Knowledge graph-oriented multi-agent collaborative question and answer system

Abstract

The invention relates to the technical field of artificial intelligence, in particular to a knowledge graph-oriented multi-agent collaborative question-answering system, which comprises an acquisition component, a standby filter, a feature analyzer, an analyzer, a controller and a collaborative regulator, by collecting the key feature parameter entity similarity and relation complexity of the knowledge graph of the target platform in the historical period, the risk tendency period of the system can be quickly distinguished. Moreover, whether the system operation meets the standard or not can be determined based on the difference value between the risk feature data factor and the preset risk feature data factor threshold value in the risk tendency period, the system operation accuracy is improved, and when the system operation does not meet the standard, the reason label for the system operation not meeting the standard and the corresponding processing mode can be rapidly determined. The system overcomes the problem of low accuracy caused by the fact that a traditional system independently pays attention to agent indexes and question and answer accuracy.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a multi-agent collaborative question-answering system oriented towards knowledge graphs. Background Technology

[0002] Multi-agent collaborative question-answering systems based on knowledge graphs have become a core architecture in the complex knowledge service field due to their powerful reasoning and interpretation capabilities. However, their operational status is highly dependent on the data quality of the underlying knowledge graph and the performance stability of each agent's service. In actual operation, the risk of system performance degradation or even failure is often caused by two factors: firstly, the dynamic evolution of the knowledge graph itself, such as increased entity ambiguity and increased relational complexity, which directly increases the noise and difficulty of querying and reasoning; secondly, the aging performance of the hardware devices supporting the agents, leading to increased processing latency and error rates. Traditional operation and maintenance methods usually monitor hardware indicators or the final question-answering effect independently, making it difficult to provide early and accurate risk attribution warnings. They often only respond passively after problems appear, lacking a systematic risk management chain from pre-event feature analysis to post-event root cause determination, resulting in low operation and maintenance efficiency and challenges to system reliability. Therefore, there is an urgent need for an intelligent monitoring method that can deeply integrate graph data features and agent operation performance, and achieve proactive risk perception and accurate attribution.

[0003] Chinese Patent Publication No. CN118797017A discloses an intelligent question-answering system based on the collaboration of a large language model and a knowledge graph. The system includes: breaking down a complex problem into multiple simpler problems; analyzing the correlation between the simpler problems and basic functions to form multiple reasoning paths; automatically extracting structured information from the simpler problems using a multi-task learning framework based on the large model to construct a knowledge graph; and constructing a cumulative reasoning learning framework based on a large logical reasoning model to iteratively verify the process results of the knowledge graph formed based on the multiple reasoning paths, thereby correcting the reasoning paths until the correct answer is derived.

[0004] Chinese Patent Publication No. CN118410145A discloses a bridge maintenance question-answering method that integrates a large-scale model and a knowledge graph. The method includes: constructing a Prompt instruction for complex bridge maintenance questions; inputting the complex question and Prompt instruction into a large-scale dialogue model incorporating bridge maintenance knowledge, and decomposing the complex question; understanding the first sub-question in the current iteration, generating a Cypher query statement, and retrieving the answer from the bridge maintenance knowledge graph; inputting the answer to the first sub-question, the second sub-question, and the Prompt instruction into the large-scale dialogue model to further decompose the second sub-question and execute the next iteration; if the further decomposition fails, rewriting the second sub-question; retrieving the corresponding answer based on the Cypher query statement for the second sub-question; and combining the answers to all sub-questions to generate the final answer to the complex question. This invention utilizes the collaboration and fusion of a large-scale dialogue model and a bridge maintenance knowledge graph to achieve question-answering for complex questions.

[0005] However, the following problems still exist in the existing technology:

[0006] Existing technologies do not consider the noise impact caused by differences in entity ambiguity and relationship complexity in question-answering systems, as well as the decrease in accuracy caused by the increase in agent latency and error rate. Summary of the Invention

[0007] To address this, the present invention provides a multi-agent collaborative question-answering system for knowledge graphs, which overcomes the noise effects caused by differences in entity ambiguity and relation complexity in existing question-answering systems, as well as the problems of low system accuracy and low efficiency caused by increased agent latency and error rates.

[0008] To achieve the above objectives, this invention provides a multi-agent collaborative question-answering system oriented towards knowledge graphs, comprising:

[0009] The data acquisition component is used to collect key feature parameters of the knowledge graph data of the target platform within a historical period, as well as risk feature data information of different intelligent agents of the target platform within a historical period.

[0010] Backup filters, including microwave filters for noise amplification of the target platform;

[0011] A feature parser, connected to the acquisition component, is used to analyze key feature representation values ​​based on the key feature parameters and to analyze risk feature data factors based on the risk feature data information.

[0012] An analyzer, connected to the acquisition component and the feature parser, is used to determine the risk propensity period and risk significance period of the question-answering system of the target platform based on the variance of the key feature representation values.

[0013] The controller, which is connected to the analyzer and the backup filter respectively, is used to determine whether the system operation meets the standard based on the difference between the risk characteristic data factor and the predetermined risk characteristic data factor threshold during the risk tendency period.

[0014] A coordinating regulator, connected to the controller, is used to determine the cause label and corresponding handling method for the system operation not meeting the standard based on the difference between the key feature characterization value and the predetermined key feature characterization threshold: increasing the risk feature data factor threshold, or turning on the microwave filter;

[0015] The key feature parameters include the entity similarity of the target platform's question-answering system and the relational complexity of the knowledge graph within the time period.

[0016] The risk characteristic data includes the latency and error rates of each agent on the target platform within a time period.

[0017] Furthermore, the relational complexity of the knowledge graph is determined based on the ratio of the number of sub-problems that a complex problem is broken down into to a predetermined threshold for the number of sub-problems.

[0018] Furthermore, the key feature representation value is determined based on the sum of the first key feature factor and the second key feature factor;

[0019] The first key feature factor is determined based on the ratio of the predetermined entity similarity to the actual entity similarity threshold of the target platform's question-answering system within the time period.

[0020] The second key feature factor is determined based on the ratio of the relational complexity of the target platform's question-answering system to a predetermined relational complexity threshold within a time period.

[0021] Furthermore, the risk tendency period is the time period during which the variance of the key feature representation value of the target platform is less than or equal to a predetermined variance threshold.

[0022] Furthermore, the risk characteristic data factor is determined based on the sum of the first risk characteristic factor and the second risk characteristic factor;

[0023] The first risk characteristic factor is determined based on the ratio of the latency rate of a single agent of the target platform to a predetermined latency rate threshold within a time period.

[0024] The second risk characteristic factor is determined based on the ratio of the error rate of a single agent of the target platform to a predetermined error rate threshold within a time period.

[0025] Preferably, the question-and-answer system operates in accordance with the standard when the difference between the risk feature data factor and the predetermined risk feature data factor threshold is less than or equal to the predetermined difference threshold.

[0026] The question-and-answer system fails to meet the standard if the difference between the risk feature data factor and the predetermined risk feature data factor threshold is greater than the predetermined difference threshold.

[0027] Preferably, the reasons why the question-and-answer system does not meet the standards include a first reason label and a second reason label;

[0028] The first cause label is that the standard operating deviation within any intelligent agent of the system leads to misjudgment of operation.

[0029] The second cause label is that the noise in the system's knowledge graph exceeds the noise threshold, causing the system to fail to meet the standards.

[0030] Preferably, the condition for the first cause label is that the difference between the key feature representation value and the predetermined key feature representation threshold within the time period is less than or equal to the predetermined difference threshold.

[0031] The condition for the second reason label is that the difference between the key feature representation value and the predetermined key feature representation threshold within the time period is less than or equal to the predetermined difference threshold.

[0032] Preferably, the processing method of the first cause label is to increase the threshold of the system risk feature data factor, and the increase is determined based on the difference between the key feature representation value and the predetermined key feature representation threshold within the time period.

[0033] The second cause label is handled by enabling the microwave filter.

[0034] Preferably, the risk significance period is the time period during which the variance of the key feature representation value of the target platform is greater than a predetermined variance threshold; the processing method is to activate an alarm signal.

[0035] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention provides a multi-agent collaborative question-answering system for knowledge graphs, including a data acquisition component, a backup filter, a feature parser, an analyzer, a controller, and a collaborative regulator. By acquiring key feature parameters of the target platform's knowledge graph—entity similarity and relation complexity—over a historical period, it can quickly distinguish the system's risk tendency cycle. Moreover, during the risk tendency cycle, it can determine whether the system operation meets the standard based on the difference between risk feature data factors and a predetermined risk feature data factor threshold, improving the accuracy of system operation. When the standard is not met, it quickly identifies the reason label for the system's non-compliance and the corresponding processing method, overcoming the problem of low accuracy caused by traditional systems that focus solely on agent indicators and question-answering accuracy. By integrating the entity similarity and relation complexity features of the knowledge graph, as well as the agent's latency and error rate for operational indicator monitoring, it achieves prediction, accurate diagnosis, and adaptive regulation of potential system risks, overcoming the limitations of single-dimensional system monitoring and improving the system's accuracy and efficiency.

[0036] In particular, by analyzing the variance changes of key characteristic parameters over historical periods, risk tendencies can be identified before system performance significantly declines. This predictive early warning mechanism based on data trends proactively mitigates system maintenance risks, greatly reducing the occurrence of system downtime leading to inefficiency.

[0037] In particular, threshold comparison and difference analysis can accurately distinguish the cause labels of system anomalies, thereby accurately identifying agent performance problems, reducing the occurrence of misjudgments of the performance degradation of the knowledge graph itself or individual agents, and accurately identifying graph complexity problems, which can distinguish between real noise and system performance problems, improve operation and maintenance efficiency, reduce the time of traditional trial and error troubleshooting, and improve system efficiency.

[0038] In particular, the system has closed-loop optimization capabilities, which can automatically trigger corresponding optimization strategies based on diagnostic results. For agent problems, it dynamically adjusts thresholds to improve system robustness; for graph noise problems, it intelligently activates backup filters to improve answer quality. This self-adjusting capability enables the system to adapt to constantly changing environments and loads, significantly reducing misjudgments or omissions caused by human intervention, and achieving autonomous optimization and continuous efficient operation of the system. Attached Figure Description

[0039] Figure 1 This is a structural block diagram of a knowledge graph-oriented multi-agent collaborative question-answering system according to an embodiment of the present invention;

[0040] Figure 2 This is a flowchart illustrating the steps of the multi-agent collaborative question-answering method for knowledge graphs according to an embodiment of the present invention.

[0041] Figure 3 A logical block diagram for determining the risk tendency cycle of the question-answering system of the target platform in an embodiment of the present invention;

[0042] Figure 4 A flowchart illustrating the steps for determining risk characteristic data factors in an embodiment of the invention. Detailed Implementation

[0043] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0044] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0045] It should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the term "connection" should be interpreted broadly. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be a connection within two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0046] Please see Figure 1 As shown, Figure 1 This is a structural block diagram of a knowledge graph-oriented multi-agent collaborative question-answering system according to an embodiment of the present invention. The embodiment provides a knowledge graph-oriented multi-agent collaborative question-answering system, including:

[0047] The data acquisition component is used to collect key feature parameters of the knowledge graph data of the target platform within a historical period, as well as risk feature data information of the target platform's intelligent agents within a historical period.

[0048] Backup filters, including microwave filters for noise amplification of the target platform;

[0049] A feature parser, connected to the acquisition component, is used to analyze key feature representation values ​​based on the key feature parameters and to analyze risk feature data factors based on the risk feature data information.

[0050] An analyzer, connected to the acquisition component and the feature parser, is used to determine the risk propensity period and risk significance period of the question-answering system of the target platform based on the variance of the key feature representation values.

[0051] The controller, which is connected to the analyzer and the backup filter respectively, is used to determine whether the system operation meets the standard based on the difference between the risk characteristic data factor and the predetermined risk characteristic data factor threshold during the risk tendency period.

[0052] A coordinating regulator, connected to the controller, is used to determine the cause label and corresponding handling method for the system operation not meeting the standard based on the difference between the key feature characterization value and the predetermined key feature characterization threshold: increasing the risk feature data factor threshold, or turning on the microwave filter;

[0053] The key feature parameters include the entity similarity of the target platform's question-answering system and the relational complexity of the knowledge graph within the time period.

[0054] The risk characteristic data includes the latency and error rates of each agent on the target platform within a time period.

[0055] This invention provides a multi-agent collaborative question-answering system for knowledge graphs, including a data acquisition component, a backup filter, a feature parser, an analyzer, a controller, and a collaborative regulator. It efficiently ensures the stable and high-quality operation of the multi-agent collaborative question-answering system for knowledge graphs. The data acquisition component accurately collects key feature parameters and risk feature data; the feature parser transforms the data into effective feature values ​​and factors; the analyzer uses variance to determine the risk cycle, enabling early warning. The controller determines whether the system meets the standards during the risk cycle, while the collaborative regulator locates the reasons for non-compliance and quickly resolves the problem by adjusting thresholds or activating the microwave filter. Overall, it achieves early risk warning, accurate judgment, and efficient repair, improving the system's accuracy and efficiency.

[0056] Please see Figure 2 The diagram shows a flowchart illustrating the steps of a knowledge graph-oriented multi-agent collaborative question-answering method according to an embodiment of the present invention. The operation steps of the knowledge graph-oriented multi-agent collaborative question-answering system in this embodiment include:

[0057] Step S1: Obtain key feature parameters from the knowledge graph data of the target platform within the historical period, and analyze key feature representation values ​​based on the key feature parameters;

[0058] Step S2: Determine the risk tendency period and risk significance period of the target platform's question-answering system based on the variance of key feature representation values ​​within the time period;

[0059] Step S3: Extract risk characteristic data information of different intelligent agents of the target platform within the historical period based on the determined risk propensity;

[0060] Step S4: Analyze the risk characteristic data factors based on the risk characteristic data information, and determine whether the system operation meets the standard based on the difference between the risk characteristic data factors and the predetermined risk characteristic data factor threshold;

[0061] When the system fails to meet the standard, the reason for the non-compliance is determined based on the difference between the key feature characterization value and the predetermined key feature characterization threshold, along with the corresponding handling method: increase the risk feature data factor threshold, or turn on the microwave filter.

[0062] The implementation example effectively improves the operational quality and risk management capabilities of a knowledge graph-based multi-agent collaborative question-answering system. From a risk warning perspective, by using key feature representation values ​​and variance analysis, it accurately identifies risk tendencies and significant cycles, proactively capturing potential system vulnerabilities and promptly detecting risk tendencies such as knowledge graph data anomalies or agent response delays. By extracting risk features and analyzing data factors, and comparing them with thresholds, it clarifies whether the system meets standards, providing insights into the root causes of excessive error rates in individual agents for optimization. When the system fails to meet standards, it determines the cause label and processing method based on feature differences, quickly resolving the problem by increasing the threshold or activating the microwave filter, thereby improving the accuracy and efficiency of system operation. The implementation example achieves early warning, accurate diagnosis, and efficient repair of system risks, improving the reliability, accuracy, and operational efficiency of the question-answering system, and better meeting user question-answering needs.

[0063] Specifically, there are no restrictions on the method for obtaining entity similarity. It can be obtained by structural comparison, by calculating the weighted sum of the similarity of entity nodes and the similarity of paths, or by comparing the first entity to be calculated with the second entity to be calculated, thereby obtaining the graph kernel similarity between the first entity and the second entity. Other forms can also be used, which will not be elaborated here.

[0064] Specifically, there is no limitation on the method of obtaining the relation complexity of the knowledge graph. Preferably, it is determined from the static structure of the knowledge graph, based on the ratio of the number of sub-problems to the predetermined threshold of the number of sub-problems. Based on the calculation method of relation connection density, the relational strength of entities is reflected by statistically analyzing the relational strength of entities, including the average number of relations of entities and the sparsity of relational distribution. The denser the relational strength and the more uneven the distribution, the higher the relational complexity. Other forms can also be used, which will not be elaborated here.

[0065] Specifically, there is no limitation on the method of obtaining the latency rate of a single agent. In this embodiment, the latency rate is determined based on the ratio of the number of times the agent's task processing time exceeds the threshold to the total number of tasks within a unit of time. The timestamp of the entire lifecycle of the agent's processing of a single task from "receive-start-end" is recorded, the actual time consumed is calculated and compared with the preset threshold, and the timeout ratio is statistically analyzed. This method is applicable to all types of agents, including query parsing agents and medical diagnosis agents. Other methods can also be used, which will not be elaborated here.

[0066] Similarly, there are no restrictions on the method of obtaining the error rate. In this example, the error rate is the proportion of times the agent's processing result does not conform to the facts of the knowledge graph or the user's needs within a unit of time. The accuracy risk is quantified by "result verification + feedback collection". In this example, the correctness of the agent's output result is verified based on the authoritative facts in the knowledge graph, and the error ratio is statistically analyzed. This is applicable to agents whose output results can be directly compared with the knowledge graph, including data query agents and entity recognition agents. Other forms can also be used, which will not be elaborated here.

[0067] It is understood that there is no limitation on the classification of intelligent agents. The intelligent agents in the example include: problem understanding intelligent agents, query planning intelligent agents, retrieval execution intelligent agents, reasoning intelligent agents, and verification intelligent agents. Other forms may also be used, which will not be elaborated here.

[0068] It is understood that the historical period interval for collecting the knowledge graph data and the risk characteristic data of the intelligent agent of the target platform in the embodiment is [4, 6], in hours. At the same time, the time period interval of the embodiment is [0.5, 2], in hours. Other historical period intervals and time period forms can also be used, which will not be elaborated here.

[0069] It is understood that the acquisition component of the embodiment has the function of automatic data acquisition, the analyzer automatically analyzes the data, and the controller and coordinator can perform corresponding debugging. At the same time, it has the functions of monitoring data query, export, and automatic backup, and can record data information of the target platform in historical periods.

[0070] It is understandable that the application scenarios of the embodiments are high-complexity knowledge-driven scenarios with extremely high requirements for question-answering accuracy and system stability, including:

[0071] In the field of financial risk control, the implementation example can monitor an intelligent query platform based on enterprise relationship graphs in real time. When the complexity of entity relationships in the graph increases sharply due to frequent mergers and acquisitions, it can accurately identify the query noise caused by this and automatically activate a backup filter to ensure the accuracy of the risk assessment report. At the same time, it can effectively distinguish this anomaly from the delay and errors caused by the aging of each intelligent agent, thus avoiding misjudgment.

[0072] In the medical diagnostic assistance platform, the system can continuously perceive the changing trends of the similarity between disease and drug entities in the medical knowledge graph. It can provide early warnings before the evolution of the graph causes confusion and delays in the diagnostic reasoning paths of each agent and a significant increase in the risk of errors. It can also intelligently adjust the fault tolerance threshold of the agents to ensure the reliable generation and efficient response of diagnostic suggestions and guarantee the continuous and stable operation of the system.

[0073] In addition, it can also be applied to other scenarios. In scenarios such as intelligent manufacturing and smart government that rely on large-scale dynamic knowledge graphs for decision support, the system can play a key role, which will not be elaborated here.

[0074] Specifically, there are no restrictions on the specific structure of the controller and the coordinator; they can all be composed of logic components, including field-programmable components, computers, or microprocessors in computers.

[0075] Specifically, the relational complexity of the knowledge graph is determined based on the ratio of the number of sub-problems to a predetermined threshold for the number of sub-problems in a complex problem.

[0076] The predetermined threshold for the number of sub-questions is determined in advance, wherein the average number of sub-questions in the knowledge graph data of the target platform's question-answering system over a certain number of historical time periods is recorded in advance, and the average number of sub-questions is determined as the threshold for the number of sub-questions.

[0077] The implementation example evaluates complexity by comparing the ratio of the number of sub-problems to a threshold, transforming abstract relational complexity into quantifiable metrics. This makes the evaluation more aligned with question-and-answer task scenarios, reducing the likelihood of traditional evaluations becoming detached from practical needs. Simultaneously, it provides precise parameters for the data acquisition components, assisting the feature parser in generating reliable feature values, enabling the analyzer to more accurately predict risk cycles, the controller to precisely determine standard threshold benchmarks, and the coordinator to address specific problems. Ultimately, this enhances the system's ability to control the difficulty of knowledge graph associations, optimizes multi-agent collaboration efficiency, and ensures question-and-answer accuracy and system stability.

[0078] Specifically, the key feature representation value is determined based on the sum of the first key feature factor and the second key feature factor;

[0079] The first key feature factor is determined based on the ratio of the predetermined entity similarity threshold of the target platform's question-answering system to the actual entity similarity within the time period.

[0080] The second key feature factor is determined based on the ratio of the relational complexity of the target platform's question-answering system to a predetermined relational complexity threshold within a time period.

[0081] The predetermined entity similarity threshold is determined in advance, wherein the average entity similarity of the target platform question answering system within a certain historical time period is recorded in advance, and the product of the average entity similarity and the accuracy coefficient is determined as the entity similarity threshold, and the accuracy coefficient is selected in the interval [0.90, 0.95].

[0082] Similarly, the predetermined relation complexity threshold is determined in advance, wherein the average relation complexity of the target platform's question-answering system over a certain number of historical time periods is recorded in advance, and the product of the average relation complexity and the deviation coefficient is determined as the relation complexity threshold, with the deviation coefficient selected within the interval [1.05, 1.15].

[0083] Key feature representation values ​​are determined by summing two-dimensional feature factors. The first key feature factor combines entity similarity, and the second relationship feature factor combines relational complexity. This transforms the core features of the knowledge graph into quantifiable indicators, reducing the possibility of misjudgments due to the one-sidedness of single feature evaluation. Accurate representation values ​​provide reliable data support for the analyzer, thereby accurately determining risk cycles; overall, it enhances the system's ability to control knowledge graph features, improves the efficiency of multi-agent collaboration, and ensures question-answering accuracy and stable system operation.

[0084] Please see Figure 3 As shown, it is a logical block diagram of the risk tendency period of the question-and-answer system of the target platform in an embodiment of the present invention. The risk tendency period is the time period during which the variance of the key feature representation value of the target platform is less than or equal to a predetermined variance threshold.

[0085] Specifically, the predetermined variance threshold is determined in advance, wherein the average variance of the key feature representation values ​​over a certain number of historical time periods is recorded in advance, and the product of the average variance and the precision coefficient is determined as the variance threshold, wherein the precision coefficient is selected within the range of [1.15, 1.25].

[0086] The implementation clearly defines risk periods using variance thresholds, coupled with alarm processing. A risk-prone period is identified when the variance of a key feature value is less than or equal to the threshold, allowing for early detection of potential risks. A significant risk period is defined when the variance exceeds the threshold, accurately identifying risks that have already materialized. When a significant risk period is triggered, an alarm signal is activated, reducing the impact of risks on system operation. Simultaneously, the clear period division allows for targeted resource allocation, enabling early prediction during risk-prone periods and rapid response during significant risk periods. Overall, this improves the system's risk identification accuracy and response efficiency, enhances the stability of multi-agent collaborative question answering, ensures the quality of knowledge graph-driven question answering services, and reduces user experience degradation or data bias caused by risks.

[0087] Please see Figure 4The diagram shown is a flowchart illustrating the steps for determining risk characteristic data factors in an embodiment of the invention. The process for determining risk characteristic data factors in the embodiment includes:

[0088] Extract the latency and error rate of a single agent on the target platform within the time period;

[0089] The ratio of the latency rate of a single agent on the target platform to a predetermined latency rate threshold within the calculation time period is determined as the first risk characteristic factor;

[0090] The ratio of the error rate of a single agent on the target platform to a predetermined error rate threshold within the calculation time period is determined as the second risk characteristic factor;

[0091] The sum of the first risk characteristic factor and the second risk characteristic factor is determined as the risk characteristic data factor.

[0092] The embodiment determines risk characteristic data factors by summing a first risk characteristic factor and a second risk characteristic factor. This quantitatively and collaboratively integrates the timeliness and accuracy indicators of the agent, reducing the one-sidedness of assessing risk with a single indicator and comprehensively reflecting the operational risks of the agent. Accurate risk characteristic data factors provide a reliable basis for the controller to determine system compliance, promptly identify agent problems, lay the foundation for subsequent collaborative regulators to formulate handling solutions, effectively improve the system's risk management capabilities, and ensure the stability and efficiency of multi-agent collaborative question answering.

[0093] The predetermined latency threshold is determined in advance, wherein the average latency rate of each agent within a certain historical time period is recorded in advance, and the product of the average latency rate and the offset system is determined as the latency threshold. The offset system is selected in the interval [0.95, 0.99].

[0094] Similarly, the predetermined error rate threshold is determined in advance, wherein the average error rate of each agent in a certain historical time period is recorded in advance, and the product of the average error rate and the error coefficient is determined as the error rate threshold, and the error system is selected in the interval [0.92, 0.97].

[0095] Specifically, the question-and-answer system operates in accordance with the standard when the difference between the risk feature data factor and the predetermined risk feature data factor threshold is less than or equal to the predetermined difference threshold.

[0096] The question-and-answer system fails to meet the standard if the difference between the risk feature data factor and the predetermined risk feature data factor threshold is greater than the predetermined difference threshold.

[0097] The threshold for risk characteristic data factors in the example is selected within the interval [2.15, 2.25].

[0098] The difference threshold in the example is selected within the interval [0.02, 0.15].

[0099] The implementation example clarifies the criteria for determining system operating standards by using a difference threshold. The difference between risk characteristic data factors and a preset threshold is used as the core judgment basis. When the difference is less than or equal to the preset threshold, the system is deemed to meet the standards; when the difference is greater than the preset threshold, it is deemed not to meet the standards. This reduces the occurrence of misjudgments common in traditional methods and improves the accuracy of system operation. By quickly and accurately identifying the system's operating status within a risk tendency cycle, it reduces misjudgments caused by minor deviations and also reduces the occurrence of missed risks due to significant deviations. For systems that meet the standards, the current collaborative mode can be maintained to ensure question-and-answer efficiency; for systems that do not meet the standards, the collaborative regulator can be triggered in a timely manner to resolve the issue by adjusting thresholds or activating filters, thereby further improving the accuracy and efficiency of system operation.

[0100] Specifically, the reasons why the question-and-answer system described in the embodiment does not meet the standards include a first reason label and a second reason label;

[0101] The first cause label is that the standard operating deviation within any intelligent agent of the system leads to misjudgment of operation.

[0102] The second cause label is that the noise in the system's knowledge graph exceeds the noise threshold, causing the system to fail to meet the standards.

[0103] The noise threshold is predetermined, wherein the average value of the knowledge graph noise over a certain number of historical time periods is recorded in advance, and the product of the average noise value and the accuracy coefficient is determined as the noise threshold. The accuracy system is selected in the interval [1.05, 1.25].

[0104] The first cause label in the embodiment is the deviation of the internal standard operation of the intelligent agent. The intelligent agent has a preset standard operation logic, including reasoning rules, data processing thresholds, and knowledge call process. When the internal standard deviates, it will cause the system to fail to meet the standard through the link of deviation transmission - reasoning distortion - result misjudgment.

[0105] The second cause label in the example is excessive noise in the knowledge graph. Knowledge graph noise includes incorrect entity associations, redundant relationships, and outdated attribute values ​​exceeding the threshold. When these exceed the threshold, they will destroy the semantic consistency of the knowledge graph and affect the accuracy and efficiency of the system through noise interference, knowledge transfer distortion, and collaborative reasoning bias.

[0106] This implementation clearly identifies two core cause labels to pinpoint system non-compliance issues. The non-compliance causes are broken down into internal standard deviation of the agent and excessive noise in the knowledge graph, reducing the ambiguity of traditional cause identification and making fault tracing more accurate. If the cause label is the first cause, the agent's operating standards or calibration parameters are optimized accordingly; if the cause label is the second cause, noise reduction measures such as microwave filters can be quickly activated. This significantly shortens problem-solving time, reduces system downtime or inefficient operation time, and improves system accuracy and efficiency.

[0107] Specifically, in this embodiment, the processing method of the first cause label is to increase the threshold of the system risk feature data factor, and the increase is determined based on the difference between the key feature representation value and the predetermined key feature representation threshold within the time period.

[0108] The second cause label is handled by enabling the microwave filter.

[0109] The predetermined key feature representation threshold is selected within the range [2.15, 2.35].

[0110] The implementation example uses a cause-label-processing method for precise matching. For the first cause label, it increases the threshold of the risk feature data factor, with the increase determined by the difference between the key feature representation value and the threshold. This reduces the occurrence of thresholds being too high or too low due to fixed adjustments, and dynamically adapts to the agent's bias level, reducing misjudgments while ensuring the system's sensitivity to risk. For the second cause label, the implementation example activates a microwave filter to quickly filter noisy data, restore the semantic consistency of the knowledge graph, and reduce continuous noise interference in multi-agent reasoning. This effectively improves the accuracy and efficiency of the system's risk management.

[0111] Specifically, the knowledge graph-oriented multi-agent collaborative question-answering system also includes an early warning device, which can determine whether to activate an alarm signal.

[0112] The significant risk period is the time period during which the variance of the key feature representation value of the target platform is greater than a predetermined variance threshold; the processing method is to activate an alarm signal.

[0113] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A multi-agent collaborative question-answering system oriented towards knowledge graphs, characterized in that, include: The data acquisition component is used to collect key feature parameters of the knowledge graph data of the target platform within a historical period, as well as risk feature data information of different intelligent agents of the target platform within a historical period. Backup filter, which is a microwave filter used to filter noise from the target platform; A feature parser, connected to the acquisition component, is used to analyze key feature representation values ​​based on the key feature parameters and to analyze risk feature data factors based on the risk feature data information. An analyzer, connected to the acquisition component and the feature parser, is used to determine the risk propensity period and risk significance period of the question-answering system of the target platform based on the variance of the key feature representation values. The controller, which is connected to the analyzer and the backup filter respectively, is used to determine whether the system operation meets the standard based on the difference between the risk characteristic data factor and the predetermined risk characteristic data factor threshold during the risk tendency period. A coordinating regulator, connected to the controller, is used to determine the cause label and corresponding handling method for the system operation not meeting the standard based on the difference between the key feature characterization value and the predetermined key feature characterization threshold: increasing the risk feature data factor threshold, or turning on the microwave filter; The key feature parameters include the entity similarity of the target platform's question-answering system and the relational complexity of the knowledge graph within the time period. The risk characteristic data includes the latency and error rates of each agent on the target platform within a time period.

2. The knowledge graph-oriented multi-agent collaborative question-answering system according to claim 1, characterized in that, The relational complexity of the knowledge graph is determined based on the ratio of the number of sub-problems that a complex problem is broken down into to a predetermined threshold for the number of sub-problems.

3. The knowledge graph-oriented multi-agent collaborative question-answering system according to claim 2, characterized in that, The key feature representation value is determined based on the sum of the first key feature factor and the second key feature factor; The first key feature factor is determined based on the ratio of the predetermined entity similarity to the actual entity similarity threshold of the target platform's question-answering system within a time period. The second key feature factor is determined based on the ratio of the relational complexity of the target platform's question-answering system to a predetermined relational complexity threshold within a time period.

4. The knowledge graph-oriented multi-agent collaborative question-answering system according to claim 1, characterized in that, The risk tendency period is the time period during which the variance of the key feature representation value of the target platform is less than or equal to a predetermined variance threshold.

5. The knowledge graph-oriented multi-agent collaborative question-answering system according to claim 1, characterized in that, The risk characteristic data factors are determined based on the sum of the first risk characteristic factor and the second risk characteristic factor; The first risk characteristic factor is determined based on the ratio of the latency rate of a single agent of the target platform to a predetermined latency rate threshold within a time period. The second risk characteristic factor is determined based on the ratio of the error rate of a single agent of the target platform to a predetermined error rate threshold within a time period.

6. The knowledge graph-oriented multi-agent collaborative question-answering system according to claim 1, characterized in that, The question-and-answer system operates in compliance with the standard if the difference between the risk feature data factor and the predetermined risk feature data factor threshold is less than or equal to the predetermined difference threshold. The question-and-answer system fails to meet the standard if the difference between the risk feature data factor and the predetermined risk feature data factor threshold is greater than the predetermined difference threshold.

7. The knowledge graph-oriented multi-agent collaborative question-answering system according to claim 6, characterized in that, The reasons why the question-and-answer system does not meet the standards include the first reason label and the second reason label; The first cause label is that the standard operating deviation within any intelligent agent of the system leads to misjudgment of operation. The second cause label is that the noise in the system's knowledge graph exceeds the noise threshold, causing the system to fail to meet the standards.

8. The knowledge graph-oriented multi-agent collaborative question-answering system according to claim 7, characterized in that, The condition for the first cause label is that the difference between the key feature representation value and the predetermined key feature representation threshold within the time period is less than or equal to the predetermined difference threshold. The condition for the second reason label is that the difference between the key feature representation value and the predetermined key feature representation threshold within the time period is greater than the predetermined difference threshold.

9. The knowledge graph-oriented multi-agent collaborative question-answering system according to claim 8, characterized in that, The first cause label is processed by increasing the threshold of the system risk feature data factor, and the increase is determined based on the difference between the key feature representation value and the predetermined key feature representation threshold within the time period. The second cause label is handled by enabling the microwave filter.

10. The knowledge graph-oriented multi-agent collaborative question-answering system according to claim 9, characterized in that, The significant risk period is the time period during which the variance of the key feature representation value of the target platform is greater than a predetermined variance threshold; the processing method is to activate an alarm signal.

Images