Search

CN-121997946-A - Multi-agent collaborative answering interaction method and system in educational scene

CN121997946ACN 121997946 ACN121997946 ACN 121997946ACN-121997946-A

Abstract

The invention provides a multi-agent collaborative answering interaction method and system in an education scene, which relate to the technical field of intelligent education and comprise the steps of distributing a problem to be answered to a plurality of agents to independently generate candidate answering contents; the method comprises the steps of identifying dispute fragments with logic conflict or knowledge contradiction through semantic alignment analysis and extracting a dispute focus, constructing a negotiation interaction flow, holding intelligent agents with different views to provide reasoning basis, evaluating each view and selecting a negotiation conclusion based on logic completeness and knowledge consistency, and fusing a non-disputed part and the negotiation conclusion to generate a consistency answer result.

Inventors

  • ZHANG WEI

Assignees

  • 北京学海云影科技有限公司

Dates

Publication Date
20260508
Application Date
20260224

Claims (10)

  1. 1. The multi-agent collaborative answering interaction method in the education scene is characterized by comprising the following steps: distributing the questions to be solved to a plurality of agents, wherein each agent independently generates candidate solution contents aiming at the questions to be solved; Semantic alignment analysis is carried out on each candidate answer content, dispute fragments with logic conflict or knowledge contradiction are identified, and dispute focuses corresponding to the dispute fragments are extracted; aiming at the dispute focus, constructing a negotiation interaction flow among the intelligent agents, in the negotiation interaction flow, the intelligent agents holding different viewpoints sequentially provide reasoning basis for supporting the viewpoints of the intelligent agents, evaluating the viewpoints based on the logic completeness and knowledge consistency of the reasoning basis, and selecting the viewpoint with the highest evaluation value as a negotiation conclusion of the dispute focus; and fusing the non-disputed part in each candidate answer content with the negotiation conclusion to generate a consistency answer result and presenting the consistency answer result to the learner.
  2. 2. The method of claim 1, wherein the step of distributing the problem to be solved to a plurality of agents, each agent independently generating candidate solution content for the problem to be solved comprises: Carrying out knowledge point decomposition and cognition level analysis on the to-be-solved problem to obtain a plurality of knowledge point units forming the to-be-solved problem and dependency relations among the knowledge point units; Selecting a plurality of agents participating in collaborative answering from an agent resource pool based on the knowledge point unit and the dependency relationship, wherein each agent in the agent resource pool has capability feature vectors for representing knowledge coverage range and historical answering quality; Distributing answering responsibility ranges for the intelligent agents according to the capability feature vectors, wherein the answering responsibility ranges limit a subset of knowledge point units which are covered by the intelligent agents; And sending a problem distribution instruction containing the problem to be solved and the answering responsibility range to corresponding agents, wherein each agent independently generates candidate solution contents.
  3. 3. The method of claim 1, wherein the steps of performing semantic alignment analysis on each candidate solution content, identifying a dispute segment for which there is a logical conflict or a knowledge conflict, and extracting a dispute focus corresponding to the dispute segment comprise: Decomposing each candidate solution content into a plurality of solution fragments bearing independent semantics, and labeling each solution fragment with the source identification of the candidate solution content to which the solution fragment belongs; Extracting expression fragments aiming at the knowledge dimension from solution fragments of different candidate solution contents based on the candidate solution content source identification aiming at each knowledge dimension in the to-be-solved problem, and forming fragment comparison groups by the expression fragments aiming at the same knowledge dimension from different candidate solution content source identifications; And carrying out logic consistency test and knowledge contradiction detection on the expression fragments in each fragment comparison group, if the reasoning preconditions among the expression fragments in the fragment comparison group are mutually exclusive or the conclusions are mutually negative, marking the expression fragments as dispute fragments, and extracting the bifurcation points which lead the reasoning preconditions to be mutually exclusive or the conclusions to be mutually negative as dispute focuses.
  4. 4. A method according to claim 3, wherein the step of performing a logical consistency check and knowledge contradiction detection on the presentation segments in the segment comparison group comprises: Extracting propositions of each expression fragment in the fragment comparison group; Carrying out logical relation deduction on propositions sets of different expression fragments in the same fragment comparison group, identifying propositions pairs with direct negative relations and marking the propositions pairs as logic conflicts, wherein the direct negative relations are that one propositions are true and the other propositions are necessarily false; For propositions which do not have direct negative relations but involve the same knowledge object, a precondition set and a conclusion set of each proposition are constructed, attribute value conflicts exist between the precondition sets or propositions pairs which cannot be established by logic between the conclusion sets are identified and marked as knowledge conflicts, wherein the attribute value conflicts are that the same knowledge attribute is endowed with different attribute values, and the fact that two conclusions cannot be established by logic simultaneously means that two conclusions cannot be established by the same conclusions under the same precondition.
  5. 5. The method of claim 1, wherein for the dispute focus, a negotiation interaction flow between agents is constructed, in which agents holding different viewpoints sequentially provide reasoning bases supporting their viewpoints, and each viewpoint is evaluated based on logical completeness and knowledge consistency of the reasoning bases, and the step of selecting the viewpoint with the highest evaluation value as a negotiation conclusion of the dispute focus comprises: Extracting a current dispute focus and a dispute fragment related to the current dispute focus, determining agents holding different views according to candidate solution content source identifiers of the dispute fragment, and determining speaking sequence; Sequentially sending an inference basis request instruction to each agent according to the speaking sequence, wherein the inference basis request instruction comprises a current dispute focus and an inference basis of an agent which has been spoken, each agent generates an inference basis, and the inference basis comprises an inference step sequence, a knowledge source and a questioning content, and the questioning content points to an inference jump position or a precondition missing position in the inference step sequence of the agent which has been spoken; The method comprises the steps of carrying out cross verification analysis on each agent reasoning basis, identifying a knowledge source superposition part, extracting knowledge nodes of mutual authentication, improving the credibility weight of the corresponding agent, counting the number of suspected reasoning jump positions and premise missing positions in each agent reasoning basis, and reducing the credibility weight of the corresponding agent; Calculating a logic completeness score according to the number of the reasoning jump positions and the premise missing positions, calculating a knowledge consistency score according to the coincidence degree of knowledge sources and the knowledge field of the problem to be solved, calculating evaluation values of all agent views according to the adjusted credibility weights, and selecting the view with the highest evaluation value as a negotiation conclusion.
  6. 6. The method of claim 5, wherein the step of performing a cross-validation analysis on each agent inference basis comprises: Extracting knowledge sources and an inference step sequence from an inference basis provided by each agent, and constructing an inference path diagram of each agent, wherein the inference steps in the inference path diagram are used as nodes, and logic deduction relations among the steps are used as directed edges; Node matching is carried out on the reasoning path diagrams of different agents, node pairs which are the same in the reasoning step content and knowledge source are identified as mutual authentication nodes, the duty ratio of the mutual authentication nodes in the reasoning path diagrams of each agent is counted, and the higher the duty ratio is, the higher the credibility weight of the agent is improved; extracting the inference jump position or the premise missing position of the to-be-challenged from the challenged content provided by each agent, marking the to-be-challenged nodes in the inference path diagram of the corresponding to the to-be-challenged agent, calculating the criticality of the to-be-challenged nodes in the inference path diagram, calculating the criticality by counting the number of follow-up inference steps depending on the nodes, and reducing the credibility weight of the agent according to the number of the to-be-challenged nodes and the criticality.
  7. 7. The method of claim 1, wherein the step of fusing non-disputed portions of each candidate solution content with the negotiation results to generate consistent answer results for presentation to a learner comprises: For each knowledge node, respectively calculating the knowledge integrity of the knowledge node in the non-disputed part and the negotiation conclusion, wherein the knowledge integrity is calculated by counting the number of associated preconditions and the number of reasoning steps, reserving the expression of one party with higher knowledge integrity and deleting the expression of the other party; Detecting logic break points of the connection positions between the non-disputed parts and the negotiation conclusion, identifying intermediate reasoning steps or preconditions missing between knowledge nodes before and after the connection positions, and extracting reasoning paths connecting the knowledge nodes before and after the connection positions from each candidate solution content as logic connection fragments; And constructing answering results according to the sequence of the non-disputed part and the negotiation conclusion, inserting the logic connection fragments into the connection positions, generating consistent answering results and presenting the consistent answering results to learners.
  8. 8. A multi-agent collaborative answering interaction system in an educational scenario for implementing the method of any of the preceding claims 1-7, comprising: the problem acquisition module is used for acquiring a to-be-solved problem proposed by a learner; The problem distribution module is used for distributing the problem to be solved to a plurality of agents, and each agent independently generates candidate solution contents aiming at the problem to be solved; The semantic alignment analysis module is used for carrying out semantic alignment analysis on each candidate answer content, identifying dispute fragments with logic conflict or knowledge contradiction, and extracting dispute focuses corresponding to the dispute fragments; The negotiation interaction module is used for constructing a negotiation interaction flow among the intelligent agents aiming at the dispute focus, in the negotiation interaction flow, the intelligent agents holding different viewpoints sequentially provide reasoning basis for supporting the viewpoints of the intelligent agents, evaluate the viewpoints based on the logic completeness and knowledge consistency of the reasoning basis, and select the viewpoint with the highest evaluation value as a negotiation conclusion of the dispute focus; And the result fusion module is used for fusing the non-disputed part in each candidate answer content with the negotiation conclusion, generating a consistency answer result and presenting the consistency answer result to the learner.
  9. 9. An electronic device, comprising: A processor; A memory for storing processor-executable instructions; Wherein the processor is configured to invoke the instructions stored in the memory to perform the method of any of claims 1 to 7.
  10. 10. A computer readable storage medium having stored thereon computer program instructions, which when executed by a processor, implement the method of any of claims 1 to 7.

Description

Multi-agent collaborative answering interaction method and system in educational scene Technical Field The invention relates to an intelligent education technology, in particular to a multi-agent collaborative answering interaction method and system in an education scene. Background The artificial intelligence technology is widely applied in the education field, and the existing intelligent answering method generally adopts a single agent to answer questions raised by learners, but has obvious limitations. The knowledge coverage of a single agent is limited, and for interdisciplinary or complex problems, it is often difficult to give comprehensive and accurate solutions, key knowledge points may be omitted or the answer content generated by the single agent in the reasoning step lacks a verification mechanism, and when the agent has deviation or reasoning error on knowledge understanding, the error content can be directly presented to a learner, so that the learning effect is influenced and even misguidance is generated. Although schemes of multi-agent cooperation also appear in the prior art, the schemes mostly adopt a simple voting mechanism or a result splicing mode, logic conflicts and knowledge contradictions between answer contents of different agents cannot be effectively identified and solved, deep analysis and verification of an agent reasoning process are lacking, so that the fused answer result still possibly contains inconsistent or wrong information, and stability and reliability of answer quality are difficult to ensure. Disclosure of Invention The embodiment of the invention provides a multi-agent collaborative answering interaction method and system in an education scene, which can solve the problems in the prior art. In a first aspect of the embodiment of the present invention, a multi-agent collaborative answer interaction method in an educational scenario is provided, including: distributing the questions to be solved to a plurality of agents, wherein each agent independently generates candidate solution contents aiming at the questions to be solved; Semantic alignment analysis is carried out on each candidate answer content, dispute fragments with logic conflict or knowledge contradiction are identified, and dispute focuses corresponding to the dispute fragments are extracted; aiming at the dispute focus, constructing a negotiation interaction flow among the intelligent agents, in the negotiation interaction flow, the intelligent agents holding different viewpoints sequentially provide reasoning basis for supporting the viewpoints of the intelligent agents, evaluating the viewpoints based on the logic completeness and knowledge consistency of the reasoning basis, and selecting the viewpoint with the highest evaluation value as a negotiation conclusion of the dispute focus; and fusing the non-disputed part in each candidate answer content with the negotiation conclusion to generate a consistency answer result and presenting the consistency answer result to the learner. The step of distributing the to-be-solved problem to a plurality of agents, wherein each agent independently generates candidate solution contents for the to-be-solved problem comprises the following steps: Carrying out knowledge point decomposition and cognition level analysis on the to-be-solved problem to obtain a plurality of knowledge point units forming the to-be-solved problem and dependency relations among the knowledge point units; Selecting a plurality of agents participating in collaborative answering from an agent resource pool based on the knowledge point unit and the dependency relationship, wherein each agent in the agent resource pool has capability feature vectors for representing knowledge coverage range and historical answering quality; Distributing answering responsibility ranges for the intelligent agents according to the capability feature vectors, wherein the answering responsibility ranges limit a subset of knowledge point units which are covered by the intelligent agents; and sending a problem distribution instruction containing the problem to be solved and the answering responsibility range to corresponding agents, wherein each agent independently generates candidate answering contents based on the answering responsibility range. The method for carrying out semantic alignment analysis on each candidate solution content, identifying the dispute fragments with logic conflict or knowledge contradiction, and extracting the dispute focuses corresponding to the dispute fragments comprises the following steps: Semantic unit segmentation is carried out on each candidate answer content, each candidate answer content is decomposed into a plurality of answer fragments bearing independent semantics, and each answer fragment is marked with the source identification of the candidate answer content to which the answer fragment belongs; Extracting expression fragments aiming at the knowledge dimension from solutio