routine descriptions, elements or blocks in the flow diagrams described herein and / or depicted in the attached figures should be understood as potentially representing modules, segments, or portions of code that include one or more executable instructions for implementing specific logical functions or elements in the routine. Alternate implementations are included within the scope of examples described herein in which elements or functions can be deleted, or implemented in differing order from as described and shown, including substantially synchronously or in reverse order, depending on the functionality involved. Such alternate implementations should be understood as falling within the scope of the examples described herein.

[0183] While the foregoing describes different embodiments, it is understood that they are by way of example only and are not limiting. Those skilled in the relevant art will appreciate that modifications can be made without departing from the spirit and scope of the application as defined by the appended claims. Therefore, the breadth and scope of the application disclosed herein should not be limited by any of the above-described exemplary embodiments, but should be defined in accordance with the following claims and their equivalents.

Claims

1. A system for assisting in reviewing documents based on an AI model, comprising: A data review auxiliary module, the data review auxiliary module includes: The pre-audit and review agent is configured to verify whether the information submitted by the merchant is complete and to extract data from the submitted information; A risk audit agent configured to verify the authenticity, consistency and risk of the data based on the extracted data; A compliance audit agent, configured to verify whether the data is legal and compliant based on the extracted data; A manual review module configured to manually review the review results from the pre-review review agent, the risk review agent, and the compliance review agent; and The routing module is configured to select an AI model for constructing a corresponding sub-agent from multiple AI models based on the audit task to be performed.

2. The system according to claim 1, wherein The construction of the corresponding sub-agent includes: The sub-agent that performs the audit task is constructed by adding prompt words, variables and templates corresponding to the audit task in the API configuration of the selected AI model.

3. The system according to claim 1, wherein The pre-qualification and review agent is also configured to automatically identify missing information and generate a list of supplementary information to be returned to the merchant.

4. The system according to claim 1, wherein: In addition to the intelligent agent, the data review auxiliary module also uses RAG and fine-tuning technology; The RAG updates audit standards in real time by connecting to external regulatory policy libraries; performs cross-validation by simultaneously retrieving multi-dimensional data; and identifies potential compliance risks by automatically comparing merchant-submitted information with the latest industry standards. The fine-tuning refers to the manual review module fine-tuning the relevant parameters in the prompt words to improve the review accuracy based on the review results of each intelligent agent.

5. The system according to claim 1, wherein: The routing module has a built-in model performance monitoring system to monitor the indicator data of each AI model in real time and store the indicator data in a log database.

6. The system according to claim 5, wherein: The routing module selects an AI model to be connected from the multiple AI models based on the characteristics of the audit task, the performance of the AI ​​model and / or the historical indicator data of the AI ​​model.

7. The system according to claim 5, wherein: The routing module is also configured to switch the pending audit tasks to other AI models when an abnormality is detected in the selected AI model.

8. A method for assisting in reviewing documents using an AI model based on the system according to any one of claims 1 to 7, comprising: Merchants submit information to the payment platform to request entry review; The pre-audit intelligent entity verifies whether the information submitted by the merchant is complete, extracts data from the submitted information, and forwards the audit results and extraction results to the manual review module; The risk audit intelligence verifies the authenticity, consistency and risk of the data based on the extracted data, and forwards the risk audit results to the manual audit module; The compliance audit agent verifies whether the data is legal and compliant based on the extracted data, and forwards the compliance audit results to the manual audit module; The manual review module receives the review results and the extraction results forwarded from the pre-review review agent, the risk review agent, and the compliance review agent, respectively, and performs manual review on them; as well as The manual review module generates a merchant review report based on the results of the manual review and returns it to the merchant.

9. The method according to claim 8, wherein Also includes: The manual review module adjusts the prompt words, parameters and templates of the sub-agent performing the corresponding review task according to the results of the manual review.

10. The method according to claim 8, wherein Also includes: When the routing module detects an abnormality in the selected AI model, it switches the pending audit task to other AI models.

Images