A medical insurance knowledge base automatic construction method based on multi-agent cooperation

Abstract

The application provides a medical insurance knowledge base automatic construction method based on multi-agent cooperation, and relates to the technical field of medical insurance knowledge base construction. The application first defines and configures multiple agents with differentiated medical insurance knowledge background and functions, and equips each agent with standard behavior. The initial query drives multi-round dialogue of the agents, collects and fuses dialogue data to extract core positions and semantic relationship graphs, constructs a collaborative decision-making model to reach a consensus and compiles it into structured medical insurance knowledge information. A third-party verification agent is introduced, a knowledge pedigree graph is constructed based on knowledge items and traceability data, the credibility is calculated through evidence theory fusion decision algorithm, conflicts are resolved, and the output suggestion is adopted. The generated knowledge base is stored and a retrieval enhanced generation question and answer system is built, user feedback is recorded, and regular new topic discussion and artificial expert intervention are combined to realize continuous learning and dynamic iteration of the knowledge base, which can automatically construct and dynamically update the structured medical insurance knowledge base to support intelligent question and answer.

Description

Technical Field

[0001] This application relates to the field of medical insurance knowledge base construction technology, and in particular to an automatic construction method for medical insurance knowledge base based on multi-agent collaboration. Background Technology

[0002] The medical insurance policy documents are complex and frequently updated. The implementation of applications such as intelligent question answering and policy interpretation urgently requires a structured, highly reliable medical insurance knowledge base that can dynamically adapt to policy changes. The core technical requirements are focused on the automated and efficient extraction of knowledge, the collaborative verification of multi-source information, and the continuous iterative updates of the knowledge base to meet the needs of large-scale construction and precise application.

[0003] The current mainstream solution for this need is an automatic medical insurance knowledge construction scheme based on a single intelligent agent and a preset rule base. By configuring medical insurance domain rule templates for the intelligent agent, key information is extracted from policy texts, treatment guidelines and other materials to generate structured knowledge items, and a simple keyword matching mechanism is used for preliminary verification.

[0004] The existing solution has obvious shortcomings: First, a single intelligent agent is difficult to cover the multi-dimensional scenarios of medical insurance knowledge, and the rigid constraints of the rule base result in poor adaptability to complex semantics and policy boundaries; Second, the lack of a multi-subject collaborative verification mechanism makes it difficult to guarantee the credibility of knowledge items and easily leads to information conflicts; Third, the lack of a linkage and iteration mechanism between user feedback and policy updates results in insufficient dynamic updating capability of the knowledge base, making it difficult to keep up with the pace of medical insurance policy adjustments. Summary of the Invention

[0005] The purpose of this application is to provide an automatic construction method for medical insurance knowledge base based on multi-agent collaboration, so as to solve the problems of insufficient adaptation to multi-dimensional scenarios of medical insurance knowledge, lack of guarantee of knowledge credibility, and lag in dynamic updates of the knowledge base in the existing technology.

[0006] To address the aforementioned technical problems, in a first aspect, this application provides a method for automatically constructing a medical insurance knowledge base based on multi-agent collaboration, comprising:

[0007] Define and configure multiple intelligent agents with differentiated medical insurance knowledge backgrounds and functions, and configure exclusive prompt words for each intelligent agent to regulate behavior patterns;

[0008] The system drives multiple agents to engage in multi-turn dialogues based on an initial query question, collects dialogue data and performs fusion analysis to extract the core positions and semantic relationship graphs of each agent.

[0009] A collaborative decision-making model is constructed based on the core positions and semantic relationship graphs of each agent. A consensus solution is reached through relational reasoning and constraint solving, and the consensus solution is compiled into structured medical insurance knowledge information.

[0010] introducing a third-party verification agent, constructing a knowledge pedigree graph based on the medical insurance knowledge information and the traceability data of the medical insurance knowledge information, calculating the credibility of the medical insurance knowledge information through an evidence theory fusion decision algorithm, automatically resolving conflicts of the medical insurance knowledge information that exist, and outputting the medical insurance knowledge information that is recommended to be adopted;

[0011] storing the medical insurance knowledge information that is recommended to be adopted to generate a medical insurance knowledge base.

[0012] Optionally, the method further comprises:

[0013] constructing a retrieval-enhanced question and answer system based on the medical insurance knowledge base, receiving user questions and intelligently responding, and recording user feedback data generated in the interactive process;

[0014] automatically identifying medical insurance knowledge information in the medical insurance knowledge base that needs to be updated based on the user feedback data, triggering a re-discussion and update process, and realizing continuous learning and dynamic iterative updating of the medical insurance knowledge base through regular initiation of new topic discussions and manual expert intervention for correction.

[0015] Optionally, the introducing a third-party verification agent, constructing a knowledge pedigree graph based on the medical insurance knowledge information and the traceability data of the medical insurance knowledge information, comprises:

[0016] introducing a third-party verification agent, accessing, by the third-party verification agent, dialogue history records and consensus generation process data of all agents as traceability data of the medical insurance knowledge information;

[0017] constructing a knowledge pedigree graph based on the medical insurance knowledge information and the traceability data of the medical insurance knowledge information, wherein nodes in the knowledge pedigree graph represent the medical insurance knowledge information, and edges represent derivative relationships or conflict relationships between entries.

[0018] Optionally, the calculating the credibility of the medical insurance knowledge information through an evidence theory fusion decision algorithm, and the automatically resolving conflicts of the medical insurance knowledge information that exist, and outputting the medical insurance knowledge information that is recommended to be adopted, comprises:

[0019] extracting evidence indicators based on the knowledge pedigree graph, wherein the evidence indicators comprise agent contribution, stance consistency, and response stability;

[0020] inputting the evidence indicators into an evidence theory fusion decision algorithm to calculate a credibility score of each medical insurance knowledge information;

[0021] identifying medical insurance knowledge information that exists in conflict according to the credibility score, and performing automatic resolution processing on the identified conflicting entries;

[0022] Output the medical insurance knowledge information recommended for adoption after conflict resolution processing.

[0023] Optionally, the collaborative decision-making model is constructed based on the core positions and semantic relationship graphs of each agent, a consensus scheme is reached through relationship reasoning and constraint solving, and the consensus scheme is compiled into structured medical insurance knowledge information, including:

[0024] Based on the core positions and semantic relationship graphs of each agent, a collaborative decision-making model is constructed, in which the core positions are mapped to decision variables, and the relationships in the semantic relationship graph are mapped to constraint conditions between decision variables.

[0025] Relationship reasoning and constraint solving are performed on the collaborative decision-making model, candidate consensus schemes that satisfy the constraint conditions are generated by adjusting the values of the decision variables, and the consensus scheme with the smallest conflict is selected from the candidate consensus schemes as the final consensus scheme.

[0026] The final consensus scheme is compiled into structured medical insurance knowledge information, and the structured format of the medical insurance knowledge information includes knowledge identification, content description, applicable scenario, and source citation fields.

[0027] Optionally, the collaborative decision-making model is constructed based on the core positions and semantic relationship graphs of each agent, including:

[0028] The core position of each agent is mapped to a decision variable in the collaborative decision-making model, and an initial state value is assigned to the decision variable.

[0029] The edge relationships connecting different core position nodes in the semantic relationship graph are analyzed, and the edge relationships are classified into conflict relationships, support relationships, or dependency relationships.

[0030] According to the classified edge relationship types, constraint conditions between corresponding decision variables are generated, the conflict relationship generates mutual exclusion constraint conditions, the support relationship generates collaborative constraint conditions, and the dependency relationship generates prerequisite constraint conditions.

[0031] All decision variables and corresponding constraint conditions are integrated into the same decision network to complete the construction of the collaborative decision-making model.

[0032] Optionally, the core positions and semantic relationship graphs of each agent are extracted by driving multiple agents to conduct multiple rounds of dialogue based on the initial query question, collecting dialogue data and performing fusion analysis, including:

[0033] The initial query question is input into all agents, triggering each agent to generate an initial response, and the dialogue rounds between agents are organized in a predetermined order.

[0034] Collect the response text generated by each agent in each round of dialogue to form a dialogue dataset;

[0035] Perform text parsing on the dialogue dataset to identify key argument sentences and viewpoint expressions in the response text of each agent as the core position of the corresponding agent;

[0036] By comparing the core positions of different agents, the overlap, difference and correlation between the positions are calculated to generate a semantic relationship graph, in which the nodes represent the core positions and the edges represent the similarity, conflict or complementary relationship between the positions.

[0037] Optionally, the text parsing on the dialogue dataset to identify key argument sentences and viewpoint expressions in the response text of each agent as the core position of the corresponding agent comprises:

[0038] Perform structured parsing on each agent's response text in the dialogue dataset to identify complete semantic units;

[0039] Based on a preset argument type library, identify the argument type of each semantic unit;

[0040] From the identified semantic units, filter out the argument type units containing core viewpoint keywords or having explicit judgment properties as key argument sentences;

[0041] Classify and integrate all key argument sentences of the same agent according to the corresponding argument type to form the core position of the corresponding agent.

[0042] Optionally, according to the confidence score, the medical insurance knowledge information with conflicts is identified, and automatic resolution processing is performed on the identified conflict items, comprising:

[0043] Compare the confidence score with a preset conflict detection threshold to filter out medical insurance knowledge information with a confidence score higher than the conflict detection threshold to form a candidate item set;

[0044] In the candidate item set, identify and mark medical insurance knowledge information with semantic conflicts to form a conflict item group;

[0045] According to the confidence score, sort the medical insurance knowledge information in the conflict item group;

[0046] Based on the sorting result, select the medical insurance knowledge information with the highest confidence score from the conflict item group as the medical insurance knowledge information recommended to be adopted after conflict resolution processing.

[0047] Optionally, the definition and configuration of a plurality of agents with differentiated medical insurance knowledge background and functions, and the configuration of exclusive prompt words for each agent to regulate the behavior mode, comprise:

[0048] define and configure a plurality of agents corresponding to different functional fields, including a medical insurance policy interpretation agent, a clinician agent, a patient agent, and an audit compliance agent;

[0049] For each agent, extract knowledge data from data sources related to the functional field to form the knowledge background of the corresponding agent, and design exclusive prompt content for each agent type;

[0050] associate the prompt content with each agent instance to ensure that the behavior pattern conforms to the preset specification.

[0051] The medical insurance knowledge base automatic construction method based on multi-agent collaboration provided in the present application can cover medical insurance knowledge scenarios from multiple dimensions by defining and configuring a plurality of agents with differentiated medical insurance knowledge backgrounds and functions and standardizing their behavior patterns; With the help of initial query-driven multi-round dialogue of agents and extraction of core positions and semantic relationship graphs, knowledge associations and different perspectives can be comprehensively captured; Through the consensus reached by the collaborative decision-making model and the compilation of structured knowledge entries, the accuracy and structuring of knowledge are ensured; The introduction of third-party verification agents to construct knowledge pedigree graphs, calculate credibility, and resolve conflicts improves the credibility and consistency of knowledge entries; The generation of a knowledge base and the construction of a question and answer system record user feedback, enabling effective application and feedback collection of knowledge; Based on feedback and regular discussions, manual intervention triggers updates, ensuring that the knowledge base can continuously learn and dynamically iterate, ultimately achieving the automatic and efficient construction of a medical insurance knowledge base, high credibility, precise and intelligent application, and dynamic adaptation to policy changes and user needs.

[0052] Further, the third-party verification agent is introduced to access the dialogue history records and consensus generation process data of all agents as traceability data of medical insurance knowledge information, to construct a knowledge pedigree graph with nodes as knowledge entries and edges as derivative or conflict relationships between entries based on these traceability data and medical insurance knowledge information, to extract agent contribution, position consistency and response stability as evidence indicators, to input the evidence theory fusion decision algorithm to calculate the credibility score of each knowledge entry, to identify and automatically resolve conflict entries, and to output the medical insurance knowledge information recommended for adoption. Through the independent intervention of the third-party verification agent and the support of the whole-process traceability data, combined with the relationship visualization of the knowledge pedigree graph and the quantitative evaluation of the evidence theory fusion decision algorithm, the credibility of the medical insurance knowledge information is accurately determined and conflicts are automatically resolved, which not only ensures the accuracy and consistency of the output knowledge entries, but also provides traceable evidence for the knowledge entries, further improving the reliability of the medical insurance knowledge base. BRIEF DESCRIPTION OF DRAWINGS

[0053] In order to make the technical scheme of the embodiments of the present application or the prior art clearer, the accompanying drawings needed in the embodiment or prior art description will be briefly introduced. Obviously, the accompanying drawings in the following description only need to be some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative work.

[0054] Figure 1 A flowchart of a medical insurance knowledge base automatic construction method based on multi-agent cooperation provided by an embodiment of the present application is shown in the figure.

[0055] Figure 2 A flowchart of a specific embodiment of a medical insurance knowledge base automatic construction method based on multi-agent cooperation provided by an embodiment of the present application is shown in the figure.

[0056] Figure 3 An interaction timing diagram of a specific embodiment of a medical insurance knowledge base automatic construction method based on multi-agent cooperation provided by an embodiment of the present application is shown in the figure.

[0057] Figure 4 A structural diagram of a medical insurance knowledge base automatic construction system based on multi-agent cooperation provided by an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0058] In order to make the technical scheme of the embodiments of the present application or the prior art clearer, the accompanying drawings needed in the embodiment or prior art description will be briefly introduced. Obviously, the accompanying drawings in the following description only need to be some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative work.

[0059] The core of the present application is to provide a medical insurance knowledge base automatic construction method based on multi-agent cooperation, and a flowchart of a specific embodiment thereof is shown in the figure. Figure 1 The method comprises the following steps.

[0060] S101, define and configure a plurality of agents with differentiated medical insurance knowledge background and functions, and configure a dedicated prompt word for each agent to regulate the behavior mode.

[0061] Optionally, step S101 can specifically include the following steps.

[0062] S1011, define and configure a plurality of agents corresponding to different functional fields, and the agents include a medical insurance policy interpretation agent, a clinical physician agent, a patient agent and an audit compliance agent.

[0063] S1012, for each agent, extract knowledge data from the data source related to the functional field to form the knowledge background of the corresponding agent, and design exclusive prompt word content for each agent type.

[0064] S1013, associate the prompt word content with each agent instance to ensure that the behavior mode conforms to the preset specification.

[0065] In the embodiment of the application, first, the work categories that need to be covered in the construction process of the medical insurance knowledge base are determined through step S1011, such as policy interpretation, clinical matching, patient consultation response, compliance verification, etc., and then the corresponding agents are defined according to these categories, and the work content of each agent is divided to ensure that the functions of different agents are not overlapped and can cover key links, such as when dealing with medical insurance drug-related knowledge construction, the policy interpretation agent is responsible for explaining the drug reimbursement policy, the clinical doctor agent is responsible for explaining the clinical application scene of the drug, the patient agent is responsible for simulating the common questions of patients about drug reimbursement, and the audit compliance agent is responsible for checking whether the drug reimbursement-related knowledge conforms to the rules.

[0066] Secondly, the data sources related to the functional field of each agent are determined through step S1012, such as the data source corresponding to the medical insurance policy interpretation agent is the medical insurance policy document, the data source corresponding to the clinical doctor agent is the clinical medical guideline, etc., and then useful knowledge data is extracted from these data sources to build a dedicated knowledge background for each agent; according to the function of each agent, exclusive prompt word content is designed, which includes role definition, behavior restriction sentence and response format requirement.

[0067] Finally, through step S1013, the prompt word is associated with the corresponding agent instance to ensure that the agent can automatically load the bound prompt word when starting to run; then during the running process of the agent, the prompt word will guide the agent to adjust the response generation process, so that the agent can produce responses strictly according to the role definition, behavior restriction and response format requirement in the prompt word when processing tasks.

[0068] S102, based on the initial query question, drive multiple agents to carry out multiple rounds of dialogue, collect dialogue data and perform fusion analysis to extract the core position of each agent and the semantic relationship graph.

[0069] Optionally, step S102 can specifically include the following steps:

[0070] S1021, input the initial query question into all agents to trigger each agent to generate an initial response, and organize the dialogue rounds between the agents according to a preset order.

[0071] The initial query question is the first medical insurance related question to start the dialogue of the agent. The multi-round dialogue is the multiple interactions of the agents around the initial query question.

[0072] S1022, collect the response texts generated by each agent in each round of dialogue to form a dialogue data set.

[0073] The dialogue data set is all the response contents generated by the agents in each round of dialogue.

[0074] S1023, text parsing is performed on the dialogue data set to identify the key argument sentences and viewpoint expressions in the response texts of each agent as the core positions of the corresponding agents.

[0075] The core position is the most critical viewpoint and argument expressed by each agent in the dialogue.

[0076] The step S1023 specifically includes the following processes: structurally parsing each agent response text in the dialogue data set to identify complete semantic units; identifying the argument type of each semantic unit based on a preset argument type library; filtering the argument type units containing core viewpoint keywords or having explicit judgment properties from the identified semantic units as key argument sentences; and classifying and integrating all key argument sentences of the same agent according to the corresponding argument type to form the core position of the corresponding agent.

[0077] S1024, by comparing the core positions of different agents, the overlap, difference and correlation between the positions are calculated to generate a semantic relationship graph, in which the nodes represent the core positions and the edges represent the similar, conflicting or complementary relationships between the positions.

[0078] The semantic relationship graph is a tool for displaying the correlation between the core positions of different agents in a graphical form.

[0079] In the embodiments of the present application, first, a specific initial query question is determined through step S1021, and this question is sent to all configured agents to trigger each agent to independently generate a first response according to its own knowledge background and function. The preset order of the agent dialogue is set in advance, such as arranging the dialogue rounds in the order of: medical insurance policy interpretation agent, clinical doctor agent, patient agent, and audit compliance agent. In each subsequent round of dialogue, the current agent will carefully check the response contents of all previous agents and generate a supplementary response by combining the relevant information according to its own function.

[0080] Secondly, step S1022 clarifies the scope of content to be collected, specifying all response texts generated by each agent in each round of dialogue. An automatic collection method is used to record the response text of each agent during the dialogue process, including the agent type, round, response time, and specific text content. All collected information is then categorized and organized according to dialogue round and agent type to form a well-structured dialogue dataset, facilitating subsequent text parsing and analysis.

[0081] Then, in step S1023, the response text of each agent in the dialogue dataset is structured and parsed, breaking down the originally coherent text into multiple complete semantic units. Referring to the preset claim type library, it is determined whether each semantic unit belongs to the category of factual statement, rule reference, opinion judgment, or appeal expression. Then, from these classified semantic units, units containing core opinion keywords, such as "reimbursement conditions," "necessary," "compliant," etc., or units with clear judgment properties are selected as key claim statements. All key claim statements of the same agent are classified and integrated according to the corresponding claim type to form the core position of the agent.

[0082] Finally, the core positions of all agents are collected in step S1024, and each agent's core position is treated as an independent analysis object. The overlap, difference and correlation between positions are calculated by text comparison. A semantic relationship graph is generated based on the calculation results. A core position is represented by a node. If there is the same content between positions, they are connected by edges labeled as similar. If there is opposite content, they are connected by edges labeled as conflicting. If there is complementary content, they are connected by edges labeled as complementary.

[0083] Specifically, the core positions of all agents are collected, and each agent's core position is broken down into several independent statements with complete meaning, serving as the basic unit of computation. Then, the overlap, difference, and correlation between the core positions of different agents are calculated, where the formula for calculating overlap is:

[0084]

[0085] The number of identical statements between core positions A and B refers to the number of independent statements in the core positions of the two agents that have completely identical meanings, and the total number of statements of core positions A / B refers to the total number of all independent statements after the core positions of the corresponding agents are broken down.

[0086] The formula for calculating the degree of difference is:

[0087]

[0088] Among them, the number of different expressions of core positions A and B refers to the number of independent expressions with completely different meanings and no relation in the core positions of the two agents.

[0089] The calculation formula of the correlation degree is:

[0090]

[0091] Among them, the number of supplementary expressions of core positions A and B refers to the number of independent expressions of one agent that can perfect the explanation of the expressions of another agent.

[0092] Finally, according to the calculated overlap degree, difference degree and correlation degree, the relationship between the core positions of different agents is judged. When the overlap degree reaches a certain degree, it is determined as a similar relationship, when the difference degree reaches a certain degree, it is determined as a conflict relationship, and when the supplementary expression accounts for a high proportion in the correlation degree, it is determined as a supplementary relationship, and then a semantic relationship graph is generated, with nodes representing the core positions of each agent and edges marked as similar, conflict or supplementary representing the corresponding relationship between the positions.

[0093] In actual application, in the process of constructing a medical insurance knowledge base, the initial query question set by the staff is "What are the conditions for B-class drugs to be included in medical insurance reimbursement?", which is input into the medical insurance policy interpretation, clinical doctors, patients, and audit compliance four agents. Three rounds of dialogue are organized in the order of policy interpretation, clinical doctors, patients, and audit compliance. The clinical doctor agent supplements the influence of the clinical application scene of B-class drugs on reimbursement in the second round of reference to the initial response of the policy interpretation agent. The audit compliance agent supplements the compliance requirements that need to be met during reimbursement in the third round of reference to the responses in the previous two rounds. The staff collects all the response texts of all agents in the three rounds of dialogue to form a dialogue data set. After processing the data set, the core positions of each agent are extracted, such as the core position of the policy interpretation agent is "B-class drug reimbursement needs to meet the insured duration and drug indication". Then the relationship between the positions is analyzed, and a semantic relationship graph with similar, conflict or supplementary relationship is generated.

[0094] The present application drives the multi-round dialogue of the agent through the initial query question and collects and analyzes the data, which can not only enable each agent to fully express its own point of view and comprehensively cover the medical insurance knowledge dimensions related to the initial question, but also clearly extract the core positions of each agent and clarify the relationship between different positions, providing comprehensive and orderly basic data support for subsequent construction of a collaborative decision-making model and formation of a consensus solution.

[0095] S103, based on the core positions of each agent and the semantic relationship graph, a collaborative decision-making model is constructed, a consensus solution is reached through relationship reasoning and constraint solving, and the consensus solution is compiled into structured medical insurance knowledge information.

[0096] Optionally, step S103 can specifically include the following steps:

[0097] S1031, constructing a collaborative decision-making model based on the core positions of each agent and the semantic relationship graph, wherein the core positions are mapped as decision variables in the collaborative decision-making model, and the relationships in the semantic relationship graph are mapped as constraint conditions between the decision variables.

[0098] The collaborative decision-making model is a tool for integrating the core positions and relationships of each agent to assist in reaching an agreed solution.

[0099] The step S1031 specifically includes the following processes: mapping the core position of each agent as a decision variable in the collaborative decision-making model, and assigning an initial state value to the decision variable; analyzing the edge relationships connecting different core position nodes in the semantic relationship graph, and classifying the edge relationships into conflict relationships, support relationships or dependency relationships; generating corresponding constraint conditions between decision variables according to the classified edge relationship types, wherein the conflict relationship generates mutual exclusion constraint conditions, the support relationship generates collaborative constraint conditions, and the dependency relationship generates prerequisite constraint conditions; integrating all decision variables and corresponding constraint conditions into the same decision network to complete the construction of the collaborative decision-making model.

[0100] S1032, performing relationship reasoning and constraint solving on the collaborative decision-making model, generating candidate consensus solutions that satisfy the constraint conditions by adjusting the values of the decision variables, and selecting the solution with the least conflict from the candidate consensus solutions as the final consensus solution.

[0101] The candidate consensus solution is a unified view generated under the model constraints, and the solution with the least conflict is the unified view with the least contradiction between positions in the candidate solution.

[0102] S1033, compiling the final consensus solution into structured medical insurance knowledge information, wherein the structured format of the medical insurance knowledge information includes knowledge identification, content description, applicable scenario and source citation fields.

[0103] The knowledge identification is a unique code used to distinguish different knowledge items, the content description is a specific textual description of the consensus solution, the applicable scenario is a specific context in which the knowledge item can be applied, and the source citation is information explaining the source of the knowledge item. The consensus solution is a unified view formed after the coordination of the core positions of each agent. The structured medical insurance knowledge information is a medical insurance knowledge unit arranged in a fixed format.

[0104] In the embodiments of the present application, first, the core position of each agent is mapped to a decision variable in the collaborative decision-making model through step S1031, and an initial state value is assigned to each variable; the edge relationship connecting different core position nodes in the semantic relationship graph is analyzed and classified into conflict, support or dependency relationship, then the corresponding decision variable constraint condition is generated according to the classified edge relationship type, the conflict relationship generates mutual exclusion constraint condition, such as variable A and variable B cannot be true at the same time, the support relationship generates collaborative constraint condition, such as variable C and variable D need to be true at the same time, the dependency relationship generates prerequisite constraint condition, such as variable E is true, variable F needs to be true first; all decision variables and corresponding constraint conditions are integrated into the same decision network to complete the construction of the collaborative decision-making model.

[0105] Secondly, through step S1032, relationship reasoning is performed on the constructed collaborative decision-making model, and the logical possibility of variable value is deduced according to the constraint conditions between the decision variables; constraint solving is performed, and multiple candidate consensus schemes satisfying all constraint conditions are generated by adjusting the values of the decision variables; finally, conflict analysis is performed on all candidate consensus schemes, the number of contradictions between different positions in each scheme is compared, and the scheme with the least contradiction is selected as the final consensus scheme.

[0106] Finally, through step S1033, the fixed format of the structured medical insurance knowledge information is determined, and the four fields including knowledge identification, content description, applicable scenario and source reference are determined; for the final consensus scheme, the corresponding content is filled in each field: the knowledge identification generates a unique code according to the preset rule, for example, in the format of “knowledge type-subject abbreviation-sequence number”; the content description directly fills in the specific expression of the final consensus scheme; the applicable scenario describes which medical insurance related judgment or consultation scenarios it can be used for combined with the theme of the consensus scheme; the source reference marks the provenance of each agent core position on which the consensus scheme is based, such as agent type and dialogue record number. Finally, the contents of the four fields are integrated to form a complete structured medical insurance knowledge information, for example, arranging the contents of each field in order to ensure the format is standard and the information is complete.

[0107] In practical applications, during the construction of the medical insurance knowledge base, staff members built a collaborative decision-making model based on the core positions and semantic relationship graphs of four agents regarding "reimbursement for medical treatment in other locations." The policy interpretation agent's positions, such as "advance registration is required," were mapped as decision variables and assigned initial values. The supporting relationships between "advance registration is required" and "registration materials must be authentic" in the semantic relationship graph were used to generate collaborative constraints. Reasoning and solving were performed on the model, generating three candidate solutions: "advance registration is required for medical treatment in other locations, materials must be authentic, and reimbursement is based on the prescribed ratio." The solution with the least conflict, "advance registration is required for medical treatment in other locations, materials must be authentic, and reimbursement ratio is appropriate to the condition," was selected as the final consensus solution. This solution was then compiled in a structured format, identified as "YK-YD-002," with the content description being the final consensus solution, applicable to the scenario of "determination of medical insurance reimbursement process for medical treatment in other locations," and the source citations indicated by the dialogue record numbers of each agent.

[0108] This application constructs a collaborative decision-making model and performs relational reasoning and constraint solving, which can effectively coordinate the core positions of each intelligent agent to form a consensus scheme with minimal conflict. The model is then compiled into structured knowledge items, which not only ensures the consistency and rationality of knowledge, but also makes medical insurance knowledge more regular, facilitating subsequent verification, storage and application.

[0109] S104. Introduce a third-party verification agent to construct a knowledge genealogy graph based on the medical insurance knowledge information and the source data of the medical insurance knowledge information. Calculate the credibility of the medical insurance knowledge information through evidence theory fusion decision algorithm, automatically resolve conflicts in medical insurance knowledge information, and output the recommended medical insurance knowledge information to be adopted.

[0110] Optionally, step S104 may specifically include the following steps:

[0111] S1041. Introduce a third-party verification agent, and access the dialogue history and consensus generation process data of all agents through the third-party verification agent as traceability data for medical insurance knowledge information.

[0112] The third-party verification agent is a virtual processing unit independent of the participating dialogue agents, responsible for verifying medical insurance knowledge information. Source data refers to historical dialogue records and consensus generation process data that can trace the origin of medical insurance knowledge information.

[0113] S1042. Construct a knowledge genealogy graph based on the medical insurance knowledge information and the source data of the medical insurance knowledge information. In the knowledge genealogy graph, nodes represent the medical insurance knowledge information, and edges represent the derivative or conflict relationships between items.

[0114] The knowledge genealogy graph is a graphical representation of medical insurance knowledge information and the relationships between items, using nodes and edges. Nodes represent knowledge items, and edges represent derivative or conflict relationships.

[0115] S1043. Extract evidence indicators based on the knowledge genealogy graph, wherein the evidence indicators include agent contribution, position consistency and response stability.

[0116] S1044. Input the evidence indicators into the evidence theory fusion decision algorithm to calculate the credibility score of each piece of medical insurance knowledge information.

[0117] The evidence-theory fusion decision algorithm is a method for calculating the credibility of knowledge items by integrating multiple evidence indicators. The credibility score is a numerical measure of the reliability of medical insurance knowledge information. The recommended medical insurance knowledge information is the knowledge unit recommended for use after verification and conflict resolution.

[0118] S1045. Identify conflicting medical insurance knowledge information based on the credibility score, and perform automatic resolution processing on the identified conflicting items.

[0119] Specifically, step S1045 includes the following process: comparing the credibility score with a preset conflict detection threshold, filtering out medical insurance knowledge information with a credibility score higher than the conflict detection threshold to form a candidate item set; identifying and marking medical insurance knowledge information with semantic conflicts in the candidate item set to form a conflict item group; sorting the medical insurance knowledge information in the conflict item group according to the credibility score; and selecting the medical insurance knowledge information with the highest credibility score from the conflict item group as the medical insurance knowledge information recommended for adoption after conflict resolution processing based on the sorting result.

[0120] S1046. Output the recommended medical insurance knowledge information after conflict resolution.

[0121] Medical insurance knowledge information Medical insurance knowledge information Medical insurance knowledge information Medical insurance knowledge information In the embodiments of this application, such as Figure 2 As shown, firstly, step S1041 identifies and introduces a third-party verification agent independent of the previously participating dialogue agents. This agent does not participate in the initial dialogue and consensus generation, but is only responsible for verifying knowledge items. Access permissions are configured for the third-party verification agent, enabling it to obtain the dialogue history of all participating dialogue agents, including the response content of each agent in each round. It can also obtain consensus generation process data, such as the core position extraction results, semantic relationship graph, constraints of the collaborative decision-making model, and candidate consensus schemes. Finally, these obtained dialogue history and consensus generation process data are integrated as source data for subsequent medical insurance knowledge information, providing a source basis for the credibility assessment of knowledge items.

[0122] Secondly, collect the medical insurance knowledge information that needs to be constructed into the knowledge pedigree graph through step S1042. These entries are structured knowledge entries generated through step S103. Secondly, sort the provenance data of these knowledge entries, analyze the relationship between each entry from the provenance data, and determine the derivation relationship if an entry is further refined or supplemented based on the content of another entry. Determine the conflict relationship if the core content of two entries is opposite and cannot exist at the same time. Finally, draw the knowledge pedigree graph with medical insurance knowledge information as the node and the derivation relationship or conflict relationship as the edge. For example, use a circular node to represent each knowledge entry, use a straight line with a "derivation" label to connect the nodes with a derivation relationship, and use a dashed line with a "conflict" label to connect the nodes with a conflict relationship.

[0123] Then, determine the medical insurance knowledge information that needs to extract evidence indicators from the knowledge pedigree graph through step S1043; extract the agent contribution degree by checking the participation of each agent in the generation process of the entry in the provenance data, such as whether to propose key points or participate in consensus discussion, etc. Evaluate the importance of each agent and obtain the agent contribution degree of the entry, then extract the consistency of the position, analyze whether the core points of the entry of each participating agent in the provenance data are consistent, if the views of most agents are unified, the consistency of the position is high, finally extract the response stability, check whether the response of each agent to the related content of the entry in multiple rounds of dialogue is stable, if the response content is not obviously contradictory, the response stability is high.

[0124] Next, determine the calculation logic of the evidence theory fusion decision algorithm through step S1044, take the agent contribution degree score, the position consistency score, and the response stability score as input evidence indicators, and the calculation formula is:

[0125]

[0126] Among them, , , The weights of the three evidence indicators are 1.

[0127] For example, if the agent contribution degree score is 0.8, the position consistency score is 0.7, and the response stability score is 0.9, and the weight is , then the credibility score is: Output the credibility score of each medical insurance knowledge information calculated, as the basis for subsequent conflict identification.

[0128] Then, in step S1045, a conflict detection threshold is set according to the knowledge quality requirement, and the credibility score of each medical insurance knowledge information is compared with the threshold to screen out entries with a score higher than the conflict detection threshold, and these entries constitute a candidate entry set; semantic analysis is performed on the entries in the candidate entry set to identify entries with mutually contradictory core contents, and these entries are marked as a conflict entry group; then, the entries in the conflict entry group are sorted from high to low according to the credibility score, and based on the sorting result, the entry with the highest credibility score in the conflict entry group is selected as the medical insurance knowledge information recommended for adoption after conflict resolution processing.

[0129] Finally, in step S1046, all results after conflict resolution processing are collected, including the high-credibility entries screened out from the conflict entry group and the conflict-free candidate entries, and final verification is performed on these entries to confirm that they have no semantic conflicts and the credibility scores meet the requirements, and the entries that pass the verification are output to form the medical insurance knowledge information recommended for adoption.

[0130] In actual application, a third-party verification agent is introduced by the staff, which has accessed all agent dialogue records and consensus generation process data about "medical reimbursement in different places"; a knowledge pedigree graph is constructed based on 5 knowledge entries such as "medical reimbursement in different places requires advance filing" and traceability data, wherein "advance filing is required" and "advance filing can be handled online" are derivative relations, and "reimbursement ratio 50%" and "reimbursement ratio 60%" are conflict relations; the agent contribution degree, stand consistency and response stability are extracted as evidence indexes, which are input into an algorithm to calculate the credibility scores of the entries; the scores are compared with a preset threshold 0.7 to screen out 4 candidate entries, and the scores 0.72 corresponding to "reimbursement ratio 50%" and the score 0.68 corresponding to "reimbursement ratio 60%" are marked as a conflict group, wherein the score 0.68 corresponding to "reimbursement ratio 60%" is lower than the threshold and is not selected, and this is only an example of a conflict scenario, and "reimbursement ratio 50%" with a higher score is selected after sorting; finally, the knowledge entries recommended for adoption are output, including "medical reimbursement in different places requires advance filing (credibility 0.83)" and "reimbursement ratio 50% (credibility 0.72)".

[0131] The present application can objectively evaluate the reliability of medical insurance knowledge information by introducing a third-party verification agent, constructing a knowledge pedigree graph using traceability data, combining evidence theory fusion decision algorithm to calculate credibility, and automatically resolving conflicts to output the recommended adoption entries, which not only ensures traceability of the source, but also improves the accuracy and consistency of the knowledge, and provides a high-quality knowledge unit for the medical insurance knowledge base.

[0132] S105, storing the medical insurance knowledge information recommended for adoption to generate a medical insurance knowledge base.

[0133] The medical insurance knowledge base is a database set storing these structured knowledge entries.

[0134] In the embodiments of the present application, first, it is confirmed through step S1051 that the medical insurance knowledge information recommended to be adopted has completed the structured processing, each item contains complete knowledge identification, content description, applicable scenario and source reference field; a suitable database, such as a relational database, is selected, the database table structure is designed according to the structured format, the columns corresponding to the item fields are set, including the knowledge identification column, the content description column, the applicable scenario column and the source reference column; then each structured knowledge item is inserted into the database table according to the field corresponding relationship, after the storage of all the recommended to be adopted items is completed, the database is indexed and optimized, such as the knowledge identification and the applicable scenario column are indexed, and the medical insurance knowledge base which can be efficiently queried is generated.

[0135] S106, constructing a retrieval enhancement generated question and answer system based on the medical insurance knowledge base, receiving user questions and intelligently answering, and recording user feedback data generated in the interactive process.

[0136] Optionally, step S106 can specifically include the following steps:

[0137] S1061, storing the medical insurance knowledge information recommended to be adopted into the database according to the structured format, and generating a medical insurance knowledge base.

[0138] S1061, constructing a retrieval enhancement generated question and answer system based on the medical insurance knowledge base, the retrieval enhancement generated question and answer system including a retrieval component and a generation component.

[0139] S1062, receiving user questions through the retrieval enhancement generated question and answer system, sequentially calling the retrieval component and the generation component, generating and returning intelligent answers.

[0140] S1063, recording the evaluation data of the user on the intelligent answers and the subsequent interactive behaviors of the user in the process of the interaction between the retrieval enhancement generated question and answer system and the user, and forming user feedback data.

[0141] Among them, the retrieval enhancement generated question and answer system is an intelligent question and answer tool combining knowledge retrieval and natural language generation functions. The user question is the input content of the user on the medical insurance problem. The intelligent answer is the natural language reply generated by the system for the user question. The user feedback data is the evaluation of the reply and the subsequent operation information generated when the user interacts with the system.

[0142] In the embodiments of the present application, first, the overall architecture of the retrieval-enhanced generation question-answering system is determined based on the generated medical insurance knowledge base in step S1061, and the function boundaries and data interaction modes of the retrieval component and the generation component are clarified; the retrieval component is developed, and technologies such as keyword matching and semantic similarity calculation are adopted to realize the function of searching for related knowledge items from the knowledge base according to user questions; then the generation component is developed, and a pre-trained language model is adopted to fine-tune it in the medical insurance field, so that it can generate fluent and accurate natural language responses based on the retrieved knowledge items; the retrieval component and the generation component are integrated into the same system, and the component calling logic is set to retrieval first and then generation, thereby completing the construction of the retrieval-enhanced generation question-answering system.

[0143] Then, the medical insurance question input by the user is received through the interactive interface of the retrieval-enhanced generation question-answering system in step S1062; the system calls the retrieval component first according to the preset logic, converts the user question into a retrieval request, and the retrieval component searches for related knowledge items from the medical insurance knowledge base according to the request; then the system calls the generation component, delivers the list of retrieved knowledge items to the generation component as input, and the generation component converts the item information into a colloquial response in accordance with the user's understanding habits based on the content of these items and in combination with natural language generation rules; finally, the system returns the generated intelligent response to the user through the interactive interface, and completes one question-answering interaction.

[0144] Finally, the feedback entry such as the satisfaction evaluation button, the follow-up input box and the related link entry is set on the interactive interface while the intelligent response is returned by the retrieval-enhanced generation question-answering system in step S1063; after the user views the response, the system records the user's operation: if the user clicks the evaluation button, the corresponding satisfaction evaluation data is recorded, such as "evaluation: satisfied" if the "satisfied" button is clicked; if the user makes a follow-up question, such as inputting "how to query the designated hospital", the follow-up content is recorded; if the user clicks the related link, the clicked link name and time are recorded; the recorded evaluation data and subsequent interaction behaviors are integrated and stored in the format of "user ID-question content-answer content-feedback data-interaction time", thereby forming the user feedback data.

[0145] In practical application, in a certain medical insurance service platform project, staff members processed 200 suggested items, such as "Advance registration is required for out-of-town medical treatment (YK-YD-001)" and "Reimbursement for Category C medical treatment items requires designated hospitals (YK-C-003)," and stored them in a structured format in a MySQL database to generate a medical insurance knowledge base. Based on this knowledge base, a retrieval-enhanced question-and-answer system was built. The retrieval component used the TF-IDF keyword matching algorithm, and the generation component used a lightweight natural language generation model. When user A entered "How to register for out-of-town medical treatment" into the system, the system called the retrieval component to find the YK-YD-001 item, and the generation component converted it into the response "Advance registration is required for out-of-town medical treatment. You can do it through the online APP or offline service window" and returned it. At the same time, the system recorded user A's "satisfaction" rating for the response and subsequent actions such as clicking the "Online APP Processing Process" link, forming user feedback data.

[0146] This application achieves centralized management and intelligent question-and-answer functions for medical insurance knowledge by storing high-quality knowledge entries to generate a medical insurance knowledge base and constructing a retrieval-enhanced question-and-answer system. It can quickly and accurately respond to users' medical insurance inquiries, and the recorded user feedback data provides direct user demand basis for subsequent updates and optimizations of the knowledge base, forming a closed loop of "knowledge storage - intelligent application - feedback collection".

[0147] S107. Based on the user feedback data, automatically identify the medical insurance knowledge information in the medical insurance knowledge base that needs to be updated, trigger a re-discussion and update process, and realize the continuous learning and dynamic iterative update of the medical insurance knowledge base by periodically launching new topic discussions and human expert intervention for correction.

[0148] Optionally, step S107 may specifically include the following steps:

[0149] S1071. Analyze the user feedback data and automatically identify the medical insurance knowledge information in the medical insurance knowledge base that needs to be updated.

[0150] Among them, the medical insurance knowledge information that needs to be updated refers to the knowledge units in the knowledge base that need to be adjusted due to user feedback or changes in the field.

[0151] S1072. For medical insurance knowledge information that needs to be updated, automatically generate an update request and trigger a multi-agent re-discussion process.

[0152] The re-discussion process is a process in which multiple agents engage in dialogue and reach a consensus again regarding the items that need to be updated.

[0153] S1073. Regularly generate new topic discussion requests based on the development dynamics in the medical insurance field, and initiate a multi-agent discussion process for the new topic.

[0154] The new topic discussion request is a new content demand dynamically generated according to the medical insurance field and needs to be discussed by the agent.

[0155] S1074, when the re-discussion process or the new topic discussion process cannot reach an effective consensus, introducing an artificial expert to make corrections.

[0156] S1075, integrating the medical insurance knowledge generated after re-discussion, new topic discussion or artificial expert correction into the medical insurance knowledge base, and completing the dynamic iterative update of the medical insurance knowledge base.

[0157] In the embodiment of the application, first, all user feedback data is collected and arranged through step S1071, and classified according to the format of “corresponding knowledge item-feedback content-feedback times”; the keyword frequency statistics and satisfaction analysis method are used to identify the items that need to be updated: for the items with a satisfaction degree lower than a preset value, directly marked as needing to be updated; for the items with a large number of follow-up questions, the keywords in the follow-up questions are extracted to determine whether the item lacks these key information, and if so, marked as needing to be updated; all marked items needing to be updated are arranged into a list to prepare for the subsequent update process.

[0158] Secondly, for each knowledge item that needs to be updated, an update request is automatically generated through step S1072, which includes the existing content of the item, user feedback specific questions and supplement key information, optimization expression and other update requirements; the update request is sent to all agents involved in the generation of the item, triggering the re-discussion process: first, the update request drives the multi-round dialogue of the agent, collects dialogue data to extract core positions, builds a collaborative decision-making model to reach a new consensus, and generates an updated knowledge item.

[0159] Then, through step S1073, a regular triggering time is set, information from official websites of the medical insurance bureau, authoritative medical platforms and other channels is crawled through network crawler technology, and the development trend of the medical insurance field is collected; the collected development trend is screened and arranged, the content with timeliness and practicality is determined as a new topic, a new topic discussion request is generated, which includes the specific content of the new trend and the discussion direction of “discussion knowledge item structure, such as reimbursement conditions, proportion, applicable scenarios”; finally, the new topic discussion request is sent to the relevant agent, and the discussion process of the new topic is started, and the new medical insurance knowledge information is generated according to the steps of steps S102-S103.

[0160] Then, the discussion process is monitored when the agent re-discusses or discusses a new topic through step S1074, and a consensus judgment condition is set: if the discussion round reaches the preset upper limit and there is still semantic conflict, or the identification rate of all agents on the core content does not reach the preset value, it is determined that an effective consensus cannot be reached; a consensus dispute report is automatically generated, and the dispute report is sent to a preset artificial expert in the medical insurance field. The expert analyzes and adjusts the dispute focus in combination with his own professional knowledge, relevant policy documents and clinical actual situation, and determines the final knowledge item content.

[0161] Finally, the new items formed by re-discussion, the items generated by new topic discussion and the items revised by artificial experts are integrated into the original medical insurance knowledge base through step S1075, replacing the old items or adding new items, and the dynamic iterative update of the knowledge base is completed.

[0162] In actual application, in the medical insurance knowledge base maintenance project, the staff analyzes user feedback data and finds that 30% of users mark "unsatisfied" and frequently ask questions for the "C-class diagnosis and treatment item reimbursement process" item, which is identified as an item that needs to be updated; the system automatically generates an update request, triggers multi-agent re-discussion, and supplements the process steps of "online submission-application-audit-reimbursement to account"; the system collects medical insurance field dynamics on the 10th of each month and finds that "new H-class drug medical insurance policy is introduced", generates a new topic request, and forms the drug reimbursement condition item after discussion by the agent, but there is a dispute on the "reimbursement time limit" and consensus cannot be reached; the artificial expert is introduced, and the expert determines the time limit as "2 years of insurance" according to the policy document; finally, the updated "C-class diagnosis and treatment process" item and the newly added "H-class drug reimbursement condition" item (including the time limit revised by the expert) are integrated into the medical insurance knowledge base, and the iterative update is completed.

[0163] The present application identifies the items that need to be updated through user feedback and triggers re-discussion, combines regular new topic discussion and artificial expert revision, realizes the continuous learning and dynamic iteration of the medical insurance knowledge base, makes the knowledge base respond to user demand changes and medical insurance field development dynamics in a timely manner, and always maintains the accuracy, completeness and timeliness of the knowledge, thereby providing reliable knowledge support for the intelligent question answering system.

[0164] The following is a complete example for steps S101-S107, as shown in Figure 3

[0165] Firstly, in the construction of the medical insurance knowledge base, four types of agents including medical insurance policy interpretation, clinical doctors, patients and audit compliance are defined and configured, knowledge backgrounds are extracted from data sources such as medical insurance policy documents and clinical guidelines, and exclusive prompt words including role definition, behavior restriction and response format are designed for each agent, such as requiring the audit compliance agent to "check information only according to official documents", and then the prompt words are bound to the corresponding agents. ​

[0166] Next, the system initiates the initial topic "What is the hospitalization deductible line for residents in City B?" triggering a multi-round dialogue: the medical insurance policy interpretation agent answers "700 yuan", the patient agent questions "social platforms say it's 650 yuan", the audit compliance agent checks the latest documents to confirm "700 yuan" and points out that the platform information is outdated. Collect dialogue data to form a dataset, analyze and extract the core positions of each agent, calculate the position relationship, and generate a semantic relationship graph with "700 yuan" and "650 yuan" as nodes, marking the conflict relationship.

[0167] Secondly, a collaborative decision-making model is constructed, the core positions are mapped to decision variables, and the conflict relationship is set as a mutual exclusion constraint. After reasoning, the audit agent selects "700 yuan" according to the official document as the final consensus, and compiles it into a structured knowledge entry: knowledge identifier "YK-B-012", content "The hospitalization deductible line for residents in City B is 700 yuan", source "B City Medical Insurance Bureau Document No. 2025".

[0168] Then, a third-party verification agent is introduced, which accesses the dialogue history and consensus data as traceability data, constructs a knowledge pedigree graph, extracts evidence indicators such as agent contribution and position consistency, and inputs the algorithm to calculate a credibility score of 0.93. Since there are no conflict entries, the suggestion is to adopt the entry.

[0169] Subsequently, the entry is stored in the database in a structured format to generate a knowledge base, and based on this, a search-enhanced question-answering system is constructed; when the user asks about the deductible line, the system retrieves the entry and generates a natural language response, while recording the user's "satisfaction" evaluation data.

[0170] Finally, analyze the feedback data, there is no need to update this entry; regularly collect medical insurance dynamics, no relevant policy adjustments are found; there is no current discussion controversy, no need for expert intervention. If there is a policy change in the future, it will trigger a re-discussion and update process, realizing the continuous iteration of the knowledge base.

[0171] Figure 4 A structural diagram of a specific implementation of a medical insurance knowledge base automatic construction system based on multi-agent collaboration provided by the embodiments of the present application, referring to Figure 4 The system can include:

[0172] The configuration module 41 is configured to define and configure a plurality of agents with differentiated medical insurance knowledge backgrounds and functions, and to configure exclusive prompt words for each agent to regulate behavior patterns;

[0173] The analysis module 42 is configured to drive multiple agents to conduct multi-round dialogue based on an initial query question, collect dialogue data and perform fusion analysis to extract the core positions of each agent and a semantic relationship graph;

[0174] The reasoning module 43 is configured to construct a collaborative decision-making model based on the core positions of each agent and the semantic relationship graph, reach a consensus scheme through relationship reasoning and constraint solving, and compile the consensus scheme into structured medical insurance knowledge information.

[0175] The verification module 44 is configured to introduce a third-party verification agent, construct a knowledge pedigree graph based on the medical insurance knowledge information and traceability data of the medical insurance knowledge information, calculate the credibility of the medical insurance knowledge information through an evidence theory fusion decision algorithm, automatically resolve conflicts of medical insurance knowledge information that exist, and output medical insurance knowledge information that is recommended to be adopted.

[0176] The generation module 45 is configured to store the medical insurance knowledge information that is recommended to be adopted to generate a medical insurance knowledge base.

[0177] The medical insurance knowledge base automatic construction system based on multi-agent collaboration of the embodiments of the present application is used to implement the foregoing medical insurance knowledge base automatic construction method based on multi-agent collaboration, and therefore the specific embodiments in the medical insurance knowledge base automatic construction system based on multi-agent collaboration can be seen from the foregoing embodiment part of the medical insurance knowledge base automatic construction method based on multi-agent collaboration. The specific embodiments can be referred to the description of the respective embodiment parts, and will not be described here again.

[0178] The medical insurance knowledge base automatic construction method and system based on multi-agent collaboration provided in the present application are described in detail above. The principles and embodiments of the present application are described by applying specific examples in this paper, and the above embodiment descriptions are only used to help understand the method of the present application and its core idea. It should be pointed out that, for ordinary skilled persons in the technical field, some improvements and modifications can be made to the present application without departing from the principles of the present application, and these improvements and modifications also fall within the protection scope of the present application.

Claims

1. A medical insurance knowledge base automatic construction method based on multi-agent cooperation, characterized in that, The method comprises the following steps: Defining and configuring a plurality of agents with differentiated medical insurance knowledge background and functions, and configuring each agent with a dedicated prompt word to regulate the behavior mode; Based on the initial query question, drive multiple agents to conduct multiple rounds of dialogue, collect dialogue data and perform fusion analysis to extract the core position and semantic relationship graph of each agent; Based on the core position and semantic relationship graph of each agent, a collaborative decision-making model is constructed, a consensus scheme is reached through relationship reasoning and constraint solving, and the consensus scheme is compiled into structured medical insurance knowledge information; Introduce a third-party verification agent, construct a knowledge pedigree graph based on the medical insurance knowledge information and the traceability data of the medical insurance knowledge information, calculate the credibility of the medical insurance knowledge information through evidence theory fusion decision algorithm, automatically resolve the conflict of the medical insurance knowledge information with conflict, and output the medical insurance knowledge information recommended for adoption; The medical insurance knowledge information recommended for adoption is stored to generate a medical insurance knowledge base; The collaborative decision-making model is constructed based on the core position and semantic relationship graph of each agent, the core position is mapped as a decision variable in the collaborative decision-making model, and the relationship in the semantic relationship graph is mapped as a constraint condition between decision variables; Relationship reasoning and constraint solving are performed on the collaborative decision-making model, a candidate consensus scheme that satisfies the constraint condition is generated by adjusting the value of the decision variable, and the candidate consensus scheme with the smallest conflict is selected as the final consensus scheme; The final consensus scheme is compiled into structured medical insurance knowledge information, and the structured format of the medical insurance knowledge information includes knowledge identification, content description, applicable scenario and source citation field; The collaborative decision-making model is constructed based on the core position and semantic relationship graph of each agent, which comprises: Map the core position of each agent to a decision variable in the collaborative decision-making model, and assign an initial state value to the decision variable; Analyze the edge relationship connecting different core position nodes in the semantic relationship graph, and classify the edge relationship into conflict relationship, support relationship or dependency relationship; According to the classified edge relationship type, generate corresponding constraint conditions between decision variables, the conflict relationship generates mutual exclusion constraint condition, the support relationship generates cooperative constraint condition, and the dependency relationship generates premise constraint condition; Integrate all decision variables and corresponding constraint conditions into the same decision network to complete the construction of the collaborative decision-making model. Further comprising:

2. The method of claim 1, wherein, Based on the medical insurance knowledge base, a retrieval enhancement generation question and answer system is constructed, which receives user questions and provides intelligent answers, and records user feedback data generated during the interaction process; Based on the user feedback data, automatically identify the medical insurance knowledge information that needs to be updated in the medical insurance knowledge base, trigger the re-discussion and update process, and realize the continuous learning and dynamic iterative update of the medical insurance knowledge base through the methods of regularly starting new topic discussion and manual expert intervention correction. ​ 3. The method of claim 1, wherein, The third-party verification agent is introduced, a knowledge pedigree graph is constructed based on the medical insurance knowledge information and the traceability data of the medical insurance knowledge information, and the knowledge pedigree graph includes: Introducing a third-party verification agent, accessing the dialogue history records and consensus generation process data of all agents through the third-party verification agent as the traceability data of the medical insurance knowledge information; A knowledge pedigree graph is constructed based on the medical insurance knowledge information and the traceability data of the medical insurance knowledge information, and the nodes in the knowledge pedigree graph represent the medical insurance knowledge information, and the edges represent the derivation relationship or conflict relationship between entries.

4. The method of claim 1, wherein, The evidence theory fusion decision algorithm is used to calculate the credibility of the medical insurance knowledge information, and the conflicting medical insurance knowledge information is automatically eliminated, and the medical insurance knowledge information that is recommended to be adopted is output, including: Based on the knowledge pedigree graph, evidence indicators are extracted, including agent contribution, stance consistency and response stability; The evidence indicators are input into the evidence theory fusion decision algorithm to calculate the credibility score of each medical insurance knowledge information; According to the credibility score, the medical insurance knowledge information with conflicts is identified, and the identified conflicting entries are automatically eliminated; The medical insurance knowledge information that is recommended to be adopted after the conflict elimination processing is output.

5. The method of claim 1, wherein, Based on the initial query problem, multiple agents are driven to carry out multiple rounds of dialogue, dialogue data is collected and fused for analysis to extract the core stance and semantic relationship graph of each agent, including: The initial query problem is input into all agents to trigger each agent to generate an initial response, and the dialogue rounds between agents are organized in a predetermined order; Collect the response text generated by each agent in each round of dialogue to form a dialogue data set; Text analysis is performed on the dialogue data set to identify key claim sentences and viewpoint expressions in the response text of each agent as the core stance of the corresponding agent; By comparing the core stances of different agents, the overlap, difference and correlation between the stances are calculated to generate a semantic relationship graph, in which the nodes represent the core stances and the edges represent the similarity, conflict or complementary relationship between the stances.

6. The method of claim 5, wherein, Text analysis is performed on the dialogue data set to identify key claim sentences and viewpoint expressions in the response text of each agent as the core stance of the corresponding agent, including: Structural analysis is performed on each agent response text in the dialogue data set to identify complete semantic units; Based on a pre-set claim type library, each semantic unit is identified by claim type; From the identified semantic units, the claim type units containing core viewpoint keywords or having explicit judgment properties are selected as key claim sentences; All key claim sentences of the same agent are classified and integrated according to the corresponding claim type to form the core stance of the corresponding agent.

7. The method of claim 4, wherein, According to the credibility score, the medical insurance knowledge information with conflicts is identified, and the identified conflicting entries are automatically eliminated, including: The credibility score is compared with a pre-set conflict detection threshold to select medical insurance knowledge information with a credibility score higher than the conflict detection threshold to form a candidate entry set; In the candidate item set, medical insurance knowledge information with semantic conflicts is identified and marked to form a conflict item group; According to the credibility score, the medical insurance knowledge information in the conflict item group is sorted; Based on the sorting result, the medical insurance knowledge information with the highest credibility score is selected from the conflict item group as the medical insurance knowledge information recommended to be adopted after conflict resolution processing.

8. The method of claim 1, wherein, The definition and configuration of a plurality of intelligent agents with differentiated medical insurance knowledge background and functions, and the configuration of exclusive prompt words for each intelligent agent to regulate behavior patterns, including: Define and configure a plurality of intelligent agents corresponding to different functional fields, including medical insurance policy interpretation intelligent agents, clinical physician intelligent agents, patient intelligent agents, and audit compliance intelligent agents; For each intelligent agent, extract knowledge data from data sources related to the functional field to form the knowledge background of the corresponding intelligent agent, and design exclusive prompt word content for each intelligent agent type; The prompt word content is associated with each intelligent agent instance to ensure that the behavior pattern conforms to the preset specification.

Images