Data processing methods and electronic devices for intelligent agents

By employing a multi-agent cross-review and multi-dimensional questioning mechanism, defects in agent response data are identified and corrected, thus solving the problem of insufficient accuracy in agent output and achieving higher data processing accuracy and stability.

CN123117329APending Publication Date: 2026-11-17WEBANK (CHINA)
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611288363.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-24
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

When dealing with complex problems, existing intelligent agents suffer from insufficient accuracy in their output results due to their own knowledge biases and reasoning blind spots.

Method used

Through cross-review and multi-dimensional questioning mechanisms among multiple agents, defects in the initial response data are identified and corrected, candidate response data is generated, and the final target response data is selected through comprehensive scoring.

Benefits of technology

It improves the accuracy of agent data processing, reduces the probability of original defects remaining, and enhances the quality and stability of the final response data.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123117329A_ABST
    Figure CN123117329A_ABST
Patent Text Reader

Abstract

This application discloses a data processing method and electronic device for intelligent agents. The method includes: acquiring initial response data generated by multiple intelligent agents for target problem data; for each intelligent agent, identifying defects in the initial response data of the intelligent agent through other intelligent agents among the multiple intelligent agents, and acquiring candidate response data obtained by the intelligent agent adjusting the initial response data based on the existing defects; and determining target response data from the candidate response data corresponding to the multiple intelligent agents respectively. The technical solution of this application can identify the current intelligent agent's response defects from the external perspective of other intelligent agents, overcome the limitations of self-examination by a single intelligent agent, detect reasoning defects that are difficult for a single intelligent agent to discover, complete the iteration of defect perception-directed correction, and improve the data processing accuracy of the intelligent agent.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence technology, specifically relating to a data processing method and electronic device for an intelligent agent. Background Technology

[0002] With the rapid development of artificial intelligence technology, intelligent agents are widely used in various complex problem-solving scenarios. Current solutions utilize intelligent agents by providing the user's problem to the agent, which then independently generates a response. However, single-agent systems are limited by their own knowledge biases and reasoning blind spots, leading to reasoning illusions and consequently, the accuracy of the final output needs improvement. Summary of the Invention

[0003] The purpose of this application is to provide a data processing method and electronic device for intelligent agents, so as to improve the accuracy of data processing for intelligent agents.

[0004] Other features and advantages of this application will become apparent from the following detailed description, or may be learned in part from practice of this application.

[0005] According to one aspect of the embodiments of this application, a data processing method for an intelligent agent is provided, comprising:

[0006] Acquire initial response data generated by multiple agents in response to the target problem data; For each agent, defects in the initial response data of the agent are identified by other agents among the plurality of agents, and candidate response data is obtained by the agent adjusting the initial response data based on the defects. The target response data is determined from the candidate response data corresponding to the multiple intelligent agents.

[0007] According to one aspect of the embodiments of this application, a data processing apparatus for an intelligent agent is provided, comprising: The initial response module is used to acquire the initial response data generated by multiple agents in response to the target question data. The adjustment module is used to identify defects in the initial response data of each agent through other agents among the plurality of agents, and to obtain candidate response data obtained by the agent adjusting the initial response data based on the defects. The determination module is used to determine the target response data from the candidate response data corresponding to the multiple intelligent agents respectively.

[0008] In one embodiment of this application, the determining module is specifically used for: For each agent, multiple review agents question the candidate response data of the agent and obtain the optimized response data obtained by the agent based on the questioning results; the questioning includes raising questions and verifying defects, and one review agent corresponds to one questioning dimension; The target response data is determined from the optimized response data corresponding to the multiple intelligent agents.

[0009] In one embodiment of this application, the determining module is specifically used for: If the query result of any of the review agents indicates that the optimized response data has defects that need to be optimized, then the optimized response data is used as the candidate response data, and the step of querying the candidate response data of the agents by multiple review agents is returned. If the query results of multiple review agents all indicate that the optimized response data does not have any defects to be optimized, or if the number of iterations reaches a preset number, then the optimized response data is taken as the optimized response data obtained by the agent.

[0010] In one embodiment of this application, the determining module is specifically used for: For each agent, the optimized response data of the agent is scored by multiple review agents, and a comprehensive score of the optimized response data is obtained based on the scores of the multiple review agents. The optimized response data with the highest overall score will be used as the target response data.

[0011] In one embodiment of this application, the review agent includes a compliance agent, which is used to perform compliance detection; the determination module is specifically used for: If the score of the compliant intelligent agent for the optimized response data of the intelligent agent is a set value, then the comprehensive score of the optimized response data is set to the set value; If the score given by the compliance agent to the optimized response data is not a set value, then the scores of other review agents are weighted and summed to obtain the comprehensive score of the optimized response data.

[0012] In one embodiment of this application, the apparatus further includes a data completion module, used for: Intent identification is performed on user question data to determine the intent uncertainty level of the user question data; the intent uncertainty level characterizes the degree to which the user intent of the user question data cannot be uniquely determined. The user question data is supplemented based on the level of intent uncertainty to obtain the target question data.

[0013] In one embodiment of this application, the data completion module is specifically used for: The original question data is subjected to intent recognition to obtain the intent probability distribution of the original question data; the intent probability distribution includes the probability that the original question data belongs to each user intent; The information entropy of the original question data is determined by the probability distribution of the intent. The level of intent uncertainty of the user question data is determined based on the relationship between the information entropy and the preset entropy value.

[0014] In one embodiment of this application, the data completion module is specifically used for: If the uncertainty level of the intent is low, then the user question data will be used as the target question data; If the uncertainty level of the intent is medium uncertainty, then the user question data is supplemented according to the interaction cost and inference risk cost corresponding to the user question data to obtain the target question data; the interaction cost includes the cost required to supplement the question data through interactive operation, and the inference risk cost includes the risk cost of generating incorrect response data based on the user question data; If the uncertainty level of the intent is high, then the user question data is completed through interactive operation to obtain the target question data.

[0015] In one embodiment of this application, the data completion module is specifically used for: The interaction cost is generated based on the number of missing information in the user question data, wherein the interaction cost is positively correlated with the number of missing information. Based on the posterior probabilities of multiple user intents obtained from intent recognition, the inference information entropy is calculated, and the inference risk cost is generated based on the inference information entropy; wherein, the posterior probability of the user intent is the probability that the user intent belongs to a real intended purpose, calculated based on historical intent recognition data. If the interaction cost is greater than the reasoning risk cost, then the user question data will be used as the target question data. If the interaction cost is less than the reasoning risk cost, then the user question data is completed through interactive operations to obtain the target question data.

[0016] According to one aspect of the embodiments of this application, a computer-readable medium is provided, on which a computer program is stored, which, when executed by a processor, implements the data processing method of the intelligent agent as described in the above technical solutions.

[0017] According to one aspect of the embodiments of this application, an electronic device is provided, the electronic device comprising: a processor; and a memory for storing executable instructions of the processor; wherein the processor executes the executable instructions to cause the electronic device to perform a data processing method for an intelligent agent as described in the above technical solutions.

[0018] According to one aspect of the embodiments of this application, a computer program product or computer program is provided, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the data processing method for an intelligent agent as described in the above technical solutions.

[0019] In the technical solution provided in this application embodiment, multiple agents generate initial response data for the target problem data. Then, for each agent, other agents identify defects in the initial response data. This allows for the identification of current agent response defects from the external perspective of other agents, overcoming the limitations of a single agent's self-examination and detecting reasoning defects that are difficult for a single agent to discover. Simultaneously, candidate response data is obtained by the agents adjusting the initial response data based on the existing defects. This means that after identifying defects, the original agents are driven to adjust their responses to obtain candidate responses, rather than simply discarding the original output. Instead, an iterative defect perception-targeted correction process is completed, retaining the reasonable parts of the original response while repairing defects, thereby improving the accuracy of the agents' data processing. Finally, the target response data is determined from the candidate response data corresponding to each of the multiple agents. This involves filtering the target response from the set of candidate responses after all iterative corrections. Since all candidate responses participating in the filtering have undergone defect identification and adjustment, the probability of original defects remaining is reduced, thus improving the accuracy of the final target response data.

[0020] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0021] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0022] Figure 1 An exemplary system architecture block diagram illustrating the application of the technical solution of this application is shown schematically.

[0023] Figure 2 A flowchart illustrating a data processing method for an intelligent agent provided in one embodiment of this application is shown schematically.

[0024] Figure 3 A flowchart illustrating a data processing method for an intelligent agent provided in one embodiment of this application is shown schematically.

[0025] Figure 4 A schematic diagram illustrating the data processing process of an intelligent agent provided in one embodiment of this application is shown.

[0026] Figure 5 A flowchart illustrating a data processing method for an intelligent agent provided in one embodiment of this application is shown schematically.

[0027] Figure 6 A schematic block diagram of the data processing apparatus for an intelligent agent provided in an embodiment of this application is shown.

[0028] Figure 7 A schematic diagram of a computer system architecture suitable for implementing the embodiments of this application is shown. Detailed Implementation

[0029] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided to make this application more comprehensive and complete, and to fully convey the concept of the exemplary embodiments to those skilled in the art.

[0030] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a thorough understanding of embodiments of this application. However, those skilled in the art will recognize that the technical solutions of this application can be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of this application.

[0031] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0032] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily need to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0033] In the embodiments of this application, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0034] It is understood that in the specific implementation of this application, customer information (such as transaction information, reconciliation data) and other related data are involved. When the above embodiments of this application are applied to specific products or technologies, customer permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0035] Figure 1 An exemplary system architecture block diagram illustrating the application of the technical solution of this application is shown schematically.

[0036] like Figure 1 As shown, the system architecture 100 may include terminal devices 110, a network 120, and a server 130. Terminal devices 110 may include smartphones, tablets, laptops, smart voice interaction devices, smart home appliances, in-vehicle terminals, etc. Server 130 may be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services. Network 120 may be a communication medium of various connection types capable of providing a communication link between terminal devices 110 and server 130, such as a wired communication link or a wireless communication link.

[0037] Depending on the implementation requirements, the system architecture in this application embodiment can have any number of terminal devices, networks, and servers. For example, server 130 can be a server group composed of multiple server devices. In addition, the technical solutions provided in this application embodiment can be applied to terminal device 110, or to server 130, or can be implemented jointly by terminal device 110 and server 130. This application does not impose any special limitations on this.

[0038] The data processing method for the intelligent agent provided in this application will be described in detail below with reference to specific implementation methods.

[0039] Figure 2 A flowchart illustrating a data processing method for an intelligent agent according to an embodiment of this application is shown schematically. This method can be implemented by a terminal device or a server, and can be implemented by an intelligent agent, which can be deployed in the terminal device or server. Figure 2 As shown, the data processing method for the intelligent agent provided in this embodiment includes steps 210 to 230, as detailed below: Step 210: Obtain the initial response data generated by multiple agents for the target question data.

[0040] An agent is an executive entity capable of understanding and generating reasoning responses to input questions, and can independently output response data. The target question data is the agent's input question containing the user's intent; it can be any form of data, such as images, text, or audio. The target question data defines the task the agent needs to complete or the response result it needs to output. For example, the target question data could be code to generate a certain function, analyze the meaning of a word, translate a passage, etc. The target question data is input to the agent, and the agent generates corresponding response data, denoted as the initial response data. In this embodiment, the agent is used to generate response data for the question; therefore, the agent can also be called a generative agent or a generative intelligence agent. This embodiment uses multiple agents to reason and respond to the same target question data, with each agent generating a corresponding initial response data.

[0041] In one embodiment, a suitable number of agents can be invoked to generate multiple initial response data corresponding to the target problem data, based on the complexity of the task and the number of possible solutions. The complexity of the task refers to the comprehensive level of the reasoning difficulty, number of constraints, solution dimensions, and boundary uncertainties of the target task indicated by the target problem data. For example, tasks can be divided into three categories: simple tasks, medium-complexity tasks, and high-complexity tasks. The higher the complexity of the task, the more agents are used. The number of possible solutions refers to the total number of differentiated, effective, and feasible solutions to the target task within a reasonable solution range. The response data output by the agents to the target problem data includes at least one solution to the problem; therefore, the initial response data can also be referred to as initial solutions. The more possible solutions there are, the more agents are used.

[0042] In one embodiment, it can be stipulated that the number of agents must be greater than or equal to the number of feasible solution dimensions in the core of the task, to ensure that each type of feasible solution is generated by an independent agent, and to avoid missing solution dimensions.

[0043] In one embodiment, a task parsing model can be used to identify the complexity level of the target task in the target problem data and predict the range of feasible solutions corresponding to the task; then, a task complexity score is calculated by combining the task complexity and the number of possible solutions, and finally, the number of agents is determined based on the complexity score. Optionally, the higher the complexity score, the more agents are required.

[0044] Step 220: For each agent, identify the defects in the agent's initial response data through other agents among multiple agents, and obtain candidate response data obtained by the agent adjusting the initial response data based on the defects.

[0045] The process iterates through each agent, without relying on its own self-checking. Instead, other agents act as external reviewers to identify defects in the agent's initial response data. These defects include factual errors, logical contradictions, missing information, reasoning flaws, off-topic responses, illusions, compliance deficiencies, risk control weaknesses, and profit-related issues—all problems that could lead to incorrect responses. The identified defects are then fed back to the agent, which, based on the defect feedback, modifies its initial response and generates revised candidate response data.

[0046] This process enables multiple agents to engage in differentiated competition through game theory. Each agent proactively leverages its strengths and avoids its weaknesses, emphasizing the unique advantages, core innovations, and implementation highlights of its own solution. At the same time, it objectively analyzes and amplifies the shortcomings and deficiencies of competitors' solutions (i.e., the initial response data of other agents), clearly highlighting the differentiated capability barrier of "what it has achieved and what competitors have not covered." Simultaneously, it learns from the advantages of competitors (i.e., the initial response data of other agents), completes individual genetic mutation and upgrades, and achieves self-evolution of its solution.

[0047] In one embodiment, step 220 includes the following steps: multiple agents engage in pairwise cross-games to actively identify illusion vulnerabilities, compliance flaws, risk control shortcomings, and profit deficiencies in their opponents' solutions, forming a list of adversarial challenges; the challenged agents repair the flaws in their effective games and eliminate the flaws in their own solutions; each agent absorbs superior strategies from competitors, completes positive optimization and upgrades of its solutions, and outputs candidate response data; after one round of peer-to-peer game, all solutions eliminate low-level illusions, make up for shortcomings, and achieve local strategy equilibrium.

[0048] In one embodiment, step 220 can be repeated multiple times. That is, after each agent repairs a defect and outputs candidate response data, this candidate response data can be used as initial response data to receive defects identified by other agents and repair them again. Multiple rounds of repair can further encourage agents to "leverage their strengths and compensate for their weaknesses," improving the accuracy of the candidate response data ultimately output by the agents. During the multiple rounds of repetition, when the number of repetitions reaches a set value, or the deviation between the candidate response data obtained from repairing defects in two consecutive rounds is less than a threshold, the multi-agent game-based competition process can be considered to have converged, and the iterative process can be stopped.

[0049] Step 230: Determine the target response data from the candidate response data corresponding to multiple agents.

[0050] After the preceding steps, each agent outputs a set of candidate response data that has completed defect correction. All candidate response data are then subjected to one or more processing methods such as screening, evaluation, and aggregation to select the final target response data as the output of the entire method.

[0051] In one embodiment, each candidate response data can be quantitatively scored based on dimensions such as content accuracy, logical completeness, question suitability, rigor of expression, and comprehensiveness using preset evaluation rules. Alternatively, a comprehensive evaluation of each candidate response data can be conducted through intelligent agent aggregation and judgment, and key information comparison and fusion. Finally, based on the quantitative scoring or comprehensive evaluation results, the optimal candidate response data is selected as the target response data.

[0052] In the technical solution provided in this application embodiment, multiple agents generate initial response data for the target problem data. Then, for each agent, other agents identify defects in the initial response data. This allows for the identification of current agent response defects from the external perspective of other agents, overcoming the limitations of a single agent's self-examination and detecting reasoning defects that are difficult for a single agent to discover. Simultaneously, candidate response data is obtained by the agents adjusting the initial response data based on the existing defects. This means that after identifying defects, the original agents are driven to adjust their responses to obtain candidate responses, rather than simply discarding the original output. Instead, an iterative defect perception-targeted correction process is completed, retaining the reasonable parts of the original response while repairing defects, thereby improving the accuracy of the agents' data processing. Finally, the target response data is determined from the candidate response data corresponding to each of the multiple agents. This involves filtering the target response from the set of candidate responses after all iterative corrections. Since all candidate responses participating in the filtering have undergone defect identification and adjustment, the probability of original defects remaining is reduced, thus improving the accuracy of the final target response data.

[0053] Figure 3A flowchart illustrating a data processing method for an intelligent agent according to an embodiment of this application is shown schematically. This embodiment is a further refinement of the above embodiment. Figure 3 As shown, this embodiment includes the following steps 310-340: Step 310: Obtain the initial response data generated by multiple agents for the target question data.

[0054] Step 320: For each agent, identify the defects in the agent's initial response data through other agents among multiple agents, and obtain candidate response data obtained by the agent adjusting the initial response data based on the defects.

[0055] Steps 310-320 can be referred to the relevant descriptions of steps 210-220 above, and will not be repeated here.

[0056] Step 330: For each agent, multiple review agents question the candidate response data of the agent and obtain the optimized response data obtained by the agent based on the questioning results. The questioning includes raising questions and verifying defects. One review agent corresponds to one questioning dimension.

[0057] The review agent is an expert-type agent, distinct from the generating agent, and is an independent review and verification entity. To differentiate between the aforementioned agent that generates response data and the review agent in this embodiment, the agent that generates response data will be referred to as the generating agent thereafter.

[0058] Each review agent corresponds to one questioning dimension. Questioning refers to raising doubts and verifying defects. By questioning the candidate response data through the review agent, it raises questions about the candidate response data and verifies whether there are defects in the candidate response data. Each review agent is responsible for a single questioning dimension, avoiding dimension confusion and redundant verification, and ensuring that the questioning is professional and precise. The process of the review agent questioning the candidate response data is similar to the defect identification process of other agents for a particular agent. The questioning results output by the review agent also include the defects existing in the candidate response data. The generating agent optimizes the candidate response data based on the questioning results, outputting more accurate optimized response data.

[0059] In one embodiment, multiple questioning dimensions matching the response verification scenario are pre-configured. Typical dimensions include factual accuracy verification, logical closure verification, content integrity verification, solution feasibility verification, compliance verification, expression rigor verification, and requirement matching degree verification. Each dimension is independently bound to a dedicated review agent, and each review agent possesses professional verification and questioning reasoning capabilities for the corresponding dimension. Then, the candidate response data corresponding to each generating agent is synchronously distributed to all dedicated review agents. Each review agent independently conducts questioning work based on its own dedicated questioning dimension. After all review agents complete the verification, they summarize the questioning content and defect verification results of each dimension to form a multi-dimensional structured questioning report, clarifying the problems, vulnerability locations, and optimization suggestions for each dimension. The questioning report is fed back to the corresponding generating agent, which responds to the questioning questions of each dimension one by one, checking and optimizing each questioning point, thereby solving the problems of single-dimensional defects and multi-dimensional cross-conflicts, completing a second deep iteration, and finally generating optimized response data with higher accuracy, fewer defects, and multi-dimensional compliance.

[0060] In one embodiment, the questioning and response data optimization processes of multiple review agents can be executed in sequence. First, the multiple review agents are sorted, and then questioning and response optimization are performed sequentially according to the sorted order. Specifically, the previous review agent questions the candidate response data, obtains the questioning result, and the generating agent optimizes the candidate response data based on the questioning result before accepting the questioning from the next review agent, and so on, until all review agents have completed their questioning.

[0061] In one embodiment, each review agent can be sorted according to rules such as the severity of the defect harm corresponding to each question dimension and the order of the inference chain.

[0062] In one embodiment, step 330 can be repeated multiple times. That is, after each agent optimizes the response and outputs optimized response data, the optimized response data can be used as candidate response data to receive multiple review agents for further questioning. Through multiple rounds of questioning, the accuracy of the optimized response data finally output by the agent can be further improved.

[0063] In one embodiment, step 330 specifically includes: if the query result of any review agent indicates that the optimized response data has defects to be optimized, then the optimized response data is used as candidate response data, and the step of querying the candidate response data of the agent by multiple review agents is returned; if the query results of multiple review agents all indicate that the optimized response data does not have defects to be optimized, or if the number of iterations reaches a preset number, then the optimized response data is used as the optimized response data obtained by the agent.

[0064] During the repeated execution of multi-review agent challenges, if the challenge result of one review agent indicates that the candidate response data still has defects that need optimization, it means that the stopping condition for repeated iteration has not been met, and the multi-review agent challenge process continues. If the challenge results of all review agents indicate that the candidate response data has no defects that need optimization, it is determined that the stopping iteration condition has been met. At this time, the optimized response data output by the generating agent is the final optimized response data output by the generating agent. In one embodiment, to avoid infinite loops, the maximum number of iterations is set to a preset number. When the number of iterations reaches the preset number, regardless of the challenge result output by the review agent, the iteration stops, and the optimized response data output by the generating agent at this time is taken as the final optimized response data output by the generating agent.

[0065] For example, taking the financial sector as an example, four expert agents—compliance, risk control, user experience, and business profitability—are introduced as review agents to form a multi-dimensional expert inquiry. Through multiple rounds of iterative screening, the professionalism of each solution is continuously improved, driving the solution to continuously evolve until global convergence. First, the four types of review agents initiate high-level adversarial inquiries from a financial professional perspective, conducting in-depth source verification of compliance red lines, risk loopholes, user experience, and business rationality. Each round of inquiry and response is a solution screening and optimization. Based on the problems that arise during the inquiry process, the solution is continuously revised, that is, the response data is optimized. When the solution can no longer be questioned or improved, or when the maximum number of iterations is reached, the expert inquiry is considered complete, reaching the optimal strategy equilibrium state.

[0066] Step 340: Determine the target response data from the optimized response data corresponding to multiple agents.

[0067] The process of optimizing the screening of response data can be referred to the screening process of candidate response data in step 230 above, and will not be repeated here.

[0068] The technical solution of this application first generates initial responses independently by multiple agents, broadening the problem-solving perspective and avoiding one-sided and singular responses. Then, the explicit defects of the initial responses are corrected through cross-review by the agents, completing the initial quality optimization. Next, a multi-dimensional review agent questioning mechanism is introduced to accurately identify and repair response defects, achieving in-depth iterative optimization of responses. Finally, multiple optimized responses are comprehensively evaluated and the best is output. By eliminating various defects layer by layer and refining the response quality step by step, the accuracy of agent responses is significantly improved, effectively enhancing the output quality and stability of complex problem solving.

[0069] In one embodiment, the target response data determination process includes: for each agent, scoring the agent's optimized response data by multiple review agents, and obtaining a comprehensive score of the optimized response data based on the scores of the multiple review agents; and taking the optimized response data with the highest comprehensive score as the target response data.

[0070] This embodiment uses multiple review agents to score the optimized response data of each generating agent based on their own dimensions. For each generating agent, there are multiple scores generated by review agents. These scores are merged to obtain a comprehensive score for the optimized response data output by that generating agent. For example, these scores can be summed, weighted summed, or averaged to obtain the comprehensive score. The comprehensive score reflects the overall accuracy of the optimized response data across various dimensions; a higher comprehensive score indicates higher overall accuracy. Therefore, the optimized response data with the highest comprehensive score is ultimately selected as the final target response data for output.

[0071] In one embodiment, the plurality of review agents include a compliance agent, which is used to perform compliance checks on the response data. The process of calculating the comprehensive score includes: if the score of the compliance agent for the optimized response data of the agent is a set value, then the comprehensive score of the optimized response data is set to the set value; if the score of the compliance agent for the optimized response data of the agent is not a set value, then the scores of the other review agents are weighted and summed to obtain the comprehensive score of the optimized response data.

[0072] The compliance check is used to verify whether the response data complies with industry standards, laws and regulations, platform rules, task constraints, and publicly available compliance guidelines. Therefore, the compliance check is of paramount importance. If the optimized response data fails the compliance check, it is unusable and is deleted without being considered as target response data. The compliance agent outputs a compliance score based on the compliance check results. If an optimized response data fails the compliance check, its compliance score is set to a preset value, such as 0. Therefore, when the compliance agent's score for the agent's optimized response data is the preset value, it is determined that the optimized response data has failed the compliance check, and its overall score is directly set to the preset value, without considering scores from other dimensions. If the compliance agent's score for the agent's optimized response data is not the preset value, it means that the optimized response data has passed the compliance check. In this case, the scores from other review agents are weighted and summed to obtain the overall score of the optimized response data. The weights of each agent in the other review agents can be preset, and the sum of the weights of all agents is 1.

[0073] For example, introducing compliance Risk control User experience Commercial revenue The system employs four review agents, each scoring each solution (optimizing response data) based on its own knowledge system and area of ​​interest. The scoring range is as follows: Compliance Dimension Expert Score (0 or 100 points, score for hard compliance determination, 0 points for non-compliance) Risk control expert scoring (0-100 points, scores for risk rationality and completeness) User experience expert rating (0-100 points, market adaptability and user-friendliness scores) Business profitability expert rating (0-100 points, business sustainability and profitability scores) The overall score S is calculated as follows:

[0074] in, , , Risk control User experience Commercial revenue The weights corresponding to these three review agents, .

[0075] The following is a specific example illustrating the implementation process of the technical solution of this application.

[0076] Target data: Designed for working professionals aged 22-35, with loan amounts of 500-5000 yuan, terms of 7-90 days, flexible repayment options, purely online risk control, and no collateral required.

[0077] S1. Three agents (generating agents) independently generate initial response data. After passing the system's basic initial screening, if there are no defects or errors in basic parameters, all of them enter the competitive game stage and output candidate response data.

[0078] Option A (Inclusive and Relaxed): Low entry barriers and weakened qualification review, focusing on user coverage, but initially there are issues of incomplete compliance evidence and hidden illusions; Option B (Sound Risk Control): Strong qualification review and strict access rules, focusing on low bad debts, initially with excessive risk control and the illusion of fabricated industry rules; Option C (Balanced and Stable): Credit rating system + dynamic interest rate, no severe illusion, balanced across all dimensions, initial weakness is the vague wording of the rules.

[0079] S2, peer competition and self-learning repair. That is, the three intelligent agents detect and repair defects among themselves.

[0080] Pairwise cross-questioning, for example: Option A questions Option C for "lacking detailed rules and only providing conceptual definitions." Option C then supplements and refines the interest rate tiering rules and outputs candidate response data. Option B questions Option A's "lack of compliance basis and overly lenient entry requirements." Option A supplements its basic risk control logic with reference to Option B and then outputs candidate response data. Option C questions Option B for "exhibiting severe illusions and fabricating industry rules." Option B then removes the illusion rules and outputs candidate response data.

[0081] S3, multiple rounds of expert questioning and solution iteration, namely multiple review intelligent agents questioning and optimization.

[0082] Create four review agents focusing on compliance, risk control, user experience, and business revenue. Each agent's context includes extensive knowledge and case studies relevant to its category. Each agent poses specialized questions to optimize response data, such as: Solution B for the business revenue review intelligent agent asks questions based on its risk control strategy to ensure loan scale. Solution B is based on the lightweight access logic of the inquiry, deletes invalid and strict rules, and outputs optimized response data. User experience review intelligent agent questioning solution C, based on the credit rating risk control strategy, requires the collection of customer information and whether the customer accepts it. Solution C adds mandatory reading and personal information use authorization agreement signing process based on questioning, and outputs optimized response data.

[0083] S4. Calculation of comprehensive score.

[0084] After the first three stages, all three solutions passed the compliance check, and the scores of each reviewing agent are shown in Table 1 below: Table 1

[0085] Overall score formula ; Overall score for Option A: 70×0.5+95×0.3+75×0.2=35+28.5+15=78.5; Overall score for Option B: 90×0.5+65×0.3+70×0.2=45+19.5+14=78.5; Overall score for Option C: 88×0.5+88×0.3+90×0.2=44+26.4+18=88.4; As can be seen, Scheme C has the highest overall score, so Scheme C will be used as the target response data output.

[0086] Figure 4 A flowchart illustrating a data processing method for an intelligent agent according to an embodiment of this application is shown schematically. This embodiment is a further refinement of the above embodiment. Figure 4 As shown, this embodiment includes the following steps 410-450: Step 410: Perform intent identification on the user question data to determine the intent uncertainty level of the user question data; the intent uncertainty level characterizes the degree to which the user intent of the user question data cannot be uniquely determined.

[0087] User question data refers to the original question data input by users. Intent recognition is performed on this data to determine the user's true intention in asking the question. User intent represents what the user wants to do and what goal they want to achieve. Identifying the user intent is crucial to determining the target task that the intelligent agent needs to perform. Based on the identified user intent, the intent uncertainty level is determined. The intent uncertainty level characterizes the degree to which the user intent cannot be uniquely determined, that is, the degree of uncertainty of the user intent.

[0088] In one embodiment, intent recognition can be performed on user question data using an intent recognition model, or it can be performed using an intelligent agent. Optionally, user intent recognition can be performed by performing semantic parsing, keyword extraction, requirement decomposition, and constraint matching on the user question data.

[0089] In one embodiment, if a user question is semantically complete, has clear requirements, a unique scenario, and is unambiguous and lacks information, it is determined to be at a low uncertainty level; if a user question has problems such as semantic ambiguity, lack of key information, multiple requirements, and unclear scenario boundaries, and there are multiple reasonable ways to interpret it, it is determined to be at a medium or high uncertainty level.

[0090] In one embodiment, the intention uncertainty level calculation process includes: performing intention identification on the original question data to obtain the intention probability distribution of the original question data; determining the information entropy of the original question data based on the intention probability distribution; and determining the intention uncertainty level of the user question data based on the relationship between the information entropy and a preset entropy value.

[0091] The intent probability distribution includes the probability that the original question data belongs to each user intent. The agent or intent recognition model outputs all matching candidate business intents based on its knowledge; these candidate intents represent possible user intents. The probability of a user intent represents the probability that the intent belongs to the user's true intent, and the sum of the probabilities of all user intents is 1.

[0092] Information entropy is calculated based on the intent probability distribution, where information entropy... The calculation method is as follows:

[0093] in which, represents the probability of the i-th user intention, and n is the total number of identified user intentions.

[0094] The information entropy represents the degree of uncertainty of the user intention. When the information entropy is greater than the preset entropy value, it indicates that the degree of uncertainty of the user intention is high, that is, the degree to which the user intention cannot be uniquely determined is high, so the intention uncertainty level can be defined as a high uncertainty level; when the information entropy is less than the preset entropy value, it indicates that the degree of uncertainty of the user intention is low, so the intention uncertainty level can be defined as a low uncertainty level. The case where the information entropy is equal to the preset entropy value can be classified into the case where the information entropy is greater than the preset entropy value or the case where the information entropy is less than the preset entropy value according to actual requirements. For subsequent numerical judgments involved, this manner can be adopted unless otherwise specified.

[0095] In an embodiment, the preset entropy values include a first entropy value and a second entropy value, and the first entropy value is less than the second entropy value. The judgment manner of the intention uncertainty level is: when the information entropy is less than or equal to the first entropy value (for example, H(X)≤0.3), it is judged as a low uncertainty level, this case can also be called low entropy certainty, that is, the lower the entropy, the higher the certainty of the user intention; when the information entropy is greater than the first entropy value and less than the second entropy value (for example, 0.3<H(X)<0.8), it is judged as a medium uncertainty level, this case can also be called medium entropy uncertainty; when the information entropy is greater than or equal to the second entropy value (for example, H(X)≥0.8), it is judged as a high uncertainty level, this case can also be called high entropy uncertainty. It can be seen that the degrees to which the user intentions corresponding to the low uncertainty level, the medium uncertainty level and the high uncertainty level cannot be uniquely determined increase sequentially.

[0096] Step 420: Perform data completion processing on the user question data according to the intention uncertainty level to obtain target question data.

[0097] Perform graded and differentiated data completion processing on the original user question data according to the intention uncertainty level, wherein the data completion processing includes supplementing missing information in the original user question data, correcting ambiguities in the original user question, and the like.

[0098] In an embodiment, the specific process of step 420 includes: if the intention uncertainty level is the low uncertainty level, taking the user question data as the target question data; if the intention uncertainty level is the medium uncertainty level, performing data completion processing on the user question data according to the interaction cost and reasoning risk cost corresponding to the user question data to obtain the target question data; if the intention uncertainty level is the high uncertainty level, performing data completion processing on the user question data through an interaction operation to obtain the target question data.

[0099] Specifically, for user question data with low uncertainty levels, the task data is relatively complete with few omissions and no additional supplementation is needed. The original user question data can be directly used as the target question data to ensure efficient task processing. When the agent infers from the user question data to generate response data, it can supplement the small amount of missing data based on its own rules.

[0100] For user question data with medium uncertainty, the appropriate completion method is determined based on interaction cost and inference risk cost. Interaction cost includes the cost of completing the question data through interactive operations, which include asking the user questions to obtain the supplementary data. Inference risk cost includes the risk cost of generating incorrect response data based on the user question data, representing the decision risk that still exists after inference is completed.

[0101] In one embodiment, the specific process of data completion based on interaction cost and inference risk cost includes: generating interaction cost based on the amount of missing information in user question data; calculating inference information entropy based on the posterior probabilities of multiple user intents obtained from intent recognition, and generating inference risk cost based on the inference information entropy; if the interaction cost is greater than the inference risk cost, then the user question data is used as the target question data; if the interaction cost is less than the inference risk cost, then the user question data is completed through interactive operations to obtain the target question data.

[0102] Calculate the interaction cost and the inference risk cost separately. If the interaction cost is greater than the inference risk cost, it means that supplementing information through human-computer interaction requires a large number of interactions, time costs, and user operation costs, resulting in extremely high overall interaction costs. In this case, instead of supplementing data based on interactive operations, the agent can directly use the user's question data as the target question data for subsequent inference, allowing it to automatically supplement data and generate response data based on its own experience and accumulated knowledge during the inference process. If the interaction cost is less than the inference risk cost, it indicates a high risk of incorrect inference results. In this case, the agent can ask the user questions through interactive operations to supplement the question data and obtain the target question data.

[0103] In this embodiment, the interaction cost is generated based on the amount of missing information in the user's question data, and the interaction cost is positively correlated with the amount of missing information. For example, the interaction cost... The calculation method is as follows:

[0104] in, The cost coefficient for a single question, for example . This indicates the number of missing information items in the user issue data.

[0105] Inference risk cost is based on inference information entropy, which characterizes the entropy value of residual uncertainty after inference, and also represents the uncertainty of the inference result. For example, inference risk cost... The calculation method is as follows:

[0106]

[0107] in, Given user question data X, the posterior probability of a user intent h is calculated based on historical intent recognition data. The posterior probability of a user intent is the probability that the user intent belongs to a true user intent. For example, if a user intent is identified as a true user intent 30% of the time in historical intent recognition data, then the posterior probability of that user intent is 30%. The smaller the value, the more unique the reasoning result and the lower the risk. The larger the value, the more divergent the inference results and the higher the risk of error. Risk weighting coefficients corresponding to business scenarios, such as financial scenarios. Reasoning about risk costs This refers to the business risk cost incurred by an intelligent agent forcibly inferring and outputting incorrect response data based on current user question data.

[0108] For user issue data with a high level of uncertainty, indicating severe data gaps, data can be completed directly through interactive operations.

[0109] In one embodiment, data completion is performed through interactive operations, including integrating user data obtained from multiple rounds of interactive questioning to generate complete target question data.

[0110] This embodiment quantifies interaction costs by measuring the amount of missing information and inference risks by measuring the entropy of intent information, thus achieving precise quantification of problem ambiguity and processing costs. Based on cost trade-offs, it adaptively selects either "direct inference" or "interactive completion" strategies, avoiding both resource waste and user experience degradation caused by high interaction costs, and mitigating response bias and demand mismatch issues caused by high inference risks. It dynamically balances processing efficiency and response accuracy, significantly reducing the defect rate of subsequent multi-agent responses from the input source, and significantly improving the adaptability, stability, and accuracy of solving complex fuzzy problems.

[0111] It should be noted that in the process of completing the problem data through interactive operation, the intention uncertainty level of the completed problem needs to be re-evaluated in step 410 based on the completed problem data until the target intention data with low entropy certainty is finally obtained, or the interaction cost is higher than the reasoning risk cost when the problem data with medium entropy uncertainty is obtained, and then proceed to the next step 430.

[0112] Step 430: Obtain the initial response data generated by multiple agents for the target question data.

[0113] Step 440: For each agent, identify the defects in the agent's initial response data through other agents among multiple agents, and obtain candidate response data obtained by the agent adjusting the initial response data based on the defects.

[0114] Step 450: Determine the target response data from the candidate response data corresponding to multiple agents.

[0115] Steps 430-450 can refer to the relevant descriptions in steps 210-230 above, and will not be repeated here.

[0116] The technical solution provided in this application embodiment performs data completion processing on the user question data based on the uncertainty level of the user question data's intent before the agent generates response data, so as to obtain the target question data with the lowest overall cost. This is equivalent to optimizing the input data of the agent, reducing the user input cost, solving the pain points of incomplete requirements and a lot of implicit knowledge, thereby improving the accuracy of the agent's input data, and further improving the accuracy of the agent's data processing.

[0117] The following is a specific example illustrating the implementation process of the technical solution of this application.

[0118] For example, Figure 5 A flowchart illustrating a data processing method for an intelligent agent according to an embodiment of this application is shown schematically, such as... Figure 5 As shown, the method includes the following procedures: S1. Input, Preprocessing, Intent Mining. The AI ​​agent first outputs all matching business candidate intents based on its knowledge. The model then uses softmax normalization to output the raw predicted probabilities of each business candidate intent, ensuring that the sum of all probabilities is 1. To avoid the occurrence of a probability of 0. Meaningless computational errors, smoothing constraints are applied to extremely low probabilities: if Injunction This ensures stable calculations.

[0119] S2. Information Entropy Uncertainty Quantification. Iterate through all intent probabilities, substitute them into the information entropy formula, and sum them term by term to finally obtain the true uncertain entropy value of the current user need (i.e., user question data). .

[0120] S3. Entropy-Adaptive Decision Rule. Based on the entropy value and the game outcome, a three-level adaptive strategy is established, completely replacing the traditional manual threshold: Low-entropy determinacy (H(X)≤0.3): The demand ambiguity is extremely low, the cost after the game approaches 0, and the complete reasoned demand is directly output; Medium-entropy uncertainty (0.3<H(X)<0.8): Risk and cost are balanced, the minimum cost is solved through game theory, and missing parameters are pertinently queried; High-entropy uncertainty (H(X)≥0.8): The residual entropy of reasoning is extremely high and the risk is uncontrollable, blind reasoning is frozen, and core information is collected uniformly.

[0121] S4. Low-entropy determinacy: Direct reasoning is performed based on user question data to obtain initial response data generated by the agent.

[0122] S5. High-entropy uncertainty: The user is directly asked a question to complete the question data, and the intent uncertainty is re-evaluated.

[0123] S6. Medium-entropy uncertainty: The operation to be performed is determined based on the interaction cost and the reasoning risk cost. If the interaction cost is higher than the reasoning risk cost, that is, the residual risk entropy after reasoning is relatively high, the user is asked a question to complete the question data, and the intent uncertainty is re-evaluated. If the interaction cost is lower than the reasoning risk cost, that is, the residual risk entropy after reasoning is relatively low, reasoning is directly performed to obtain initial response data generated by the agent.

[0124] For example, user question data: help me design a small-sum loan product; fixed algorithm parameters: single query cost α=0.1, risk weight coefficient β=1. The process is executed as follows: Step 1: Preprocess the user question data. Clean invalid semantics, standardize the demand as "develop a small-sum loan financial product", splice historical business data, and lock the scenario suitable for short-term small-sum credit.

[0125] Step 2: Quantify the information entropy uncertainty.

[0126] The current demand can correspond to multiple product directions: consumer loan, working capital loan, large-amount small-sum loan, long-term short-term product, and the intent distribution is scattered.

[0127] Combined with the above implementation steps, the actual entropy value of the current demand is calculated: the 4 types of intent probabilities output by the model are: 0.35 for small-sum working capital loan, 0.33 for consumer installment loan, 0.18 for business loan, and 0.14 for emergency loan, with the sum of probabilities being 1. Substitute into the formula for calculation:

[0128] The preset entropy values are set to 0.3 and 0.8, and it is finally determined that the user intent is of high-entropy uncertainty.

[0129] Step 3: Generate demand assumptions through abductive deduction.

[0130] The algorithm generates two sets of potential product hypotheses based on the business scenario: h1: General-purpose small-amount consumer loans for ordinary users; h2: Short-term working capital loans for office workers.

[0131] Step 4: Risk and cost game.

[0132] After normalizing the probabilities of the two sets of hypotheses following the inference: general consumer loans 0.52, and working capital loans for salaried workers 0.48, the probability distributions are extremely close, while the inference results diverge. Substituting into the residual entropy formula for calculation:

[0133] 1. Reasoning risk cost calculation:

[0134] 2. Question interaction cost calculation: Currently, four core pieces of information are missing: audience, credit limit, term, and risk control.

[0135] Questioning Cost Much smaller than the risk cost .

[0136] Step 5: Ask a question with the lowest cost.

[0137] AI merges multiple missing pieces of information and completes information collection with a single question (minimizing costs): Please tell me the target user group, loan amount range, loan term, and risk control requirements for this microloan product, and I will generate complete and feasible product design requirements for you.

[0138] Step 6: User completes information and iterates again.

[0139] The user added: "Targeting office workers aged 22-35, with loan amounts of 500-5000 yuan, terms of 7-90 days, flexible borrowing and repayment, purely online risk control, and no collateral required." Secondary entropy calculation: Demand is uniquely determined. (Low entropy, no uncertainty), reasoning about residual entropy The risk is extremely low, and the cost of the game is close to zero.

[0140] Step 7: Output the complete request, that is, output the target problem data.

[0141] Develop microloan financial products with the following requirements: targeting office workers aged 22-35, with loan amounts of 500-5000 yuan, terms of 7-90 days, flexible repayment options, and purely online risk control with no collateral required.

[0142] It should be noted that although the steps of the method in this application are described in a specific order in the accompanying drawings, this does not require or imply that the steps must be performed in that specific order, or that all the steps shown must be performed to achieve the desired result. Additional or alternative steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0143] The following describes an embodiment of the apparatus of this application, which can be used to execute the data processing method of the intelligent agent in the above embodiments of this application. Figure 6 A schematic block diagram of the data processing apparatus for an intelligent agent provided in an embodiment of this application is shown. Figure 6 As shown, the data processing device for an intelligent agent provided in this application embodiment includes: Initial response module 610 is used to acquire initial response data generated by multiple agents in response to the target question data; The adjustment module 620 is used to identify defects in the initial response data of each agent through other agents among the plurality of agents, and to obtain candidate response data obtained by the agent adjusting the initial response data based on the defects. The determining module 630 is used to determine the target response data from the candidate response data corresponding to the multiple intelligent agents respectively.

[0144] In one embodiment of this application, the determining module 630 is specifically used for: For each agent, multiple review agents question the candidate response data of the agent and obtain the optimized response data obtained by the agent based on the questioning results; the questioning includes raising questions and verifying defects, and one review agent corresponds to one questioning dimension; The target response data is determined from the optimized response data corresponding to the multiple intelligent agents.

[0145] In one embodiment of this application, the determining module 630 is specifically used for: If the query result of any of the review agents indicates that the optimized response data has defects that need to be optimized, then the optimized response data is used as the candidate response data, and the step of querying the candidate response data of the agents by multiple review agents is returned. If the query results of multiple review agents all indicate that the optimized response data does not have any defects to be optimized, or if the number of iterations reaches a preset number, then the optimized response data is taken as the optimized response data obtained by the agent.

[0146] In one embodiment of this application, the determining module 630 is specifically used for: For each agent, the optimized response data of the agent is scored by multiple review agents, and a comprehensive score of the optimized response data is obtained based on the scores of the multiple review agents. The optimized response data with the highest overall score will be used as the target response data.

[0147] In one embodiment of this application, the review agent includes a compliance agent, which is used to perform compliance detection; the determination module 630 is specifically used for: If the score of the compliant intelligent agent for the optimized response data of the intelligent agent is a set value, then the comprehensive score of the optimized response data is set to the set value; If the score given by the compliance agent to the optimized response data is not a set value, then the scores of other review agents are weighted and summed to obtain the comprehensive score of the optimized response data.

[0148] In one embodiment of this application, the apparatus further includes a data completion module, used for: Intent identification is performed on user question data to determine the intent uncertainty level of the user question data; the intent uncertainty level characterizes the degree to which the user intent of the user question data cannot be uniquely determined. The user question data is supplemented based on the level of intent uncertainty to obtain the target question data.

[0149] In one embodiment of this application, the data completion module is specifically used for: The original question data is subjected to intent recognition to obtain the intent probability distribution of the original question data; the intent probability distribution includes the probability that the original question data belongs to each user intent; The information entropy of the original question data is determined by the probability distribution of the intent. The level of intent uncertainty of the user question data is determined based on the relationship between the information entropy and the preset entropy value.

[0150] In one embodiment of this application, the data completion module is specifically used for: If the uncertainty level of the intent is low, then the user question data will be used as the target question data; If the uncertainty level of the intent is medium uncertainty, then the user question data is supplemented according to the interaction cost and inference risk cost corresponding to the user question data to obtain the target question data; the interaction cost includes the cost required to supplement the question data through interactive operation, and the inference risk cost includes the risk cost of generating incorrect response data based on the user question data; If the uncertainty level of the intent is high, then the user question data is completed through interactive operation to obtain the target question data.

[0151] In one embodiment of this application, the data completion module is specifically used for: The interaction cost is generated based on the number of missing information in the user question data, wherein the interaction cost is positively correlated with the number of missing information. Based on the posterior probabilities of multiple user intents obtained from intent recognition, the inference information entropy is calculated, and the inference risk cost is generated based on the inference information entropy; wherein, the posterior probability of the user intent is the probability that the user intent belongs to a real intended purpose, calculated based on historical intent recognition data. If the interaction cost is greater than the reasoning risk cost, then the user question data will be used as the target question data. If the interaction cost is less than the reasoning risk cost, then the user question data is completed through interactive operations to obtain the target question data.

[0152] The specific details of the data processing device for the intelligent agent provided in the various embodiments of this application have been described in detail in the corresponding method embodiments, and will not be repeated here.

[0153] Figure 7 A schematic block diagram of a computer system architecture for implementing an electronic device according to embodiments of the present application is shown.

[0154] It should be noted that, Figure 7 The computer system 700 of the electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0155] like Figure 7As shown, the computer system 700 includes a central processing unit (CPU) 701, which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) 702 or programs loaded from storage section 708 into random access memory (RAM) 703. The RAM 703 also stores various programs and data required for system operation. The CPU 701, ROM 702, and RAM 703 are interconnected via a bus 704. An input / output interface 705 (I / O interface) is also connected to the bus 704.

[0156] The following components are connected to the input / output interface 705: an input section 706 including a keyboard, mouse, etc.; an output section 707 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 708 including a hard disk, etc.; and a communication section 709 including a network interface card such as a local area network card, modem, etc. The communication section 709 performs communication processing via a network such as the Internet. A drive 710 is also connected to the input / output interface 705 as needed. A removable medium 711, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 710 as needed so that computer programs read from it can be installed into the storage section 708 as needed.

[0157] Specifically, according to embodiments of this application, the processes described in the various method flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 709, and / or installed from removable medium 711. When the computer program is executed by central processing unit 701, it performs various functions defined in the system of this application.

[0158] It should be noted that the computer-readable medium shown in the embodiments of this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a 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. In this application, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such transmitted data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wired, etc., or any suitable combination thereof.

[0159] 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.

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

[0161] Through the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, touch terminal, or network device, etc.) to execute the methods according to the embodiments of this application.

[0162] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

[0163] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A data processing method for an intelligent agent, characterized in that, include: Acquire initial response data generated by multiple agents in response to the target problem data; For each agent, defects in the initial response data of the agent are identified by other agents among the plurality of agents, and candidate response data is obtained by the agent adjusting the initial response data based on the defects. The target response data is determined from the candidate response data corresponding to the multiple intelligent agents.

2. The data processing method for an intelligent agent according to claim 1, characterized in that, Determining the target response data from the candidate response data corresponding to the multiple intelligent agents includes: For each agent, multiple review agents question the candidate response data of the agent and obtain the optimized response data obtained by the agent based on the questioning results; the questioning includes raising questions and verifying defects, and one review agent corresponds to one questioning dimension; The target response data is determined from the optimized response data corresponding to the multiple intelligent agents.

3. The data processing method for an intelligent agent according to claim 2, characterized in that, Multiple review agents challenge the candidate response data of the agent, and the optimized response data obtained by the agent based on the challenge results is obtained, including: If the query result of any of the review agents indicates that the optimized response data has defects that need to be optimized, then the optimized response data is used as the candidate response data, and the step of querying the candidate response data of the agents by multiple review agents is returned. If the query results of multiple review agents all indicate that the optimized response data does not have any defects to be optimized, or if the number of iterations reaches a preset number, then the optimized response data is taken as the optimized response data obtained by the agent.

4. The data processing method for an intelligent agent according to claim 2, characterized in that, Determining the target response data from the optimized response data corresponding to each of the multiple intelligent agents includes: For each agent, the optimized response data of the agent is scored by multiple review agents, and a comprehensive score of the optimized response data is obtained based on the scores of the multiple review agents. The optimized response data with the highest overall score will be used as the target response data.

5. The data processing method for an intelligent agent according to claim 4, characterized in that, The review agent includes a compliance agent, which performs compliance checks; a comprehensive score for the optimized response data is obtained based on the scores corresponding to multiple review agents, including: If the score of the compliant intelligent agent for the optimized response data of the intelligent agent is a set value, then the comprehensive score of the optimized response data is set to the set value; If the score given by the compliance agent to the optimized response data is not a set value, then the scores of other review agents are weighted and summed to obtain the comprehensive score of the optimized response data.

6. The data processing method for an intelligent agent according to claim 1, characterized in that, Before acquiring the initial response data generated by multiple agents for the target question data, the method further includes: Intent identification is performed on user question data to determine the intent uncertainty level of the user question data; the intent uncertainty level characterizes the degree to which the user intent of the user question data cannot be uniquely determined. The user question data is supplemented based on the level of intent uncertainty to obtain the target question data.

7. The data processing method for an intelligent agent according to claim 6, characterized in that, Perform intent identification on user question data to determine the level of intent uncertainty of the user question data, including: The original question data is subjected to intent recognition to obtain the intent probability distribution of the original question data; the intent probability distribution includes the probability that the original question data belongs to each user intent; The information entropy of the original question data is determined by the probability distribution of the intent. The level of intent uncertainty of the user question data is determined based on the relationship between the information entropy and the preset entropy value.

8. The data processing method for an intelligent agent according to claim 6, characterized in that, Based on the stated level of intent uncertainty, the user question data is augmented to obtain the target question data, including: If the uncertainty level of the intent is low, then the user question data will be used as the target question data; If the uncertainty level of the intent is medium uncertainty, then the user question data is processed to complete the data based on the interaction cost and inference risk cost corresponding to the user question data to obtain the target question data; the interaction cost includes the cost required to complete the question data through interactive operations, and the inference risk cost includes the risk cost of generating incorrect response data based on the user question data; If the uncertainty level of the intent is high, then the user question data is completed through interactive operation to obtain the target question data.

9. The data processing method for an intelligent agent according to claim 8, characterized in that, Based on the interaction cost and inference risk cost corresponding to the user question data, data completion processing is performed on the user question data to obtain the target question data, including: The interaction cost is generated based on the number of missing information in the user question data, wherein the interaction cost is positively correlated with the number of missing information. Based on the posterior probabilities of multiple user intents obtained from intent recognition, the inference information entropy is calculated, and the inference risk cost is generated based on the inference information entropy; wherein, the posterior probability of the user intent is the probability that the user intent belongs to a real intended purpose, calculated based on historical intent recognition data. If the interaction cost is greater than the reasoning risk cost, then the user question data will be used as the target question data. If the interaction cost is less than the reasoning risk cost, then the user question data is completed through interactive operations to obtain the target question data.

10. An electronic device, characterized in that, include: processor; as well as Memory for storing the executable instructions of the processor; The processor is configured to execute the executable instructions to enable the electronic device to implement the data processing method of the intelligent agent as described in any one of claims 1 to 9.