Data processing method based on multi-agent collaboration and related device

Abstract

The invention provides a data processing method based on multi-agent collaboration and a related device, and the method comprises the steps: a parent agent calculates the decision weight of each sub-agent according to user question data, and distributes a data processing task and a task execution rule to each sub-agent according to the decision weight, the child agents execute the data processing task according to the task execution rule and return data processing results to the parent agent, and the parent agent fuses the data processing results of all the child agents to obtain target reply content. Thus, the matching degree of the sub-agents and the user demands is quantified through the decision weight, the limitation that tasks are distributed only according to functions in the prior art is broken through, the matching precision of the tasks and the agent capabilities is improved, meanwhile, differentiated task execution rules are configured for the sub-agents based on the decision weight, and the task execution efficiency is improved. And the overall processing efficiency and reliability of multi-agent collaboration are improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, specifically to a data processing method and related apparatus based on multi-agent collaboration. Background Technology

[0002] An agent is an intelligent entity that can autonomously perceive its environment, make decisions, and execute tasks. Limited by the functional boundaries and knowledge reserves of a single agent, data processing mechanisms based on multi-agent collaboration have become the mainstream technical approach in complex scenarios. Especially in intelligent dialogue scenarios in the financial field, this mechanism can solve domain-specific problems caused by massive heterogeneous data and complex terminology to a certain extent.

[0003] However, under the existing mechanism, the task allocation of the parent agent to the child agent relies on simple functional matching, that is, the task is allocated according to the function of the child agent. There is a lack of quantitative evaluation of the degree of matching between the child agent and the user's needs, which leads to inaccurate task allocation and affects the overall processing efficiency and reliability of multi-agent collaboration. Summary of the Invention

[0004] This application provides a data processing method and related apparatus based on multi-agent collaboration, aiming to improve the overall processing efficiency and reliability of multi-agent collaboration by quantifying the degree of matching between sub-agents and user needs, thereby achieving precise configuration of data processing tasks and task execution rules.

[0005] In a first aspect, embodiments of this application provide a data processing method based on multi-agent collaboration, applied to an electronic device, the electronic device including a parent agent and multiple child agents, the method comprising: The parent agent calculates the decision weight of each child agent based on the user's question data. The decision weight is used to characterize the degree of matching between the child agent and the user's needs. The parent agent assigns data processing tasks and task execution rules to each child agent according to the decision weights. The task execution rules are used to indicate the constraints on the child agents when performing data processing tasks. The multiple sub-agents execute the data processing task according to the task execution rules and return the data processing result to the parent agent; The parent agent fuses the data processing results of each child agent to obtain the target response content.

[0006] Secondly, embodiments of this application provide a data processing device based on multi-agent collaboration, applied to an electronic device, the electronic device including a parent agent and multiple child agents, the device comprising: The computing unit is used to enable the parent agent to calculate the decision weight of each child agent based on the user's query data. The decision weight is used to characterize the degree of matching between the child agent and the user's needs. The allocation unit is used to enable the parent agent to allocate data processing tasks and task execution rules to each child agent according to the decision weights, wherein the task execution rules are used to indicate the constraints on the child agents when performing data processing tasks. An execution unit is configured to cause the plurality of sub-intelligent agents to execute the data processing task according to the task execution rules, and return the data processing result to the parent intelligent agent; The fusion unit is used to fuse the data processing results of each child agent by the parent agent to obtain the target response content.

[0007] Thirdly, embodiments of this application provide an electronic device including a processor, a memory, and one or more programs, the one or more programs being stored in the memory and configured to be executed by the processor, the programs including instructions for performing steps in the method as described in the first aspect of this application.

[0008] Fourthly, embodiments of this application provide a computer-readable storage medium having a computer program or instructions stored thereon, wherein the computer program or instructions, when executed by a processor, implement the steps of the method described in the first aspect of this application.

[0009] As can be seen from the embodiments of this application, the parent agent calculates the decision weight of each child agent based on the user's query data to quantify the degree of matching between the child agent and the user's needs. Then, based on the decision weight, it assigns data processing tasks and task execution rules to each child agent. The task execution rules are used to indicate the constraints on the child agent when performing the data processing tasks. The child agent executes the data processing tasks according to the task execution rules and returns the data processing results to the parent agent. The parent agent merges the data processing results of each child agent to obtain the target response content. In this way, by quantifying the degree of matching between the child agent and the user's needs through decision weight, it overcomes the limitation of assigning tasks only according to functions in the prior art, improves the matching accuracy between tasks and agent capabilities, and at the same time, configures differentiated task execution rules to child agents based on decision weight, thereby improving the overall processing efficiency and reliability of multi-agent collaboration. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is an architectural block diagram of an intelligent dialogue system provided in an embodiment of this application; Figure 2 This is an architectural block diagram of another intelligent dialogue system provided in the embodiments of this application; Figure 3 This is a block diagram of an intelligent agent deployment for an electronic device provided in an embodiment of this application; Figure 4 This is a flowchart illustrating a data processing method based on multi-agent collaboration provided in an embodiment of this application; Figure 5 This is a structural block diagram of a data processing device based on multi-agent collaboration provided in an embodiment of this application; Figure 6 This is a structural block diagram of another data processing device based on multi-agent collaboration provided in the embodiments of this application; Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0012] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0013] The terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0014] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0015] In related technologies, data processing mechanisms based on multi-agent collaboration typically involve a parent agent (or master agent) performing semantic understanding and intent recognition, then assigning data processing tasks to each sub-agent according to their core functions. For example, instructing a data retrieval agent to retrieve target data from a target data source, and / or instructing a document parsing agent to read and analyze document content. However, when faced with complex requirements, the parent agent only assigns data processing tasks based on function, without considering the matching degree between user needs and each sub-agent. Under the constraint of computing resources, sub-agents with high matching degrees may not be able to explore deeply, while sub-agents with low matching degrees may undertake redundant work, affecting the overall processing efficiency and reliability of multi-agent collaboration. Especially in intelligent dialogue scenarios in the financial field, the user needs implied in user questions often place extremely stringent requirements on the accuracy, reliability, and response efficiency of the data. Existing data processing mechanisms based on multi-agent collaboration cannot effectively support intelligent dialogue scenarios in the financial field.

[0016] To address the aforementioned issues, embodiments of this application provide a data processing method and related apparatus based on multi-agent collaboration.

[0017] The system architecture involved in the embodiments of this application is described below.

[0018] In one embodiment, such as Figure 1 As shown, the intelligent dialogue system 10 includes a first electronic device 11 and a second electronic device 12, which are communicatively connected. The first electronic device 11 acts as a client, and the second electronic device 12 acts as a server. A user performs a question operation on the first electronic device 11, which converts the question operation into user question data and sends it to the second electronic device 12. The second electronic device 12 then executes the multi-agent collaborative data processing method as described in the embodiments of this application.

[0019] In another embodiment, such as Figure 2 As shown, the intelligent dialogue system 10 includes only a first electronic device 11. The user performs a question operation on the first electronic device 11, and the first electronic device 11 converts the question operation into user question data and executes the data processing method based on multi-agent collaboration as described in the embodiments of this application.

[0020] Furthermore, such as Figure 3As shown, the electronic device 30 includes a parent agent 31 and multiple child agents 32. Both the parent agent 31 and the child agents 32 are functional modules or computing entities running on the electronic device 30, collectively forming a multi-agent collaborative processing architecture. The parent agent 31, as the core management unit of the collaborative architecture, is responsible for coordinating the entire data processing process, including receiving user query data, calculating the decision weights of each child agent 32, allocating data processing tasks and task execution rules, and fusing the processing results returned by the child agents 32. The multiple child agents 32 are functional execution units, each possessing specific data processing capabilities. Under the scheduling of the parent agent 31, they must complete their corresponding data processing work according to the allocated tasks and execution rules, and feed the results back to the parent agent 31. In this embodiment, the child agents 32 can be implemented as data retrieval agents, document analysis agents, and knowledge question-answering agents. The data retrieval agent is used to retrieve target fields from target data sources, the document analysis agent is used to read and analyze the content of target documents, extracting viewpoints or conclusions, and the knowledge question-answering agent is used to search for explanations or associations of target terms from a knowledge graph.

[0021] Specifically, the electronic device 30 may be either the first electronic device 11 or the second electronic device 12.

[0022] The following describes a data processing method based on multi-agent collaboration provided by an embodiment of this application.

[0023] Please see Figure 4 , Figure 4 This is a flowchart illustrating a data processing method based on multi-agent collaboration provided in an embodiment of this application. Figure 4 As shown, the method includes: S401, the parent agent calculates the decision weight of each child agent based on the user's question data.

[0024] The user-generated question data includes text data and behavioral data. The text data refers to the text content entered by the user, and the behavioral data refers to other user actions besides text input, such as document upload, image upload, and function selection. This is not a single, definitive definition.

[0025] The decision weights are used to characterize the degree of matching between the sub-agents and the user's needs. Specifically, the parent agent calculates the decision weights of each sub-agent by quantifying the degree of matching between each sub-agent and the user's needs. For example, if the user's query operations include uploading a document and inputting the text "summarize the main points of this document," then the decision weights of each sub-agent can be calculated based on quantitative evaluation methods after intent recognition. It is understandable that in this example, compared to other sub-agents, the document parsing agent has a higher degree of matching with the user's needs, i.e., a higher corresponding decision weight.

[0026] S402, the parent agent assigns data processing tasks and task execution rules to each child agent according to the decision weights.

[0027] In this context, data processing tasks refer to the specific work assigned by the parent agent to the child agent based on decision weights and user needs, tailored to the child agent's capabilities. For example, an instruction data retrieval agent might retrieve target fields from a target data source and limit the depth and breadth of its retrieval through decision weights, thereby balancing the efficiency and resources of a single child agent when performing data processing tasks.

[0028] The task execution rules are used to indicate the constraints on sub-agents when performing data processing tasks. For example, they constrain the system resources that sub-agents can schedule when performing data processing tasks. The degree of constraint depends on the decision weights, thereby configuring differentiated task execution rules for each sub-agent and improving the overall processing efficiency and reliability of multi-agent collaboration.

[0029] S403, the plurality of sub-intelligent agents execute the data processing task according to the task execution rules, and return the data processing result to the parent intelligent agent.

[0030] S404, the parent agent fuses the data processing results of each child agent to obtain the target response content.

[0031] The parent agent performs correlation analysis on the data processing results of different child agents, links logically complementary data processing results, removes or corrects duplicate or conflicting results, and organizes the integrated information into text content that conforms to natural language expression habits and meets user needs, thus obtaining the target response content. Furthermore, the parent agent can adaptively adjust the presentation of data processing results in the target response content based on decision weights to accurately match the user's direct needs.

[0032] As can be seen, in this embodiment, the parent agent calculates the decision weight of each child agent based on the user's query data to quantify the degree of matching between the child agent and the user's needs. Then, based on the decision weight, it assigns data processing tasks and task execution rules to each child agent. The task execution rules indicate the constraints on the child agent when performing the data processing tasks. The child agents execute the data processing tasks according to the task execution rules and return the data processing results to the parent agent. The parent agent merges the data processing results of each child agent to obtain the target response. Thus, by quantifying the degree of matching between the child agent and the user's needs through decision weights, it overcomes the limitation of assigning tasks solely based on function in existing technologies, improves the matching accuracy between tasks and agent capabilities, and enhances the overall processing efficiency and reliability of multi-agent collaboration by configuring differentiated task execution rules for the child agents based on decision weights.

[0033] In one possible example, the parent agent calculates the decision weight of each child agent based on user question data, including: the parent agent performs word segmentation on the user question data to obtain multiple keywords; the parent agent extracts features from the user question data to obtain multiple demand features; the parent agent determines the child agent and bonus value corresponding to each demand feature based on preset feature matching rules, the feature matching rules including the correspondence between demand features, associated child agents, and bonus values; for each child agent, the multiple keywords are matched with the keyword library corresponding to the child agent to calculate the first matching score of the child agent; and, based on the sum of the basic score and bonus value of the child agent, a second matching score of the child agent is calculated; and, based on the first matching score and the second matching score, the decision weight of the child agent is calculated.

[0034] In this example, the parent agent calculates the child agent's first matching score through keyword matching and the child agent's second matching score through feature matching, thereby quantifying the degree of matching between the child agent and the user's needs and calculating the decision weight.

[0035] Specifically, in terms of keyword matching, the parent agent calls a pre-trained word segmentation model to segment the user's query data, obtaining multiple keywords with independent semantics. The word segmentation model is trained on financial corpora, which avoids splitting specialized financial terms and improves segmentation accuracy. Then, keyword matching is performed on each sub-agent, matching multiple keywords with the sub-agent's corresponding keyword library and calculating a first matching score. Each sub-agent is configured with a keyword library corresponding to its function. For example, the keyword library for the data retrieval agent includes indicator field keywords such as "net profit," "profit margin," and "asset-liability ratio"; the keyword library for the document analysis agent includes document and analysis keywords such as "document," "document," "annual report," "annual report," "interpretation," and "chapter"; and the keyword library for the knowledge question answering agent includes knowledge query keywords such as "definition," "meaning," "terminology," and "explanation." Furthermore, when calculating the first matching score, the frequency of words that precisely match the keyword library of the sub-agent is first counted. For keywords that do not precisely match, the semantic similarity between each keyword and the keyword library of the sub-agent is calculated through a pre-trained word vector model. Finally, the word frequency and semantic similarity are weighted, summed, and normalized to obtain the first matching score corresponding to the sub-agent.

[0036] Specifically, in terms of feature matching, the parent agent calls a pre-trained feature extraction model to extract features from the user's query data, obtaining multiple requirement features. The feature extraction model is trained based on a preset rule template, resulting in output requirement features that include four categories: interaction behavior features, time element features, and terminology density features. Interaction behavior features characterize user actions, such as document uploads, image uploads, and clicks; time element features characterize time markers in the user's input text, such as "latest," "current," and "third quarter of 2024"; terminology density features characterize the frequency of financial terminology in the user's input text. Then, based on preset feature matching rules, the corresponding sub-agent and its score are determined for each requirement feature. For example, if document upload is present in the interaction behavior features, the associated sub-agent is identified as a document parsing agent, with a first preset score, which can be further differentiated based on the type of uploaded document. Similarly, if the text content contains at least three terms, the associated sub-agent is identified as a knowledge-based question-answering agent, with a second preset score, which can be further differentiated based on the actual terminology density. Furthermore, a base score is set for each sub-agent, and after adding it to the sum of the bonus scores, the result is normalized to obtain the second matching score.

[0037] Finally, the first and second matching scores are dynamically weighted and fused to obtain the decision weight. Specifically, the decision weight = α × first matching score + (1-α) × second matching score, where α is the fusion coefficient, which can be dynamically adjusted based on the task type complexity of the user query data. For example, for simple tasks, such as pure data queries, α = 0.7, meaning the decision weight is determined primarily based on keyword matching; for complex tasks, such as mixed query analysis combining literature and data, α = 0.3, meaning the decision weight is determined primarily based on feature matching, further improving accuracy.

[0038] As can be seen, in this example, the parent agent calculates the first matching score of each child agent through keyword matching, and the second matching score of each child agent through demand feature matching. Then, the decision weight of each child agent is calculated, and the degree of matching between each child agent and the user's needs is comprehensively evaluated. This provides data support for subsequent task allocation and rule configuration, and improves the overall processing efficiency and result reliability of multi-agent collaboration.

[0039] In one possible example, the task execution rules include exception handling rules and normal execution rules. The exception handling rules are used to indicate the constraints when a sub-agent returns a data processing result under an exception scenario. The normal execution rules are used to indicate the constraints when a sub-agent does not execute a data processing task under the exception scenario. The exception scenario refers to a scenario that affects the normal execution of the data processing task. The parent agent assigns data processing tasks and task execution rules to each sub-agent according to the decision weights, including: the parent agent assigning a first data processing task and a second data processing task to a first sub-agent, and assigning the first data processing task and a second data processing task to a second sub-agent. The data processing task is as follows: the first sub-agent refers to a sub-agent whose decision weight is not less than a first threshold; the second sub-agent refers to a sub-agent whose decision weight is less than the first threshold; the first data processing task refers to the necessary processing steps to meet user needs; the second data processing task refers to the related processing steps to assist in understanding the response content; the parent agent assigns a first normal execution rule to the first sub-agent and a second normal execution rule to the second sub-agent, wherein the constraint strength of the first normal execution rule is higher than the constraint strength of the second normal execution rule; the parent agent assigns the anomaly response rule to all sub-agents.

[0040] In this model, the parent agent assigns data processing tasks and task execution rules to each child agent based on decision weights, achieving more accurate demand matching and improving the overall processing efficiency of multi-agent collaboration. In this example, the child agent with a decision weight not less than a first threshold is the first child agent, and the child agent with a decision weight less than the first threshold is the second child agent. Compared to the second child agent, the first child agent can perform deeper exploration when executing data processing tasks and is subject to stricter rule constraints to ensure the reliability of the output results.

[0041] Specifically, in allocating data processing tasks, a first data processing task and a second data processing task are assigned to the first sub-agent. This means that the first sub-agent is instructed not only to perform the necessary processing steps to meet the user's needs, but also to perform related processing steps to aid in understanding the response content. For example, when the data retrieval agent acts as the first sub-agent, it not only needs to retrieve the indicator fields required to answer the user's question from the target data source, but can also further retrieve other data of the same type or from the same period to conduct comparative analysis of those indicator fields. Furthermore, a first data processing task is assigned to the second sub-agent, meaning that the second sub-agent is instructed to perform only the necessary processing steps to meet the user's needs, without performing redundant work unrelated to the necessary processing steps, thus achieving resource balancing based on demand matching.

[0042] In this example, the task execution rules include exception handling rules and normal execution rules, which respectively constrain the sub-agent's handling methods in abnormal and normal scenarios. Abnormal scenarios refer to situations that affect the normal execution of data processing tasks, such as data loss, data errors, or system failures. Exception handling rules instruct the constraints on the sub-agent when returning data processing results in abnormal scenarios, ensuring the reliability and traceability of the output results as much as possible while avoiding output interruption. Normal execution rules instruct the constraints on the sub-agent when it does not execute data processing tasks in abnormal scenarios, ensuring that the reliability of the sub-agent's output results is commensurate with the importance of the assigned task.

[0043] Specifically, regarding the allocation of task execution rules, a first normal execution rule is assigned to the first sub-agent, and a second normal execution rule is assigned to the second sub-agent. The constraint strength of the first normal execution rule is higher than that of the second normal execution rule. This constraint strength can be reflected in several aspects, such as resource quotas: compared to the second sub-agent, the first sub-agent can have a higher proportion of computing power quota to match the importance of the data processing tasks it undertakes; or, in terms of verification strategies, compared to the second sub-agent, the first sub-agent can have more complex verification strategies to ensure higher credibility of its output. Furthermore, anomaly response rules are assigned to all sub-agents to endow them with the ability to cope with abnormal scenarios, further improving the overall credibility of the output results.

[0044] As can be seen in this example, the parent agent assigns different data processing tasks and differentiated task execution rules to each child agent based on decision weights, in order to adapt the degree of matching between the child agents and user needs. This allows child agents whose functions are more in line with user needs to have more resources to conduct deeper exploration when performing data processing tasks under differentiated rule constraints, thereby improving the efficiency and reliability of data processing. At the same time, it empowers all child agents to cope with abnormal scenarios, further improving the overall credibility of the output results.

[0045] In one possible example, the normal execution rule includes a verification rule, which refers to the verification strategy of the sub-agent for its data processing results. The parent agent assigns a first normal execution rule to the first sub-agent and a second normal execution rule to the second sub-agent, including: the parent agent assigning a first verification rule to the first sub-agent, which instructs the first sub-agent to verify its data processing results from multiple verification dimensions; and the parent agent assigning a second verification rule to the second sub-agent, which instructs the second sub-agent to verify its data processing results from a single verification dimension. The number of verification dimensions is used to characterize the coverage of risk prevention and control for the sub-agent's data processing results.

[0046] In the absence of abnormal scenarios, the parent agent assigns differentiated normal execution rules to the first and second child agents based on decision weights to adapt to the importance of the tasks undertaken by different child agents. These rules include verification rules, which are the verification strategies used by the child agents to verify their data processing results in order to improve the credibility of their output results. The more complex the verification strategy, the higher the credibility of its output results. Therefore, it is necessary to assign relatively complex verification rules to the first child agent with higher decision weights and relatively simple verification rules to the second child agent with lower decision weights to achieve a dynamic balance between resources and effectiveness.

[0047] Specifically, in this example, the parent agent assigns a first verification rule to the first child agent, instructing the first child agent to verify its data processing results from multiple verification dimensions. Similarly, a second verification rule is assigned to the second child agent, instructing the second child agent to verify its data processing results from a single verification dimension. The number of verification dimensions characterizes the coverage of risk prevention and control for the child agent's data processing results; more verification dimensions mean more comprehensive risk prevention and control coverage. The multiple verification dimensions included in the first verification rule include, but are not limited to, data source authority verification, data logic consistency verification, and timeliness verification. This avoids contamination by non-authoritative data, contradictions within the data itself, and misleading information from expired data, comprehensively covering risks to the data processing results across multiple dimensions. The single verification dimension included in the second verification rule is a preset basic consistency verification, such as verifying only the consistency of the content source without verifying its authority, simplifying the verification strategy for non-core tasks. Furthermore, in addition to the number of verification dimensions, the first verification rule can also include multiple levels of verification depth, that is, adopt multi-level progressive verification. After completing the first-level verification, if the result meets the preset conditions, such as the numerical fluctuation exceeding the reasonable range, it will enter the second-level verification, such as cross-data source cross-verification, until the verification passes or the anomaly is confirmed, which further improves the credibility of the content output by the first sub-agent.

[0048] As can be seen, in this example, the parent agent assigns differentiated verification rules to the child agents based on decision weights to adapt to the importance of the tasks undertaken by different child agents. More complex verification strategies are used to constrain child agents with high decision weights, while simpler verification strategies are used to constrain child agents with low decision weights, thereby further improving the overall processing efficiency and reliability of multi-agent collaboration.

[0049] In one possible example, the abnormal scenario includes a data abnormal scenario, which refers to an abnormal scenario in which a sub-agent cannot obtain data from the target data source. The parent agent assigns the abnormal response rules to all sub-agents, including: the parent agent assigns data abnormal response rules to all sub-agents, and the data abnormal response rules are used to instruct the sub-agents to extract inferred data from associated data sources as data processing results. The associated data source refers to a data source that has a data association or logical association with the target data source and is in a non-abnormal state.

[0050] In abnormal scenarios, the sub-agent may fail to execute data processing tasks normally, leading to task interruption or reduced reliability of output results. Specifically, in data anomaly scenarios, the sub-agent may be unable to obtain data from the target data source. For example, an anomaly in the target data table may prevent the data retrieval agent from obtaining the target field from the table, or an anomaly in the target knowledge base may prevent the knowledge answering agent from obtaining the knowledge node from the knowledge base.

[0051] In this example, the exception handling rules assigned by the parent agent to all child agents include data exception handling rules, which instruct the child agents to extract inferred data from associated data sources as data processing results. Associated data sources refer to data sources that have data or logical association with the target data source and are in a non-abnormal state.

[0052] For example, if a user's target data is "Company A's gross profit margin in the third quarter of 2024," and the target data source, such as "Company A's financial statements for the third quarter of 2024," is inaccessible due to an anomaly, the data retrieval agent will extract and extrapolate data from related data sources as the data processing result, according to the data anomaly handling rules. For instance, the related data sources might include: Company A's financial statements for the first and second quarters of 2024, which belong to the same financial data system as the target data source for that year, have a direct data relationship, and are in normal status; Company A's financial statements for the third quarter of 2023, which are historical data from the same quarter as the target data source, have a logical relationship, reflect the quarterly gross profit margin fluctuation pattern, and are in normal status; and Company C's financial statements for the third quarter of 2024, which have an industry logical relationship with the target data source, and industry data can be used as a benchmark for gross profit margin, and are in normal status. In some embodiments, the data retrieval agent performs the following operations according to the data anomaly handling rules: First, it extracts gross profit margin data from Company A's financial statements for the first and second quarters of 2024 and calculates the average gross profit margin for the first two quarters of the year; second, it extracts gross profit margin data from Company A's financial statements for the third quarter of 2023 and, combined with historical data, obtains the range of fluctuation of Company A's third quarter compared to the average of the first two quarters; finally, it performs cross-validation with Company C's gross profit margin for the third quarter of 2024 to obtain the estimated data on Company A's gross profit margin for the third quarter of 2024, outputs the estimated data as the data processing result, and labels the data as estimated data and associated data source.

[0053] As can be seen in this example, the parent agent assigns data anomaly handling rules to all child agents, enabling the child agents to still extract inferred data from related data sources as data processing results even in data anomaly scenarios. This not only ensures the continuity of data processing tasks but also ensures the traceability of results by clearly labeling the source and logic, avoiding task interruption or ambiguity of result credibility due to missing data.

[0054] In one possible example, the parent agent fuses the data processing results of each child agent to obtain the target response content, including: in response to the absence of the abnormal scenario, calculating the output content ratio of each child agent according to the decision weight; and fusing the data processing results of each child agent according to the output content ratio to obtain the target response content.

[0055] In the absence of abnormal scenarios, the sub-agents can normally acquire data, execute data processing tasks according to normal execution rules, and return data processing results. The parent agent then fuses these results to obtain the target response. Specifically, the parent agent calculates the proportion of each sub-agent's output content based on its decision weights. Then, based on this proportion, it determines the length of each data processing result within the target response. After extracting core information and deduplicating the data processing results from each sub-agent, they are fused according to their allocated length proportions to output the target response content that conforms to natural language expression habits.

[0056] As can be seen in this example, when no abnormal scenarios exist, the parent agent calculates the proportion of each child agent's output content based on the decision weights, and then merges the data processing results of each child agent according to the proportion of output content to obtain the target response content. This ensures that the core data and key analyses most relevant to the user's needs occupy the majority of the response, reducing interference from irrelevant information and improving the overall user experience.

[0057] In one possible example, the parent agent fuses the data processing results of each child agent to obtain the target response content, including: in response to the existence of the abnormal scenario, identifying abnormal child agents based on the data processing results, wherein the abnormal child agent refers to a child agent affected by the abnormal scenario; reducing the decision weight of the abnormal child agent based on the degree of abnormal impact, and proportionally increasing the decision weight of other child agents; calculating the output content ratio of each child agent based on the adjusted decision weight; and fusing the data processing results of each child agent based on the output content ratio to obtain the target response content.

[0058] In this process, when an abnormal scenario occurs, the affected sub-agents return data processing results according to the abnormal response rules. The parent agent identifies the abnormal sub-agent based on the abnormal markers within these results. Simultaneously, as the reliability of the abnormal sub-agents' outputs decreases, the parent agent reduces the decision weight of the abnormal sub-agents based on the degree of impact of the abnormality, while proportionally increasing the decision weights of other sub-agents. The parent agent then calculates the proportion of output content based on the adjusted decision weights and merges them to obtain the target response content. This avoids unreliable inferences from the response being overly represented, reducing the interference of abnormal scenarios on the accuracy of the output results.

[0059] The degree of abnormal impact is used to characterize the degree of interference of abnormal scenarios on abnormal sub-agents. Taking a data abnormality scenario as an example, based on preset abnormality judgment rules, the degree of abnormal impact can be determined according to the proportion of abnormal data. For example, when the data is completely missing, the degree of abnormal impact is severe, and the corresponding weight adjustment range is the largest. When the data is partially missing, the degree of abnormal impact is moderate, and the corresponding weight adjustment range is relatively small.

[0060] As can be seen in this example, when there is an abnormal scenario, the parent agent reduces the decision weight of the abnormal child agent and increases the decision weight of other child agents in proportion, so as to avoid unreliable inference data accounting for too high a proportion in the response, so that the effective information of the normal child agent can be fully utilized and the fault tolerance of multi-agent collaboration can be improved.

[0061] For examples consistent with the above embodiments, please refer to... Figure 5 , Figure 5 This is a structural block diagram of a data processing device based on multi-agent collaboration provided in an embodiment of this application. The data processing device 50 based on multi-agent collaboration includes: a calculation unit 501, used to enable the parent agent to calculate the decision weight of each child agent based on user query data, wherein the decision weight is used to characterize the degree of matching between the child agent and the user's needs; an allocation unit 502, used to enable the parent agent to allocate data processing tasks and task execution rules to each child agent according to the decision weights, wherein the task execution rules are used to indicate the constraints of the child agent when executing the data processing tasks; an execution unit 503, used to enable the multiple child agents to execute the data processing tasks according to the task execution rules and return the data processing results to the parent agent; and a fusion unit 504, used to enable the parent agent to fuse the data processing results of each child agent to obtain the target response content.

[0062] In one possible example, regarding the calculation of the decision weight of each child agent by the parent agent based on the user question data, the calculation unit 501 is specifically configured to: perform word segmentation on the user question data to obtain multiple keywords; extract features from the user question data to obtain multiple demand features; determine the child agent and bonus value corresponding to each demand feature based on a preset feature matching rule, wherein the feature matching rule includes the correspondence between demand features, associated child agents, and bonus values; for each child agent, match the multiple keywords with the keyword library corresponding to the child agent to calculate the first matching score of the child agent; and calculate the second matching score of the child agent based on the sum of the basic score and bonus value of the child agent; and calculate the decision weight of the child agent based on the first matching score and the second matching score.

[0063] In one possible example, the task execution rules include exception handling rules and normal execution rules. The exception handling rules are used to indicate the constraints when a sub-agent returns a data processing result under an exception scenario, and the normal execution rules are used to indicate the constraints when a sub-agent does not execute a data processing task under the exception scenario. An exception scenario refers to a scenario that affects the normal execution of a data processing task. Regarding the parent agent allocating data processing tasks and task execution rules to each sub-agent according to the decision weights, the allocation unit 502 is specifically used to: cause the parent agent to allocate a first data processing task and a second data processing task to the first sub-agent, and... The first data processing task is assigned to the second sub-agent, where the first sub-agent is one whose decision weight is not less than a first threshold, and the second sub-agent is one whose decision weight is less than the first threshold. The first data processing task refers to the necessary processing steps to meet user needs, and the second data processing task refers to the related processing steps to assist in understanding the response content. A first normal execution rule is assigned to the first sub-agent, and a second normal execution rule is assigned to the second sub-agent, where the constraint strength of the first normal execution rule is higher than that of the second normal execution rule. The anomaly response rule is assigned to all sub-agents.

[0064] In one possible example, the normal execution rule includes a verification rule, which refers to the verification strategy of the sub-agent for its data processing results. Regarding the allocation of the first normal execution rule to the first sub-agent by the parent agent and the allocation of the second normal execution rule to the second sub-agent, the allocation unit 502 is specifically configured to: allocate the first verification rule to the first sub-agent, the first verification rule instructing the first sub-agent to verify its data processing results from multiple verification dimensions; and allocate the second verification rule to the second sub-agent, the second verification rule instructing the second sub-agent to verify its data processing results from a single verification dimension; wherein the number of verification dimensions characterizes the coverage of risk prevention and control for the sub-agent's data processing results.

[0065] In one possible example, the abnormal scenario includes a data abnormal scenario, which refers to an abnormal scenario in which the sub-agent cannot obtain data from the target data source. In terms of the parent agent allocating the abnormal response rules to all sub-agents, the allocation unit 502 is specifically used to: enable the parent agent to allocate data abnormal response rules to all sub-agents. The data abnormal response rules are used to instruct the sub-agents to extract inferred data from associated data sources as data processing results. The associated data source refers to a data source that has a data association or logical association with the target data source and is in a non-abnormal state.

[0066] In one possible example, regarding the fusion of the data processing results of each child agent by the parent agent to obtain the target response content, the fusion unit 504 is specifically configured to: in response to the absence of the abnormal scenario, calculate the output content ratio of each child agent according to the decision weight; and fuse the data processing results of each child agent according to the output content ratio to obtain the target response content.

[0067] In one possible example, regarding the fusion of the data processing results of each child agent by the parent agent to obtain the target response content, the fusion unit 504 is specifically configured to: in response to the existence of the abnormal scenario, determine the abnormal child agent based on the data processing results, wherein the abnormal child agent refers to the child agent affected by the abnormal scenario; and reduce the decision weight of the abnormal child agent based on the degree of abnormal impact, and proportionally increase the decision weight of other child agents; and calculate the output content ratio of each child agent according to the adjusted decision weight; and fuse the data processing results of each child agent according to the output content ratio to obtain the target response content.

[0068] It is understood that since the method embodiments and the device embodiments are different presentations of the same technical concept, the content of the method embodiment section in this application should be adapted to the device embodiment section in a synchronous manner, and will not be repeated here.

[0069] When using integrated units, such as Figure 6 As shown, Figure 6 This is a structural block diagram of another data processing device based on multi-agent collaboration provided in the embodiments of this application. Figure 6 The data processing device 50 based on multi-agent collaboration includes a processing module 52 and a communication module 51. The processing module 52 controls and manages the actions of the data processing device, for example, executing the steps of the computing unit 501, the allocation unit 502, the execution unit 503, and the fusion unit 504, and / or other processes for executing the techniques described herein. The communication module 51 supports interaction between the data processing device and other devices. Figure 6 As shown, the data processing device based on multi-agent collaboration may further include a storage module 53, which is used to store the program code and data of the data processing device based on multi-agent collaboration.

[0070] All relevant content in each scenario involved in the above method embodiments can be referenced from the functional descriptions of the corresponding functional modules, and will not be repeated here. The above-mentioned data processing device 50 based on multi-agent collaboration can execute the above-mentioned... Figure 4 The data processing method based on multi-agent collaboration is shown.

[0071] Based on the description of the above method and device embodiments, please refer to... Figure 7 , Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 7 The illustrated electronic device includes a memory 701, a processor 702, a communication interface 703, and a bus 704. The memory 701, processor 702, and communication interface 703 are interconnected via the bus 704. Specifically, the electronic device may refer to the first electronic device 11 or the second electronic device 12 in the above embodiments.

[0072] The memory 701 may be a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM).

[0073] The memory 701 can store programs. When the program stored in the memory 701 is executed by the processor 702, the processor 702 and the communication interface 703 are used to execute the various steps of the data processing method based on multi-agent cooperation in the embodiments of this application.

[0074] The processor 702 may be a general-purpose central processing unit (CPU), microprocessor, application specific integrated circuit (ASIC), graphics processing unit (GPU), or one or more integrated circuits, used to execute relevant programs to achieve the functions required by the units in the electronic device of this application embodiment, or to execute the data processing method based on multi-agent cooperation of this application method embodiment.

[0075] The processor 702 can also be an integrated circuit chip with signal processing capabilities. In implementation, each step of the multi-agent collaborative data processing method of this application can be completed by the integrated logic circuits in the hardware of the processor 702 or by instructions in software form. The processor 702 can also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in the memory 701. The processor 702 reads the information in the memory 701 and, in conjunction with its hardware, performs the functions required by the units included in the electronic device of this application embodiment, or executes the data processing method based on multi-agent collaboration of this application method embodiment.

[0076] Communication interface 703 uses transceiver devices, such as, but not limited to, transceivers, to enable communication between electronic devices and other devices or communication networks. For example, data can be acquired through communication interface 703.

[0077] Bus 704 may include a pathway for transmitting information between various components of an electronic device (e.g., memory 701, processor 702, communication interface 703).

[0078] It should be noted that, although Figure 7 The illustrated electronic device only shows the memory 701, processor 702, and communication interface 703. However, those skilled in the art should understand that in specific implementations, the electronic device may also include other components necessary for normal operation. Furthermore, depending on specific needs, those skilled in the art should understand that the electronic device may also include hardware components for implementing other additional functions. Moreover, those skilled in the art should understand that the electronic device may only include the components necessary for implementing the embodiments of this application, and may not necessarily include... Figure 7 All the devices shown.

[0079] This application also provides a computer storage medium storing a computer program / instructions thereon, which, when executed by a processor, implements some or all of the steps of any of the methods described in the above method embodiments.

[0080] In the several embodiments provided in this application, it should be understood that the disclosed methods and apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for example, the division of units is merely a logical functional division, and there may be other division methods in actual implementation; for example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between devices or units, and may be electrical, mechanical, or other forms.

[0081] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0082] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. This computer program product includes one or more computer instructions. When these computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in or transmitted through a computer-readable storage medium. The computer instructions can be transmitted from one website, computer, electronic device, or data center to another website, computer, electronic device, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as an electronic device or data center that integrates one or more available media. The available medium can be a read-only memory, or a random access memory, or a magnetic medium, such as a floppy disk, hard disk, magnetic tape, magnetic disk, or an optical medium, such as a digital universal optical disc, or a semiconductor medium, such as a solid-state drive.

[0083] The above description is merely a specific implementation of the embodiments of this application, but the protection scope of the embodiments of this application is not limited thereto. Any changes or substitutions within the technical scope disclosed in the embodiments of this application should be covered within the protection scope of the embodiments of this application. Therefore, the protection scope of the embodiments of this application should be determined by the protection scope of the claims.

[0084] The device embodiments described above are merely illustrative. The units and modules described as separate components may or may not be physically separate. Furthermore, some or all of the units and modules can be selected to achieve the purpose of this embodiment, depending on actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0085] While this application discloses the above information, it is not limited thereto. Any person skilled in the art can easily conceive of variations or substitutions without departing from the spirit and scope of this application, and can make various alterations and modifications, including combinations of the different functions and implementation steps described above, as well as software and hardware implementation methods, all of which are within the protection scope of this application.

Claims

1. A data processing method based on multi-agent cooperation, characterized in that, The method is applied to an electronic device including a parent agent and a plurality of child agents, and the method comprises: The parent agent calculates a decision weight of each child agent according to user question data, the decision weight being used to represent a matching degree of the child agent with user demand; The parent agent allocates a data processing task and a task execution rule to each child agent according to the decision weight, the task execution rule being used to indicate a constraint condition of the child agent when executing the data processing task; The plurality of child agents execute the data processing task according to the task execution rule and return a data processing result to the parent agent; The parent agent fuses the data processing result of each child agent to obtain target reply content.

2. The method of claim 1, wherein, The parent agent calculates a decision weight of each child agent according to user question data, comprising: The parent agent performs word segmentation processing on the user question data to obtain a plurality of keywords; The parent agent performs feature extraction on the user question data to obtain a plurality of demand features; The parent agent determines a child agent corresponding to each demand feature and a score value based on a preset feature matching rule, the feature matching rule comprising a corresponding relationship among the demand feature, the associated child agent and the score value; For each child agent, the plurality of keywords and a keyword library corresponding to the child agent are matched to calculate a first matching score of the child agent, and a second matching score of the child agent is calculated according to a basic score value and a total score value of the child agent, and a decision weight of the child agent is calculated according to the first matching score and the second matching score.

3. The method according to claim 1 or 2, characterized in that, The task execution rule comprises an abnormality coping rule and a normal execution rule, the abnormality coping rule being used to indicate a constraint condition of the child agent when returning the data processing result in an abnormal scenario, and the normal execution rule being used to indicate a constraint condition of the child agent when not executing the data processing task in the abnormal scenario, the abnormal scenario being a scenario affecting normal execution of the data processing task, and the parent agent allocating the data processing task and the task execution rule to each child agent according to the decision weight, comprising: The parent agent allocates a first data processing task and a second data processing task to a first child agent and allocates the first data processing task to a second child agent, the first child agent being a child agent with a decision weight not less than a first threshold value, the second child agent being a child agent with a decision weight less than the first threshold value, the first data processing task being a necessary processing step meeting user demand, and the second data processing task being an associated processing step assisting in understanding reply content; The parent agent allocates a first normal execution rule to the first child agent and allocates a second normal execution rule to the second child agent, a constraint strength of the first normal execution rule being higher than a constraint strength of the second normal execution rule; The parent agent allocates the abnormality coping rule to all child agents.

4. The method of claim 3, wherein, The normal execution rule includes a verification rule, the verification rule is a verification strategy of the child agent for the data processing result, the parent agent assigns a first normal execution rule to the first child agent and a second normal execution rule to the second child agent, and the method comprises the following steps of: The parent agent assigns a first verification rule to the first child agent, and the first verification rule is used to instruct the first child agent to verify the data processing result from multiple verification dimensions; The parent agent assigns a second verification rule to the second child agent, and the second verification rule is used to instruct the second child agent to verify the data processing result from a single verification dimension; The number of verification dimensions is used to represent the coverage of the child agent for the risk prevention and control of the data processing result.

5. The method of claim 3, wherein, The abnormal scenario includes a data abnormal scenario, the data abnormal scenario is an abnormal scenario that the child agent cannot obtain data from a target data source, and the parent agent assigns the abnormal coping rule to all child agents, and the method comprises the following steps of: The parent agent assigns a data abnormal coping rule to all child agents, and the data abnormal coping rule is used to instruct the child agent to extract calculated data from an associated data source as the data processing result, and the associated data source is a data source that has data association or logical association with the target data source and is in a non-abnormal state.

6. The method of claim 5, wherein, The parent agent fuses the data processing result of each child agent to obtain the target reply content, and the method comprises the following steps of: In response to the absence of the abnormal scenario, calculating the output content proportion of each child agent according to the decision weight; and fusing the data processing result of each child agent according to the output content proportion to obtain the target reply content.

7. The method of claim 5, wherein, The parent agent fuses the data processing result of each child agent to obtain the target reply content, and the method comprises the following steps of: In response to the presence of the abnormal scenario, determining an abnormal child agent based on the data processing result, the abnormal child agent is a child agent affected by the abnormal scenario; and reducing the decision weight of the abnormal child agent based on the abnormal influence degree, and increasing the decision weight of other child agents by comparison; and calculating the output content proportion of each child agent according to the adjusted decision weight; and fusing the data processing result of each child agent according to the output content proportion to obtain the target reply content.

8. A data processing apparatus based on multi-agent collaboration, characterized by, The application is applied to an electronic device, the electronic device comprises a parent agent and a plurality of child agents, and the device comprises: A calculation unit is configured to cause the parent agent to calculate a decision weight of each child agent according to user question data, and the decision weight is used to represent the matching degree of the child agent and the user demand; An assignment unit is configured to cause the parent agent to assign a data processing task and a task execution rule to each child agent according to the decision weight, and the task execution rule is used to instruct the constraint condition of the child agent when executing the data processing task; An execution unit is configured to cause the plurality of child agents to execute the data processing task according to the task execution rule, and return the data processing result to the parent agent; A fusion unit is configured to fuse the data processing results of each child agent by the parent agent to obtain a target reply content.

9. An electronic device, comprising: A computer program product comprising a processor, a memory and one or more programs stored in the memory and configured to be executed by the processor, the programs comprising instructions for performing the steps of the method according to any one of claims 1-7.

10. A computer readable storage medium having stored thereon computer programs / instructions, characterized in that, The computer program / instructions, when executed by a processor, implement the steps of the method according to any one of claims 1-7.

Images