A credit risk assessment method based on multi-agent argumentation collaboration

Abstract

The application discloses a credit risk assessment method based on multi-agent debate cooperation, relates to the technical field of financial technology and artificial intelligence, and comprises the following steps: collecting multi-source heterogeneous data of a borrower and constructing a structured portrait vector; an optimistic assessment agent and a prudent assessment agent generate an initial supporting credit granting argument and an initial opposing credit granting argument respectively; the two parties carry out multi-round confrontation debate on weak links in the argument of the other party, calculate the opinion divergence and the argument comprehensive quality score after each round of debate, and perform compliance review; after meeting the convergence condition, the final risk score is calculated in combination with historical debate records and a soft compliance constraint set, the risk score is mapped into a risk level, and a five-level tree reasoning chain containing the risk level, the argument, the evidence and the data source field is output. The application can break through the limitation of a single model perspective, realize traceable and auditable confrontation risk assessment, effectively control the algorithm cost while improving the identification accuracy, and meet the regulatory compliance requirements.

Description

Technical Field

[0001] This invention relates to the fields of financial technology and artificial intelligence, and in particular to a credit risk assessment method based on multi-agent debate-based collaboration. Background Technology

[0002] Existing credit risk assessment technologies have mainly gone through three stages: expert rule engines, statistical scoring cards, and machine learning. Expert rule engines rely on IF-THEN rule sets, which have shortcomings such as limited coverage dimensions and difficulty in resolving rule conflicts; statistical scoring cards, represented by logistic regression, are limited by linear assumptions and have difficulty capturing non-linear interaction relationships between features; although machine learning and deep learning stages have significantly improved prediction accuracy, they have "black box" problems and cannot meet the requirements of financial regulation for the interpretability and auditability of decisions.

[0003] Based on this, existing AI risk control solutions still face the following prominent problems: First, the perspective of a single model is limited, and there is a lack of multi-angle examination of risk factors; second, it is difficult to coordinate the contradictions between the perspectives of multi-dimensional judgments; third, multi-model integration methods (such as Bagging, Boosting, and Stacking) only form "soft consensus" based on voting or weighted average, and lack the "hard confrontation" process between models; fourth, the decision-making process is not traceable, and it is difficult to provide a complete reasoning link from raw data to the final decision.

[0004] Although the Industrial and Commercial Bank of China's "Smart Loan" has made initial explorations into the application of multi-agent systems in credit scenarios, and the academic TradingAgents framework has verified the superiority of the bull-bear debate model in investment decision-making, a systematic technical solution for applying a debate-based multi-agent architecture to credit risk control remains lacking. Therefore, there is an urgent need for a credit risk assessment method that can integrate multi-perspective adversarial debate, adaptive convergence, and traceable reasoning to meet the requirements of financial businesses for comprehensive risk identification, interpretable decision-making, and regulatory compliance. Summary of the Invention

[0005] The technical problem to be solved by this invention is to address the shortcomings of existing technologies, specifically by providing a credit risk assessment method based on multi-agent debate-based collaboration, as detailed below: 1) In a first aspect, the present invention provides a credit risk assessment method based on multi-agent debate-based collaboration, the specific technical solution of which is as follows: S1. Collect multi-source heterogeneous data of borrowers through a unified API gateway, standardize the multi-source heterogeneous data into a structured profile vector of the borrower, and determine the initial support credit argument data and the initial opposition credit argument data based on the structured profile vector. S2. Analyze the initial opposing credit argument data and the initial supporting credit argument data, and combine the structured profile vector and external knowledge base to determine the supporting credit argument data and opposing credit argument data in the current period, and update the initial opposing credit argument data and the initial supporting credit argument data to the historical debate record. S3. Based on the opposing credence argument data, supporting credence argument data, historical debate records, structured profile vector, and external knowledge base of the current period, determine the target supporting credence argument data and target opposing credence argument data for the next period, update the target supporting credence argument data and target opposing credence argument data to the historical debate records, and calculate the target viewpoint divergence degree and target comprehensive quality score of the target supporting credence argument data and target opposing credence argument data; S4, the target support credit argument data, the target opposition credit argument data, and the updated historical debate records in the next cycle are used as the opposition credit argument data, the support credit argument data, and the historical debate records in the current cycle of S3. S3 is repeated until the target opinion divergence degree or the target comprehensive quality score meets the preset conditions. Combining the historical debate records, the preset soft compliance constraint set, and the target comprehensive quality score, the final risk score is calculated. The risk level of the borrower is determined based on the final risk score, and the credit risk management plan for the borrower is determined based on the risk level.

[0006] The beneficial effects of the credit risk assessment method based on multi-agent debate-based collaboration provided by this invention are as follows: This invention collects multi-source heterogeneous data of borrowers through a unified API gateway and constructs structured profile vectors. Optimistic and prudent assessment parties engage in multiple rounds of adversarial debate based on these profile vectors and an external knowledge base. Each round updates historical debate records and calculates the degree of disagreement and overall quality score. The debate terminates based on adaptive convergence conditions. This invention addresses the limitations of single-model perspectives, the difficulty in reconciling contradictions among multi-dimensional risk assessments, the lack of adversarial verification processes in traditional integration methods, and the untraceability of the decision-making reasoning chain. It achieves a comprehensive improvement in risk identification, enhanced interpretability and auditability of decisions, and increased accuracy in default prediction. Simultaneously, by embedding a set of soft compliance constraints into the adjudication process, it enhances compliance risk control capabilities and shortens approval timelines.

[0007] Based on the above solution, the present invention can be further improved as follows.

[0008] Furthermore, the process S1 specifically includes: The multi-source heterogeneous data of the borrower is obtained through a unified API gateway. Missing value imputation, outlier detection and format unification are performed on each data source in the multi-source heterogeneous data. The processed fields of each data source are encoded into a structured dictionary according to the preset profile dimensions. The structured dictionary includes at least: original field values, derived indicator values, data timestamps and data source identifiers. All structured dictionaries corresponding to all data sources are aggregated into a structured profile vector of the borrower; the structured profile vector is input into a pre-trained optimistic assessment agent, which generates initial credit support argument data. The initial credit support argument data includes: a first textual claim, a first evidence list, and a first confidence score. Each item in the first evidence list is associated with at least one specific data source field in the multi-source heterogeneous data. The structured profile vector is input into a pre-trained prudent assessment agent, which generates initial objection argument data. The initial objection argument data includes: a second textual claim, a second list of evidence, and a second confidence score. Each item in the second list of evidence is associated with at least one specific data source field in the multi-source heterogeneous data.

[0009] Furthermore, the process S2 specifically includes: The initial support and opposition arguments are stored as the first record in the historical debate record. The second evidence list and second confidence level value in the initial opposition credit argument data are analyzed to identify the first link in the initial opposition credit argument data where the confidence level is lower than a preset threshold or the evidence is missing. Based on the first link where the evidence is missing, the related fields are retrieved from the structured profile vector, and the first related data is retrieved from the positive evidence knowledge base in the external knowledge base. Based on the retrieval results, the support credit argument data for the current period is generated. The first evidence list and first confidence value in the initial support credit argument data are analyzed, and the second link in the initial support credit argument data with a confidence value lower than a preset threshold or missing evidence is identified. Based on the second link with missing evidence, the related fields are retrieved from the structured profile vector, and the second related data is retrieved from the negative evidence knowledge base in the external knowledge base. Based on the retrieval results, the opposition credit argument data for the current period is generated. The supporting credit argument data in the current period includes: a third text claim, a third evidence list, and a third confidence level value. The opposing credit argument data in the current period includes: a fourth text claim, a fourth evidence list, and a fourth confidence level value. Each item in the third evidence list and the fourth evidence list is associated with at least one specific data source field in the multi-source heterogeneous data.

[0010] Furthermore, the S3 process specifically includes: The opposing credit argument data and the supporting credit argument data for the current period are appended and stored in the historical debate record; The fourth evidence list and fourth confidence level value in the current period's opposing credit argument data are analyzed to identify the third link in the current period's opposing credit argument data where the confidence level is lower than a preset threshold or the evidence is missing. Based on the third link where the evidence is missing, the related fields are retrieved from the structured profile vector, and the third related data is retrieved from the positive evidence knowledge base in the external knowledge base. Based on the retrieval results, the target supporting credit argument data for the next period is generated. The third evidence list and third confidence level value in the supporting credit argument data of the current period are analyzed. The fourth link in the supporting credit argument data of the current period with a confidence level lower than a preset threshold or missing evidence is identified. Based on the fourth link with missing evidence, the related fields are retrieved from the structured profile vector, and the fourth related data is retrieved from the negative evidence knowledge base in the external knowledge base. Based on the retrieval results, the target opposing credit argument data for the next period is generated. The data supporting the credit granting of the target and the data opposing the credit granting of the target are added and stored in the historical debate record; Extract the first semantic vector corresponding to the target support credit argument data, extract the second semantic vector corresponding to the target opposition credit argument data, calculate the cosine distance between the first semantic vector and the second semantic vector, and use the cosine distance as the target opinion divergence degree; Obtain the data source authority weight, logical rigor score, and novelty score of the evidence cited in the data supporting the credibility of the target and the data opposing the credibility of the target. Calculate the weighted sum of the data source authority weight, the logical rigor score, and the novelty score, and use the weighted sum as the overall quality score of the target.

[0011] Furthermore, the process of S4 specifically includes: Set a round counter and set its initial value to 1; assign the target support credit argument data in the next cycle to the support credit argument data in the current cycle, assign the target opposition credit argument data in the next cycle to the opposition credit argument data in the current cycle, and assign the updated historical debate record to the historical debate record in the current cycle. Repeat step S3. After each execution of step S3, increment the value of the round counter by 1, and obtain the new target opinion divergence degree and the new target comprehensive quality score obtained after executing step S3. At the same time, based on the structured profile vector and the current period's support credit argument data and opposition credit argument data, check whether any item in the hard rule set is triggered. If any item is triggered, interrupt the repeated execution and directly output the rejection decision as the risk level of the borrower. If no item is triggered, generate a soft compliance constraint set and update the preset soft compliance constraint set. If the degree of disagreement of the new target viewpoint obtained at the moment is lower than the preset disagreement threshold, or if the difference between the comprehensive quality score of the new target obtained at the moment and the comprehensive quality score of the target obtained in the previous round is lower than the preset improvement threshold, or if the value of the debate round counter has reached the preset maximum number of rounds, the repeated execution will be terminated. After terminating repeated execution, the target comprehensive quality score obtained from the last execution of S3 is used as the final comprehensive quality score. All supporting credit argument data and all opposing credit argument data stored in the historical debate record, as well as each soft constraint item in the preset soft compliance constraint set, are used. The final comprehensive quality score is multiplied by the first sum of the preset optimistic weight and prudent weight to obtain the first product. The sum of the influence weights of each soft constraint item is multiplied by the preset compliance weight to obtain the second product. The first product and the second product are added together to obtain the final risk score. The final risk score is mapped to the corresponding risk level according to the preset risk rating mapping table; a five-level tree-structured reasoning chain containing risk rating, adjudication basis, debate rounds, argumentation items, evidence items and data source fields is generated according to the risk level, and the five-level tree-structured reasoning chain is output as the credit risk management plan for the borrower.

[0012] 2) Secondly, the present invention also provides a credit risk assessment system based on multi-agent debate-style collaboration, the specific technical solution of which is as follows: The construction module is used to: collect multi-source heterogeneous data of borrowers through a unified API gateway, standardize the multi-source heterogeneous data into a structured profile vector of the borrower, and determine the initial support credit argument data and the initial opposition credit argument data based on the structured profile vector; The determination module is used to: analyze the initial opposing credit argument data and the initial supporting credit argument data, and combine the structured profile vector and the external knowledge base to determine the supporting credit argument data and opposing credit argument data in the current period, and update the initial opposing credit argument data and the initial supporting credit argument data to the historical debate record; The calculation module is used to: determine the target supporting credence argument data and the target opposing credence argument data for the next period based on the opposing credence argument data, the supporting credence argument data, the historical debate records, the structured profile vector, and the external knowledge base in the current period; update the target supporting credence argument data and the target opposing credence argument data to the historical debate records; and calculate the target opinion divergence degree and the target comprehensive quality score of the target supporting credence argument data and the target opposing credence argument data. The repeat module is used to: take the target support credit argument data, the target opposition credit argument data, and the updated historical debate records in the next cycle as the opposition credit argument data, the support credit argument data, and the historical debate records in the current cycle of the calculation module, repeat the calculation module, until the target opinion divergence degree or the target comprehensive quality score meets the preset conditions, combine the historical debate records, the preset soft compliance constraint set, and the target comprehensive quality score to calculate the final risk score, determine the risk level of the borrower based on the final risk score, and determine the credit risk management plan for the borrower based on the risk level.

[0013] Based on the above solution, the present invention can be further improved as follows.

[0014] Furthermore, the building module is specifically used for: The multi-source heterogeneous data of the borrower is obtained through a unified API gateway. Missing value imputation, outlier detection and format unification are performed on each data source in the multi-source heterogeneous data. The processed fields of each data source are encoded into a structured dictionary according to the preset profile dimensions. The structured dictionary includes at least: original field values, derived indicator values, data timestamps and data source identifiers. All structured dictionaries corresponding to all data sources are aggregated into a structured profile vector of the borrower; the structured profile vector is input into a pre-trained optimistic assessment agent, which generates initial credit support argument data. The initial credit support argument data includes: a first textual claim, a first evidence list, and a first confidence score. Each item in the first evidence list is associated with at least one specific data source field in the multi-source heterogeneous data. The structured profile vector is input into a pre-trained prudent assessment agent, which generates initial objection argument data. The initial objection argument data includes: a second textual claim, a second list of evidence, and a second confidence score. Each item in the second list of evidence is associated with at least one specific data source field in the multi-source heterogeneous data.

[0015] Furthermore, the determining module is specifically used for: The initial support and opposition arguments are stored as the first record in the historical debate record. The second evidence list and second confidence level value in the initial opposition credit argument data are analyzed to identify the first link in the initial opposition credit argument data where the confidence level is lower than a preset threshold or the evidence is missing. Based on the first link where the evidence is missing, the related fields are retrieved from the structured profile vector, and the first related data is retrieved from the positive evidence knowledge base in the external knowledge base. Based on the retrieval results, the support credit argument data for the current period is generated. The first evidence list and first confidence value in the initial support credit argument data are analyzed, and the second link in the initial support credit argument data with a confidence value lower than a preset threshold or missing evidence is identified. Based on the second link with missing evidence, the related fields are retrieved from the structured profile vector, and the second related data is retrieved from the negative evidence knowledge base in the external knowledge base. Based on the retrieval results, the opposition credit argument data for the current period is generated. The supporting credit argument data in the current period includes: a third text claim, a third evidence list, and a third confidence level value. The opposing credit argument data in the current period includes: a fourth text claim, a fourth evidence list, and a fourth confidence level value. Each item in the third evidence list and the fourth evidence list is associated with at least one specific data source field in the multi-source heterogeneous data.

[0016] Furthermore, the computing module is specifically used for: The opposing credit argument data and the supporting credit argument data for the current period are appended and stored in the historical debate record; The fourth evidence list and fourth confidence level value in the current period's opposing credit argument data are analyzed to identify the third link in the current period's opposing credit argument data where the confidence level is lower than a preset threshold or the evidence is missing. Based on the third link where the evidence is missing, the related fields are retrieved from the structured profile vector, and the third related data is retrieved from the positive evidence knowledge base in the external knowledge base. Based on the retrieval results, the target supporting credit argument data for the next period is generated. The third evidence list and third confidence level value in the supporting credit argument data of the current period are analyzed. The fourth link in the supporting credit argument data of the current period with a confidence level lower than a preset threshold or missing evidence is identified. Based on the fourth link with missing evidence, the related fields are retrieved from the structured profile vector, and the fourth related data is retrieved from the negative evidence knowledge base in the external knowledge base. Based on the retrieval results, the target opposing credit argument data for the next period is generated. The data supporting the credit granting of the target and the data opposing the credit granting of the target are added and stored in the historical debate record; Extract the first semantic vector corresponding to the target support credit argument data, extract the second semantic vector corresponding to the target opposition credit argument data, calculate the cosine distance between the first semantic vector and the second semantic vector, and use the cosine distance as the target opinion divergence degree; Obtain the data source authority weight, logical rigor score, and novelty score of the evidence cited in the data supporting the credibility of the target and the data opposing the credibility of the target. Calculate the weighted sum of the data source authority weight, the logical rigor score, and the novelty score, and use the weighted sum as the overall quality score of the target.

[0017] Furthermore, the repeating module is specifically used for: Set a round counter and set its initial value to 1; assign the target support credit argument data in the next cycle to the support credit argument data in the current cycle, assign the target opposition credit argument data in the next cycle to the opposition credit argument data in the current cycle, and assign the updated historical debate record to the historical debate record in the current cycle. The calculation module is repeated. After each execution of the calculation module, the value of the round counter is incremented by 1, and the new target opinion divergence degree and new target comprehensive quality score obtained after executing the calculation module are obtained. At the same time, based on the structured profile vector and the current period's support credit argument data and opposition credit argument data, it is detected whether any item in the hard rule set is triggered. If any item is triggered, the repeated execution is interrupted and a rejection decision is directly output as the risk level of the borrower. If no item is triggered, a soft compliance constraint set is generated and the preset soft compliance constraint set is updated. If the degree of disagreement of the new target viewpoint obtained at the moment is lower than the preset disagreement threshold, or if the difference between the comprehensive quality score of the new target obtained at the moment and the comprehensive quality score of the target obtained in the previous round is lower than the preset improvement threshold, or if the value of the debate round counter has reached the preset maximum number of rounds, the repeated execution will be terminated. After terminating repeated execution, the target comprehensive quality score obtained from the last execution of the calculation module is used as the final comprehensive quality score. All supporting credit argument data and all opposing credit argument data stored in the historical debate record, as well as each soft constraint item in the preset soft compliance constraint set, are used. The final comprehensive quality score is multiplied by the first sum of the preset optimistic weight and prudent weight to obtain the first product. The sum of the influence weights of each soft constraint item is multiplied by the preset compliance weight to obtain the second product. The first product and the second product are added together to obtain the final risk score. The final risk score is mapped to the corresponding risk level according to the preset risk rating mapping table; a five-level tree-structured reasoning chain containing risk rating, adjudication basis, debate rounds, argumentation items, evidence items and data source fields is generated according to the risk level, and the five-level tree-structured reasoning chain is output as the credit risk management plan for the borrower.

[0018] 3) In a third aspect, the present invention also provides an electronic device, the electronic device including a processor coupled to a memory, the memory storing at least one computer program, the at least one computer program being loaded and executed by the processor to enable the electronic device to perform any of the above methods.

[0019] 4) In a fourth aspect, the present invention also provides a computer-readable storage medium storing at least one computer program, which is loaded and executed by a processor to enable a computer to implement any of the above methods.

[0020] It should be noted that the beneficial effects of the technical solutions of the second to fourth aspects of the present invention and their corresponding possible implementations can be found in the above description of the technical effects of the first aspect and its corresponding possible implementations, and will not be repeated here. Attached Figure Description

[0021] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a flowchart illustrating a credit risk assessment method based on multi-agent debate-based collaboration, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the model architecture of a credit risk assessment method based on multi-agent debate-based collaboration according to an embodiment of the present invention. Figure 3 This is a time-series diagram of the multi-agent debate process in a credit risk assessment method based on multi-agent debate-based collaboration, according to an embodiment of the present invention. Figure 4This is a schematic diagram of the debate convergence determination algorithm for a credit risk assessment method based on multi-agent debate-based collaboration, according to an embodiment of the present invention. Figure 5 This is a schematic diagram of scenario-adaptive agent configuration for a credit risk assessment method based on multi-agent debate-style collaboration, according to an embodiment of the present invention. Figure 6 This is a schematic diagram of a traceable reasoning chain structure for a credit risk assessment method based on multi-agent debate-style collaboration, according to an embodiment of the present invention. Figure 7 This is a schematic diagram of the embedded constraint process of a compliance agent in a credit risk assessment method based on multi-agent debate-style collaboration, according to an embodiment of the present invention. Figure 8 This is a structural framework diagram of an electronic device according to the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0023] like Figures 1 to 7 As shown in the figure, a credit risk assessment method based on multi-agent debate-based collaboration according to an embodiment of the present invention includes the following steps: S1. Collect multi-source heterogeneous data of borrowers through a unified API gateway, standardize the multi-source heterogeneous data into a structured profile vector of the borrower, and determine the initial support credit argument data and the initial opposition credit argument data based on the structured profile vector. S2. Analyze the initial opposing credit argument data and the initial supporting credit argument data, and combine the structured profile vector and external knowledge base to determine the supporting credit argument data and opposing credit argument data in the current period, and update the initial opposing credit argument data and the initial supporting credit argument data to the historical debate record. S3. Based on the opposing credence argument data, supporting credence argument data, historical debate records, structured profile vector, and external knowledge base of the current period, determine the target supporting credence argument data and target opposing credence argument data for the next period, update the target supporting credence argument data and target opposing credence argument data to the historical debate records, and calculate the target viewpoint divergence degree and target comprehensive quality score of the target supporting credence argument data and target opposing credence argument data; S4, the target support credit argument data, the target opposition credit argument data, and the updated historical debate records in the next cycle are used as the opposition credit argument data, the support credit argument data, and the historical debate records in the current cycle of S3. S3 is repeated until the target opinion divergence degree or the target comprehensive quality score meets the preset conditions. Combining the historical debate records, the preset soft compliance constraint set, and the target comprehensive quality score, the final risk score is calculated. The risk level of the borrower is determined based on the final risk score, and the credit risk management plan for the borrower is determined based on the risk level.

[0024] The beneficial effects of the credit risk assessment method based on multi-agent debate-based collaboration provided by this invention are as follows: This invention collects multi-source heterogeneous data of borrowers through a unified API gateway and constructs structured profile vectors. Optimistic and prudent assessment parties engage in multiple rounds of adversarial debate based on these profile vectors and an external knowledge base. Each round updates historical debate records and calculates the degree of disagreement and overall quality score. The debate terminates based on adaptive convergence conditions. This invention addresses the limitations of single-model perspectives, the difficulty in reconciling contradictions among multi-dimensional risk assessments, the lack of adversarial verification processes in traditional integration methods, and the untraceability of the decision-making reasoning chain. It achieves a comprehensive improvement in risk identification, enhanced interpretability and auditability of decisions, and increased accuracy in default prediction. Simultaneously, by embedding a set of soft compliance constraints into the adjudication process, it enhances compliance risk control capabilities and shortens approval timelines.

[0025] For ease of understanding, the technical terms used in this application are explained in a standardized manner: Unified Application Programming Interface Gateway (API Gateway): This gateway is used to centrally manage and schedule access requests from multiple data source interfaces, enabling parallel data acquisition and protocol conversion.

[0026] Borrowing entities: Natural persons or legal entities applying for credit, such as small and micro business owners, individual consumers, or upstream and downstream enterprises in the supply chain.

[0027] Multi-source heterogeneous data: Original data originating from different information systems and differing in data structure, format, and semantics, including: credit data, tax data, business registration data, bank transaction data, industry prosperity data, and public opinion data.

[0028] Standardization: The process of converting raw data into a unified format, unified unit, and unified encoding, including missing value imputation, outlier detection, and format normalization.

[0029] Structured profile vector: Encodes the borrower's multi-dimensional feature information into a fixed-length numerical vector, with each dimension corresponding to a preset profile feature (such as income stability, debt ratio, industry risk index, etc.).

[0030] Initial support for credit argument data: The first structured argument for credit is generated by the Optimistic Evaluation Agent (OEA) before the debate begins, including the first textual claim, the first list of evidence, and the first confidence score.

[0031] Initial objection to credit argument data: The first structured argument against credit or request for higher access generated by the Cautious Evaluation Agent (CEA) before the debate begins, including the second textual claim, the second list of evidence, and the second confidence level value.

[0032] External knowledge base: A pre-stored collection of domain knowledge, specifically divided into a positive evidence knowledge base (denoted as K_positive) and a negative evidence knowledge base (denoted as K_cautious), which respectively provide background facts and industry rules supporting and opposing credit.

[0033] The Positive Evidence Knowledge Base is a pre-built collection of facts, patterns, and case information that support credit granting decisions. It provides optimistic assessment agents with the knowledge support for "positive arguments," enabling them to construct arguments supporting credit granting from positive dimensions such as enterprise growth, industry opportunities, operational advantages, and customer loyalty. This includes, but is not limited to: positive indicators of macroeconomic and industry prosperity: industry average growth rate, median industry profitability, upward trend in industry prosperity, emerging market opportunities, etc. For example, "the software industry's average revenue growth rate over the past three years is 15%" and "the industry is currently in an expansion phase." Successful business case studies: anonymized cases of companies in the same industry or of similar size that have successfully obtained credit and achieved healthy development under similar financial conditions, used for comparative argumentation. For example, "a small technology company with annual revenue of 50 million and one M1 delinquency received an A+ rating from a bank." Positive financial patterns: reference ranges of financial indicators reflecting healthy enterprise operation, such as a current ratio greater than 1.5, a debt-to-equity ratio less than 60%, and a revenue growth rate exceeding the industry average. These are used to determine whether the borrower's indicators are within a "safe range." Common sense about customer loyalty and supply chain stability: such as "the average cooperation period of upstream and downstream suppliers of core enterprises exceeds 3 years and can be regarded as stable" and "a customer repurchase rate of more than 50% indicates that the product has a market".

[0034] The Negative Evidence Knowledge Base is a pre-built collection of risk facts, risk patterns, and cautionary cases that are unfavorable to credit decisions. This knowledge base provides prudential assessment agents with the knowledge support for "reverse argumentation," enabling them to construct arguments against credit granting or demand higher access standards from negative dimensions such as financial risk, industry downturn, related risks, and liquidity pressure. This includes, but is not limited to: Industry risk warning information: industry overcapacity, price wars, technological substitution, environmental protection-related production restrictions, and downward trends in industry prosperity. Financial risk characteristic database: reflecting typical characteristics and thresholds of corporate financial deterioration, such as "a quick ratio below 0.5 indicates severe insufficiency in short-term solvency," and "accounts receivable turnover days exceeding twice the industry average indicate difficulty in collection." Related risk cases: de-sensitized cases of risk events such as corporate asset transfers through related-party transactions and broken guarantee chains.

[0035] Historical debate records: A collection of all arguments supporting and opposing credence, stored in order of debate rounds. Each record contains the full text of the argument, the timestamp of its generation, and the round it belongs to.

[0036] Current cycle: The current round in a multi-round debate, corresponding to the k-th round of the claims (k≥1).

[0037] Next cycle: The subsequent round of the current cycle, corresponding to the (k+1)th round of the claim.

[0038] Target Supporting Credentials Data: Supporting credentials generated by the optimistic assessment agent in the next cycle, including third text claims, third evidence lists, and third confidence scores (in the current cycle of S2, supporting credentials data includes third text claims, etc.; in S3, target supporting credentials data also adopts a similar structure, but to distinguish the time sequence, it is still called target supporting credentials data).

[0039] Targeted opposition to creditworthiness data: Opposition to creditworthiness generated by the prudent assessment agent in the next cycle, including the fourth textual claim, the fourth list of evidence, and the fourth confidence level.

[0040] Target viewpoint divergence: The cosine distance, calculated from the semantic vectors of the target's support and opposition arguments in the same round, is used to quantify the degree of difference between the two viewpoints.

[0041] Overall Quality Score for Target: A weighted sum calculated from the data sources of authoritativeness, logical rigor, and novelty scores of the evidence cited in the supporting and opposing arguments for the target, used to assess the reliability and depth of the argument.

[0042] Preset soft compliance constraint set: A set of constraint items dynamically generated by the Compliance Review Agent (CRA) when no hard rules are triggered. Each soft constraint item contains a constraint type field and an impact weight field, describing compliance items that do not meet the hard violation standards but need to be included in risk consideration (e.g., "the proportion of related-party transactions is high but not exceeded").

[0043] Final risk score: A weighted score calculated by combining the argument quality scores from the entire debate history with the impact weights of soft compliance constraints. The score typically ranges from 0 to 1.

[0044] Risk level: Based on the credit risk level mapped by the final risk score, a seven-tier system is adopted: AAA (very low risk), AA (low risk), A (low to medium risk), BBB (medium risk), BB (high to medium risk), B (high risk), and Reject.

[0045] Credit risk management solution: The final output decision results include risk rating and a five-level tree-structured reasoning chain, which can be used to guide subsequent operations such as credit approval, interest rate pricing, and credit limit adjustment.

[0046] Optimistic Evaluation Agent (OEA) is an intelligent agent driven by a large language model. It has a built-in positive evidence knowledge base K_positive and is responsible for generating arguments that support credit from positive dimensions such as enterprise growth and industry opportunities.

[0047] Cautious Evaluation Agent (CEA): This is an intelligent agent driven by a large language model. It has a built-in negative evidence knowledge base K_cautious and is responsible for generating arguments against credit granting or requiring higher access standards from negative dimensions such as financial risk and industry downturn.

[0048] Compliance Review Agent (CRA) is an intelligent agent driven by a hybrid rule engine and a large language model. It has built-in regulatory rule sets such as the "Interim Measures for the Administration of Internet Loans by Commercial Banks" and the "Measures for the Administration of Anti-Money Laundering by Banking Financial Institutions" to detect hard compliance warning lines and output soft constraints.

[0049] Data Acquisition Agent (DAA): This agent is responsible for concurrently calling multiple data source interfaces through a unified API gateway to perform data acquisition and initial cleaning.

[0050] Adjudication Agent (AA): After the debate, the agent is responsible for synthesizing the arguments and compliance results of all parties, calculating the final risk score, and generating a traceable inference chain.

[0051] Furthermore, the textual claims discussed in this paper (including the first textual claim, the second textual claim, etc.) are complete sentences or paragraphs expressed in natural language with a clear opinion bias, directly conveying the assessment agent's core judgment on whether the borrower should receive credit. The textual claim is the concluding statement of the entire argument, and all subsequent evidence serves to support it.

[0052] Similarly, the evidence list is an array (list structure) consisting of multiple evidence items, each of which is a specific factual basis supporting the text's claims. Each piece of evidence must be associated with at least one specific data source field (i.e., a field or derived indicator in the original multi-source heterogeneous data) to ensure the verifiability and traceability of the evidence.

[0053] The confidence score is a floating-point number between 0 and 1, used to quantify the degree of certainty or confidence that the assessment agent has in the arguments it generates (including textual claims and lists of evidence). A higher score indicates that the agent is more confident in the correctness of the argument, while a lower score indicates that the agent has greater uncertainty about the argument or relies on weaker evidence.

[0054] In this scheme, the specific implementation process of S1 includes: S101. Parallel collection of multi-source heterogeneous data through a unified API gateway. The Data Acquisition Agent (DAA) pre-configures an interface routing table for the unified API gateway. This table records the access addresses, authentication tokens, and data formats for the credit reporting interface, transaction log system interface, and Wind industry data interface. Upon receiving a credit application request, the DAA triggers parallel calls, simultaneously sending data query requests to the aforementioned multiple interfaces. Each interface returns data in its original format: Credit reporting data includes credit records, overdue records, and query records; tax data includes VAT declaration amounts, income tax declaration amounts, and growth rates; business registration data includes years of establishment, registered capital, and change records; judicial data includes the number and types of litigation cases; transaction log data includes average daily settlement flow and corporate revenue-to-expenditure ratio; and industry data includes a business climate index.

[0055] S102. Standardize multi-source heterogeneous data. The DAA performs the following operations on the data returned from each data source: Missing value imputation: for numeric fields, the median is used; for categorical fields, the mode is used; and for time series fields, linear interpolation is used. Outlier detection: values ​​exceeding the normal range are identified and removed based on the three-standard-deviation principle, and the removed outliers are replaced with the field mean. Format standardization: all date fields are converted to "YYYY-MM-DD" format, all amount fields are converted to floating-point numbers in ten thousand yuan, and all full-width characters in text fields are converted to half-width characters. Each processed data source is encoded into a structured dictionary containing at least four key-value pairs: original field values ​​(retaining the untransformed original data), derived indicator values ​​(e.g., the growth rate over the past 12 months calculated from the original data, the debt-to-equity ratio, etc.), data timestamp (UTC time of data acquisition), and data source identifier (e.g., "PBOC_CREDIT" represents a bank's credit reporting system, and "TAX" represents taxation).

[0056] S103, Aggregate and generate structured image vectors All structured dictionaries corresponding to data sources are concatenated and normalized according to preset profile dimensions. The preset profile dimensions are set according to actual conditions and are not specifically limited here, but for ease of understanding, the following example is provided: The preset profile dimensions include 7 major categories and 28 sub-dimensions: credit category (including overdue number, credit utilization rate, query frequency), tax category (including revenue scale, revenue growth rate, tax credit rating), business category (including establishment year, registered capital, change frequency), judicial category (including litigation number, amount involved), transaction category (including average daily turnover, income-expenditure ratio, volatility coefficient), industry category (including prosperity index), and public opinion category (including number of negative news and sentiment score). The value of each sub-dimension is mapped to the [0,1] interval through min-max normalization, and finally forms a 28-dimensional floating-point vector, denoted as the structured profile vector P.

[0057] S104. Generate initial support credit argument data and initial opposition credit argument data. The structured profile vector P is input into the pre-trained Optimistic Evaluation Agent (OEA) model and the Cautious Evaluation Agent (CEA) model, respectively.

[0058] The optimistic assessment agent employs a large language model based on the Transformer architecture, fine-tuned on a financial domain corpus. It generates initial support data for credit arguments using a structured profile vector P and a built-in positive evidence knowledge base as input. This initial support data is a JSON structure containing three fields: a first textual claim (e.g., "The company's revenue has grown by 18% in the past 12 months, it belongs to an encouraged industry, and approval is recommended"), a first evidence list (containing multiple pieces of evidence, each presented in the format of "data source field name: specific value", e.g., "Tax_Revenue Growth Rate: 18%", "Industry_Tendency: Encouraged"), and a first confidence score (a floating-point number between 0 and 1, e.g., 0.85, representing the model's confidence level in the claim). Each piece of evidence in the first evidence list is associated with at least one specific data source field from the multi-source heterogeneous data through a field mapping table (e.g., "Tax_Revenue Growth Rate" corresponds to the growth rate calculated from the "cumulative value of VAT declarations in the past 12 months" in the tax declaration data).

[0059] Similarly, the Critical Assessment Agent (CEA) model receives the same structured profile vector P and a built-in negative evidence knowledge base to generate initial objection data. This initial objection data also contains three fields: a second textual claim (e.g., "The company has M1 delinquency records in the past 24 months, and the credit balance and application amount totaling 17 million yuan have a low coverage ratio to sales revenue; it is recommended to raise the admission requirements or reject the application"), a second evidence list (e.g., "Credit Report_M1 Delinquency Count: 1 time", "Transaction_Credit Balance to Sales Revenue Ratio: 0.32"), and a second confidence score (e.g., 0.78). Each piece of evidence in the second evidence list is also associated with a specific data source field.

[0060] By collecting multi-source data in parallel through a unified API gateway, the heterogeneous data is standardized into structured profile vectors. Pre-trained optimistic and prudent assessment agents are then used to generate initial supporting and opposing credit arguments, respectively. This achieves a comprehensive and multi-faceted initial characterization of the borrower, providing a structured starting point for subsequent adversarial debate. Compared to traditional single-model approaches, this step forces the construction of initial viewpoints from both positive and negative perspectives, avoiding assessment biases caused by the lack of a clear initial perspective.

[0061] In this scheme, the specific implementation process of S2 includes: S201, Store the initial arguments to the historical debate record. Create an empty historical debate record table, with the round as the primary key. Each row contains the round number, supporting arguments (including textual claims, evidence lists, and confidence levels), opposing arguments, and a timestamp. Store the initial supporting arguments (including the first textual claim, the first evidence list, and the first confidence level value) and the initial opposing arguments (including the second textual claim, the second evidence list, and the second confidence level value) generated in S104 as the record for round 0 in this table. Initialize the current period index k to 1.

[0062] S202, Optimistically assess the agent's analysis of initial opposing credit arguments and generate supporting credit arguments for the current period. The Optimistic Assessment Agent (OEA) is invoked, taking into account the structured profile vector P, the list of second evidence and the second confidence level value from the initial objection argument data, and historical debate records (currently only including round 0 records). The OEA parses the list of second evidence, identifying evidence items with a confidence level below a preset threshold (e.g., 0.6), or key fields that are clearly missing from the evidence list (e.g., the initial objection argument mentions "liquidity pressure" but does not cite any bank transaction data). These identified weaknesses are called the first weaknesses. Based on the first weaknesses, the OEA retrieves relevant fields from the structured profile vector P: if the weakness involves "liquidity pressure," it retrieves sub-dimensions such as "average daily settlement flow," "corporate income-to-expenditure ratio," and "volatility coefficient" from the transaction flow profile; simultaneously, it retrieves supplementary first-related data from the positive evidence knowledge base, such as retrieving "general liquidity level in the software industry." Based on the above search results, OEA generates supporting credit argument data for the current period. This supporting credit argument data includes third-party textual claims (e.g., "Despite a single M1 delinquency, the company's average daily cash flow over the past six months has remained stable at over RMB XX,000, and the revenue-to-expenditure ratio of 'a' indicates ample operating cash flow and controllable liquidity risk"), a list of third-party evidence (containing specific values ​​extracted from the profile vector and external information referenced from the knowledge base), and a third-party confidence score (e.g., 0.82). Each piece of evidence in the third-party evidence list is associated with a specific data source field in the multi-source heterogeneous data (e.g., "Transaction_Average Daily Settlement Cash Flow" corresponds to a field in the bank's transaction flow system).

[0063] S203, Carefully evaluate the initial supporting credit argument in Agent analysis and generate opposing credit arguments for the current period. The Critical Assessment Agent (CEA) is invoked, taking into account the structured profile vector P, the first evidence list and first confidence level value from the initial supporting credential argument data, and historical debate records. The CEA parses the first evidence list, identifying evidence items with a confidence level below a preset threshold (e.g., 0.6), or logical loopholes in the evidence (e.g., the initial supporting argument emphasizes "industry prosperity index 0.72" but fails to analyze the trend of prosperity index changes). These weak links are referred to as the second link (corresponding to "second link" in claim 3). Based on the second link, the CEA retrieves relevant fields from the structured profile vector P (e.g., retrieves the month-on-month change value of the industry prosperity index over the past three months) and supplements second relevant data from the negative evidence knowledge base (e.g., retrieves "cases of marginal deterioration in the software industry prosperity index" and "policy tightening warning signals"). Based on the search results, CEA generates opposition data for the current period. This opposition data includes a fourth textual claim (e.g., "Although the industry prosperity index of 0.72 is higher than the average, it has decreased by 4% month-on-month in the past three months, indicating that the industry has entered a downward trend. Coupled with the company's history of single overdue payments, the credit risk is rising"), a fourth evidence list (including "Industry_Prosperity Index Month-on-Month Change: -4%" extracted from the profile vector), and a fourth confidence score (e.g., 0.75). Each piece of evidence in the fourth evidence list is also associated with a specific data source field.

[0064] S204, add the argument data for the current period to the historical debate record. The supporting argument data (third text claim, third evidence list, third confidence level value) generated in S202 and the opposing argument data (fourth text claim, fourth evidence list, fourth confidence level value) generated in S203 are stored as the first round record in the historical debate record table.

[0065] By having optimistic and prudent assessment agents respectively refute and supplement the opposing side's initial arguments, a first round of "hard confrontation" debate was achieved. Compared to the "soft consensus" in existing integration methods, this step forces both parties to provide evidence-based rebuttals based on specific data source fields, significantly improving the depth and traceability of risk arguments. Simultaneously, the continuous accumulation of historical debate records provides a complete data foundation for subsequent rulings and the generation of traceable reasoning chains.

[0066] In this solution, the specific implementation process of S3 includes: S301, append the argument data for the current period to the historical debate record. Upon entering the k-th round (k≥1) of the debate, the supporting and opposing arguments generated in the previous round (k-1 round) are appended to the historical debate record. This operation is a repetition of S204, ensuring that the historical record fully reflects all rounds.

[0067] S302, Generate target support credit justification data for the next cycle. The Optimistic Evaluation Agent (OEA) is invoked, taking into account the structured profile vector P, the list of fourth evidence and the fourth confidence level value from the opposing credence argument data of the current period (round k-1) (i.e., opposing credence arguments generated by S203 or the previous round S304; it should be noted that the reason for mentioning the previous round S304 here is: when step S3 is executed for the first time, the opposing credence argument data at this time is the opposing credence argument data of S203; if S4 has been executed and S3 is executed again, then the opposing credence argument data is the opposing credence argument data generated by the previous round S304), and all historical debate records. The OEA analyzes the opposing credence argument data of the current period, identifying the links where the confidence level is lower than a preset threshold (e.g., 0.6) or where evidence is missing, which are called the third links. In the third stage, OEA retrieves relevant fields from the structured profile vector P (e.g., if the opposing argument points to "accounts receivable turnover days being too long," then it retrieves the "accounts receivable turnover days" field from the transaction profile), and retrieves third-level relevant data from the positive evidence knowledge base (e.g., "average discount rate in the accounts receivable factoring market" and "core enterprise payment credit records"). Based on the search results, OEA generates target support credit argument data for the next cycle (round k). This target support credit argument data is structurally consistent with the aforementioned support credit argument data, including textual claims, a list of evidence, and confidence scores. Each piece of evidence in the evidence list is associated with a specific data source field.

[0068] S303, Generate target opposition creditworthiness argument data for the next cycle. The CEA (Conscious Evaluation Agent) is invoked, taking into account the structured profile vector P, the list of third evidence and third confidence levels from the supporting credible arguments data of the current period (round k-1) (i.e., the supporting credible arguments generated in S202 or the previous round S303), and all historical debate records. The CEA analyzes the supporting credible arguments data of the current period, identifying stages where the confidence level is below a preset threshold or where evidence is missing, termed the fourth stage. Based on the fourth stage, the CEA retrieves relevant fields from the structured profile vector P and fourth-related data from the negative evidence knowledge base (e.g., retrieving "increased trend of delinquency rates in the same industry" and "regulatory penalty cases"). Based on the search results, the CEA generates the target opposing credible arguments data for the next period (round k), which also includes textual claims, a list of evidence, confidence levels, and specific data source fields related to the evidence.

[0069] S304, Add target argument data to historical debate records The target support and credit argument data generated in S302 and the target opposition and credit argument data generated in S303 are added as records of round k and stored in the historical debate record table.

[0070] S305, Calculate the degree of divergence of the target viewpoint. The first semantic vector corresponding to the target supporting credit argument data and the second semantic vector corresponding to the target opposing credit argument data are extracted. Specifically, a pre-trained financial domain language model (such as FinBERT, a BERT variant pre-trained for financial text) is used to jointly encode the textual claims and evidence lists in the argument data: the textual claims of the argument and the textual descriptions of each piece of evidence in the evidence list are concatenated into a string, which is then input into the language model. The output vector corresponding to the [CLS] marker of the last layer of the model is taken as the semantic vector of the argument. For the target supporting credit argument data, this vector is called the first semantic vector; for the target opposing credit argument data, this vector is called the second semantic vector.

[0071] Calculate the cosine distance between the first semantic vector and the second semantic vector. The formula for calculating the cosine distance is as follows: divide the dot product of the two vectors by the product of their magnitudes to obtain the cosine similarity; then subtract this cosine similarity from 1 to get the cosine distance. This cosine distance is the target opinion divergence degree D_k. The divergence degree D_k ranges from [0, 2], with a larger value indicating a greater difference in opinions between the two sides.

[0072] S306, Calculate the overall quality score of the target. For both the data supporting and opposing the creditworthiness of the target, we iterate through each piece of evidence in its evidence list and obtain three scoring dimensions: Data Source Authority Weight: A weight value is assigned based on the type of data source associated with the evidence. For example: credit data has a weight of 1.0, business registration data has a weight of 0.9, industry transaction data has a weight of 0.8, judicial data has a weight of 0.7, industry sentiment data has a weight of 0.6, and third-party public opinion data has a weight of 0.3. The authority weight of each piece of evidence is directly taken from the preset value of the corresponding data source.

[0073] Logical Rigor Score: This score is output synchronously by the Agent when generating arguments, and ranges from 0 to 1. It is calculated by the Agent's internal logical consistency assessment module, based on whether the argument satisfies rules such as "evidence supports the claim" and "refutation targets the opponent's core points."

[0074] Novelty score: Calculated by the Agent by comparing the degree of overlap between the evidence cited in the current argument and evidence that has appeared in the historical debate record. If a piece of evidence has never appeared in the historical record, the novelty score is 1.0; if it appears repeatedly, it is calculated as 1 / (number of occurrences + 1). The novelty score for a single piece of evidence ranges from 0 to 1.

[0075] For each round, the overall target quality score is calculated as follows: The weighted sum of the three scores mentioned above for all evidence in that round is calculated, where the weights for evidence reliability (α), logical rigor (β), and novelty (γ) satisfy α + β + γ = 1, for example, α = 0.5, β = 0.3, γ = 0.2 (these weights can be pre-configured). Specifically, for each piece of evidence e, its individual quality score Q_e = α × authority weight + β × logical rigor score + γ × novelty score. The arithmetic mean of the individual quality scores for all evidence in that round is taken to obtain the overall target quality score Q_k for that round.

[0076] This step automates the iterative process of multi-round adversarial debate. In each round, an optimistic assessment agent and a prudent assessment agent provide evidence to refute the weaknesses in the opponent's latest arguments, continuously deepening the understanding of the borrower's risk characteristics. Simultaneously, by calculating the degree of disagreement (cosine distance of semantic vectors) and a comprehensive quality score (evidence reliability + logical rigor + novelty), quantitative evidence is provided for debate convergence, avoiding infinite loops and effectively controlling computational costs.

[0077] In this scheme, the specific implementation process of S4 includes: S401, Initialize the round counter and set the loop variable. Set up a round counter, initializing it to 1. This round counter records the round number of the currently ongoing k-th round of debate. Simultaneously, assign the target support credibility argument data for the next cycle (round 1) obtained in S3 to the support credibility argument data for the current cycle, assign the target opposition credibility argument data for the next cycle to the opposition credibility argument data for the current cycle, and assign the updated historical debate record (which now includes rounds 0 and 1) to the historical debate record for the current cycle.

[0078] S402, Repeat S3 and embed compliance review Enter the loop and repeat S3. After each execution of S3 (i.e., after completing one round of debate), perform the following operations: Increment the value of the round counter by 1.

[0079] Obtain the new target opinion divergence degree (denoted as D_counter) and the new target comprehensive quality score (denoted as Q_counter) obtained after this execution of S3.

[0080] The Compliance Review Agent (CRA) is invoked. The CRA receives a structured profile vector P and supporting and opposing credit argument data for the current period (i.e., the two arguments generated in the latest round). The CRA has a built-in set of hard rules, R_hard, which includes typical rules such as: Rule HR_001 "Borrower involved in irregular financing list", Rule HR_002 "Number of overdue payments ≥ 3 in the past 24 months", Rule HR_003 "Industry belongs to prohibited credit industries" (e.g., high-pollution, high-energy-consumption restricted categories), and Rule HR_004 "Related party transaction amount exceeds 50% of net assets and cannot provide genuine trade background", etc. The CRA checks each item in the current data and argument content to see if it triggers any of the above items. If any item is triggered, the loop is immediately interrupted, and the process jumps to S404 to directly output a rejection decision. If no hard rules are triggered, the CRA further generates a set of soft compliance constraints, C. The set of soft compliance constraints is a list, where each element is a soft constraint item, containing a constraint type field (e.g., "high level of related-party transactions," "non-compliant watchlist," etc.) and an impact weight field (values ​​from 0 to 1, indicating the degree of negative impact of the constraint on the final rating). The generated set of soft constraints is then updated into the preset set of soft compliance constraints (initially an empty set).

[0081] S403, Determining the Conditions for Termination of Debate After each round of compliance review, determine whether any one of the following three termination conditions is met: Condition 1: The target viewpoint divergence degree D_counter in the current round is lower than the preset divergence threshold θ_D. θ_D is usually set to 0.15, indicating that the differences in viewpoints between the two sides are very small and tend towards consensus.

[0082] Condition 2: The difference between the target overall quality score Q_counter in the current round and the target overall quality score Q_{counter-1} in the previous round is lower than the preset improvement threshold θ_Q. θ_Q is usually set to 0.02, indicating that the improvement in argument quality is less than 2%, and the marginal benefit of continuing the debate is very low.

[0083] Condition 3: The round counter reaches the preset maximum round number K_max. For micro and small enterprise loan products, K_max is set to 5; for personal consumer credit, K_max is set to 3; and for supply chain finance, K_max is set to 4.

[0084] If any of the above conditions are met, the repeated execution loop will terminate and the process will proceed to the adjudication phase in S405. If the conditions are not met, the supporting credit argument data, opposing credit argument data, and historical debate records for the current cycle will be updated to the latest values, and then the process will jump back to S402 to continue the next round of debate.

[0085] S404, hard compliance rejection (branch interruption) If any hard rule in S402 is triggered, all debate immediately ceases, and the decision agent (AA) directly outputs "Reject" as the borrower's risk level, noting the triggered rule number and the corresponding data source field in the inference chain. For example, "Hard rule HR_002 triggered: 4 overdue payments in the past 24 months, data sourced from the 'Overdue Payments' field of a certain bank's credit report." This rejection decision does not require weighted scoring calculation.

[0086] S405, Final Risk Score Calculation After the debate is properly concluded (i.e., any of the conditions in S403 are met), the adjudication agent performs the following calculations: Obtain the target overall quality score obtained from the last execution of S3, and denot it as the final overall quality score Q_final (i.e. Q_counter).

[0087] Retrieves all supporting and opposing arguments from the historical debate log, covering all rounds. Note that the historical debate log includes all arguments from round 0 (the initial round) and rounds 1 through 'counter'.

[0088] Obtain the preset set of soft compliance constraints (a set updated multiple times) and extract the influence weights of all soft constraint items in it.

[0089] Calculate the first product: Multiply the final comprehensive quality score Q_final by the first sum of the preset optimistic weight ω_O and prudent weight ω_C (i.e., ω_O + ω_C). The values ​​of the optimistic weight and prudent weight are dynamically loaded according to the type of credit product. For example, in micro and small enterprise loans, ω_O = 0.30, ω_C = 0.45, and the first sum is 0.75. First product = Q_final × (ω_O + ω_C).

[0090] Calculate the second product: Multiply the sum of the impact weights of each soft constraint by the preset compliance weight ω_compliance. Let there be m soft constraints, each with an impact weight w_i (i=1..m), then the sum of the impact weights is sum(w_i). The second product = sum(w_i) × ω_compliance. ω_compliance is also preset according to the product type; for example, in micro and small enterprise loans, ω_compliance = 0.25.

[0091] The final risk score, Score_final, is calculated as the sum of the first and second products. Note that sum(w_i) in the second product is typically small (e.g., 0.1 to 0.5), and the overall Score_final ranges from 0 to 1.

[0092] S406, Risk Level Mapping The final risk score is mapped to the corresponding risk level based on a pre-defined risk rating mapping table. An example mapping table is shown below (it can be adjusted according to business needs): Score_final ≥ 0.85 → AAA; 0.75 ≤ Score_final<0.85 → AA; 0.65 ≤ Score_final<0.75 → A; 0.55 ≤ Score_final<0.65 → BBB; 0.45 ≤ Score_final<0.55 → BB; 0.30 ≤ Score_final<0.45 → B; Score_final<0.30 → Reject.

[0093] S407, Generate a five-level tree-structured reasoning chain as a credit risk management solution. The adjudication agent generates a five-level tree-structured reasoning chain based on the above process, as follows: Level 1 (Risk Rating): Outputs the risk level (e.g., "A") determined in step S406.

[0094] Level 2 (Basis for Adjudication): Outputs the calculation formula for the final risk score and the values ​​of each weight, while also listing the soft constraints that are referenced.

[0095] Level 3 (Debate Rounds): List all round numbers in chronological order and indicate which round triggered the termination condition (e.g., "Round 2, divergence 0.12 < 0.15, terminate debate").

[0096] Level 4 (Argument Items): For each round, output the textual claims and confidence scores generated by the optimistic assessment agent and the prudent assessment agent, respectively.

[0097] Level 5 (Evidence Item → Data Source Field): For each piece of evidence in each argument, expand to display the specific content of the evidence and its associated data source field (including data source name, field path, and original value). For example, "Evidence 'Revenue growth rate of 18% in the past 12 months' → Data source: Official website of a certain entity, Field: Value-added tax declaration amount_cumulative_past 12 months, calculated growth rate of 18%".

[0098] The aforementioned five-level tree-structured reasoning chain, along with risk levels, is encapsulated into a credit risk management solution and output to the approval system or user interface. This solution can be directly used for credit approval decisions, regulatory audit follow-up, and post-loan risk management.

[0099] The compliance review agent's hard-rule veto mechanism mitigates compliance risks at the architectural level. Even if the arguments of both the optimistic assessment agent and the prudent assessment agent deviate from regulatory requirements, the compliance review agent can still interrupt the process and output a rejection decision. The final output's five-level tree-like reasoning chain achieves complete traceability from the final rating to the original data source fields. Furthermore, by influencing the weighted score through a set of soft compliance constraints, risk factors that do not meet the hard violation standards but still require attention are incorporated into the decision-making process, achieving a two-layer compliance defense line of "hard constraint flow interruption + soft constraint weight adjustment".

[0100] Furthermore, the process S1 specifically includes: The multi-source heterogeneous data of the borrower is obtained through a unified API gateway. Missing value imputation, outlier detection and format unification are performed on each data source in the multi-source heterogeneous data. The processed fields of each data source are encoded into a structured dictionary according to the preset profile dimensions. The structured dictionary includes at least: original field values, derived indicator values, data timestamps and data source identifiers. All structured dictionaries corresponding to all data sources are aggregated into a structured profile vector of the borrower; the structured profile vector is input into a pre-trained optimistic assessment agent, which generates initial credit support argument data. The initial credit support argument data includes: a first textual claim, a first evidence list, and a first confidence score. Each item in the first evidence list is associated with at least one specific data source field in the multi-source heterogeneous data. The structured profile vector is input into a pre-trained prudent assessment agent, which generates initial objection argument data. The initial objection argument data includes: a second textual claim, a second list of evidence, and a second confidence score. Each item in the second list of evidence is associated with at least one specific data source field in the multi-source heterogeneous data.

[0101] Furthermore, the process S2 specifically includes: The initial support and opposition arguments are stored as the first record in the historical debate record. The second evidence list and second confidence level value in the initial opposition credit argument data are analyzed to identify the first link in the initial opposition credit argument data where the confidence level is lower than a preset threshold or the evidence is missing. Based on the first link where the evidence is missing, the related fields are retrieved from the structured profile vector, and the first related data is retrieved from the positive evidence knowledge base in the external knowledge base. Based on the retrieval results, the support credit argument data for the current period is generated. The first evidence list and first confidence value in the initial support credit argument data are analyzed, and the second link in the initial support credit argument data with a confidence value lower than a preset threshold or missing evidence is identified. Based on the second link with missing evidence, the related fields are retrieved from the structured profile vector, and the second related data is retrieved from the negative evidence knowledge base in the external knowledge base. Based on the retrieval results, the opposition credit argument data for the current period is generated. The supporting credit argument data in the current period includes: a third text claim, a third evidence list, and a third confidence level value. The opposing credit argument data in the current period includes: a fourth text claim, a fourth evidence list, and a fourth confidence level value. Each item in the third evidence list and the fourth evidence list is associated with at least one specific data source field in the multi-source heterogeneous data.

[0102] Furthermore, the S3 process specifically includes: The opposing credit argument data and the supporting credit argument data for the current period are appended and stored in the historical debate record; The fourth evidence list and fourth confidence level value in the current period's opposing credit argument data are analyzed to identify the third link in the current period's opposing credit argument data where the confidence level is lower than a preset threshold or the evidence is missing. Based on the third link where the evidence is missing, the related fields are retrieved from the structured profile vector, and the third related data is retrieved from the positive evidence knowledge base in the external knowledge base. Based on the retrieval results, the target supporting credit argument data for the next period is generated. The third evidence list and third confidence level value in the supporting credit argument data of the current period are analyzed. The fourth link in the supporting credit argument data of the current period with a confidence level lower than a preset threshold or missing evidence is identified. Based on the fourth link with missing evidence, the related fields are retrieved from the structured profile vector, and the fourth related data is retrieved from the negative evidence knowledge base in the external knowledge base. Based on the retrieval results, the target opposing credit argument data for the next period is generated. The data supporting the credit granting of the target and the data opposing the credit granting of the target are added and stored in the historical debate record; Extract the first semantic vector corresponding to the target support credit argument data, extract the second semantic vector corresponding to the target opposition credit argument data, calculate the cosine distance between the first semantic vector and the second semantic vector, and use the cosine distance as the target opinion divergence degree; Obtain the data source authority weight, logical rigor score, and novelty score of the evidence cited in the data supporting the credibility of the target and the data opposing the credibility of the target. Calculate the weighted sum of the data source authority weight, the logical rigor score, and the novelty score, and use the weighted sum as the overall quality score of the target.

[0103] Furthermore, the process of S4 specifically includes: Set a round counter and set its initial value to 1; assign the target support credit argument data in the next cycle to the support credit argument data in the current cycle, assign the target opposition credit argument data in the next cycle to the opposition credit argument data in the current cycle, and assign the updated historical debate record to the historical debate record in the current cycle. Repeat step S3. After each execution of step S3, increment the value of the round counter by 1, and obtain the new target opinion divergence degree and the new target comprehensive quality score obtained after executing step S3. At the same time, based on the structured profile vector and the current period's support credit argument data and opposition credit argument data, check whether any item in the hard rule set is triggered. If any item is triggered, interrupt the repeated execution and directly output the rejection decision as the risk level of the borrower. If no item is triggered, generate a soft compliance constraint set and update the preset soft compliance constraint set. If the degree of disagreement of the new target viewpoint obtained at the moment is lower than the preset disagreement threshold, or if the difference between the comprehensive quality score of the new target obtained at the moment and the comprehensive quality score of the target obtained in the previous round is lower than the preset improvement threshold, or if the value of the debate round counter has reached the preset maximum number of rounds, the repeated execution will be terminated. After terminating repeated execution, the target comprehensive quality score obtained from the last execution of S3 is used as the final comprehensive quality score. All supporting credit argument data and all opposing credit argument data stored in the historical debate record, as well as each soft constraint item in the preset soft compliance constraint set, are used. The final comprehensive quality score is multiplied by the first sum of the preset optimistic weight and prudent weight to obtain the first product. The sum of the influence weights of each soft constraint item is multiplied by the preset compliance weight to obtain the second product. The first product and the second product are added together to obtain the final risk score. The final risk score is mapped to the corresponding risk level according to the preset risk rating mapping table; a five-level tree-structured reasoning chain containing risk rating, adjudication basis, debate rounds, argumentation items, evidence items and data source fields is generated according to the risk level, and the five-level tree-structured reasoning chain is output as the credit risk management plan for the borrower.

[0104] This invention also provides a credit risk assessment system based on multi-agent debate-based collaboration, the specific technical solution of which is as follows: The construction module is used to: collect multi-source heterogeneous data of borrowers through a unified API gateway, standardize the multi-source heterogeneous data into a structured profile vector of the borrower, and determine the initial support credit argument data and the initial opposition credit argument data based on the structured profile vector; The determination module is used to: analyze the initial opposing credit argument data and the initial supporting credit argument data, and combine the structured profile vector and the external knowledge base to determine the supporting credit argument data and opposing credit argument data in the current period, and update the initial opposing credit argument data and the initial supporting credit argument data to the historical debate record; The calculation module is used to: determine the target supporting credence argument data and the target opposing credence argument data for the next period based on the opposing credence argument data, the supporting credence argument data, the historical debate records, the structured profile vector, and the external knowledge base in the current period; update the target supporting credence argument data and the target opposing credence argument data to the historical debate records; and calculate the target opinion divergence degree and the target comprehensive quality score of the target supporting credence argument data and the target opposing credence argument data. The repeat module is used to: take the target support credit argument data, the target opposition credit argument data, and the updated historical debate records in the next cycle as the opposition credit argument data, the support credit argument data, and the historical debate records in the current cycle of the calculation module, repeat the calculation module, until the target opinion divergence degree or the target comprehensive quality score meets the preset conditions, combine the historical debate records, the preset soft compliance constraint set, and the target comprehensive quality score to calculate the final risk score, determine the risk level of the borrower based on the final risk score, and determine the credit risk management plan for the borrower based on the risk level.

[0105] It should be noted that the beneficial effects of the credit risk assessment system based on multi-agent debate-based collaboration provided in the above embodiments are the same as those of the credit risk assessment method based on multi-agent debate-based collaboration described above, and will not be repeated here. Furthermore, the system provided in the above embodiments is only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the system can be divided into different functional modules according to the actual situation to complete all or part of the functions described above. In addition, the system and method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process is detailed in the method embodiments, and will not be repeated here.

[0106] like Figure 8 As shown, an electronic device 300 according to an embodiment of the present invention includes a processor 320 coupled to a memory 310. The memory 310 stores at least one computer program 330, which is loaded and executed by the processor 320 to enable the electronic device 300 to implement any of the above-mentioned methods. Specifically: The electronic device 300 can vary considerably due to differences in configuration or performance. It may include one or more processors 320 (Central Processing Units, CPUs) and one or more memories 310. The memories 310 store at least one computer program 330, which is loaded and executed by the processors 320 to enable the electronic device 300 to implement the multi-agent debate-based collaborative credit risk assessment method provided in the above embodiments. Of course, the electronic device 300 may also have wired or wireless network interfaces, a keyboard, and input / output interfaces for input and output. It may also include other components for implementing device functions, which will not be elaborated upon here.

[0107] An embodiment of the present invention provides a computer-readable storage medium storing at least one computer program, which is loaded and executed by a processor to enable a computer to implement any of the above-described methods.

[0108] Alternatively, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), magnetic tape, a floppy disk, and an optical data storage device, etc.

[0109] In an exemplary embodiment, a computer program product or computer program is also provided, which includes computer instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the electronic device to perform any of the methods described above.

[0110] It should be noted that the terms "first" and "second" in the specification and claims of this application are used to distinguish similar objects and represent a limitation on a specific order or sequence. Where appropriate, the order of use for similar objects can be interchanged so that the embodiments of this application described herein can be implemented in an order other than that shown or described.

[0111] Those skilled in the art will recognize that this invention can be implemented as a system, method, or computer program product. Therefore, this disclosure can be specifically implemented in the following forms: it can be entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software, generally referred to herein as a "circuit," "module," or "system." Furthermore, in some embodiments, the invention can also be implemented as a computer program product contained in one or more computer-readable media, which includes computer-readable program code.

[0112] Any combination of one or more computer-readable media may be used. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device.

[0113] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A credit risk assessment method based on multi-agent debate-based collaboration, characterized in that, include: S1. Collect multi-source heterogeneous data of borrowers through a unified API gateway, standardize the multi-source heterogeneous data into a structured profile vector of the borrower, and determine the initial support credit argument data and the initial opposition credit argument data based on the structured profile vector. S2. Analyze the initial opposing credit argument data and the initial supporting credit argument data, and combine the structured profile vector and external knowledge base to determine the supporting credit argument data and opposing credit argument data in the current period, and update the initial opposing credit argument data and the initial supporting credit argument data to the historical debate record. S3. Based on the opposing credence argument data, supporting credence argument data, historical debate records, structured profile vector, and external knowledge base of the current period, determine the target supporting credence argument data and target opposing credence argument data for the next period, update the target supporting credence argument data and target opposing credence argument data to the historical debate records, and calculate the target viewpoint divergence degree and target comprehensive quality score of the target supporting credence argument data and target opposing credence argument data; S4, the target support credit argument data, the target opposition credit argument data, and the updated historical debate records in the next cycle are used as the opposition credit argument data, the support credit argument data, and the historical debate records in the current cycle of S3. S3 is repeated until the target opinion divergence degree or the target comprehensive quality score meets the preset conditions. Combining the historical debate records, the preset soft compliance constraint set, and the target comprehensive quality score, the final risk score is calculated. The risk level of the borrower is determined based on the final risk score, and the credit risk management plan for the borrower is determined based on the risk level.

2. The credit risk assessment method based on multi-agent debate-style collaboration according to claim 1, characterized in that, The process S1 specifically includes: The multi-source heterogeneous data of the borrower is obtained through a unified API gateway. Missing value imputation, outlier detection and format unification are performed on each data source in the multi-source heterogeneous data. The processed fields of each data source are encoded into a structured dictionary according to the preset profile dimensions. The structured dictionary includes at least: original field values, derived indicator values, data timestamps and data source identifiers. All structured dictionaries corresponding to all data sources are aggregated into a structured profile vector of the borrower; the structured profile vector is input into a pre-trained optimistic assessment agent, which generates initial credit support argument data. The initial credit support argument data includes: a first textual claim, a first evidence list, and a first confidence score. Each item in the first evidence list is associated with at least one specific data source field in the multi-source heterogeneous data. The structured profile vector is input into a pre-trained prudent assessment agent, which generates initial objection argument data. The initial objection argument data includes: a second textual claim, a second list of evidence, and a second confidence score. Each item in the second list of evidence is associated with at least one specific data source field in the multi-source heterogeneous data.

3. The credit risk assessment method based on multi-agent debate-based collaboration according to claim 1, characterized in that, The process S2 specifically includes: The initial support and opposition arguments are stored as the first record in the historical debate record. The second evidence list and second confidence level value in the initial opposition credit argument data are analyzed to identify the first link in the initial opposition credit argument data where the confidence level is lower than a preset threshold or the evidence is missing. Based on the first link where the evidence is missing, the related fields are retrieved from the structured profile vector, and the first related data is retrieved from the positive evidence knowledge base in the external knowledge base. Based on the retrieval results, the support credit argument data for the current period is generated. The first evidence list and first confidence value in the initial support credit argument data are analyzed, and the second link in the initial support credit argument data with a confidence value lower than a preset threshold or missing evidence is identified. Based on the second link with missing evidence, the related fields are retrieved from the structured profile vector, and the second related data is retrieved from the negative evidence knowledge base in the external knowledge base. Based on the retrieval results, the opposition credit argument data for the current period is generated. The supporting credit argument data in the current period includes: a third text claim, a third evidence list, and a third confidence level value. The opposing credit argument data in the current period includes: a fourth text claim, a fourth evidence list, and a fourth confidence level value. Each item in the third evidence list and the fourth evidence list is associated with at least one specific data source field in the multi-source heterogeneous data.

4. The credit risk assessment method based on multi-agent debate-based collaboration according to claim 1, characterized in that, The S3 process specifically includes: The opposing credit argument data and the supporting credit argument data for the current period are appended and stored in the historical debate record; The fourth evidence list and fourth confidence level value in the current period's opposing credit argument data are analyzed to identify the third link in the current period's opposing credit argument data where the confidence level is lower than a preset threshold or the evidence is missing. Based on the third link where the evidence is missing, the related fields are retrieved from the structured profile vector, and the third related data is retrieved from the positive evidence knowledge base in the external knowledge base. Based on the retrieval results, the target supporting credit argument data for the next period is generated. The third evidence list and third confidence level value in the supporting credit argument data of the current period are analyzed. The fourth link in the supporting credit argument data of the current period with a confidence level lower than a preset threshold or missing evidence is identified. Based on the fourth link with missing evidence, the related fields are retrieved from the structured profile vector, and the fourth related data is retrieved from the negative evidence knowledge base in the external knowledge base. Based on the retrieval results, the target opposing credit argument data for the next period is generated. The data supporting the credit granting of the target and the data opposing the credit granting of the target are added and stored in the historical debate record; Extract the first semantic vector corresponding to the target support credit argument data, extract the second semantic vector corresponding to the target opposition credit argument data, calculate the cosine distance between the first semantic vector and the second semantic vector, and use the cosine distance as the target opinion divergence degree; Obtain the data source authority weight, logical rigor score, and novelty score of the evidence cited in the data supporting the credibility of the target and the data opposing the credibility of the target. Calculate the weighted sum of the data source authority weight, the logical rigor score, and the novelty score, and use the weighted sum as the overall quality score of the target.

5. The credit risk assessment method based on multi-agent debate-style collaboration according to claim 1, characterized in that, The S4 process specifically includes: Set a round counter and set its initial value to 1; assign the target support credit argument data in the next cycle to the support credit argument data in the current cycle, assign the target opposition credit argument data in the next cycle to the opposition credit argument data in the current cycle, and assign the updated historical debate record to the historical debate record in the current cycle. Repeat step S3. After each execution of step S3, increment the value of the round counter by 1, and obtain the new target opinion divergence degree and the new target comprehensive quality score obtained after executing step S3. At the same time, based on the structured profile vector and the current period's support credit argument data and opposition credit argument data, check whether any item in the hard rule set is triggered. If any item is triggered, interrupt the repeated execution and directly output the rejection decision as the risk level of the borrower. If no item is triggered, generate a soft compliance constraint set and update the preset soft compliance constraint set. If the degree of disagreement of the new target viewpoint obtained at the moment is lower than the preset disagreement threshold, or if the difference between the comprehensive quality score of the new target obtained at the moment and the comprehensive quality score of the target obtained in the previous round is lower than the preset improvement threshold, or if the value of the debate round counter has reached the preset maximum number of rounds, the repeated execution will be terminated. After terminating repeated execution, the target comprehensive quality score obtained from the last execution of S3 is used as the final comprehensive quality score. All supporting credit argument data and all opposing credit argument data stored in the historical debate record, as well as each soft constraint item in the preset soft compliance constraint set, are used. The final comprehensive quality score is multiplied by the first sum of the preset optimistic weight and prudent weight to obtain the first product. The sum of the influence weights of each soft constraint item is multiplied by the preset compliance weight to obtain the second product. The first product and the second product are added together to obtain the final risk score. The final risk score is mapped to the corresponding risk level according to the preset risk rating mapping table; a five-level tree-structured reasoning chain containing risk rating, adjudication basis, debate rounds, argumentation items, evidence items and data source fields is generated according to the risk level, and the five-level tree-structured reasoning chain is output as the credit risk management plan for the borrower.

6. A credit risk assessment system based on multi-agent debate-style collaboration, characterized in that, include: The construction module is used to: collect multi-source heterogeneous data of borrowers through a unified API gateway, standardize the multi-source heterogeneous data into a structured profile vector of the borrower, and determine the initial support credit argument data and the initial opposition credit argument data based on the structured profile vector; The determination module is used to: analyze the initial opposing credit argument data and the initial supporting credit argument data, and combine the structured profile vector and the external knowledge base to determine the supporting credit argument data and opposing credit argument data in the current period, and update the initial opposing credit argument data and the initial supporting credit argument data to the historical debate record; The calculation module is used to: determine the target supporting credence argument data and the target opposing credence argument data for the next period based on the opposing credence argument data, the supporting credence argument data, the historical debate records, the structured profile vector, and the external knowledge base in the current period; update the target supporting credence argument data and the target opposing credence argument data to the historical debate records; and calculate the target opinion divergence degree and the target comprehensive quality score of the target supporting credence argument data and the target opposing credence argument data. The repeat module is used to: take the target support credit argument data, the target opposition credit argument data, and the updated historical debate records in the next cycle as the opposition credit argument data, the support credit argument data, and the historical debate records in the current cycle of the calculation module, repeat the calculation module, until the target opinion divergence degree or the target comprehensive quality score meets the preset conditions, combine the historical debate records, the preset soft compliance constraint set, and the target comprehensive quality score to calculate the final risk score, determine the risk level of the borrower based on the final risk score, and determine the credit risk management plan for the borrower based on the risk level.

7. A credit risk assessment system based on multi-agent debate-based collaboration according to claim 6, characterized in that, The building module is specifically used for: The multi-source heterogeneous data of the borrower is obtained through a unified API gateway. Missing value imputation, outlier detection and format unification are performed on each data source in the multi-source heterogeneous data. The processed fields of each data source are encoded into a structured dictionary according to the preset profile dimensions. The structured dictionary includes at least: original field values, derived indicator values, data timestamps and data source identifiers. All structured dictionaries corresponding to all data sources are aggregated into a structured profile vector of the borrower; the structured profile vector is input into a pre-trained optimistic assessment agent, which generates initial credit support argument data. The initial credit support argument data includes: a first textual claim, a first evidence list, and a first confidence score. Each item in the first evidence list is associated with at least one specific data source field in the multi-source heterogeneous data. The structured profile vector is input into a pre-trained prudent assessment agent, which generates initial objection argument data. The initial objection argument data includes: a second textual claim, a second list of evidence, and a second confidence score. Each item in the second list of evidence is associated with at least one specific data source field in the multi-source heterogeneous data.

8. A credit risk assessment system based on multi-agent debate-based collaboration according to claim 6, characterized in that, The determining module is specifically used for: The initial support and opposition arguments are stored as the first record in the historical debate record. The second evidence list and second confidence level value in the initial opposition credit argument data are analyzed to identify the first link in the initial opposition credit argument data where the confidence level is lower than a preset threshold or the evidence is missing. Based on the first link where the evidence is missing, the related fields are retrieved from the structured profile vector, and the first related data is retrieved from the positive evidence knowledge base in the external knowledge base. Based on the retrieval results, the support credit argument data for the current period is generated. The first evidence list and first confidence value in the initial support credit argument data are analyzed, and the second link in the initial support credit argument data with a confidence value lower than a preset threshold or missing evidence is identified. Based on the second link with missing evidence, the related fields are retrieved from the structured profile vector, and the second related data is retrieved from the negative evidence knowledge base in the external knowledge base. Based on the retrieval results, the opposition credit argument data for the current period is generated. The supporting credit argument data in the current period includes: a third text claim, a third evidence list, and a third confidence level value. The opposing credit argument data in the current period includes: a fourth text claim, a fourth evidence list, and a fourth confidence level value. Each item in the third evidence list and the fourth evidence list is associated with at least one specific data source field in the multi-source heterogeneous data.

9. An electronic device, characterized in that, The electronic device includes a processor coupled to a memory storing at least one computer program, which is loaded and executed by the processor to enable the electronic device to perform the method as described in any one of claims 1 to 5.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one computer program, which is loaded and executed by a processor to enable the computer to perform the method as described in any one of claims 1 to 5.

Images