Intelligent payment service auditing system, method, device and equipment

Abstract

The invention discloses an intelligent payment service auditing system, method, device and equipment, belongs to the technical field of data processing, and aims to improve the auditing efficiency and accuracy of payment services. The method comprises the following steps: receiving material information of a to-be-audited payment service sent by a client; calling an agent matched with each auditing dimension according to a plurality of auditing dimensions involved in the material information, and sending the material information to each agent in parallel; receiving auditing sub-results generated by each agent for independent analysis of the auditing dimension matched with the agent; when it is detected that divergence exists between the audit sub-results, a debate instruction is issued to each agent so as to control the agents to conduct debate; and receiving the review sub-result of each agent after debate correction, and generating a review result based on each corrected review sub-result.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to an intelligent payment business review system, method, apparatus and equipment. Background Technology

[0002] With the rapid development of the insurance industry, the volume of insurance payouts for various insurance products such as medical insurance, accident insurance, and critical illness insurance has continued to rise, and the payout scenarios have become increasingly complex and diverse. In the traditional insurance payout process, users need to upload multiple application materials, including medical bills, diagnostic certificates, and identity documents. The completeness, authenticity, and compliance of these materials are then manually reviewed item by item to ensure accuracy. For example, checking whether the invoice amount matches the medical record and whether the examination and treatment dates are within the policy's validity period.

[0003] However, with the surge in the number of insurance claim applications, the manual review method has revealed significant shortcomings: on the one hand, it is inefficient and cannot meet the needs of large-scale, high-concurrency business; on the other hand, since the review standards rely on personal experience, there are problems such as strong subjectivity and poor consistency, which can easily lead to misjudgment or omission, thereby affecting customer experience and risk control quality.

[0004] To improve automation, intelligent payment review systems based on rule matching or machine learning have emerged. For example, some systems use optical character recognition (OCR) technology to extract image information of invoices uploaded by users and combine it with preset business rules for review to determine whether to approve insurance payments. However, these methods still have significant shortcomings: rule-based systems are limited by fixed rule sets, making it difficult to cover the ever-evolving new insurance payment scenarios and exhibiting weak generalization ability; while machine learning-based methods, although possessing certain learning capabilities, heavily rely on large amounts of high-quality labeled data. When faced with complex situations such as ambiguous data, missing information, or multimodal heterogeneous data, the accuracy of recognition and reasoning decreases significantly, easily leading to erroneous decisions.

[0005] Therefore, the review methods for insurance-related payment transactions in related technologies suffer from low review efficiency and insufficient accuracy. Summary of the Invention

[0006] This application provides an intelligent payment business review system, method, apparatus, and equipment to improve the review efficiency and accuracy of payment business.

[0007] In a first aspect, embodiments of this application provide an intelligent payment business review system, the system comprising: a client, a server, and multiple intelligent agents, wherein: The client is used to send the material information of the payment transaction to be reviewed to the server; The server is configured to: invoke an agent matching each audit dimension based on the multiple audit dimensions involved in the material information; send the material information to each agent in parallel; receive audit sub-results generated independently by each agent for the audit dimension matching itself; when a discrepancy is detected between the audit sub-results, issue a debate command to each agent to control the debate between the agents; receive the audit sub-results corrected by the debate from each agent; generate an audit result based on the corrected audit sub-results; and return the audit result to the client. Each agent is configured to analyze the material information based on the review dimensions matched by the agent, and send the obtained review sub-results to the server; after receiving the debate instruction sent by the server, it participates in the debate and sends the review sub-results corrected by the debate to the server.

[0008] Secondly, embodiments of this application provide a smart payment business review method, the method comprising: Receive material information for payment transactions pending review from the client; Based on the multiple review dimensions involved in the material information, the intelligent agent matching each review dimension is invoked, and the material information is sent to each intelligent agent in parallel; Receive the audit sub-results generated independently by each intelligent agent based on the audit dimensions that match itself; When a discrepancy is detected between the audit sub-results, a debate instruction is issued to each agent to control the debate between the agents. Receive the review sub-results of each intelligent agent after debate and correction, and generate the review result based on the corrected review sub-results.

[0009] Thirdly, embodiments of this application provide an intelligent payment business review device, the device comprising: The first receiving module is used to receive material information for payment transactions pending review sent by the client; The calling module is used to call the intelligent agent that matches each of the multiple review dimensions involved in the material information, and send the material information to each intelligent agent in parallel; The second receiving module is used to receive the audit sub-results generated independently by each intelligent agent based on the audit dimension that matches itself; The debate module is used to issue debate instructions to each agent when a disagreement is detected between the review sub-results, so as to control the debate between the agents. The generation module is used to receive the review sub-results of each intelligent agent after debate and correction, and generate the review result based on the corrected review sub-results.

[0010] Fourthly, embodiments of this application provide an electronic device, including: at least one processor, and a memory communicatively connected to the at least one processor, wherein: The memory stores a computer program that can be executed by at least one processor, which enables the at least one processor to perform the above-described intelligent payment business review method.

[0011] Fifthly, embodiments of this application provide a storage medium in which the electronic device can execute the above-described intelligent payment service review method when the computer program in the storage medium is executed by the processor of an electronic device.

[0012] Sixthly, embodiments of this application provide a computer program product that, when executed by an electronic device, enables the electronic device to perform the aforementioned intelligent payment service review method.

[0013] The beneficial effects of this application are as follows: The intelligent payment business review system provided in this application includes: a client, a server, and multiple intelligent agents. The client sends material information of the payment business to be reviewed to the server. The server calls intelligent agents matching each review dimension based on the multiple review dimensions involved in the material information, sends the material information to each intelligent agent in parallel, and receives review sub-results generated independently by each intelligent agent for its matching review dimension. When a discrepancy is detected between the review sub-results, a debate instruction is issued to each intelligent agent to control the debate between them, and the server receives the revised review sub-results from each intelligent agent after the debate. An review result is generated based on the revised review sub-results and returned to the client. Each intelligent agent analyzes the material information based on the review dimension matched to it and sends the resulting review sub-results to the server. After receiving the debate instruction from the server, it participates in the debate and sends the revised review sub-results to the server. In this embodiment, the server, based on multiple review dimensions involved in the material information, calls upon agents matching each review dimension and sends the material information to each agent in parallel. Each agent independently executes analysis tasks based on its matched review dimension, thus automating and parallelizing the review process. This effectively solves the problems of slow processing speed and difficulty in handling large-scale payment applications in traditional manual review or single-model systems, significantly improving system throughput and response efficiency. Furthermore, each agent specializes in a specific review dimension, avoiding subjective biases caused by experience differences in manual review, ensuring consistent review standards across different cases, and significantly improving the objectivity and repeatability of review results. Moreover, by introducing a debate and correction mechanism between agents: when there are disagreements among the review sub-results, the server triggers information interaction and logical negotiation between agents, enabling each agent to dynamically correct its conclusions based on contextual evidence from other dimensions. This mechanism simulates an expert review process, achieving cross-dimensional reasoning fusion and significantly reducing the risk of misjudgment and omission. This application embodiment constructs an intelligent review architecture of "multi-agent collaboration + dynamic debate correction", which significantly improves the accuracy, adaptability and reliability of payment business review while ensuring high efficiency.

[0014] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0015] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A schematic diagram of the structure of an intelligent payment business review system provided in this application embodiment; Figure 2 A flowchart of an intelligent payment business review method provided in this application embodiment; Figure 3 This application provides a schematic diagram of the review process of an intelligent payment business review system. Figure 4 A schematic diagram of the structure of an intelligent payment business review device provided in an embodiment of this application; Figure 5 This is a schematic diagram of the hardware structure of an electronic device for implementing an intelligent payment business review method, provided as an embodiment of this application. Detailed Implementation

[0016] To improve the efficiency and accuracy of payment review, this application provides an intelligent payment review system, method, apparatus, and equipment.

[0017] The preferred embodiments of this application are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit this application. Furthermore, the embodiments and features in the embodiments of this application can be combined with each other without conflict.

[0018] It should be noted that the terms "first," "second," etc., used in the description of the embodiments of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. In the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship. Furthermore, the terms "comprising" and "having," 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 is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products, or devices.

[0019] The acquisition, transmission, storage, and use of data in this application all comply with relevant national laws and regulations.

[0020] The following description, in conjunction with the accompanying drawings, illustrates some preferred embodiments of this application.

[0021] See Figure 1 , Figure 1 This is a schematic diagram of the structure of an intelligent payment business review system provided in an embodiment of this application. The intelligent payment business review system includes: a client 100, a server 110, and multiple intelligent agents 120, wherein the multiple intelligent agents 120 include a liability determination intelligent agent 1201, a disclosure matter intelligent agent 1202, a liability exemption intelligent agent 1203, a special agreement intelligent agent 1204, and an evidence verification intelligent agent 1205.

[0022] Client 100 is used to send the material information of the payment business to be reviewed to server 110.

[0023] Server 110 is used to call the intelligent agent 120 matching each audit dimension based on the multiple audit dimensions involved in the material information, and send the material information to each intelligent agent in parallel. It receives the audit sub-results generated by each intelligent agent independently for the audit dimension matching itself. When a discrepancy is detected between the audit sub-results, it issues a debate instruction to each intelligent agent to control the debate between the intelligent agents, and receives the audit sub-results corrected by each intelligent agent after the debate. Based on the corrected audit sub-results, it generates the audit result and returns the audit result to client 100.

[0024] Each of the multiple agents 120 is used to analyze the material information based on the review dimension matched by the agent, and send the review sub-results obtained from the analysis to the server 110; after receiving the debate instruction sent by the server 110, it participates in the debate and sends the review sub-results corrected by the debate to the server 110.

[0025] In some embodiments, the material information includes contract information for the payment business associated with the identity information of the payment applicant and evidence data associated with the payment business. The server 110 is used to identify the review dimensions involved in the material information based on the contract information and evidence data. The review dimensions include at least two of the following: scope of liability, disclosure matters, exemption from liability, special agreements, and factual evidence. Based on each identified review dimension, the server calls the intelligent agent that matches each identified review dimension from the preset mapping relationship between review dimensions and intelligent agents.

[0026] In some embodiments, if each audit dimension includes a scope of responsibility, each intelligent agent includes a responsibility determination intelligent agent 1201; the responsibility determination intelligent agent 1201 is used to parse each scope of protection clause in the contract information, match the evidence data with each scope of protection clause, and output the corresponding audit sub-result based on the matching result. If each audit dimension includes disclosure items, each intelligent agent includes disclosure item intelligent agent 1202; disclosure item intelligent agent 1202 is used to parse the disclosure items filled in by the applicant in the contract information, perform causal correlation analysis on the evidence data and the disclosure items, and output the corresponding audit sub-results based on the causal correlation analysis results; If each audit dimension includes liability exemption, each intelligent agent includes liability exemption intelligent agent 1203; the liability exemption intelligent agent is used to parse the exemption conditions in the contract information, determine whether the material information meets any exemption condition, and output the corresponding audit sub-result based on the judgment result; If each audit dimension includes special provisions, each intelligent agent includes special provisions intelligent agent 1204; the special provisions intelligent agent is used to parse the special provisions clauses in the contract information, determine whether the material information violates any special provisions, and output the corresponding audit sub-results based on the judgment results; If each audit dimension includes factual evidence, each intelligent agent includes an evidence verification intelligent agent 1205. The evidence verification intelligent agent 1205 is used to parse the multimodal evidence content in the evidence data, extract key factual elements, perform consistency verification between different key factual elements, and output the corresponding audit sub-results based on the consistency verification results.

[0027] In some embodiments, each audit sub-result includes a judgment direction and a judgment basis; the server 110 is used for: If any of the review sub-results contains a review sub-result with an isolated judgment direction, a debate instruction using a focused debate mechanism is issued to each agent to organize multiple rounds of debate on the judgment basis of the target agent, and to control the target agent to correct the corresponding review sub-results during the multiple rounds of debate. An isolated judgment direction means that each review sub-result belongs to two different judgment directions, and when the number of review sub-results corresponding to the two different judgment directions is different, the number of review sub-results corresponding to the two different judgment directions is one. The target agent is the agent corresponding to the review sub-result with an isolated judgment direction. Otherwise, issue debate instructions to each agent using a consultative debate mechanism to organize multiple rounds of debate around the points of contention, and control the agents to revise their respective audit sub-results through evidence exchange and logical alignment.

[0028] In some embodiments, each audit sub-result further includes a confidence score; when issuing debate instructions employing a focused debate mechanism to each agent, server 110 is used to: The organization involves multiple rounds of debate among various intelligent agents regarding the judgment criteria of the target intelligent agent, until the focused debate ends. The current review sub-result of the target intelligent agent is then used as the revised review sub-result. Each round of debate involves the following steps: Each agent is instructed to make a statement in turn, including the direction of the decision, the basis for the decision, and the corresponding confidence score. Based on the statement, other intelligent agents are instructed to question the target intelligent agent's judgment basis in turn, and to provide rebuttal evidence to refute the target intelligent agent's judgment basis. The target agent is instructed to supplement supporting evidence to support its judgment criteria based on the content of the inquiry and the rebuttal evidence. During the debate, the target agent is monitored for confidence scores based on the questioning content, rebuttal evidence, and supporting evidence updates. If the change in confidence scores exceeds a preset first threshold, the target agent is controlled to re-evaluate the corresponding judgment direction, judgment basis, and confidence scores, and the corresponding review sub-results are corrected.

[0029] In some embodiments, each audit sub-result further includes a confidence score; the decision direction includes supporting payment, opposing payment, or uncertainty; when issuing debate instructions using a negotiated debate mechanism to each agent, the server 110 is used to: Based on the review sub-results of each agent, each agent is divided into at least one of the following groups: support group, opposition group, or neutral group. Multiple rounds of debate are organized among the agents in each group, focusing on the points of contention, until the conditions for ending the debate through negotiation are met. The current review sub-results of each agent are then used as the corresponding revised review sub-results. Each round of debate involves the following steps: Each group of intelligent agents is controlled to make a statement in turn within the group. The statement includes the direction of the judgment, the basis for the judgment, and the corresponding confidence score. Through group debate, a consensus review sub-result is generated. Based on the consensus of each group, the sub-results are reviewed, and the agents of each group are organized to raise questions to each other and supplement the evidence to refute the judgment of the other group. Each group of agents is instructed to supplement the supporting evidence for their judgment based on the questions and rebuttals raised by other groups. During the debate, the confidence scores of each agent based on the question content, rebuttal evidence, and supporting evidence are monitored. If the change in the confidence score of any agent after the update exceeds the preset second threshold, the agent is controlled to re-evaluate the corresponding judgment direction, judgment basis, and confidence score, and the corresponding review sub-result is corrected.

[0030] In some embodiments, each audit sub-result further includes a confidence score; the decision direction includes supporting payment, opposing payment, or uncertainty; the server 110 is also used for: If there are discrepancies among the revised audit results of the various agents, at least one of the following decision strategies will be executed: The confidence scores corresponding to the revised audit sub-results of each agent are weighted and fused to generate the audit result. Calculate the difference between the sum of confidence scores of the supporting agents and the sum of confidence scores of the opposing agents. If the difference is greater than a preset difference, the audit result of the group with the higher sum of confidence scores will be used as the audit result. If the difference is not greater than the preset difference, the payment business will be submitted for manual review.

[0031] It should be noted that in this embodiment, the server 110 and multiple intelligent agents 120 constitute a master-slave collaborative audit architecture, with clear functional division and tight coupling in interaction. Furthermore, each intelligent agent (such as the responsibility determination intelligent agent 1201, the disclosure intelligent agent 1202, etc.) can be integrated within the server as its functional modules; each intelligent agent can also be deployed on distributed computing nodes and communicate with the server 110 via a network. Regardless of the deployment method, the server 110 always assumes the responsibilities of process scheduling, conflict arbitration, and final decision-making, while the intelligent agents only provide analysis and reasoning capabilities within their respective areas of expertise.

[0032] It should be noted that in this embodiment, each agent is not a general dialogue model, but a specialized inference engine based on a Large Language Model (LLM) that is fine-tuned for a specific review dimension. Specifically, each agent, based on a pre-trained LLM, further uses domain-specific high-quality corpora for supervised fine-tuning of training data, including but not limited to: payment business contract terms (such as structured texts such as coverage, disclaimers, and special agreements), historical payment review cases (including material information, review conclusions, and expert comments), authoritative industry guidelines (such as clinical treatment guidelines, equipment maintenance standards, and service acceptance criteria), and relevant regulatory provisions and compliance requirements (such as information disclosure rules in the financial / health service sector). Through the above training, each agent can not only accurately understand the semantics of contract language and evidence materials, but also perform multi-step logical reasoning within its assigned review dimension.

[0033] It should be noted that the above application scenarios are shown only to facilitate understanding of the spirit and principles of this application, and the implementation methods of this application are not limited in any way. On the contrary, the implementation methods of this application can be applied to any applicable scenario.

[0034] See Figure 2 , Figure 2 The flowchart illustrates an intelligent payment service review method provided in this application embodiment. This intelligent payment service review method can be applied to the aforementioned server 110 and specifically includes the following steps: In step 201, the material information of the payment business to be reviewed is received from the client.

[0035] In practice, the benefit payment can be an insurance benefit payment from various insurance products such as medical insurance, accident insurance, and critical illness insurance. The supporting documents can include contract information related to the benefit payment application and evidentiary data associated with the benefit payment application. Contract information can include policy information, benefit application information (policyholder / insured's name and identity information), and policy terms (scope of liability, disclosures, exclusions, and special provisions). The evidentiary data associated with the benefit payment will vary depending on the type of benefit payment. For example, when the benefit payment is for medical insurance, the evidentiary data can be medical bills, diagnostic certificates, and medical records. When the benefit payment is for accident insurance, the evidentiary data can be accident certificates, medical diagnostic certificates, accident photos / videos, etc. When the benefit payment is for critical illness insurance, the evidentiary data can be diagnostic reports, imaging data, and pathology reports, etc.

[0036] In practice, the payment business can also be annuity payment business, death benefit payment business, disability benefit payment business, hospitalization allowance payment business, lump-sum payment business for specific diseases, long-term care insurance payment business, or other payment business that triggers insurance liability based on contractual agreement. This application does not limit this.

[0037] In step 202, based on the multiple review dimensions involved in the material information, the intelligent agent matching each review dimension is invoked, and the material information is sent to each intelligent agent in parallel.

[0038] In practice, since the material information may include contract information and evidence data, the server can identify the review dimensions involved in the material information based on the contract information and evidence data. The review dimensions include at least two of the following: scope of responsibility, disclosure matters, exemption from liability, special agreements, and factual evidence. Then, based on the identified review dimensions, the server can call the intelligent agent that matches the identified review dimensions from the preset mapping relationship between review dimensions and intelligent agents.

[0039] For example, the preset mapping relationship between review dimensions and intelligent agents is configured as follows: The scope of liability corresponds to the liability determination AI agent 1201, the disclosure matters correspond to the disclosure matters AI agent 1202, the exemption of liability corresponds to the exemption of liability AI agent 1203, the special agreement corresponds to the special agreement AI agent 1204, and the factual evidence corresponds to the evidence verification AI agent 1205.

[0040] Then, the server distributes the material information (including contract information and evidence data) in parallel to the aforementioned invoked intelligent agents to initiate a multi-dimensional collaborative review process.

[0041] In this way, by dynamically calling the corresponding intelligent agents on demand, the system avoids redundant calculations for irrelevant review dimensions, which improves processing efficiency and ensures the accuracy of review coverage.

[0042] In step 203, the audit sub-results generated by each agent through independent analysis of the audit dimensions that match itself are received.

[0043] Each review sub-result may include a judgment direction, judgment basis, and confidence score. The judgment direction may include supporting payment, opposing payment, or uncertainty of outcome. The confidence score is used to characterize the agent's confidence and reliability in the given judgment direction and judgment basis.

[0044] In practice, the liability determination intelligent agent is used to parse each coverage clause in the contract information, match the evidence data with each coverage clause, and output the corresponding audit sub-results based on the matching results.

[0045] For example, when the reimbursement is for medical insurance, the contract information stipulates that "the coverage includes outpatient and inpatient treatment costs caused by accidents." The liability determination agent extracts the reason for the visit as "traffic accident" from the evidence data, matches it with medical records and invoice items, and confirms that the cost belongs to treatment related to accidental injury. The output review sub-result may include: "Determination direction: Support payment, determination basis: the reason for the visit matches the 'accidental injury' clause in the coverage, confidence score: 90%".

[0046] In practice, the disclosure agent is used to parse the disclosure items filled in by the applicant in the contract information, perform causal correlation analysis on the evidence data and the disclosure items, and output the corresponding audit sub-results based on the causal correlation analysis results.

[0047] The disclosure items refer to the prior statements or notifications given to the applicant during the contract signing process, which may include health disclosures, past medical history records, occupational categories, etc.

[0048] For example, if an applicant for insurance includes "no history of hypertension" in the disclosure information provided during the application process, but the medical examination report or records in the evidence data show that they were diagnosed with "grade 2 hypertension" before applying for insurance, the disclosure agent will identify this contradiction and determine that there is a failure to disclose information truthfully, which may affect the causal relationship of the insured event. The agent will then output a review sub-result that may include: "Judgment direction: Oppose payment; Judgment basis: There is a substantial discrepancy between the evidence data and the health status disclosed at the time of application; Confidence score: 85%."

[0049] In practice, the liability exemption agent is used to parse the exemption conditions in the contract information, determine whether the material information meets any exemption condition, and output the corresponding audit sub-result based on the judgment result.

[0050] Among them, exemption clauses refer to clauses explicitly stipulated in the payment service contract that exempt the party obligated to make payments from liability under specific circumstances. Examples include: the applicant's drunk driving, the application involving a pre-existing condition before the contract took effect, the application being submitted within the waiting period stipulated in the contract, or the relevant services not being completed at the institution specified in the contract.

[0051] For example, the traffic accident report in the materials states that "the insured's blood alcohol content is 85mg / 100ml," while the contract's exclusion clause explicitly states that "accidents caused by driving a motor vehicle after drinking alcohol are not covered by insurance." Based on this, the liability exclusion agent determines that the exclusion condition is met, and the output review sub-result may include: "Judgment direction: Oppose payment; Judgment basis: The accident meets the exclusion conditions for drunk driving stipulated in the contract; Confidence score: 92%."

[0052] In practice, a special intelligent agent is designated to parse the special clauses in the contract information, determine whether the material information violates any special agreement, and output the corresponding audit sub-results based on the judgment results.

[0053] For example, the contract information may contain a special clause stating "only surgical treatment costs will be reimbursed," while the applicant's submitted expense list only includes medication costs. The special agreement agent can identify this violation and output a review sub-result that may include: "Judgment direction: Oppose payment; Judgment basis: The expense list does not conform to the surgical treatment costs specified in the special agreement; Confidence score: 95%."

[0054] In practice, the evidence verification agent is used to parse the multimodal evidence content in the evidence data, extract key factual elements, perform consistency verification between different key factual elements, and output corresponding audit sub-results based on the consistency verification results.

[0055] Multimodal evidence can include text data, invoice images, video recordings, etc. Key factual elements can include time, amount, expense details, diagnosis category, etc.

[0056] For example, the medical record in the evidence data shows "hospitalization period from March 1st to March 10th, 2025", but the inpatient invoice date is "March 15th, 2025", and the expense details include "bed fee for 15 days". The evidence verification agent finds a contradiction between the time and expense logic, and the output audit sub-result may include: "Judgment direction: Result uncertain, Judgment basis: The hospitalization duration in the medical record is inconsistent with the number of days of bed fee on the invoice, further verification is required, confidence score: 70%".

[0057] In this way, each agent independently executes analysis tasks based on its matched review dimensions, achieving automation and parallelization of the review process. This effectively solves the problems of slow processing speed and inability to handle large-scale payment applications in traditional manual review or single-model systems, significantly improving system throughput and response efficiency. Furthermore, each agent specializes in a specific review dimension, making judgments based on structured rules, knowledge graphs, or trained models. This avoids subjective biases caused by experience differences in manual review, ensuring consistent review standards across different cases and significantly improving the objectivity and repeatability of review results.

[0058] In step 204, when a discrepancy is detected between the audit sub-results, a debate instruction is issued to each agent to control the debate between the agents.

[0059] In practice, the system receives the review sub-results sent by each agent and determines whether the judgment directions in each sub-result are consistent. If they are consistent, the consistent judgment direction is used as the judgment direction of the review result, and the judgment criteria in each sub-result are combined to form the judgment criteria for the review result. The review result is then returned to the client. If the judgment directions of each sub-result are not completely consistent, it is determined that there is a disagreement between the sub-results. In this case, a debate command is issued to each agent to control the debate between them.

[0060] Considering that the degree of cognitive conflict and resolution paths reflected by different types of review disagreements vary significantly—for example, when only one agent disagrees with the majority of other agents, the problem usually stems from the agent's misunderstanding of evidence or incorrect application of rules, and should be addressed through targeted inquiries to encourage self-correction; while when multiple agents form equally matched opposing viewpoints, it indicates deep-seated factual ambiguity or disagreements in rule interpretation, requiring consensus through multilateral negotiation—in this embodiment of the application, when disagreements exist among the review sub-results, the server can dynamically select an appropriate debate mechanism based on the distribution characteristics of the disagreements among the review sub-results.

[0061] In this embodiment of the application, the debate mechanism mainly includes a focused debate mechanism and a consultative debate mechanism.

[0062] Scenario 1: If there are review sub-results with isolated judgment directions among the review sub-results, the server issues a debate instruction to each agent using a focused debate mechanism to organize multiple rounds of debate on the judgment basis of the target agent, and controls the target agent to correct the corresponding review sub-results during the multiple rounds of debate.

[0063] Among them, isolated judgment direction refers to a judgment direction in which each audit sub-result belongs to two different judgment directions and the number of audit sub-results corresponding to the two different judgment directions is different, and the number of audit sub-results corresponding to the two different judgment directions is one; the target agent is the agent corresponding to the audit sub-result with isolated judgment direction.

[0064] For example, the server invoked 5 smart agents, and the audit sub-results returned by each smart agent were as follows: the agent for determining liability: supported payment; the agent for disclosing matters: supported payment; the agent for exempting liability: supported payment; the agent for special agreement: supported payment; and the agent for verifying evidence: opposed payment.

[0065] At this point, the review sub-results involve two decision directions: "Support payment" (4) and "Oppose payment" (1). Since the "Oppose payment" direction corresponds to only one review sub-result, it meets the definition of "isolated decision direction." The server identifies the evidence verification agent as the target agent. The server believes that this isolated opinion is likely due to the target agent's partial misinterpretation of the material information, deviation in rule application, or incorrect evidence association, rather than a fundamental factual dispute. Therefore, instead of initiating full consultation, targeted inquiries and evidence feedback should be used to help the agent reassess its conclusion. Therefore, the server issues a focused debate instruction to all agents.

[0066] In practice, the server can organize multiple rounds of debate among the agents regarding the judgment criteria of the target agent until the conditions for ending the focused debate are met, and then use the current review sub-result of the target agent as the revised review sub-result.

[0067] Among them, the conditions for ending the focused debate can be reaching a preset number of debate rounds, such as 3 rounds, or the change in the confidence score of the target agent is less than a preset threshold, that is, the confidence score tends to stabilize, or the judgment direction of the review sub-results of each agent is consistent. This application does not limit this.

[0068] In practice, during each round of debate, the server controls the interaction of each intelligent agent according to the following steps.

[0069] Step 1: Presentation Phase

[0070] Each agent is instructed to make a statement in turn, including the direction of the decision, the basis for the decision, and the corresponding confidence score.

[0071] For example, taking the above scenario as an example, the server calls 5 smart agents, and the audit sub-results returned by each smart agent are as follows: Liability determination smart agent: supports payment; Disclosure smart agent: supports payment; Liability exemption smart agent: supports payment; Special agreement smart agent: supports payment; Evidence verification smart agent: opposes payment.

[0072] Assume the following: Liability determination agent: Determination direction: Support payment; Determination basis: "The reason for the visit was a traffic accident, which falls within the 'accidental injury' coverage stipulated in the contract"; Confidence score: 0.92 (92%). Disclosure agent: Determination direction: Support payment; Determination basis: "The history of lumbar spine disease was not inquired at the time of insurance application, and this accident is unrelated to the pre-existing condition"; Confidence score: 0.88 (88%). ... Evidence verification agent: Determination direction: Oppose payment; Determination basis: "The hospital invoice date is March 15, 2025, but the medical record shows the discharge date as March 10, which is a time discrepancy"; Confidence score: 0.75 (75%).

[0073] Step Two: Questioning and Rebuttal Phase

[0074] Based on the statement, other agents are instructed to question the target agent's judgment criteria in turn, and to provide rebuttal evidence to refute the target agent's judgment criteria.

[0075] Among them, the target intelligent agent is the evidence verification intelligent agent. Therefore, the liability determination intelligent agent, the disclosure intelligent agent, the liability exemption intelligent agent, and the special agreement intelligent agent can successively question the judgment basis of the evidence verification intelligent agent and supplement the rebuttal evidence to refute the judgment basis of the target intelligent agent.

[0076] For example, the liability determination agent questioned the evidence verification agent's judgment criteria: "Have you considered the possibility of delays in hospital payment settlement? According to data from the medical insurance system interface, this hospital often submits invoices within 5 days." The agent then provided additional rebuttal evidence: statistics on the time difference between invoices and medical records for similar cases at the hospital over the past 3 months (average delay of 3.2 days). The special agreement stipulates that the intelligent agent questioned the evidence verification intelligent agent's judgment criteria: "The special agreement in the contract does not require 'the invoice and medical record dates to be completely consistent,' only 'the fact of the visit is true.' Is there any other evidence to prove that the visit did not occur?" The agent further provided rebuttal evidence: If the authenticity is denied solely because the invoice date is later than the discharge date, it would lead to a large number of compliant cases being wrongly rejected, which is not in line with industry practice.

[0077] Step 3: Response and Evidence Presentation Stage

[0078] The target agent is instructed to supplement its judgment with supporting evidence based on the content of the inquiry and the rebuttal evidence.

[0079] For example, based on the questioning content and rebuttal evidence, the evidence verification agent retrieves the "Expense Settlement Explanation" issued by the hospital submitted by the applicant for payment (which has not been analyzed in detail before), which states: "Due to the delay in the approval process, the invoice will be issued on the 5th day after discharge"; at the same time, it calls the medical knowledge base to confirm: "Delay in the settlement of inpatient expenses is a common operation and does not constitute fraudulent medical treatment"; therefore, it supplements the supporting evidence: after investigation, the hospital has provided an explanation for the settlement delay, and the system record is consistent with the medical record, so the time discrepancy can be reasonably explained.

[0080] Step 4: Confidence Monitoring and Dynamic Correction Phase

[0081] During the debate, the target agent is monitored for confidence scores based on the questioning content, rebuttal evidence, and supporting evidence updates. If the change in confidence scores exceeds a preset first threshold, the target agent is controlled to re-evaluate the corresponding judgment direction, judgment basis, and confidence scores, and the corresponding review sub-results are corrected.

[0082] For example, during the debate, although the evidence verification agent did not change its judgment direction, the confidence score based on the questioning content, rebuttal evidence, and supporting evidence updates dropped from 0.75 to 0.60. The server detected that the change in the confidence score (0.15) exceeded the preset first threshold (e.g., 0.1). The server determined that the evidence verification agent's cognition had been significantly shaken, triggering a correction process. The server then controlled the evidence verification agent to re-evaluate the corresponding judgment direction, judgment basis, and confidence score, and corrected the corresponding review sub-result. For example, the review sub-result corrected by the evidence verification agent in this round of debate is: Judgment direction: support payment; Judgment basis: "The original time contradiction has been clarified by the hospital, and the key factual elements are consistent"; Confidence score: 0.82.

[0083] At this point, the decision direction of each agent's review sub-result is consistent. After the conditions for ending the focused debate are met, the current review sub-result of the target agent is taken as the corrected review sub-result.

[0084] Scenario 2: If the review sub-results do not fall under the isolated direction of Scenario 1 (i.e., do not meet the condition of "only one agent holding a different opinion"), then a debate instruction using a consultative debate mechanism is issued to each agent to organize multiple rounds of debate around the points of contention, and to control the agents to revise their respective review sub-results through evidence exchange and logical alignment.

[0085] For example, the server invoked 5 smart agents, and the audit sub-results returned by each smart agent were as follows: Liability determination smart agent: supports payment; Disclosure smart agent: supports payment; Liability exemption smart agent: opposes payment; Special agreement smart agent: opposes payment; Evidence verification smart agent: uncertain.

[0086] At this point, the review sub-results involve three decision directions: "Support payment" (2), "Oppose payment" (2), and "Conclusion uncertain" (1). Since there is no isolated decision direction with only one outcome, and the supporting and opposing opinions are evenly matched, it indicates that there are deep disagreements in the system on multiple dimensions such as the boundary of responsibility, the effectiveness of disclosure, and the application of clauses. A single agent cannot dominate the conclusion. Therefore, the server issues a debate instruction for a negotiated debate mechanism to all agents.

[0087] In practice, the server can divide each agent into at least one of the following groups based on the review sub-results of each agent: support group, opposition group, or neutral group. The server can then organize multiple rounds of debate among the agents in each group around the points of contention until the conditions for ending the debate through negotiation are met. After that, the current review sub-results of each agent can be used as the corresponding revised review sub-results.

[0088] The server can classify each agent into at least one of the following groups: support group, opposition group, or neutral group, based on the decision direction and confidence score of each agent's review sub-result. For example, if an agent's decision direction is "support payment", it will be classified into the support group; if the decision direction is "oppose payment", it will be classified into the opposition group; if the decision direction is "the result is uncertain", or its confidence score is lower than the preset confidence threshold (e.g., 0.6), even if there is a clear direction, it will be regarded as having insufficient confidence in the judgment and classified into the neutral group.

[0089] In practice, the conditions for ending the negotiation and debate can be reaching a preset number of debate rounds, such as 3 or 5, or the decision direction of the review sub-results of each agent is consistent. This application does not limit this.

[0090] In practice, during each round of debate, the server controls the interaction of each intelligent agent according to the following steps.

[0091] Assumptions: The supporting group consists of: the intelligence agent for determining liability and the intelligence agent for disclosing matters; the opposing group consists of: the intelligence agent for exempting liability and the intelligence agent for special agreements; the neutral group consists of: the intelligence agent for verifying evidence.

[0092] Step 1: Group Presentations and Consensus Generation

[0093] Each group of intelligent agents is controlled to make a statement in turn within the group. The statement includes the direction of the judgment, the basis for the judgment, and the corresponding confidence score. Through group debate, a consensus review sub-result is generated.

[0094] For example, in the support group: the liability determination agent states: "The accident was an accidental fall, which is within the scope of coverage, with a confidence level of 0.90"; the disclosure agent states: "The failure to disclose lumbar spine disease is unrelated to this accident and does not constitute a reason for refusing payment, with a confidence level of 0.85."

[0095] After group debate, the following consensus review result was generated: "Judgment direction: Support payment, Judgment basis: The accident was accidental and there were no related pre-existing conditions, consensus confidence level: 0.88".

[0096] In the opposing group: the agent for exemption from liability stated: "If the incident occurred during the waiting period, liability should be waived, confidence level 0.80"; the agent for special agreement stated: "If the applicant did not undergo surgery, in violation of the agreement, only the cost of surgical treatment will be paid, confidence level 0.75".

[0097] Following group debate, the following consensus review sub-result was generated: "Judgment Direction: Oppose payment; Judgment Basis: Existence of waiting period exemption and breach of agreement to only pay surgical treatment costs; Consensus Confidence Level: 0.78"

[0098] Neutral Group: The evidence verification agent stated: "The medical records and images are consistent, but the invoice date is questionable, with a confidence level of 0.50."

[0099] Step Two: Cross-Group Questioning and Submission of Rebuttal Evidence

[0100] Based on the consensus of each group, the sub-results are reviewed, and the agents of each group are organized to raise questions to each other and provide rebuttal evidence to refute the judgment of the other group.

[0101] For example, the support group could first raise questions to the opposition group: Question 1: "Please specify which clause in the contract stipulates a 'waiting period applicable to unforeseen events'?" Counter-evidence 1: The contract clause includes the statement that "medical expenses caused by accidental injury are not subject to the waiting period."

[0102] Question 2: "The application for this project is for emergency debridement and plaster fixation, which falls under the exemption of 'necessary emergency treatment' in the special provisions of the contract. It does not mean that non-surgical treatment is always refused payment."

[0103] Counter-evidence 2: The corresponding clauses in the special agreement.

[0104] Next, the opposition group questioned the support group: Question 1: "If the applicant concealed a herniated lumbar disc, which is a potential cause of the fall (such as gait instability), is it still considered that there is 'no causal relationship'?" Counter-evidence 1: Citing a certain orthopedic clinical pathway guideline: "Degenerative spinal disease can increase the risk of falls."

[0105] Question 2: "Does the determination of accidental death rely on third-party evidence? Currently, there is only the self-report, which is not strong enough as evidence."

[0106] Counter-evidence 2: Point out that the materials lack accident scene records or eyewitness testimonies.

[0107] Finally, the neutral group questioned the opposing and supporting groups: "Is there a 'confirmation of the nature of the injury' issued by the hospital? Is there a financial explanation for the delayed invoice date?"

[0108] Step 3: Each group responds and provides supplementary supporting evidence.

[0109] Each group of agents is instructed to supplement the supporting evidence for their judgment based on the questions and rebuttals raised by other groups.

[0110] For example, the support group responded to the inquiry and provided supplementary supporting evidence: they submitted the chief complaint record in the emergency room medical record: "The patient stated that he 'slipped and fell while walking on flat ground'", and the imaging report did not show any signs of pathological fracture; at the same time, they provided the clause in the contract that "if there is no medical evidence to show that there is a direct causal relationship between the pre-existing condition and the current injury, payment shall not be refused on the grounds of failure to inform", and provided a statement of the cause of the injury stamped by the hospital's emergency department, confirming that it was "injury caused by non-pathological external force".

[0111] The opposition group responded to the inquiry and provided supplementary supporting evidence: After review, it was found that accidents are indeed not subject to the waiting period, and the reason for exemption was withdrawn. It was also discovered that the appendix did indeed contain an exception clause stating that "emergency non-surgical treatment is payable up to 2000 yuan."

[0112] The neutral group responded that the hospital's explanation for the delayed issuance of the invoice, the confirmed medical records, imaging data, and the detailed expense list were consistent.

[0113] Step 4: During the debate, monitor the confidence scores of each agent based on the question content, rebuttal evidence, and supporting evidence. If the change in the confidence score of any agent after the update exceeds the preset second threshold, control the agent to re-evaluate the corresponding judgment direction, judgment basis, and confidence score, and correct the corresponding review sub-result.

[0114] For example, if the second threshold is preset to 0.2, during the debate, if the confidence score of the liability exemption agent is updated from 0.8 to 0.3, and the change exceeds 0.2, then the liability exemption agent is controlled to re-evaluate the corresponding judgment direction, judgment basis, and confidence score. Assume that the liability exemption agent finally outputs the judgment direction: support payment, judgment basis: "waiver of waiting period for accidental injury, and no other valid reasons for exemption", and confidence score: 0.82.

[0115] During the debate, if the confidence score of the specially agreed-upon agent is updated from 0.75 to 0.45, and the change exceeds 0.2, then the specially agreed-upon agent will re-evaluate the corresponding judgment direction, judgment basis, and confidence score. Assume that the specially agreed-upon agent outputs the judgment direction as: support payment, judgment basis as: "the exception clause of 'emergency non-surgical treatment can be paid within 2000 yuan' is found", and confidence score as: 0.75.

[0116] During the debate, if the confidence level of the evidence verification agent changes from 0.50 to 0.75, and the change exceeds 0.2, then the evidence verification agent is controlled to re-evaluate the corresponding judgment direction, judgment basis, and confidence score. Assume that the evidence verification agent outputs the judgment direction: support payment, judgment basis: "key factual elements are consistent, and time differences have been clarified", and confidence score: 0.75.

[0117] At this point, all five agents output "Support payment". The server detects that the decision direction is consistent, which meets the conditions for ending the negotiation debate (such as "the decision direction of each agent is consistent"). The process terminates, and the current review sub-result of each agent is used as the corrected review sub-result.

[0118] Thus, by introducing an inter-agent debate and correction mechanism, when there are disagreements among the review sub-results, the server triggers information exchange and logical negotiation between the agents, enabling each agent to dynamically revise its conclusions based on contextual evidence from other dimensions. This mechanism simulates the expert review process, achieving cross-dimensional reasoning fusion and significantly reducing the risk of misjudgment and omission.

[0119] In step 205, the review sub-results of each agent after debate are received, and the review result is generated based on the revised review sub-results.

[0120] In practice, after multiple rounds of debate, there may still be inconsistencies in the direction of the judgment. However, in order to avoid circular and invalid debates, when the preset debate end condition is reached, such as reaching the number of debates, the debate process is controlled to end. At this time, if there are still discrepancies between the review sub-results after the corrections of each agent, at least one of the following decision strategies can be executed.

[0121] Decision Strategy 1: Weight and fuse the confidence scores corresponding to the revised audit sub-results of each agent to generate the audit result.

[0122] For example, configure preset weights for each agent (such as based on the importance of its review dimension or historical performance). First, set the confidence score for "support payment" to be positive and the confidence score for "oppose payment" to be negative. Then, multiply the assigned confidence score by the corresponding weight. If the weighted total score is greater than zero, output "support payment"; otherwise, output "oppose payment".

[0123] Decision Strategy 2: Calculate the difference between the sum of confidence scores of the supporting agents and the sum of confidence scores of the opposing agents. If the difference is greater than the preset difference, the audit sub-result of the group with the higher sum of confidence scores will be used as the audit result. If the difference is not greater than the preset difference, the payment business will be submitted for manual review.

[0124] For example, if the results after the debate are: support group: confidence scores of 0.80, 0.90, and 0.8, with a total of 2.50; opposition group: confidence scores of 0.75 and 0.75, with a total of 1.50, and the difference between the totals is greater than the preset difference of 0.5, then the audit sub-result (support payment) of the group with the higher total confidence score (support group) will be taken as the audit result.

[0125] For example, if the results of the debate are as follows: the confidence scores for the support group are 0.80 and 0.90 respectively, with a total of 1.70; and the confidence scores for the opposition group are 0.75 and 0.80 respectively, with a total of 1.55, and the difference between the totals is no greater than the preset difference of 0.5, then it is determined that the disagreement is still highly entrenched, and the payment transaction will be automatically submitted for manual review.

[0126] In practice, the audit results may also include a complete chain of reasoning and an interpretability report, providing a clear and auditable decision-making path for the quality inspection conclusions.

[0127] To further improve the accuracy of each agent's review, the results of manual review or the actual payment decision can be used as feedback to optimize the reasoning ability of each agent; through reinforcement learning or contrastive learning, the persuasiveness and accuracy of each agent in the debate can be improved.

[0128] This application embodiment constructs an intelligent review architecture of "multi-agent collaboration + dynamic debate correction", which significantly improves the accuracy, adaptability and reliability of payment business review while ensuring high efficiency.

[0129] The following section uses the payment-based business as an example of critical illness insurance business to provide a detailed introduction to the review process of the intelligent payment-based business review system for this application.

[0130] See Figure 3 , Figure 3 This is a schematic diagram of the review process of an intelligent payment business review system provided in an embodiment of this application. The review process includes four parts: data access and preprocessing, intelligent agent construction, debate and adjudication, and feedback loop.

[0131] Data Access and Preprocessing: The review system acquires the material information of critical illness insurance business to be reviewed, including the insured's basic information, insurance contract text, and medical records. Medical images are preprocessed, key features are extracted using a medical image recognition model, and aligned with the text report to generate structured medical evidence.

[0132] Intelligent Agent Construction: Based on the multiple review dimensions involved in the material information, the review system constructs multiple review intelligent agents, including an agent for liability determination, an agent for disclosure matters, an agent for liability exemption, an agent for special agreements, and an agent for evidence verification. Each intelligent agent is fine-tuned based on a large language model, with training data including insurance terms, payment cases, medical guidelines, and regulatory provisions, enabling it to possess domain expertise and reasoning capabilities.

[0133] Simultaneously, a host agent is constructed to send material information to each review agent in parallel and receive review sub-results generated by each review agent based on the review dimensions that match itself, including the judgment direction, judgment basis, and confidence score.

[0134] Debate and Adjudication: The review system detects whether the judgment directions of each review sub-result are consistent. If they are consistent and the confidence score meets the preset score, the system generates a review result based on each review sub-result and outputs a review report, including the final judgment, the complete reasoning chain, and an interpretability report, for manual review or regulatory audit. If they are inconsistent, a multi-round dialogic reasoning debate mechanism is triggered. The host agent organizes each review agent to use multi-round dialogic reasoning, including stating positions, questioning and refuting, supplementing evidence and persuading, reaching consensus or voting. The system judges whether a consensus has been reached based on the debate process. If a consensus is reached, a consensus conclusion is generated; otherwise, the host agent makes a final adjudication based on the review sub-results of each review agent, such as through weighted fusion or super arbitration. If a decision still cannot be made, the system automatically submits the case for manual review and attaches a "heatmap of disputed points".

[0135] Feedback loop: The results of manual review or the actual payment decision are used as feedback to optimize the reasoning ability of each agent; through reinforcement learning or contrastive learning, the persuasiveness and accuracy of the agents in the debate are improved, and the case library is updated.

[0136] In this way, by constructing multiple specialized intelligent agents, each independently analyzing from dimensions such as scope of liability, disclosure matters, exemption from liability, special agreements, and factual evidence, and combining them with a multi-round debate mechanism, the limitations of a single model or rule engine in complex claims scenarios are effectively addressed, significantly reducing the misjudgment rate. Employing a multi-round dialogic reasoning mechanism, each intelligent agent details its reasoning basis, cited clauses, and evidence during the debate, generating a complete reasoning chain and debate summary, providing a clear and auditable decision-making path for quality inspection conclusions. By extracting key features of medical images through a medical image recognition model and aligning them with text reports, structured medical evidence is constructed, achieving deep understanding and integration of cross-modal information and improving the analytical capabilities of medical evidence. Introducing multi-round debate, evidence supplementation, and position correction mechanisms, intelligent agents can adjust their conclusions based on new evidence during the debate process or reach a final consensus through weighted voting, enhancing the system's intelligence and self-consistency in handling cases with ambiguous boundaries or high controversy. Through an end-to-end automated quality inspection process, the workload of manual review is reduced, while the reasoning capabilities of each intelligent agent are continuously optimized through a feedback learning module, further improving the system's automation level and long-term performance. Each agent is built based on a Large Language Model (LLM) and fine-tuned using insurance terms, payment cases, and medical knowledge to acquire domain expertise and reasoning capabilities, enabling them to adapt to the complex terms and diverse medical evidence of different critical illness insurance products. A microservice architecture can be adopted, with each agent running as an API. The system controls the flow through a state machine, reducing coupling between system modules and improving reasoning efficiency and deployment convenience.

[0137] Based on the same technical concept, this application also provides an intelligent payment business review device. The principle of the intelligent payment business review device in solving the problem is similar to that of the above-mentioned intelligent payment business review method. Therefore, the implementation of the intelligent payment business review device can refer to the implementation of the intelligent payment business review method, and the repeated parts will not be described again.

[0138] Figure 4 A schematic diagram of the structure of an intelligent payment business review device provided in this application embodiment includes: The first receiving module 401 is used to receive material information of payment transactions pending review sent by the client; The calling module 402 is used to call the intelligent agent matching each review dimension according to the multiple review dimensions involved in the material information, and send the material information to each intelligent agent in parallel; The second receiving module 403 is used to receive the audit sub-results generated independently by each intelligent agent based on the audit dimension that matches itself; The debate module 404 is used to issue debate instructions to each agent when a disagreement is detected between the review sub-results, so as to control the debate between the agents. The generation module 405 is used to receive the review sub-results of each intelligent agent after debate and correction, and generate the review result based on the corrected review sub-results.

[0139] In some embodiments, the material information includes contract information for the payment transaction associated with the identity information of the payment applicant and evidence data associated with the payment transaction; The calling module 402 is specifically used to identify the review dimensions involved in the material information based on the contract information and the evidence data; the review dimensions include at least two of the following: scope of liability, disclosure matters, exemption from liability, special agreements, and factual evidence; Based on the identified review dimensions, the system calls upon the intelligent agents that match the identified review dimensions from the preset mapping relationship between review dimensions and intelligent agents.

[0140] In some embodiments, each audit sub-result includes a judgment direction and a judgment basis; The debate module 404 is specifically used to issue a debate instruction using a focused debate mechanism to each agent if there is an isolated judgment direction among the various review sub-results. This organizes the agents to conduct multiple rounds of debate on the judgment criteria of the target agent, and controls the target agent to correct the corresponding review sub-results during the multiple rounds of debate. Isolated judgment direction refers to a judgment direction in which the number of review sub-results corresponding to the two different judgment directions is one, even when the number of review sub-results corresponding to the two different judgment directions is different. The target agent is the agent corresponding to the review sub-result with the isolated judgment direction. Otherwise, a debate instruction using a negotiated debate mechanism is issued to each agent to organize multiple rounds of debate around the points of contention, and to control the agents to revise their respective audit sub-results through evidence exchange and logical alignment.

[0141] In some embodiments, each audit sub-result also includes a confidence score; When a debate instruction employing a focused debate mechanism is issued to each of the intelligent agents, the debate module 404 is specifically used to organize the intelligent agents to conduct multiple rounds of debate on the judgment criteria of the target intelligent agent until the focused debate termination condition is met, and then the current review sub-result of the target intelligent agent is used as the corrected review sub-result; wherein, each round of debate process executes the following steps: Each agent is instructed to make a statement in turn, including the direction of the decision, the basis for the decision, and the corresponding confidence score. Based on the stated content, the other intelligent agents are instructed to sequentially question the judgment basis of the target intelligent agent and supplement rebuttal evidence to refute the judgment basis of the target intelligent agent; The target agent is instructed to supplement supporting evidence to support its judgment criteria, based on the questioning content and the rebuttal evidence. During the debate, the confidence score of the target agent is updated based on the questioning content, the rebuttal evidence, and the supporting evidence. If the change in the confidence score exceeds a preset first threshold, the target agent is controlled to re-evaluate the corresponding judgment direction, judgment basis, and confidence score, and the corresponding review sub-result is corrected.

[0142] In some embodiments, each audit sub-result further includes a confidence score; the determination direction includes supporting payment, opposing payment, or uncertainty of outcome; when a debate instruction employing a negotiated debate mechanism is issued to each agent, the debate module 404 is specifically used for: Based on the review sub-results of each agent, each agent is divided into at least one of the following groups: support group, opposition group, or neutral group. Multiple rounds of debate are organized among the agents in each group around the points of contention until the negotiation and debate end conditions are met. The current review sub-results of each agent are then used as the corresponding revised review sub-results. Each round of debate involves the following steps: Each group of intelligent agents is controlled to make a statement in turn within the group. The statement includes the direction of the judgment, the basis for the judgment, and the corresponding confidence score. Through group debate, a consensus review sub-result is generated. Based on the consensus of each group, the sub-results are reviewed, and the agents of each group are organized to raise questions to each other and supplement the evidence to refute the judgment of the other group. Each group of agents is instructed to supplement supporting evidence for the judgment criteria of each group of agents based on the questions and rebuttals raised by other groups; During the debate, the confidence scores of each agent are monitored based on the questioning content, the rebuttal evidence, and the supporting evidence. If the change in the confidence score of any agent after the update exceeds a preset second threshold, the agent is controlled to re-evaluate the corresponding judgment direction, judgment basis, and confidence score, and the corresponding review sub-result is corrected.

[0143] In some embodiments, each audit sub-result further includes a confidence score; the determination direction includes supporting payment, opposing payment, or uncertainty; the generation module 405 is specifically used for: If there are discrepancies among the revised audit sub-results of the various agents, at least one of the following decision strategies will be executed: The confidence scores corresponding to the corrected audit sub-results of each intelligent agent are weighted and fused to generate the audit result. Calculate the difference between the sum of confidence scores of the supporting group agents and the sum of confidence scores of the opposing group agents. If the difference is greater than a preset difference, the audit sub-result of the group with the higher sum of confidence scores is taken as the audit result; if the difference is not greater than the preset difference, the payment transaction is submitted for manual review.

[0144] The module division in this application embodiment is illustrative and only represents one logical functional division. In actual implementation, other division methods are possible. Furthermore, the functional modules in each embodiment of this application can be integrated into a single processor, exist as separate physical entities, or be integrated into a single module. Coupling between modules can be achieved through interfaces, typically electrical communication interfaces, but mechanical interfaces or other types of interfaces are also possible. Therefore, modules described as separate components may or may not be physically separate; they can be located in one place or distributed across different locations on the same or different devices. The integrated modules described above can be implemented in hardware or as software functional modules.

[0145] Having introduced the smart payment business review method and apparatus according to exemplary embodiments of this application, we will now introduce an electronic device according to another exemplary embodiment of this application.

[0146] The following reference Figure 5 To describe an electronic device 130 implemented according to this embodiment of the present application. Figure 5 The electronic device 130 shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0147] like Figure 5 As shown, the electronic device 130 is presented in the form of a general electronic device. The components of the electronic device 130 may include, but are not limited to: at least one processor 131, at least one memory 132, and a bus 133 connecting different system components (including memory 132 and processor 131).

[0148] Bus 133 represents one or more of several bus structures, including a memory bus or memory controller, peripheral bus, processor, or local bus using any of the various bus structures.

[0149] The memory 132 may include a readable medium in the form of volatile memory, such as random access memory (RAM) 1321 and / or cache memory 1322, and may further include read-only memory (ROM) 1323.

[0150] The memory 132 may also include a program / utility 1325 having a set (at least one) of program modules 1324, including but not limited to: an operating system, one or more application programs, other program modules, and program data, each or some combination of these examples may include an implementation of a network environment.

[0151] Electronic device 130 can also communicate with one or more external devices 134 (e.g., keyboard, pointing device, etc.), and with one or more devices that enable a user to interact with electronic device 130, and / or with any device that enables electronic device 130 to communicate with one or more other electronic devices (e.g., router, modem, etc.). This communication can be performed via input / output (I / O) interface 135. Furthermore, electronic device 130 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 136. As shown, network adapter 136 communicates with other modules used in electronic device 130 via bus 133. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with electronic device 130, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0152] In an exemplary embodiment, a storage medium is also provided, which enables the electronic device to execute the aforementioned smart payment service review method when the computer program in the storage medium is executed by the processor of the electronic device. Optionally, the storage medium may be a non-transitory computer-readable storage medium, such as a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device.

[0153] In an exemplary embodiment, the electronic device of this application may include at least one processor and a memory communicatively connected to the at least one processor, wherein the memory stores a computer program executable by the at least one processor, and when the computer program is executed by the at least one processor, the at least one processor may perform the steps of any smart payment business review method provided in the embodiments of this application.

[0154] In an exemplary embodiment, a computer program product is also provided, which, when executed by an electronic device, enables the electronic device to implement any of the exemplary methods provided in this application.

[0155] Furthermore, computer program products may employ any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: electrical connections having one or more wires, portable disks, hard disks, RAM, ROM, erasable programmable read-only memory (EPROM), flash memory, optical fiber, compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0156] The program product used for smart payment business review in this application embodiment can be a CD-ROM and include program code, and can run on a computing device. However, the program product of this application is not limited to this. In this document, the readable storage medium can be any tangible medium that contains or stores a program, which can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0157] A readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying readable program code. This propagated data signal may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable signal medium may also be any readable medium other than a readable storage medium, capable of sending, propagating, or transmitting a program for use by or in conjunction with an instruction execution system, apparatus, or device.

[0158] The program code contained on the readable medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, radio frequency (RF), or any suitable combination thereof.

[0159] Program code for performing the operations of this application can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can execute entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, such as a Local Area Network (LAN) or a Wide Area Network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0160] It should be noted that although several units or sub-units of the device have been mentioned in the detailed description above, this division is merely exemplary and not mandatory. In fact, according to embodiments of this application, the features and functions of two or more units described above can be embodied in one unit. Conversely, the features and functions of one unit described above can be further divided and embodied by multiple units.

[0161] Furthermore, although the operations of the method of this application are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0162] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0163] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0164] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0165] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0166] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0167] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. An intelligent payor business audit system, characterized by, The system comprises a client, a server and a plurality of agents, wherein: The client is configured to send material information of a payment service to be audited to the server; The server is configured to call an agent matched with each audit dimension according to a plurality of audit dimensions involved in the material information, and send the material information to each agent in parallel, receive audit sub-results independently generated by each agent for the audit dimension matched with the agent, issue a debate instruction to each agent to control debate between the agents when detecting disagreement between the audit sub-results, receive the audit sub-results revised by the agents after the debate, generate an audit result based on the revised audit sub-results, and return the audit result to the client; Each agent is configured to analyze the material information based on the audit dimension matched with the agent, send an audit sub-result obtained by the analysis to the server, participate in the debate after receiving the debate instruction sent by the server, and send the audit sub-result revised by the debate to the server.

2. The system of claim 1, wherein, The material information comprises contract information of the payment service associated with identity information of a payment applicant and evidence data associated with the payment service; The server is configured to identify the audit dimensions involved in the material information according to the contract information and the evidence data, wherein the audit dimensions comprise at least two of the following: responsibility range, disclosure matter, responsibility exemption, special agreement and factual evidence; The agents matched with the identified audit dimensions are called from a preset mapping relationship between the audit dimensions and the agents.

3. The system of claim 2, wherein: If the responsibility range is included in the audit dimensions, the agents comprise a responsibility judgment agent, and the responsibility judgment agent is configured to parse each protection range clause in the contract information, match the evidence data with each protection range clause, and output a corresponding audit sub-result according to a matching result; If the disclosure matter is included in the audit dimensions, the agents comprise a disclosure matter agent, and the disclosure matter agent is configured to parse the disclosure matter filled in the contract information by the payment applicant, perform causal association analysis on the evidence data and the disclosure matter, and output a corresponding audit sub-result according to a result of the causal association analysis; If the responsibility exemption is included in the audit dimensions, the agents comprise a responsibility exemption agent, and the responsibility exemption agent is configured to parse exemption conditions in the contract information, determine whether the material information satisfies any exemption condition, and output a corresponding audit sub-result according to a determination result; If the special agreement is included in the audit dimensions, the agents comprise a special agreement agent, and the special agreement agent is configured to parse special agreement clauses in the contract information, determine whether the material information violates any special agreement, and output a corresponding audit sub-result according to a determination result. If the fact evidence is included in each of the audit dimensions, the agents include an evidence checking agent; the evidence checking agent is configured to analyze multi-modal evidence content in the evidence data, extract key fact elements, perform consistency checking between different key fact elements, and output corresponding audit sub-results according to the consistency checking results.

4. The system of claim 1, wherein, Each audit sub-result includes a judgment direction and a judgment basis; the server is configured to: If there is an audit sub-result with a judgment direction isolated among the audit sub-results, the server issues a debate instruction using a focused debate mechanism to the agents to organize the agents to perform multi-round debates on the judgment basis of a target agent, and controls the target agent to modify the corresponding audit sub-result in the multi-round debate process; the judgment direction isolated refers to that the audit sub-results belong to two different judgment directions, and the number of audit sub-results corresponding to the two different judgment directions is different, and the number of audit sub-results corresponding to one of the two different judgment directions is one; the target agent is the agent corresponding to the audit sub-result with the judgment direction isolated; Otherwise, the server issues a debate instruction using a consultation debate mechanism to the agents to organize the agents to perform multi-round debates around a controversial focus, and controls the agents to modify their own audit sub-results through evidence exchange and logical alignment.

5. The system of claim 4, wherein, Each audit sub-result further includes a confidence score; when the server issues the debate instruction using the focused debate mechanism to the agents, the server is configured to: organize the agents to perform multi-round debates on the judgment basis of a target agent until a focused debate end condition is reached, and then take the current audit sub-result of the target agent as a modified audit sub-result; wherein each round of debate process performs the following steps: instruct each agent to make a statement in turn, and the statement content includes a judgment direction, a judgment basis, and a corresponding confidence score; based on the statement content, instruct the other agents to make inquiries on the judgment basis of the target agent in turn, and to supplement refutation evidence to refute the judgment basis of the target agent; instruct the target agent to supplement support evidence to support the judgment basis of the target agent according to the inquiry content and the refutation evidence; monitor the confidence score updated by the target agent based on the inquiry content, the refutation evidence, and the support evidence during the debate process, and if the change amplitude of the confidence score exceeds a preset first threshold, control the target agent to re-evaluate the corresponding judgment direction, judgment basis, and confidence score, and modify the corresponding audit sub-result.

6. The system of claim 4, wherein, Each audit sub-result further includes a confidence score; the judgment direction includes support payment, opposition to payment, or uncertain result; when the server issues the debate instruction using the consultation debate mechanism to the agents, the server is configured to: According to the audit sub-results of the agents, the agents are divided into at least one of a support group, an opposition group or a neutral group, a plurality of rounds of debates are organized among the agents of the groups around a controversial focus until a negotiation debate end condition is reached, and the current audit sub-results of the agents are taken as corresponding revised audit sub-results; wherein each round of debate process respectively performs the following steps: The agents of each group are controlled to make statements in turn within the group, the statement content including a judgment direction, a judgment basis and a corresponding confidence score, and a group consensus audit sub-result is generated through group debate; Based on the group consensus audit sub-results of each group, the agents of each group are organized to propose inquiries to each other and refutation evidence to supplement refutation of the judgment basis of the other group; Each agent of a group is instructed to supplement support evidence to support the judgment basis of the agent according to the inquiry content and refutation evidence proposed by the other group; During the debate process, the confidence scores of the agents updated based on the inquiry content, the refutation evidence and the support evidence are monitored, and if the change amplitude of the updated confidence score of any agent exceeds a preset second threshold, the agent is controlled to re-evaluate the corresponding judgment direction, judgment basis and confidence score, and revise the corresponding audit sub-result.

7. The system of claim 4, wherein, Each audit sub-result further includes a confidence score; the judgment direction includes support for payment, opposition to payment or uncertain result; and the server is further configured to: If there is a difference between the revised audit sub-results of the agents, at least one of the following decision strategies is performed: The confidence scores corresponding to the revised audit sub-results of the agents are weighted and fused to generate an audit result; The difference between the sum of the confidence scores of the agents of the support group and the sum of the confidence scores of the agents of the opposition group is calculated, and if the difference is greater than a preset difference value, the audit sub-result of the group with the higher sum of the confidence scores is taken as the audit result; If the difference is not greater than the preset difference value, the payment business is submitted for manual review.

8. An intelligent payor service auditing method, characterized by, The method comprises: receiving material information of a payment business to be audited sent by a client; according to a plurality of audit dimensions involved in the material information, calling an agent matched with each audit dimension, and sending the material information to each agent in parallel; receiving audit sub-results generated by each agent independently analyzing an audit dimension matched with the agent; when detecting that there is a difference between the audit sub-results, issuing a debate instruction to each agent to control the agents to debate with each other; receiving revised audit sub-results of the agents after the debate, and generating an audit result based on the revised audit sub-results.

9. An intelligent payment transaction verification device, characterized in that, The apparatus comprises: a first receiving module configured to receive material information of a payment business to be audited sent by a client; a calling module configured to, according to a plurality of audit dimensions involved in the material information, call an agent matched with each audit dimension, and send the material information to each agent in parallel; a second receiving module configured to receive audit sub-results generated by each agent independently analyzing an audit dimension matched with the agent; The debate module is configured to issue a debate instruction to the agents to control the agents to debate with each other when it is detected that there is a disagreement between the audit sub-results. The generation module is configured to receive the audit sub-results revised by the agents through the debate, and generate an audit result based on the revised audit sub-results.

10. An electronic device, comprising: The method comprises the following steps: at least one processor, and a memory connected with the at least one processor in communication, The memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor to enable the at least one processor to perform the method of claim 8.

Images