Risk assessment method, device, electronic equipment, medium and program product

Abstract

The application provides a risk assessment method and device, electronic equipment, medium and program product, which can be applied to the field of artificial intelligence technology. The method comprises the following steps: obtaining evaluation data of a to-be-evaluated object; generating corresponding preliminary evaluation conclusions by using a plurality of first agents, wherein the plurality of first agents evaluate based on different risk dimensions; determining the association relationship between the preliminary evaluation conclusions; generating at least one adversarial challenge based on the association relationship, wherein the adversarial challenge at least comprises a challenge point and a target preliminary evaluation conclusion identifier associated with the challenge point; for each adversarial challenge, using the target preliminary evaluation conclusion identifier to output a response conclusion by using a target first agent, wherein the response conclusion comprises at least one supplementary information, and the supplementary information is used to form a response to the challenge point; and generating a risk evaluation result based on the preliminary evaluation conclusion, the adversarial challenge and the response conclusion.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and more specifically to a risk assessment method, apparatus, device, medium, and program product. Background Technology

[0002] Traditional credit approval and risk management require comprehensive verification of identity information, operational and financial materials. Due to the multiple data sources, inconsistent standards, frequent updates, and the presence of omissions and noise, manual review needs to complete cross-data source comparisons and selections within a short period of time. This process is time-consuming, has limited throughput, and is prone to subjective bias due to different reviewers' understanding of the weight of evidence, affecting the consistency of approval and the stability of risk control.

[0003] To improve efficiency and standardization, some solutions have introduced artificial intelligence to assist decision-making, but the following problems still exist: First, the model directly outputs approval, rejection, or risk scores, making it difficult to clearly explain the evidence and reasoning path corresponding to the conclusion, and thus difficult to support review and auditing; Second, multiple models often run independently and are simply aggregated in parallel, lacking deep collaboration, which can easily cover up conflicting conclusions and make it difficult to organize supplementary verification for evidence gaps; Third, the collaboration mechanism often pursues stable convergence, which can easily lead to consensus bias, thereby weakening the identification of low-probability, high-loss risk signals and reducing decision robustness. Summary of the Invention

[0004] In view of the above problems, embodiments of this application provide a risk assessment method, apparatus, device, medium, and program product.

[0005] According to a first aspect of this application, a risk assessment method is provided, the method comprising: acquiring assessment data of an object to be assessed; generating corresponding preliminary assessment conclusions based on the assessment data using multiple first agents, wherein the multiple first agents assess based on different risk dimensions; determining the correlation between the preliminary assessment conclusions; generating at least one adversarial query based on the correlation, the adversarial query including at least a question point and a target preliminary assessment conclusion identifier associated with the question point; for each adversarial query, outputting a response conclusion using a target first agent corresponding to the target preliminary assessment conclusion identifier, wherein the response conclusion includes at least one piece of supplementary information, the supplementary information being used to form a response to the question point; and generating a risk assessment result based on the preliminary assessment conclusions, the adversarial queries, and the response conclusions.

[0006] According to an embodiment of this application, determining the correlation between the preliminary assessment conclusions includes: using a second intelligent agent to perform structured parsing of the preliminary assessment conclusions, extracting risk elements, conclusion confidence levels, and evidence information identifiers corresponding to each risk dimension; and using the second intelligent agent to calculate the correlation between the preliminary assessment conclusions based on the risk elements, the conclusion confidence levels, and the evidence information identifiers, wherein the correlation includes at least one of conflict relationships, evidence gap relationships, and uncertainty relationships.

[0007] According to an embodiment of this application, generating at least one set of adversarial questions based on the association includes: using a second agent to identify at least one question trigger from the association, the question trigger including at least one of the conflict element corresponding to the conflict relationship, the gap evidence item corresponding to the gap evidence relationship, and the low confidence element corresponding to the uncertainty relationship; and for each question trigger, using the second agent to generate the adversarial question and determine the question priority of the adversarial question, the question priority being determined at least based on the risk impact degree corresponding to the question trigger and the conclusion confidence degree.

[0008] According to an embodiment of this application, the method further includes: activating a second agent to generate the adversarial query based on a preset review script or session state using a master agent; and routing the adversarial query to a target first agent that generates the target preliminary evaluation conclusion using the master agent based on the target preliminary evaluation conclusion identifier, thereby triggering the target first agent to output the response conclusion.

[0009] According to an embodiment of this application, the step of outputting a response conclusion for each adversarial query using the target's preliminary evaluation conclusion identifier includes: based on the query point, using the target's first intelligent agent to perform a supplementary information acquisition operation to generate a candidate response, and determining a supplementary information identifier associated with the candidate response; using the master intelligent agent to perform differential verification between the supplementary information identifier and the evidence information identifier cited in the target's preliminary evaluation conclusion; and if the differential verification passes, outputting the candidate response as the response conclusion.

[0010] According to an embodiment of this application, if differential verification fails, the master control agent sends a supplementary evidence collection instruction to the target first agent. The supplementary evidence collection instruction is used to trigger the target first agent to re-execute the supplementary information acquisition operation to obtain the updated candidate response generated by the target first agent. Based on the updated candidate response, a new supplementary information identifier is determined, and differential verification is re-performed based on the new supplementary information identifier until differential verification passes.

[0011] According to an embodiment of this application, the method further includes: in response to obtaining the response conclusion, using a second agent to determine a response sufficiency determination result based on the coverage of the question points and the conflict resolution degree of the association relationship by the response conclusion; and if the response sufficiency determination result indicates that the response is insufficient, using the second agent to generate a follow-up question, and routing the follow-up question to the corresponding target first agent via the master agent to trigger the generation of an additional response conclusion.

[0012] According to an embodiment of this application, the method further includes: generating session transcription information from the preliminary evaluation conclusion, the adversarial inquiry, the follow-up inquiry, and the additional response conclusion; arranging the session transcription information in a temporal sequence based on the routing order of the master agent, and associating each session transcription information with a participating role identifier; and presenting the session transcription information and the participating role identifier in a streaming output mode or a one-time display mode.

[0013] A second aspect of this application provides a risk assessment apparatus, comprising: an initial assessment module, configured to: acquire assessment data of an object to be assessed, and generate corresponding preliminary assessment conclusions based on the assessment data using multiple first agents, wherein the multiple first agents assess based on different risk dimensions; an adversarial inquiry module, configured to: determine the correlation between the preliminary assessment conclusions, and generate at least one adversarial inquiry based on the correlation, wherein the adversarial inquiry includes at least an inquiry point and a target preliminary assessment conclusion identifier associated with the inquiry point; a supplementary module, configured to: for each adversarial inquiry, output a response conclusion using a target first agent corresponding to the target preliminary assessment conclusion identifier, wherein the response conclusion includes at least one piece of supplementary information, the supplementary information being used to form a response to the inquiry point; and a final assessment module, configured to: generate a risk assessment result based on the preliminary assessment conclusions, the adversarial inquiry, and the response conclusion.

[0014] A third aspect of this application provides an electronic device comprising: one or more processors; and a memory for storing one or more computer programs, wherein the one or more processors execute the one or more computer programs to implement the steps of the method described above.

[0015] A fourth aspect of this application also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0016] The fifth aspect of this application also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described method.

[0017] According to embodiments of this application, by introducing multiple first agents to generate preliminary assessment conclusions according to different risk dimensions, and generating adversarial queries based on the correlation between the preliminary conclusions, the queries are further directed to the target first agent corresponding to the target preliminary assessment conclusion, forcing it to introduce supplementary information in its response conclusion to form a response to the query points. This achieves cross-validation and counter-evidence checks of risk conclusions within a closed loop of initial reporting—querying—response—summarizing. Compared to traditional single-model or multi-model simple aggregation methods, embodiments of this application can transform the risk assessment process from a black-box output into a traceable chain of reasoning and evidence, significantly enhancing the interpretability and auditability of the conclusions. Simultaneously, the adversarial query mechanism can proactively expose conflicts, evidence gaps, and low-confidence elements in the preliminary conclusions, suppressing consensus bias caused by multi-agent consensus convergence, reducing the possibility of ignoring low-probability, high-loss risk signals, improving the robustness and credibility of the risk assessment results, and providing clearer focus points for the formulation of subsequent risk control measures. Attached Figure Description

[0018] The above-mentioned contents, other objects, features and advantages of this application will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0019] Figure 1 A schematic diagram illustrating an example environment in which the methods according to embodiments of this application can be applied;

[0020] Figure 2 A flowchart illustrating a risk assessment method according to an embodiment of this application is shown schematically.

[0021] Figure 3 The flowchart illustrates a method for generating at least one set of adversarial queries based on association relationships, according to some exemplary embodiments of this application.

[0022] Figure 4 The flowchart illustrates a method for arranging and controlling a review process according to some exemplary embodiments of this application.

[0023] Figure 5 A schematic diagram illustrating the structure of a risk assessment apparatus according to an embodiment of this application is shown; and

[0024] Figure 6 A block diagram schematically illustrates an electronic device suitable for implementing a risk assessment method according to an embodiment of this application. Detailed Implementation

[0025] The embodiments of this application will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of this application. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of this application for ease of explanation. However, it will be apparent that one or more embodiments may be implemented without these specific details. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of this application.

[0026] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0027] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0028] First, the technical terms used in this article will be explained and clarified.

[0029] Intelligent agents are computational entities capable of performing analysis and reasoning based on input data and outputting results. They can be implemented using rules, models, or a combination of both. Intelligent agents can have independent input parsing, feature construction, reasoning, and output structures, and can be triggered to execute specific tasks within a process.

[0030] First agent: refers to a set of agents that assess a specific risk dimension and output preliminary assessment conclusions. Multiple first agents are responsible for different risk dimensions, forming parallel or quasi-parallel preliminary judgments to cover multiple types of risk points.

[0031] Second agent: refers to an agent used to perform structured analysis and determine correlations of preliminary multi-dimensional assessment conclusions, and can further generate adversarial queries. In some embodiments, the second agent can also play the role of an adversarial agent, that is, generate queries with the goal of revealing conflicts, gaps, and uncertainties.

[0032] Master control agent: refers to the control entity used to orchestrate the review process, maintain the session state, activate the second agent, and route queries to the target first agent. The master control agent can advance the process according to the preset review script or session state, ensuring that the timing of queries and responses is consistent, the target is clear, and traceability is guaranteed.

[0033] Adversarial questioning refers to questioning generated by a second agent based on correlations, used to challenge weaknesses in the initial assessment conclusions of the target. Adversarial questioning must include at least the questioning point and the identifier of the initial assessment conclusions of the target, used to clarify the content of the questioning and the responsible party. Adversarial questioning emphasizes problem-driven cross-verification, rather than simple summarization.

[0034] Points of contention: These refer to the specific issues requiring explanation, supplementary evidence, or clarification in adversarial questioning. They may target conflicting elements, gaps in evidence, or low-confidence elements. Points of contention may include information such as the type of supplementary information required, the scope of verification, or the time window to improve responsiveness.

[0035] Traditional credit approval and risk management typically require comprehensive verification and judgment of the applicant's identity information, business and financial materials, credit history and transaction behavior, collateral information, and compliance requirements. Due to the diverse sources of materials, inconsistent field definitions, frequent information updates, and the presence of omissions and noise, manual review often needs to complete the comparison, interpretation, and selection across data sources within a short period of time. This results in lengthy processes, limited throughput, and different reviewers assigning different weights to the same evidence, which can easily lead to subjective biases and affect the consistency of approval and the stability of risk control.

[0036] To improve efficiency and standardization, some solutions have gradually introduced AI-assisted decision-making systems. However, existing technologies still reveal a series of key issues affecting credibility and risk coverage during engineering implementation.

[0037] First, insufficient interpretability and lack of audit traceability remain prominent issues. Existing systems often employ complex machine learning or deep learning models to jointly model high-dimensional features and directly output approval, rejection, or risk scores. Such outputs typically rely primarily on statistical correlation, making it difficult to clearly present the correspondence between risk conclusions and specific evidence. They also struggle to explain which key factors trigger the risk, how these factors interact, and how the model makes choices when evidence is missing or contradictory. Due to the lack of verifiable reasoning paths and chains of evidence, reviewers find it difficult to effectively verify boundary samples, anomalous samples, or controversial samples. Furthermore, in scenarios such as regulatory audits, internal audits, and post-loan accountability, it is difficult to establish verifiable records of decision-making basis, making the system output untrustworthy and unacceptable.

[0038] Secondly, the lack of coordination caused by multi-model parallelism is a common problem. To compensate for the limitations of a single model, some systems introduce multiple models, such as credit risk and compliance review, to run in parallel. However, these models often process their own data domains and feature sets independently, ultimately forming conclusions through score weighting, threshold voting, or report listing. This parallel analysis has two typical drawbacks: First, it is difficult to detect implicit conflicts and dependencies between cross-model conclusions in a timely manner. For example, if the credit risk model gives a low risk but the anti-anomaly model finds a strong correlation anomaly, the system may mask the contradiction if it only aggregates scores, failing to trigger further verification. Second, there is a lack of a collaborative mechanism to fill evidence gaps. When key materials on which a conclusion depends for a certain dimension are missing, the field definitions are inconsistent, or the data credibility is insufficient, the system can usually only reduce the score or output an uncertainty prompt, unable to further organize cross-dimensional supplementary evidence collection and cross-validation, thus the final conclusion remains at the level of result splicing.

[0039] Third, consensus bias is prone to occur in collaborative decision-making, leading to insufficient coverage of low-probability, high-loss risks. Existing collaborative solutions often aim for stable convergence, reducing output volatility through consistency constraints, majority agreement, or averaging strategies. However, in credit risk scenarios, significant risk signals are often characterized by low frequency, weak correlation, and cross-dimensional dispersion. These may manifest as minor contradictions between materials, local anomalies in transaction behavior, sparse group characteristics of correlations, or boundary triggering of compliance conditions. If the collaborative mechanism lacks proactive questioning and counter-evidence, the system tends to reach optimistic conclusions when most dimensions show normality, causing key anomalies to be diluted or ignored. At the same time, when multiple dimensions give similar conclusions, the system lacks the ability to stress test weak links in the conclusions and cannot reverse-verify risk judgments, thereby reducing decision robustness.

[0040] Furthermore, existing systems generally face engineering challenges arising from the complexity of input evidence: duplication, omissions, delays, and discrepancies exist between different data sources, and the authenticity and consistency of materials require repeated verification; there are conditional dependencies and information transmission between risk dimensions, for example, operational anomalies may be linked to cash flow anomalies, guarantee relationships to related-party transactions, and personal liabilities to corporate cash flow, all forming a risk chain. Without an organizational mechanism oriented towards these relationships, the system struggles to aggregate scattered evidence into a verifiable logical chain, and it also struggles to determine which risk dimension should bear the responsibility for explanation, what supplementary information is needed, and where that supplementary information should come from when contradictions are discovered. Consequently, even with the introduction of multiple models or agents, the system may still fail to form a systematic process for identifying, questioning, and clarifying key risk points, making it difficult to meet the requirements of transparency, verifiability, and robustness in high-compliance and high-prudence scenarios.

[0041] Based on this, embodiments of this application provide a risk assessment method, the method comprising: acquiring assessment data of the object to be assessed; generating corresponding preliminary assessment conclusions using multiple first agents, wherein the multiple first agents assess based on different risk dimensions; determining the correlation between the preliminary assessment conclusions; generating at least one adversarial query based on the correlation, wherein the adversarial query includes at least a question point and a target preliminary assessment conclusion identifier associated with the question point; for each adversarial query, outputting a response conclusion using the target first agent corresponding to the target preliminary assessment conclusion identifier, wherein the response conclusion includes at least one piece of supplementary information, the supplementary information being used to form a response to the question point; and generating a risk assessment result based on the preliminary assessment conclusion, the adversarial query, and the response conclusion. According to embodiments of this application, by introducing multiple first agents to generate preliminary assessment conclusions according to different risk dimensions, and generating adversarial queries based on the correlation between the preliminary conclusions, the queries are further directed to the target first agent corresponding to the target preliminary assessment conclusion, forcing it to introduce supplementary information in its response conclusion to form a response to the question point, thereby achieving cross-validation and counter-verification of risk conclusions in a closed loop of initial reporting—questioning—response—summarizing. Compared to traditional single-model or multi-model simple aggregation methods, the embodiments of this application can transform the risk assessment process from a black-box output into a traceable chain of reasoning and evidence, significantly enhancing the interpretability and auditability of the conclusions. At the same time, the adversarial questioning mechanism can proactively expose conflicts, evidence gaps, and low-confidence elements in the preliminary conclusions, suppress consensus bias caused by multi-agent consensus convergence, reduce the possibility of low-probability, high-loss risk signals being ignored, improve the robustness and credibility of the risk assessment results, and provide clearer focus points for the formulation of subsequent risk control measures.

[0042] It should be noted that the risk assessment methods, apparatus, devices, media, and program products defined in this application can be used in the fields of artificial intelligence technology and fintech, and can also be used in a variety of other fields besides artificial intelligence technology and fintech. The application fields of the risk assessment methods, apparatus, devices, media, and program products provided in the embodiments of this application are not limited.

[0043] In the technical solution of this application, the user information (including but not limited to user personal information, user image information, user device information, such as location information) and data (including but not limited to data used for analysis, stored data, and displayed data) involved are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of related data all comply with relevant laws, regulations, and standards, take necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation entry points for users to choose to authorize or refuse.

[0044] In scenarios where personal information is used for automated decision-making, the methods, devices, and systems provided in this application all provide users with corresponding operation entry points for users to choose to agree to or reject the automated decision results; if the user chooses to reject, the process enters the expert decision-making process.

[0045] Figure 1 A schematic diagram of an example environment in which the method according to an embodiment of this application can be applied is shown. In this example environment 100, an application 125 is installed on a terminal device 110. A user 140 can interact with the application 125 via the terminal device 110 and / or an attached device of the terminal device 110.

[0046] In some embodiments, application 125 can be downloaded and installed on terminal device 110. In some embodiments, application 125 can also be accessed in other ways, such as through a web page. Figure 1 In environment 100, in response to application 125 being launched, terminal device 110 can display the interface 150 of application 125.

[0047] In some embodiments, terminal device 110 can communicate with server 130 to provide services to application 125. Terminal device 110 can be any type of mobile terminal, fixed terminal, or portable terminal, including mobile phones, desktop computers, laptop computers, notebook computers, netbook computers, tablet computers, media computers, multimedia tablets, personal communication system (PCS) devices, personal navigation devices, personal digital assistants (PDAs), audio / video players, digital cameras / camcorders, television receivers, radio receivers, e-book devices, gaming devices, or any combination thereof, including accessories and peripherals of these devices or any combination thereof. In some embodiments, terminal device 110 can also support any type of user-facing interface. Application 125 can be various types of computing systems / servers capable of providing computing power, including but not limited to mainframes, edge computing nodes, and computing devices in cloud environments.

[0048] In some embodiments, application 125 can provide interaction capabilities with an intelligent agent. Application 125 may include an application specifically designed to provide services to an intelligent agent, or an application integrated with an intelligent agent. Although Figure 1 The image shows a single application, but in reality, multiple applications can be installed on the terminal device 110.

[0049] In embodiments of this application, multiple intelligent agents 160 can be deployed locally on terminal device 110 or remotely. In the case of remote deployment, terminal device 110 can directly invoke the intelligent agents, or it can invoke the intelligent agents via server 130. Exemplarily, intelligent agents 160 may have intelligent dialogue and task processing capabilities. Terminal device 110 provides an interface 150 that can present interactions with intelligent agents 160. In interface 150, user 140 can initiate task requests to intelligent agents 160 by inputting natural language (e.g., text input or voice input). Optionally, user 140 can upload online or offline file dialogues to instruct intelligent agents 160 to assist in completing various tasks.

[0050] In the embodiments of this application, during interaction with user 140, intelligent agent 160 can respond to user 140's requests and handle tasks instructed by the user. In some embodiments, during task processing, intelligent agent 160 can invoke one or more tools 165 to assist in task execution and the provision of task results as needed. These tools 165 can be any type of tool, such as text generation tools, file reading tools, information search tools, online or offline databases, image processing tools, chart generation tools, web page creation tools, etc.

[0051] In some embodiments, environment 100 may further include a management node for multiple agents 160, which can interact with the agents 160. In some examples, the management node may, in response to a task request from user 140, determine the task requirements corresponding to the task request. The management node may then, based on the task requirements, assign the task request to the agent 160 that matches the task requirements, requesting that agent 160 to perform the task. In other examples, the management node may also determine an execution plan for the task based on the task requirements. The execution plan may indicate one or more subtasks required to complete the task. The management node may assign these one or more subtasks to one or more agents 160, which will then execute their respective subtasks. Regarding the management node, in some examples, the management node may be implemented by one of the multiple agents 160. In other examples, the management node may be implemented by a machine learning model, such as a language model.

[0052] In some embodiments, agent 160 may be constructed based on one or more machine learning models. In some embodiments, the machine learning model on which agent 160 is based may include at least a language model, such as a large language model. In some embodiments, the machine learning model on which agent 160 is based may include a multimodal model capable of handling multiple modal inputs, such as text input, visual input (e.g., images, videos), audio input, etc. These machine learning models may include content-generating models capable of generating corresponding outputs based on model inputs. In some embodiments, the machine learning model may receive text-modal model inputs (e.g., natural language and / or machine language) and / or non-text-modal model inputs (e.g., images, speech, videos, etc.), and may obtain corresponding model outputs based on model inputs and prompts, thereby completing the task execution.

[0053] It should be understood that the structure and function of the various elements in environment 100 are described for illustrative purposes only and are not intended to limit the scope of this application in any way.

[0054] The following will be based on Figure 1 The described scene, through Figures 2-4 The risk assessment method of the disclosed embodiments is described in detail.

[0055] Figure 2 A flowchart illustrating a risk assessment method according to an embodiment of this application is shown.

[0056] like Figure 2 As shown, the risk assessment method 200 of this embodiment includes steps S210 to S240.

[0057] In step S210, the evaluation data of the object to be evaluated is obtained, and multiple first intelligent agents are used to generate corresponding preliminary evaluation conclusions based on the evaluation data. Among them, the multiple first intelligent agents evaluate based on different risk dimensions.

[0058] In the embodiments of this application, the assessment data can come from a multi-source information set generated by the entity being assessed during the credit application, loan management, or post-loan review process. For example, it can include basic materials submitted by the applicant, as well as historical transaction and behavioral records accumulated within financial institutions. The assessment data can also further incorporate external information from publicly available channels, such as enterprise registration information, judicial documents, administrative penalty information, and industry prosperity information, to cover the assessment needs of different risk dimensions. To adapt to the format differences of different data sources, the assessment data can undergo field mapping, format standardization, missing data completion prompts, time alignment, and noise filtering before entering the intelligent agent for processing, providing a usable input basis for subsequent analysis of various risk dimensions.

[0059] It should be noted that in the technical solution of this application, the user information (including but not limited to user personal information, user image information, user device information, such as location information) and data (including but not limited to data used for analysis, stored data, and displayed data) involved are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of the relevant data all comply with relevant laws, regulations, and standards, take necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation entry points for users to choose to authorize or refuse.

[0060] In scenarios where personal information is used for automated decision-making, the methods, devices, and systems provided in this application all provide users with corresponding operation entry points for users to choose to agree to or reject the automated decision results; if the user chooses to reject, the process enters the expert decision-making process.

[0061] In some embodiments, multiple first agents can conduct assessments targeting different risk dimensions. Each first agent receives the same assessment data or a subset thereof and generates preliminary assessment conclusions from its own risk perspective. For example, a first agent for credit risk can build judgments based on repayment ability and willingness, such as combining income and expenditure structure, asset and liability structure, cash flow stability, historical delinquency behavior, credit usage intensity, and debt ratio trends, to form preliminary conclusions on risk level, key influencing factors, and anomaly indicators requiring attention. First agents for suspicious transactions and anomalies can focus on identity consistency, behavioral consistency, and abnormal relationships, such as comparing logical consistency between materials, verifying reasonable ranges for key fields, detecting abnormal behavior patterns in the application process, identifying group association characteristics and suspicious transaction chains, to form preliminary conclusions on anomaly type, anomaly intensity, and material points requiring further verification. The first intelligent agent in the compliance dimension can focus on product access, usage restrictions, industry policies and prudential rules. For example, it can determine whether the application purpose complies with credit regulations, whether the collateral meets the access requirements, whether the guarantee and related transactions trigger restrictive conditions, and whether the necessary materials required for credit approval are met, forming a preliminary conclusion on compliance risk points and supplementary compliance supporting materials.

[0062] In this embodiment, the first agent in the credit risk dimension can adopt a processing structure consisting of a data access and normalization component, a feature construction component, a risk scoring and interpretation component, an evidence citation component, and a conclusion generation component. The data access and normalization component reads transaction records, repayment records, credit usage, credit entries, and asset / liability and income / expenditure information from the assessment data, and performs field mapping, time alignment, outlier handling, and missing value alerts. The feature construction component generates an interpretable set of risk features, which may include cash flow coverage ratio, debt ratio and its rate of change, income stability, expenditure concentration, number of overdue payments and distance to the most recent overdue payment, credit usage intensity, and monthly net inflow fluctuations. The risk scoring and interpretation component inputs the risk features into the risk prediction model to obtain the risk level or risk score, and simultaneously outputs the contribution of influencing factors or an interpretable summary of reasons. The evidence citation component binds evidence information identifiers to key risk features, such as binding transaction statistics windows, credit entry numbers, or repayment record intervals. The conclusion generation component generates preliminary assessment conclusion text or structured conclusions based on the risk score, influencing factors, and evidence citations. During training, a dataset based on historical credit samples can be constructed. The labels can come from events such as overdue payments, defaults, write-offs, extensions, credit limit reductions, and early settlements. Supervision labels are generated according to a preset observation window. The model can use gradient boosting trees, logistic regression, or neural networks. The training objectives can include binary or multi-class risk prediction loss, segmented calibration loss, and monotonicity constraint regularization terms, so that key features maintain a reasonable direction in response to risk changes.

[0063] In this embodiment, the first intelligent agent for suspicious transactions and anomalies can adopt a processing structure consisting of a material consistency analysis component, an anomaly pattern detection component, a correlation analysis component, an evidence aggregation and ranking component, and a conclusion generation component. The material consistency analysis component performs consistency comparisons on identity materials, contact information, etc., and outputs consistency scores and a list of conflicting elements. The anomaly pattern detection component extracts anomaly features from transaction behavior, application behavior, and account behavior, such as high-frequency applications in a short period, abnormal combinations of key fields, and mutation points inconsistent with historical behavior, and outputs anomaly intensity in conjunction with an anomaly detection model. The correlation analysis component abstracts entities, accounts, device identifiers, and counterparties as nodes, and transactions, shared information, guarantees, or related information as edges, constructing a relationship graph and calculating cluster features and critical path features. The evidence aggregation and ranking component deduplicates, merges, and ranks the aforementioned multi-source anomaly clues, with ranking criteria including anomaly intensity, evidence coverage, and similarity to historical risk events. The conclusion generation component outputs the anomaly type, anomaly intensity, and material points requiring further verification, and binds evidence information identifiers to each anomaly clue. During training, a training set containing confirmed abnormal events and normal samples can be constructed. The model can adopt unsupervised / semi-supervised models such as binary classification models, isolated forests, or autoencoders, as well as graph representation learning models for relation graphs. The training objectives can include classification loss, ranking loss, and contrastive learning loss to improve the ability to identify a small number of high-risk samples. To reduce false alarms, stratified thresholds and cost-sensitive assessments can be introduced in the validation stage to prioritize the recall of high-impact abnormal clues and keep false alarms within an acceptable range.

[0064] In this embodiment, the first intelligent agent for the compliance dimension can adopt a processing structure consisting of a rule and policy loading component, a requirement extraction and matching component, a boundary condition determination component, an evidence chain organization component, and a conclusion generation component. The rule and policy loading component loads product access rules, usage restrictions, material list requirements, etc., and expresses the rules as an executable set of conditions. The requirement extraction and matching component extracts requirement fields corresponding to the rule conditions from the assessment data, such as usage descriptions, industry categories, related transaction elements, etc., and matches them one by one with the condition set. The boundary condition determination component handles uncertain, missing, or boundary situations, such as unclear usage descriptions, ambiguous industry classifications, missing materials, or conflicting information, and outputs a compliance risk level and a list of requirements to be supplemented. The evidence chain organization component binds evidence information identifiers to each triggered rule condition and forms a traceable matching record. The conclusion generation component outputs compliance risk points, a list of missing materials, and supplementary suggestions. The first intelligent agent for compliance can be based on rule-based reasoning or can also incorporate a text understanding model for classifying and extracting unstructured fields such as usage descriptions and industry descriptions. Training data can come from usage classification, industry classification, and compliance review records in historical approvals, and labels can be classification results or rule matching results. The model training objective can include classification loss and extraction loss, and hard examples can be enhanced through manual review samples.

[0065] In the embodiments of this application, in order to ensure that the first intelligent agent can stably produce traceable preliminary evaluation conclusions, each first intelligent agent can introduce a set of general tools and resources, and record evidence information identifiers when calling them to support subsequent inquiries and responses.

[0066] In some embodiments, the division of multiple first agents is not limited to credit risk and compliance, but can be extended to more dimensions based on the institution's business characteristics. For corporate credit scenarios, a first agent focusing on operations and supply chain dimensions can be added. This agent focuses on order stability, customer concentration, supplier concentration, payment cycle, inventory turnover, upstream and downstream dependence, and contract performance, outputting preliminary conclusions on operational volatility risk and supply chain disruption risk. For collateralized lending scenarios, a first agent focusing on collateral valuation can be added. This agent focuses on the integrity of collateral ownership, valuation rationality, liquidity, disposal cycle, double-collateralization risk, and regional liquidity differences, outputting preliminary conclusions on collateral coverage and disposal risk. For personal consumer loans or business loans, a first agent focusing on behavior and debt dimensions can be added. This agent focuses on consumption structure, borrowing behavior, debt expansion speed, income volatility and stability, and consistency of fund usage, outputting preliminary conclusions on excessive debt and deviation from intended use risk. Through these extensions, assessment data can be interpreted from a more granular risk dimension, thereby improving the adaptability of preliminary assessment conclusions to complex scenarios.

[0067] In some embodiments, the preliminary assessment conclusion may include necessary information for subsequent processing, enabling later steps to identify relationships between conclusions and trigger further processing. The preliminary assessment conclusion may include a risk level or risk score, a list of key risk elements, supporting evidence information, a brief description of key evidence, and uncertainty warnings. For example, the credit risk dimension may indicate insufficient cash flow coverage and increased liabilities, with the basis indicated by cash flow statistics and liability ledgers for a specific period; the anomaly dimension may indicate inconsistencies between materials and abnormal correlations, with the basis indicated by identity verification and correlation analysis results; the compliance dimension may indicate insufficient proof of purpose and pending verification of access conditions, with the basis indicated by rule condition comparison and material checklist inspection results. By including key elements and supporting evidence information in the preliminary conclusion, it facilitates subsequent determination of relationships between different conclusions and allows for rapid identification of data sources when supplementary materials or verification are required.

[0068] In the embodiments of this application, different first intelligent agents can generate preliminary evaluation conclusions in parallel to meet the needs of real-time approval or batch review. In the real-time approval scenario, multiple first intelligent agents can simultaneously start analysis after receiving evaluation data, output their respective conclusions within a preset time window, and then be uniformly summarized and processed by subsequent processes. In the batch review scenario, evaluation objects can be queued according to business priority. Different first intelligent agents can batch process and generate conclusions for different objects and write the conclusions into a unified evaluation record for subsequent steps to read. For scenarios with large amounts of data, the evaluation data can also be segmented, such as segmenting transaction records by time window and segmenting contracts and invoices by business document, so that the first intelligent agents first generate partial conclusions for each segment and then summarize the partial conclusions to form a preliminary evaluation conclusion.

[0069] In step S220, the correlation between the preliminary assessment conclusions is determined, and at least one adversarial question is generated based on the correlation. The adversarial question includes at least a question point and a target preliminary assessment conclusion identifier associated with the question point.

[0070] In the embodiments of this application, the correlation between preliminary assessment conclusions is used to characterize whether the outputs of different risk dimensions are consistent, contradictory, lacking in evidence, or excessively uncertain, thereby providing a basis for generating adversarial inquiries. The determination of the correlation can be based on a comprehensive judgment of the risk elements, risk levels or scores, evidence information identifiers, and uncertainty warnings carried in the preliminary assessment conclusions. For example, if the credit dimension gives a low-risk conclusion and emphasizes stable cash flow, while the anomaly dimension simultaneously indicates abnormal transaction behavior and a strong correlation, then there is a significant tension between the two in the direction of risk judgment, which can be identified as a conflicting relationship. Similarly, if the compliance dimension indicates insufficient documentation for the intended use, while the credit dimension gives an acceptable conclusion but does not cover the verification of the authenticity of the intended use, this can be identified as an evidence gap relationship. Furthermore, if the collateral valuation dimension indicates large fluctuations in the valuation range and indicates insufficient confidence, this can be identified as an uncertain relationship.

[0071] In some embodiments, the determination of relationships can employ a multi-layered approach combining rules and metrics. The first layer can perform rapid matching based on conclusion types, such as mapping conclusions to risk level ranges and comparing whether the differences in risk levels across different dimensions exceed a threshold; if the threshold is exceeded, a conflict relationship is identified. The second layer can align risk elements, such as unifying elements extracted from different dimensions into a unified element dictionary and then determining whether the same element presents opposite directions in different conclusions; for example, if one conclusion considers cash flow sufficient while another considers it tight, then the element is in conflict. The third layer can assess evidence coverage; for example, if the intersection between the evidence set upon which a conclusion depends and the key evidence set of another conclusion is too small or lacks necessary evidence types, then an evidence gap is identified. The fourth layer can identify uncertainty indicators or confidence levels; for example, if a conclusion is marked as having high uncertainty or a confidence level below a threshold, then an uncertainty relationship is formed with other conclusions. Through multi-layered assessment, both computational efficiency and the ability to capture more granular contradictions and gaps can be achieved.

[0072] In the embodiments of this application, the generation of adversarial inquiries can be centered on relationships. Each adversarial inquiry includes at least an inquiry point and a target preliminary assessment conclusion identifier. The target preliminary assessment conclusion identifier points to the source and responsibility dimension of the conclusion being questioned, while the inquiry point clarifies the risk factors, evidentiary gaps, or contradictions that need clarification. In conflict scenarios, the inquiry point can point to the source of differences in conflicting elements, such as requiring an explanation of why a low-risk assessment is still made despite the existence of an abnormal transaction chain, or requiring an explanation of whether the conclusion over-relies on historical financial performance while ignoring recent fluctuations. In evidence gap scenarios, the inquiry point can point to the type of missing evidence, such as requiring supplementary verification of the authenticity of the purpose, supplementary verification results of the ownership certificate of the pledged collateral, or supplementary performance evidence of key contracts. In uncertainty scenarios, the inquiry point can point to the source and impact of uncertainty, such as requiring an explanation of the reasons for valuation fluctuations or requiring a conservative handling strategy in the presence of uncertainty. By binding the inquiry point to the target conclusion, general questions can be avoided, making the inquiry specific and responsive.

[0073] In the embodiments of this application, the relationship and adversarial questioning can also cover the combination of multiple conclusions, and are not limited to pairwise relationships. For complex credit objects, different dimensions may form a chain of contradictions. For example, the operational dimension shows increased order fluctuations, the credit dimension shows rising liabilities, the anomaly dimension shows increased related-party transactions, and the compliance dimension shows insufficient documentation on the purpose of use. In this case, the relationship can be identified as a multi-dimensional coupling relationship, and adversarial questions can be generated around the common risk chain. For example, it may require an explanation of whether the operational fluctuations have led to a tight cash flow and triggered abnormal cash flow, or require supplementary verification of the purpose of use to ensure consistency with the related-party transaction chain. By combining multiple conclusions in the questioning, the questions can be organized from the perspective of the risk chain, avoiding cyclical verification within a single dimension, thus better meeting the comprehensive judgment needs of the actual review process.

[0074] In step S230, for each adversarial query, the target first agent is identified by the target preliminary evaluation conclusion and outputs a response conclusion, wherein the response conclusion includes at least one piece of supplementary information, which is used to form a response to the question.

[0075] In the embodiments of this application, after receiving an adversarial inquiry, the target first agent can review its existing judgments around the inquiry points, prioritize the identification of risk factors and sources of evidence related to the inquiry points, and generate a response conclusion based on this. The response conclusion not only provides a direct response to the inquiry points but can also introduce supplementary information to form a more complete chain of explanation. Supplementary information can be reflected in newly added material verification results, re-statistical analysis results of existing data, re-estimation results of time windows for key elements, or further clarification results of correlations. In this way, the response conclusion can make the basis behind the preliminary assessment conclusion explicit, enabling examiners to understand why the conclusion was reached, whether the conclusion still holds under questioning pressure, and whether the conclusion needs to be revised after introducing new information.

[0076] In the embodiments of this application, supplementary information can be obtained from the reuse of the evaluation data. The target first intelligent agent can reselect the data range and calculation caliber according to the question points, for example, expanding the statistical window of transaction flow from the past three months to the past year, or refining income and expenditure statistics from total indicators to categorical indicators to identify whether there are structural fluctuations; it can also re-analyze the original materials, for example, extracting payment terms, default clauses and performance node information from the contract text to determine whether the change in the repayment cycle is consistent with the contract terms. For collateralized credit scenarios, the target first intelligent agent can review the valuation basis of the collateral, for example, recalculating the discount rate, disposal cycle assumption or regional liquidity parameters, and write the reviewed valuation range as supplementary information into the response conclusion. Through internal extraction and recalculation, the quality of the response can be improved without relying on externally added materials, which is especially suitable for situations where the data is relatively complete but the explanatory chain is insufficient.

[0077] In the embodiments of this application, supplementary information can also come from cross-dimensional collaboration. When generating a response conclusion, the target first agent can request other first agents to provide supplementary conclusions or supplementary evidence clues related to the question point, thereby forming multi-dimensional evidence. For example, when the question point requires an explanation of the contradiction between the low-risk conclusion of the credit dimension and the high-abnormality conclusion of the anomaly dimension, the target first agent of the credit dimension can request the first agent of the anomaly dimension to provide the specific source and key triggering conditions of the anomaly elements, and combine its own analysis of cash flow and debt structure to determine whether the anomaly elements will change the assessment of repayment ability and willingness; when the question point focuses on the authenticity of the use and compliance boundaries, the target first agent of the compliance dimension can request the first agent of the operations dimension to provide supplementary conclusions on order and contract performance evidence, so as to form a cross-explanation of the rationality of the use. Through cross-dimensional collaboration, the response conclusion can reflect the causal relationship and dependency between different risk dimensions, avoid a single dimension talking to itself, and improve the overall consistency of explanation.

[0078] In some embodiments, supplementary information to the response conclusion may also include further verification results of the consistency of evidence. The target agent can conduct consistency comparisons on key materials involved in the inquiry, such as comparing differences in revenue figures between tax returns and financial statements, comparing the matching degree between contract amounts and recorded cash flow, comparing the correspondence between the list of pledged or mortgaged assets and ownership certificates, and comparing the consistency between the company's external guarantees and debt ledgers. The verification results can be included as supplementary information in the response conclusion to explain whether the basis of the preliminary assessment conclusion is reliable, whether there are any contradictions that require attention, and the degree of impact of the contradictions on the risk assessment.

[0079] In the embodiments of this application, when the target first intelligent agent outputs a response conclusion, it can establish a clear association between supplementary information and the points of inquiry to support subsequent backtracking of the explanation chain. The supplementary information may carry material number, data source identifier, time interval identifier, or verification record identifier, enabling reviewers to quickly locate the source of the supplementary information.

[0080] In step S240, a risk assessment result is generated based on the preliminary assessment conclusions, adversarial questioning, and response conclusions.

[0081] In the embodiments of this application, the risk assessment result is generated based on a comprehensive analysis of the preliminary assessment conclusion, the adversarial inquiry, and the response conclusion. The risk assessment result not only reflects the final risk judgment but also embodies the clarified contradictions, filled evidentiary gaps, and uncertainties that still require continued attention during the formation of the risk judgment. For elements targeted by the adversarial inquiry and confirmed as significant risks by the response conclusion, the risk assessment result can increase their impact weight or upgrade the risk level. For elements whose impact is proven to be limited or within a reasonable range by the response conclusion, the risk assessment result can maintain the original conclusion and indicate its explanatory basis. For elements that remain uncertain after the response conclusion, the risk assessment result can adopt a more conservative strategy, such as increasing the prudence coefficient, setting additional verification requirements, or outputting more stringent follow-up monitoring recommendations.

[0082] In some embodiments, risk assessment results can be generated using a hierarchical aggregation approach. The first layer generates conclusions within each dimension, creating risk summaries for credit, anomaly, compliance, operational, and collateral dimensions. These summaries may include key risk factors, sources of crucial evidence, confidence levels of the conclusions, and necessary supplementary verification materials. The second layer generates consistent cross-dimensional conclusions, using the conflicting relationships and evidentiary gaps revealed by adversarial inquiries as core inputs for cross-dimensional fusion. Conflicts are resolved or gaps are filled by supplementary information provided in the response conclusions, resulting in a comprehensive cross-dimensional risk assessment. The third layer generates actionable conclusions, mapping the comprehensive risk assessment to business actions, such as recommendations to approve, reject, conditionally approve, reduce the amount, or increase collateral, while providing the triggering basis and key concerns for each action. Through hierarchical aggregation, the risk assessment results retain the professional perspectives of each dimension while forming a unified comprehensive judgment, facilitating understanding and execution by reviewers.

[0083] In the embodiments of this application, the risk assessment results may include risk levels, risk scores, or risk labels, with the specific form chosen according to business needs. Risk levels can be divided into preset tiers, such as low risk, medium risk, and high risk; risk scores can be normalized according to preset intervals for easy comparison with historical samples; risk labels can be used to describe risk types, such as insufficient repayment ability, abnormal transactions, questionable use, and high uncertainty regarding collateral. The risk assessment results may also include a conclusion confidence level, reflecting the completeness of the current evidence chain and the degree of conflict resolution. The conclusion confidence level can be obtained based on the coverage of the response conclusions, the consistency of evidence sources, data quality assessment results, and the resolution rate of adversarial inquiries. For example, if all key conflicts raised in adversarial inquiries have been explained through supplementary information and evidence alignment completed, the conclusion confidence level can be increased; if key materials are still missing or conflicts cannot be resolved, the conclusion confidence level can be decreased, and the risk assessment results will indicate the need for further verification or a conservative strategy.

[0084] In some embodiments, risk assessment results may carry an index of evidentiary information associated with risk elements to support reviewers in quickly tracing back to the relevant evidence. This evidentiary information index may point to material numbers, data source identifiers, time interval identifiers, or verification record identifiers, and establish a correspondence with the risk elements. For example, for cash flow fluctuation risk elements, it can be associated with a recent period of cash flow statistics index and a collection cycle calculation index; for related-party transaction risk elements, it can be associated with a related-party relationship analysis index and a key transaction path index; for usage compliance risk elements, it can be associated with a usage proof material index and a rule condition comparison index; for collateral disposal risk elements, it can be associated with a valuation calculation index and an ownership verification index. Through this index organization method, risk assessment results can simultaneously meet the needs of automated decision output and manual review, improving the convenience of interpretation and auditing.

[0085] In the embodiments of this application, for complex objects, the risk assessment results can provide conclusion fragments for multiple risk chains before forming a summary conclusion. For example, for corporate credit recipients, conclusions can be formed for the operational chain, funding chain, related chain, and compliance chain, with each chain containing key risk elements and evidence indexes; the summary conclusion forms a final recommendation based on the risk level and mutual influence of each chain. For individual credit recipients, conclusions can be formed for income stability, debt expansion, consumption behavior, and consistency of use, before forming a summary conclusion. Through chain-based organization, the risk assessment results can better express the transmission path of risk across different dimensions, improve the transparency and understandability of comprehensive judgment, and also facilitate the setting of monitoring and handling strategies according to the chain in subsequent management.

[0086] According to embodiments of this application, multiple first agents generate preliminary assessment conclusions in parallel based on different risk dimensions, ensuring simultaneous coverage of concerns such as credit, anomalies, compliance, and operations, reducing information omissions caused by a single perspective. Adversarial queries are generated based on the relationships between preliminary assessment conclusions, proactively exposing conflicts, evidentiary gaps, and high uncertainties between conclusions, shifting the assessment focus from result aggregation to problem-driven cross-verification. Questions are directed to the first agent that generated the target conclusion, outputting a response conclusion containing supplementary information, providing new evidence and traceable explanations for the conclusion formation process, avoiding simple restatements and invalid interpretations. Compared to traditional single-model output or multi-model parallel scoring schemes, embodiments of this application improve the interpretability and audit traceability of risk assessment, reduce the failure to detect low-probability, high-loss risks caused by consensus bias, and enhance the robustness and credibility of overall decision-making.

[0087] The risk assessment method of this application will be specifically described below by way of preferred embodiments.

[0088] To address the challenge of timely identification of conflicting conclusions, evidence gaps, and high uncertainty points after parallel assessments across multiple risk dimensions, this application's embodiments employ a second intelligent agent to perform structured parsing of the preliminary assessment conclusions output by each risk dimension. This extracts risk elements, conclusion confidence levels, and evidence information identifiers. Based on this information, correlations are calculated, including at least one of conflict relationships, evidence gap relationships, and uncertainty relationships. This provides an executable trigger for generating adversarial queries. After determining the correlations, the second intelligent agent can also be used as an adversarial agent to generate adversarial queries based on these correlations and drive subsequent questioning and response processes. This provides an operable trigger and processing path for the generation and execution of adversarial queries.

[0089] In this embodiment, the processing structure of the second intelligent agent may include a conclusion parsing component, an element normalization component, an evidence alignment component, a relationship determination component, a question generation component, and a result output component. The conclusion parsing component parses the preliminary assessment conclusion into a unified set of fields and extracts risk elements and conclusion confidence levels; the element normalization component maps elements of different risk dimensions to a unified element dictionary and performs name-disambiguation; the evidence alignment component merges evidence information identifiers and performs consistency checks to form evidence coverage features; the relationship determination component determines conflict relationships, evidence gap relationships, and uncertainty relationships based on differences in element direction, conclusion confidence levels, and evidence coverage; the question generation component generates adversarial questions based on the determined relationships and configures target preliminary assessment conclusion identifiers and question points for the questions; the result output component generates relationship records and adversarial question records, and labels triggering elements and evidence indexes to support backtracking.

[0090] The training process of the second intelligent agent can be based on historical review records or manually reviewed samples to construct training data. The training data includes preliminary assessment conclusions across multiple risk dimensions, along with corresponding conflict, gap, and uncertainty annotations. It can further include adversarial question samples and their validity annotations from manual assessments. The model can employ joint training of sequence labeling and classification to improve the accuracy of conclusion parsing and relationship determination. It can also be combined with ranking or selection tasks generated by the questions for multi-task training to enhance the targeting of adversarial questions. Simultaneously, the element dictionary, thesaurus, and judgment thresholds are continuously updated through difficult example replay. Furthermore, the second intelligent agent can incorporate text parsing and structured extraction tools, element dictionaries and thesauruses, evidence identification indexing and alignment tools, rule execution tools, and log recording tools to bind and store triggering evidence and processing results when determining relationships and generating adversarial questions.

[0091] Through the above scheme, the second agent can not only automatically determine cross-dimensional relationships and discover contradictions and insufficient evidence after multi-agent parallel evaluation, but also generate targeted adversarial questions as an adversarial agent, reducing the burden of manual screening and improving the effectiveness of question generation, thereby enhancing the interpretability, auditability and robustness of the risk assessment process.

[0092] Figure 3 The flowchart illustrates a method for generating at least one set of adversarial queries based on association relationships, according to some exemplary embodiments of this application.

[0093] like Figure 3 As shown, the method for generating at least one set of adversarial queries based on association relationships includes steps S310 to S320.

[0094] In step S310, the second agent is used to identify at least one challenge trigger from the association relationship. The challenge trigger includes at least one of the following: conflict element corresponding to conflict relationship, gap evidence item corresponding to gap evidence relationship, and low confidence element corresponding to uncertainty relationship.

[0095] In the embodiments of this application, the challenge trigger can be composed of different types of information: when the relationship is conflicting, the trigger can be a conflicting element, such as the same risk element showing opposite directions in different preliminary assessment conclusions or the risk level difference exceeding a preset threshold; when the relationship is a gap in evidence, the trigger can be a gap in evidence, such as the missing material type required for a risk judgment, missing key fields, or insufficient coverage of evidence identifiers; when the relationship is uncertain, the trigger can be a low-confidence element, such as the confidence level of the preliminary assessment conclusion being lower than a threshold, large fluctuations in key elements, or insufficient credibility indicated by data quality assessment. The second intelligent agent can standardize the trigger into a unified description structure, which at least includes the trigger type, the associated target preliminary assessment conclusion identifier, the risk element or evidence identifier corresponding to the trigger, and the time window or data range in which the trigger occurs, so that it can be clearly pointed to when generating subsequent challenges.

[0096] In step S320, for each challenge trigger, the second agent generates an adversarial challenge and determines the challenge priority of the adversarial challenge. The challenge priority is determined based at least on the risk impact and conclusion confidence level corresponding to the challenge trigger.

[0097] The second agent can select different question generation strategies based on the type of trigger: for conflicting elements, the question can request an explanation of the source of the conflict, an explanation of the impact of the conflicting elements on the risk conclusion, and supplementary evidence; for gap evidence items, the question can request supplementary evidence of a specified type or provide alternative evidence and explain its credibility; for low-confidence elements, the question can request the source of uncertainty, supplementary data range, or the adoption of more conservative judgment criteria. The question priority can be determined based on the risk impact and the confidence level of the conclusion. The risk impact can be combined with factors such as the potential loss scale, business importance, or regulatory attention to form a comprehensive measure, while the confidence level of the conclusion can be derived from the confidence field of the preliminary assessment conclusion itself or inferred from evidence coverage and data quality.

[0098] By identifying conflicting elements, gap evidence items, and low-confidence elements, inconsistencies, insufficient evidence, and sources of uncertainty in cross-dimensional conclusions can be systematically exposed, reducing omissions and subjective differences caused by manual screening. Prioritizing inquiries allows for resource focus on risk impact, ensuring high-impact risk points receive priority clarification and supplementary evidence during the review process, reducing the possibility of overlooking low-probability, high-loss risks. Simultaneously, the targeted generation of adversarial inquiries helps form a traceable chain of questions and evidence, providing a clear logical thread for subsequent responses and the final risk assessment results, thereby enhancing the interpretability, auditability, and robustness of the overall decision-making process.

[0099] In scenarios involving parallel assessments across multiple risk dimensions, even with multiple preliminary assessment conclusions and adversarial inquiries, issues such as difficulty in uniformly arranging the review process, unclear inquiries, and lack of temporal control over inquiries and responses may arise. This can lead to adversarial inquiries failing to consistently fall within the responsibility dimension, impacting cross-validation efficiency and the traceability of conclusions. Particularly when the same assessment involves multiple materials, multiple risk clues, and multiple rounds of clarification, the lack of unified session state management and routing control can easily result in repeated or incorrect inquiries, or disordered response order. This weakens the adversarial mechanism's ability to reveal conflicts and evidentiary gaps and hinders the formation of a verifiable review process record. To address these issues, in the embodiments of this application, a master control agent can be introduced to arrange and control the review process.

[0100] Figure 4 The flowchart illustrating a method for arranging and controlling a review process according to some exemplary embodiments of this application is shown schematically.

[0101] like Figure 4 As shown, the method for arranging and controlling the review process includes steps S410 to S420.

[0102] In step S410, based on a preset review script or session state, the master agent activates the second agent to generate adversarial queries.

[0103] In the embodiments of this application, the master control agent is used to orchestrate the review process and coordinate the invocation of various agents. Its structure may include a session state management unit, a script and strategy execution unit, a task scheduling and routing unit, a constraint and verification unit, an evidence and log orchestration unit, and a result aggregation unit. The session state management unit is used to maintain the identifier of the object to be evaluated, the current round, the set of generated preliminary evaluation conclusions, the set of triggered inquiries, and the set of responses, and records the time sequence of each routing. The script and strategy execution unit is used to load a preset review script or select the next action according to the session state. The script may describe the rules for first generating a preliminary evaluation conclusion and then activating the second agent, then routing inquiries according to the target preliminary evaluation conclusion identifier, and advancing to the next round when a follow-up question is triggered. The task scheduling and routing unit is used to deliver adversarial or follow-up inquiries to the target first agent that generated the target preliminary evaluation conclusion, and collect its output candidate responses, response conclusions, or additional response conclusions. The constraint and verification unit is used to perform differential verification and necessary format consistency verification. Differential verification is used to compare the supplementary information identifier with the evidence information identifier cited in the target preliminary evaluation conclusion to ensure that the response introduces new evidence.

[0104] The training process of the master control agent can be based on historical approval process records, review records, and manual review samples to construct training data. The training data includes session state sequences, expected actions in different states, and annotations of routing targets and termination timings. Training objectives can include three types of tasks: next action selection, routing target selection, and termination determination. Multi-task learning improves the orchestration stability in complex sessions. At the same time, cost-sensitive objectives can be introduced to prioritize the routing of high-impact, high-risk inquiries and allow them to receive more rounds of processing.

[0105] In step S420, based on the target preliminary evaluation conclusion identifier, the master control agent routes the adversarial query to the target first agent that generated the target preliminary evaluation conclusion, triggering the target first agent to output a response conclusion.

[0106] In the embodiments of this application, the master intelligent agent can activate a second intelligent agent based on a preset review script or session state, enabling the second intelligent agent to generate adversarial queries at appropriate times and ensuring consistency between the generation of queries, the output of preliminary evaluation conclusions, and the generation of subsequent responses. Subsequently, the master intelligent agent can route the adversarial queries to the target first intelligent agent that generated the target preliminary evaluation conclusions, based on the target preliminary evaluation conclusion identifier carried by the adversarial queries, thereby assigning the queries to the risk dimension responsible for generating the conclusions. After being triggered, the target first intelligent agent outputs a response conclusion to clarify, supplement, or correct the query points, forming a one-to-one processing link between the adversarial queries and the response conclusions. Through the review script or session state, the master intelligent agent can also maintain the query rounds, processing queues, and round progression rules to prevent the adversarial query process from getting out of control or disordered.

[0107] According to embodiments of this application, the activation and routing mechanism of the master control agent enables adversarial queries to have clear triggering conditions and responsibility attribution, avoiding query generalization, misdirection, or duplicate delivery, and improving query processing efficiency; the session state-driven orchestration control enables multiple rounds of queries and responses to proceed in a traceable sequence, enhancing the traceability and auditability of the review process; at the same time, the query is directed to the first agent that generates the target conclusion, which can prompt that dimension to conduct a targeted review of its own conclusion and supplement the evidence, thereby more effectively resolving cross-dimensional conflicts, filling evidence gaps, and reducing uncertainty, ultimately improving the robustness and credibility of the risk assessment results.

[0108] During multi-risk dimension review processes, the primary agent, upon receiving adversarial inquiries, often tends to restate existing preliminary assessment conclusions or repeatedly cite the same evidence. This results in responses lacking incremental information, failing to effectively address the points of contention, and struggling to fill evidentiary gaps. Furthermore, different agents may use inconsistent granularity and scope in citing evidence, making it difficult for reviewers to determine whether the response truly introduces new evidence. This, in turn, undermines the effectiveness of the adversarial inquiry mechanism in enhancing interpretability and robustness. Especially when conflicting or uncertain relationships are triggered, if the response remains within the scope of existing evidence, the inquiry chain easily falls into circular interpretation, failing to provide substantive clarification of key risk points.

[0109] To address the aforementioned issues, in the embodiments of this application, for each adversarial query, the target first agent can perform supplementary information acquisition operations around the query point to generate candidate responses, and determine the supplementary information identifier associated with the candidate responses during the generation process. Supplementary information acquisition operations may include re-extracting and recalculating existing evaluation data, re-analyzing key material points, or re-estimating the time window of key elements, thus providing candidate responses with new sources of evidence. Subsequently, the master agent performs differential verification between the supplementary information identifier and the evidence information identifier already cited in the target's preliminary evaluation conclusion. Differential verification is used to determine whether the candidate response introduces new information different from the preliminary evaluation conclusion. When the differential verification passes, the candidate response is output as the response conclusion, thereby ensuring that each response has verifiable incremental characteristics at the evidence level and forms a clear correspondence with the query point.

[0110] By introducing supplementary information acquisition operations during the response phase, the target agent can be prompted to conduct secondary verification and supplementary evidence for the points of inquiry, thereby improving the relevance and information content of the response. Through the differential verification mechanism, the response can be computably constrained to include new evidence, avoiding vague restatements and repeated citations of evidence, and reducing the probability of ineffective loops in the process of resisting inquiries. At the same time, the differential results of supplementary information identifiers and evidence information identifiers can be used to form a clear record of evidence increments, enhancing the traceability and auditability of the review process, making the final risk assessment results more complete and robust in the interpretation chain, and increasing the reviewers' trust in the system output.

[0111] In the embodiments of this application, when the master control agent performs differential verification on the supplementary information identifier and the evidence information identifier already cited in the preliminary assessment conclusion and finds that the differential verification fails, it indicates that the candidate response generated by the target first agent still mainly relies on the existing evidence scope and is difficult to form a response with incremental evidence to the questioning points. Therefore, the master control agent can send a supplementary evidence collection instruction to the target first agent. The supplementary evidence collection instruction is used to trigger the target first agent to re-execute the supplementary information acquisition operation to obtain an updated candidate response. The supplementary information acquisition operation may include expanding the data time window to re-statistically analyze key indicators, re-analyzing the materials to extract previously unused key points, re-evaluating key elements using different calibers, and requesting supplementary conclusions or supplementary evidence information from other risk dimensions when necessary. The target first agent simultaneously determines the newly added supplementary information identifier when generating the updated candidate response. The master control agent performs differential verification again based on the newly added supplementary information identifier, and repeatedly sends the supplementary evidence collection instruction if the differential verification still fails, until the differential verification passes and the updated candidate response is confirmed as the response conclusion output.

[0112] Through the above scheme, on the one hand, the supplementary evidence collection instruction and the re-execution of the supplementary information acquisition operation form a closed-loop control, transforming the response generation from a one-time output into an iterative evidence collection process, which can continuously fill evidence gaps and enhance the explanatory power of the questioning points; on the other hand, differential verification, as an iteration termination condition, ensures that the final output response conclusion has verifiable new information at the evidence level, avoiding the candidate response repeatedly restating existing conclusions and causing the questioning chain to stagnate; in addition, the iterative process of updating candidate responses and adding supplementary information identifiers can form a clear evidence increment trajectory, which helps to improve the traceability and auditability of the review process, thereby enhancing the robustness and credibility of the final risk assessment results.

[0113] During multi-risk dimension review processes, even if the target agent has already output a response, there may still be issues such as the response only covering some elements of the questioning points, failing to address key evidentiary gaps, or failing to effectively resolve conflicts between preliminary assessment conclusions. Such insufficient responses will keep adversarial inquiries at a superficial level, failing to provide substantive clarification of key risk points, thus leading to persistent conflicts and uncertainty, ultimately affecting the reliability and robustness of the risk assessment results. Furthermore, without an objective mechanism for judging the quality of responses, the need for follow-up inquiries often relies solely on human experience, easily resulting in insufficient or excessive follow-up questioning, reducing review efficiency and introducing new subjective differences.

[0114] To address the aforementioned issues, in the embodiments of this application, a second intelligent agent can perform a response sufficiency determination after obtaining the response conclusion. The second intelligent agent can determine the response sufficiency determination result based on the coverage of the question points and the degree of conflict resolution of the related relationships in the response conclusion. Coverage can reflect whether the response conclusion fully addresses the risk elements, evidence items, and key explanation requirements involved in the question points, such as whether supplementary information corresponding to the question points is provided, whether the source of the supplementary information is clearly stated, and whether explanations are given for changes in key elements; conflict resolution can reflect whether the response conclusion alleviates or eliminates the original conflict relationship, such as whether judgments on the same element from different dimensions are unified to the same standard, whether there are still opposite conclusions after evidence alignment, and whether missing evidence is filled to change the intensity of the conflict. When the sufficiency of the response is determined to be insufficient, the second agent generates follow-up questions. Follow-up questions can target uncovered sub-questions, missing evidence items, or unresolved conflicting elements. The master agent routes the follow-up questions to the corresponding target first agent to trigger the generation of additional response conclusions, allowing the review process to proceed to the next round of clarification and supplementary evidence.

[0115] By introducing a sufficiency assessment mechanism based on coverage and conflict resolution, the quality of response conclusions can be evaluated in an actionable manner, reducing omissions and biases caused by relying solely on experience. When the response is insufficient, follow-up questions are automatically generated and additional responses are triggered through targeted routing, expanding adversarial questioning from a single round of Q&A to a controllable multi-round clarification process. This helps to continuously fill evidence gaps, gradually resolve cross-dimensional conflicts, and reduce uncertainty. At the same time, the multi-round follow-up questions and additional responses form a clear chain of questioning and a trajectory of evidence increment, enhancing the traceability and auditability of the review process, thereby improving the credibility and robustness of the final risk assessment results.

[0116] In the embodiments of this application, to facilitate examiners' understanding of the risk assessment process and support post-approval review, preliminary assessment conclusions, adversarial inquiries, follow-up inquiries, and additional response conclusions can be uniformly generated into session transcript information. The session transcript information can be stored in a structured record format, preserving the content of each output, the corresponding target conclusion identifier, the triggered inquiry points, and supplementary information identifiers, while retaining the generation time or processing round, thus transforming the risk assessment process from scattered conclusions into a continuous process record. By organizing the session transcript information, examiners can see how the initial judgment for each risk dimension was questioned, how supplementary evidence was provided, and how conflicts and gaps in evidence were gradually clarified, thereby reducing the difficulty in understanding the decision-making basis based solely on the final conclusion.

[0117] In some embodiments, the master agent can maintain routing order and session state when routing adversarial and follow-up inquiries, thus enabling temporal orchestration of session transcription information based on the routing order. Temporal orchestration can be based on routing events, arranging each inquiry sent, each response generated, each follow-up triggered, and each additional response output in the actual order of occurrence, and associating each session transcription with a participating role identifier. The participating role identifier can be used to distinguish different output sources, such as the master agent, the second agent, and the first agent for each risk dimension, thereby enabling reviewers to clearly identify the responsible party and context of each piece of information. The combination of temporal orchestration and role identifiers forms a traceable logical chain, allowing reviewers to quickly locate the response corresponding to a specific inquiry point, the supplementary evidence corresponding to a conflict element, and why a particular round of follow-up inquiries was triggered.

[0118] In the embodiments of this application, the presentation method of the session transcription information can be either streaming output or one-time display, depending on the usage scenario. In real-time review scenarios, a streaming output method can be used to present the session transcription information one by one in the order of arrangement, and the interface can be updated in real time when each piece of information is generated, so that reviewers can follow the review process synchronously and promptly identify risk points that need to be focused on. In batch review or post-audit scenarios, a one-time display method can be used to load the complete session transcription information and display it in segments by role or aggregated by round, so that reviewers can quickly browse the entire process and jump to key nodes. Regardless of the presentation method used, the participating role identifiers can be used to visually distinguish and filter the content, such as only viewing the questions generated by the second agent, only viewing the responses of the first agent for a certain risk dimension, or only viewing the follow-up questions and additional responses, to improve reading efficiency.

[0119] In some embodiments, the session transcript information can be further used to generate a summary index, enabling efficient location of key content even during lengthy sessions. The summary index can be generated around risk factors, questioning topics, evidence information identifiers, or conflict resolution results, and establish a reference relationship with the original session transcript information. Reviewers can use the summary index to quickly locate question and response segments related to a specific risk factor, or to pinpoint the moment when an evidence gap was filled and the source of supplementary information. This summary index does not alter the integrity of the session transcript information; it serves only as a navigational aid, facilitating improved usability while maintaining traceability. For scenarios requiring review notes, the risk assessment results and their corresponding key session segments can be permanently saved, allowing subsequent reviews to simultaneously view the conclusions and process evidence.

[0120] Through the above solution, on the one hand, preliminary assessment conclusions, adversarial inquiries, follow-up inquiries, and additional response conclusions can be uniformly converted into conversation transcription information and arranged chronologically according to the routing order, giving the risk assessment process a clear context and attribution of responsibility, significantly enhancing interpretability and traceability; on the other hand, the identification of participating roles enables reviewers to quickly distinguish the contributions and responsibilities of different agents, reducing understanding costs and improving review efficiency; in addition, the switchable presentation mode of streaming output or one-time display can adapt to different use cases such as real-time review and post-audit, enabling the system to support both interactive review experiences and the retention of process evidence required for compliance audits, thereby improving the credibility and audit usability of the overall risk assessment results.

[0121] Corresponding to the risk assessment method described above, embodiments of this application also provide a risk assessment device.

[0122] Figure 5 A schematic block diagram of a risk assessment apparatus according to an embodiment of this application is shown.

[0123] like Figure 5 As shown, the risk assessment device 500 of this embodiment includes an initial assessment module 510, an adversarial inquiry module 520, a supplementary module 530, and a final assessment module 540.

[0124] The initial evaluation module 510 can be used to acquire evaluation data of the object to be evaluated, and utilize multiple first agents to generate corresponding preliminary evaluation conclusions based on the evaluation data. These multiple first agents perform evaluations based on different risk dimensions. In one embodiment, the initial evaluation module 510 can be used to execute the operation S210 described above, which will not be repeated here.

[0125] The adversarial questioning module 520 can be used to determine the correlation between preliminary evaluation conclusions and generate at least one adversarial question based on the correlation. The adversarial question includes at least a question point and a target preliminary evaluation conclusion identifier associated with the question point. In one embodiment, the adversarial questioning module 520 can be used to perform the operation S220 described above, which will not be repeated here.

[0126] The supplementary module 530 can be used to output a response conclusion for each adversarial query, using the preliminary evaluation conclusion of the target to identify the corresponding target first agent. The response conclusion includes at least one piece of supplementary information, which is used to form a response to the query point. In one embodiment, the supplementary module 530 can be used to perform the operation S230 described above, which will not be repeated here.

[0127] The final evaluation module 540 can be used to generate a risk assessment result based on the preliminary evaluation conclusions, adversarial inquiries, and response conclusions. In one embodiment, the final evaluation module 540 can be used to perform the operation S240 described above, which will not be repeated here.

[0128] According to an embodiment of this application, the adversarial query module 520 can also be used to perform structured analysis of the preliminary assessment conclusion using a second intelligent agent, extract risk elements, conclusion confidence levels, and evidence information identifiers corresponding to each risk dimension; and use the second intelligent agent to calculate the correlation between the preliminary assessment conclusions based on the risk elements, conclusion confidence levels, and evidence information identifiers, wherein the correlation includes at least one of conflict relationships, evidence gap relationships, and uncertainty relationships.

[0129] According to an embodiment of this application, the adversarial questioning module 520 can also be used to identify at least one questioning trigger from the association relationship using a second agent. The questioning trigger includes at least one of the conflict elements corresponding to the conflict relationship, the gap evidence items corresponding to the gap evidence relationship, and the low confidence elements corresponding to the uncertainty relationship. For each questioning trigger, the second agent generates an adversarial question and determines the questioning priority of the adversarial question. The questioning priority is determined based at least on the risk impact and conclusion confidence corresponding to the questioning trigger.

[0130] According to embodiments of this application, the risk assessment device 500 may further include a main control module. The main control module can be used to activate a second agent to generate adversarial queries based on a preset review script or session state; and, based on a target preliminary assessment conclusion identifier, to route the adversarial queries to a target first agent that generated the target preliminary assessment conclusion, triggering the target first agent to output a response conclusion.

[0131] According to an embodiment of this application, the supplementary module 530 can also be used to generate candidate responses based on the challenge points by using the target first intelligent agent to perform supplementary information acquisition operations, and to determine the supplementary information identifier associated with the candidate responses; to use the master intelligent agent to perform differential verification between the supplementary information identifier and the evidence information identifier already cited in the target's preliminary evaluation conclusion; and, if the differential verification passes, to output the candidate response as the response conclusion. If the differential verification fails, the master intelligent agent sends a supplementary evidence collection instruction to the target first intelligent agent, which triggers the target first intelligent agent to re-execute the supplementary information acquisition operation to obtain the updated candidate response generated by the target first intelligent agent; to determine the newly added supplementary information identifier based on the updated candidate response, and to re-perform differential verification based on the newly added supplementary information identifier until the differential verification passes.

[0132] According to an embodiment of this application, the supplementary module 530 can also be used to determine the sufficiency of the response based on the coverage of the question points and the degree of conflict resolution of the association relationship by the second agent in response to obtaining the response conclusion; and if the response sufficiency result indicates that the response is insufficient, generate a follow-up question by the second agent and route the follow-up question to the corresponding target first agent through the master agent to trigger the generation of additional response conclusion.

[0133] According to embodiments of this application, the risk assessment device 500 may further include a display module. The display module can be used to generate session transcription information from preliminary assessment conclusions, adversarial queries, follow-up queries, and additional response conclusions; to chronologically arrange the session transcription information based on the routing order of the master agent, and to associate each session transcription information with a participating role identifier; and to present the session transcription information and participating role identifiers in a streaming output mode or a one-time display mode.

[0134] According to embodiments of this application, any multiple modules among the initial evaluation module 510, the adversarial query module 520, the supplementary module 530, and the final evaluation module 540 can be combined into one module, or any one of these modules can be split into multiple modules. Alternatively, at least part of the functionality of one or more of these modules can be combined with at least part of the functionality of other modules and implemented in one module. According to embodiments of this application, at least one of the initial evaluation module 510, the adversarial query module 520, the supplementary module 530, and the final evaluation module 540 can be at least partially implemented as hardware circuitry, such as a field-programmable gate array, a programmable logic array, a system-on-a-chip, a system-on-a-substrate, a system-on-package, an application-specific integrated circuit, or implemented in hardware or firmware by any other reasonable means of integrating or packaging the circuitry, or implemented in software, hardware, or firmware, or in any suitable combination of any of these three implementation methods. Alternatively, at least one of the initial evaluation module 510, the adversarial query module 520, the supplementary module 530, and the final evaluation module 540 may be implemented at least partially as a computer program module that can perform corresponding functions when the computer program module is run.

[0135] Figure 6 A block diagram schematically illustrates an electronic device suitable for implementing a risk assessment method according to an embodiment of this application.

[0136] like Figure 6 As shown, an electronic device 600 according to an embodiment of this application includes a processor 601, which can perform various appropriate actions and processes according to a program stored in a read-only memory 602 or a program loaded from a storage portion 608 into a random access memory 603. The processor 601 may include, for example, a general-purpose microprocessor, an instruction set processor and / or an associated chipset and / or a dedicated microprocessor. The processor 601 may also include onboard memory for caching purposes. The processor 601 may include a single processing unit or multiple processing units for executing different steps of the method flow according to an embodiment of this application.

[0137] Random access memory 603 stores various programs and data required for the operation of electronic device 600. Processor 601, read-only memory 602, and random access memory 603 are interconnected via bus 604. Processor 601 executes various steps of the method flow according to embodiments of this application by executing programs in read-only memory 602 and / or random access memory 603. It should be noted that the programs may also be stored in one or more memories other than read-only memory 602 and random access memory 603. Processor 601 may also execute various steps of the method flow according to embodiments of this application by executing programs stored in said one or more memories.

[0138] According to embodiments of this application, the electronic device 600 may further include an input / output interface 605, which is also connected to a bus 604. The electronic device 600 may also include one or more of the following components connected to the input / output interface 605: an input section 606 including a keyboard, mouse, etc.; an output section 607 including a cathode ray tube, liquid crystal display, etc., and a speaker, etc.; a storage section 608 including a hard disk, etc.; and a communication section 609 including a network interface card, such as a local area network card, modem, etc. The communication section 609 performs communication processing via a network such as the Internet. A drive 610 is also connected to the input / output interface 605 as needed. A removable medium 611, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 610 as needed so that computer programs read from it can be installed into the storage section 608 as needed.

[0139] Embodiments of this application also provide a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of this application.

[0140] According to embodiments of this application, the computer-readable storage medium can be a non-volatile computer-readable storage medium, such as including but not limited to: portable computer disks, hard disks, random access memory, read-only memory, erasable programmable read-only memory, portable compact disk read-only memory, optical storage devices, magnetic storage devices, or any suitable combination thereof. In embodiments of this application, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to embodiments of this application, the computer-readable storage medium may include the read-only memory 602 described above, and / or random access memory 603, and / or one or more memories other than read-only memory 602 and random access memory 603.

[0141] Embodiments of this application also include a computer program product comprising a computer program containing program code for performing the methods shown in the flowchart. When the computer program product is run on a computer system, the program code is used to cause the computer system to implement the methods provided in the embodiments of this application.

[0142] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and downloaded and installed via the communication section 609, and / or installed from the removable medium 611. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0143] In embodiments of this application, the computer program can be downloaded and installed from a network via communication section 609, and / or installed from removable medium 611. When the computer program is executed by processor 601, it performs the functions defined in the system of embodiments of this application. According to embodiments of this application, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0144] According to embodiments of this application, program code for executing the computer programs provided in the embodiments of this application can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. The program code can be executed entirely on the user's computing device, partially on the user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0145] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0146] Those skilled in the art will understand that the features described in the various embodiments of this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, the features described in the various embodiments of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application. All such combinations and / or combinations fall within the scope of this application.

Claims

1. A risk assessment method, characterized in that, The method includes: The evaluation data of the object to be evaluated is obtained, and multiple first intelligent agents are used to generate corresponding preliminary evaluation conclusions based on the evaluation data, wherein the multiple first intelligent agents perform the evaluation based on different risk dimensions. Determine the correlation between the preliminary assessment conclusions, and generate at least one adversarial question based on the correlation. The adversarial question includes at least a question point and a target preliminary assessment conclusion identifier associated with the question point. For each adversarial query, the target first agent, using the preliminary evaluation conclusion of the target, outputs a response conclusion, wherein the response conclusion includes at least one piece of supplementary information, which is used to form a response to the query point; and A risk assessment result is generated based on the preliminary assessment conclusions, the adversarial questioning, and the response conclusions.

2. The method according to claim 1, characterized in that, Determining the correlation between the preliminary assessment conclusions includes: The preliminary assessment conclusions are structured and analyzed using a second intelligent agent to extract risk elements, conclusion confidence levels, and evidence information identifiers corresponding to each risk dimension; and The second intelligent agent calculates the correlation between the preliminary assessment conclusion based on the risk factors, the confidence level of the conclusion, and the evidence information identifier. The correlation includes at least one of conflict relationship, evidence gap relationship, and uncertainty relationship.

3. The method according to claim 2, characterized in that, The generation of at least one set of adversarial queries based on the association includes: The second agent identifies at least one challenge trigger from the association, the challenge trigger including at least one of the following: a conflict element corresponding to the conflict relationship, a gap evidence item corresponding to the gap evidence relationship, and a low confidence element corresponding to the uncertainty relationship; and For each challenge trigger, the second agent generates the adversarial challenge and determines the challenge priority of the adversarial challenge. The challenge priority is determined based at least on the risk impact of the challenge trigger and the confidence level of the conclusion.

4. The method according to any one of claims 1 to 3, characterized in that, The method further includes: Based on a preset review script or session state, the master agent activates a second agent to generate the adversarial query; and Based on the target preliminary assessment conclusion identifier, the master control agent routes the adversarial query to the target first agent that generated the target preliminary assessment conclusion, triggering the target first agent to output the response conclusion.

5. The method according to claim 4, characterized in that, For each adversarial query, the target first agent outputs a response conclusion using the preliminary target evaluation conclusion, including: Based on the questioning point, the target first intelligent agent performs a supplementary information acquisition operation to generate a candidate response and determines the supplementary information identifier associated with the candidate response; The controlling agent performs differential verification between the supplementary information identifier and the evidence information identifier cited in the preliminary assessment conclusion of the target; and If the differential verification passes, the candidate response will be output as the response conclusion.

6. The method according to claim 5, characterized in that, If the differential verification fails, the master control agent sends a supplementary evidence collection instruction to the target first agent. The supplementary evidence collection instruction is used to trigger the target first agent to re-execute the supplementary information acquisition operation to obtain the updated candidate response generated by the target first agent. Based on the updated candidate response, a new supplementary information identifier is determined, and differential verification is re-performed based on the new supplementary information identifier until the differential verification passes.

7. The method according to claim 1 or 5, characterized in that, The method further includes: In response to obtaining the response conclusion, a second agent determines the sufficiency of the response based on the coverage of the question points and the degree of conflict resolution of the association based on the response conclusion; and If the response sufficiency determination result indicates that the response is insufficient, the second agent generates a follow-up question, and the master agent routes the follow-up question to the corresponding target first agent to trigger the generation of an additional response conclusion.

8. The method according to claim 7, characterized in that, The method further includes: The preliminary assessment conclusions, the adversarial questioning, the follow-up questioning, and the additional response conclusions are used to generate session transcription information. The session transcription information is temporally orchestrated based on the routing order of the master intelligent agent, and a participating role identifier is associated with each session transcription information; and The session transcription information and the participant role identifiers are presented in either a streaming output or a one-time display mode.

9. A risk assessment device, characterized in that, The device includes: The initial assessment module is used to: acquire assessment data of the object to be assessed, and generate corresponding preliminary assessment conclusions based on the assessment data using multiple first intelligent agents, wherein the multiple first intelligent agents perform assessments based on different risk dimensions; The adversarial questioning module is used to: determine the correlation between the preliminary evaluation conclusions, and generate at least one adversarial question based on the correlation, wherein the adversarial question includes at least a question point and a target preliminary evaluation conclusion identifier associated with the question point; The supplementary module is configured to: for each adversarial query, use the preliminary evaluation conclusion of the target to identify the corresponding target first agent and output a response conclusion, wherein the response conclusion includes at least one piece of supplementary information, the supplementary information being used to form a response to the query point; and The final evaluation module is used to generate a risk assessment result based on the preliminary evaluation conclusion, the adversarial inquiry, and the response conclusion.

10. An electronic device, comprising: One or more processors; Memory, used to store one or more computer programs. The characteristic feature is that the one or more processors execute the one or more computer programs to implement the steps of the method according to any one of claims 1 to 8.

11. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 8.

12. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 8.

Images