Multi-agent cooperative question answering interaction method and system in education scene

Abstract

The invention provides a multi-agent cooperative question answering interaction method and system in an education scene, and relates to the technical field of intelligent education, and the method comprises the steps: distributing a to-be-answered question to a plurality of agents, and independently generating candidate answering contents; analyzing and identifying dispute fragments with logic conflicts or knowledge conflicts through semantic alignment, and extracting dispute focuses; constructing a negotiation interaction process, providing a reasoning basis for intelligent agents with different viewpoints, evaluating each viewpoint based on logic completeness and knowledge consistency, and selecting a negotiation conclusion; and fusing the non-dispute part with the negotiation conclusion to generate a consistent question answering result.

Description

Technical Field

[0001] This invention relates to smart education technology, and more particularly to a method and system for multi-agent collaborative Q&A interaction in educational settings. Background Technology

[0002] The widespread application of artificial intelligence (AI) technology in education, particularly in intelligent question-answering methods, typically employs a single AI agent to respond to learners' questions. However, this approach has significant limitations. A single AI agent has limited knowledge coverage and often struggles to provide comprehensive and accurate answers to interdisciplinary or complex problems, potentially omitting key knowledge points or reasoning steps. The answers generated by a single AI agent lack verification mechanisms; when the agent's understanding of knowledge is flawed or its reasoning is incorrect, the erroneous content is directly presented to the learner, impacting learning outcomes and even causing confusion. While multi-agent collaborative solutions have emerged, these often rely on simple voting mechanisms or result aggregation, failing to effectively identify and resolve logical conflicts and knowledge contradictions between different agents' answers. They also lack in-depth analysis and verification of the agents' reasoning processes, resulting in inconsistent or erroneous information in the merged answers, making it difficult to guarantee the stability and reliability of the question-answering quality. Summary of the Invention

[0003] This invention provides a method and system for multi-agent collaborative Q&A interaction in educational scenarios, which can solve the problems in the prior art.

[0004] A first aspect of this invention provides a multi-agent collaborative Q&A interaction method in an educational setting, comprising:

[0005] Obtain the unanswered questions raised by learners; distribute the unanswered questions to multiple agents, and each agent independently generates candidate solutions to the unanswered questions;

[0006] Semantic alignment analysis is performed on each candidate solution to identify disputed segments with logical conflicts or knowledge contradictions, and the focus of the dispute corresponding to the disputed segments is extracted.

[0007] For the aforementioned point of contention, a negotiation and interaction process between intelligent agents is constructed. In this process, intelligent agents with different viewpoints provide reasoning evidence to support their viewpoints in turn. The viewpoints are evaluated based on the logical completeness and knowledge consistency of the reasoning evidence, and the viewpoint with the highest evaluation value is selected as the negotiation conclusion for the point of contention.

[0008] The uncontroversial parts of each candidate answer are integrated with the negotiated conclusion to generate a consistent answer result, which is then presented to the learner.

[0009] The step of distributing the unsolved question to multiple agents, with each agent independently generating candidate solutions for the unsolved question, includes:

[0010] Obtain the unanswered questions raised by learners; perform knowledge point decomposition and cognitive hierarchy analysis on the unanswered questions to obtain multiple knowledge point units that constitute the unanswered questions and the dependencies between each knowledge point unit;

[0011] Multiple agents are selected from the agent resource pool to participate in collaborative Q&A based on the knowledge point units and the dependencies. Each agent in the agent resource pool has a feature vector that represents its knowledge coverage and historical Q&A quality.

[0012] Based on the capability feature vector, each agent is assigned a scope of responsibility for answering questions, and the scope of responsibility for answering questions limits the subset of knowledge point units that each agent should cover.

[0013] A question distribution instruction containing the question to be answered and the scope of responsibility for answering the question is sent to the corresponding intelligent agent, and each intelligent agent independently generates candidate answer content based on the scope of responsibility for answering the question.

[0014] The steps of performing semantic alignment analysis on each candidate solution, identifying disputed segments with logical conflicts or knowledge contradictions, and extracting the points of contention corresponding to the disputed segments include:

[0015] The content of each candidate answer is segmented into semantic units, and each candidate answer is decomposed into multiple answer fragments carrying independent semantics. Each answer fragment is labeled with the source identifier of the candidate answer content to which it belongs.

[0016] For each knowledge dimension in the question to be answered, based on the source identifier of the candidate answer content, extract the expression fragments for that knowledge dimension from the answer fragments of different candidate answer content, and form a fragment comparison group by combining the expression fragments from different candidate answer content source identifiers for the same knowledge dimension.

[0017] Logical consistency and knowledge contradiction detection are performed on the statement fragments in each comparison group. If there are mutually exclusive reasoning premises or mutually contradictory conclusions between the statement fragments in the comparison group, the statement fragment is marked as a disputed fragment, and the divergence points that cause the mutually exclusive reasoning premises or mutually contradictory conclusions are extracted as the focus of the dispute.

[0018] The steps for performing logical consistency checks and knowledge contradiction detection on the statement fragments in each comparison group include:

[0019] Extract the set of propositions for each statement segment in the segment comparison group;

[0020] Logical relationship deduction is performed on the set of propositions with different statements within the same segment comparison group. Proposition pairs with direct negation relationship are identified and marked as logical conflicts. The direct negation relationship means that if one proposition is true, the other proposition must be false.

[0021] For propositions that do not have a direct negation relationship but involve the same knowledge object, a set of premises and a set of conclusions for each proposition are constructed. Proposition pairs that have conflicting attribute values ​​between their premise sets or whose logical conclusions cannot be simultaneously true are identified and marked as knowledge contradictions. The conflicting attribute values ​​refer to the same knowledge attribute being assigned different attribute values. The logical incompatibility refers to the two conclusions not being true at the same time under the same premises.

[0022] To address the points of contention, a negotiation and interaction process is constructed between intelligent agents. In this process, agents holding different viewpoints sequentially provide reasoning supporting their arguments. Each viewpoint is evaluated based on the logical completeness and knowledge consistency of its reasoning, and the viewpoint with the highest evaluation value is selected as the negotiation conclusion for that point of contention. The steps include:

[0023] Extract the current point of contention and its associated contention segments, and determine the agents holding different viewpoints and their speaking order based on the source identifiers of the candidate solutions to the contention segments.

[0024] According to the speaking order, the reasoning basis request instruction is sent to each agent in sequence. The reasoning basis request instruction contains the current point of contention and the reasoning basis of the speaking agents. Each agent generates reasoning basis, which includes a reasoning step sequence, knowledge source and question content. The question content points to the reasoning jump position or the missing premise position in the reasoning step sequence of the speaking agents.

[0025] Cross-validation analysis is performed on the reasoning basis of each agent to identify the overlapping parts of knowledge sources and extract mutually corroborating knowledge nodes to improve the credibility weight of the corresponding agent; the number of questioned reasoning jump positions and missing premise positions in the reasoning basis of each agent is counted to reduce the credibility weight of the corresponding agent.

[0026] The logical completeness score is calculated based on the number of reasoning jumps and missing premises. The knowledge consistency score is calculated based on the overlap between the knowledge source and the knowledge domain of the question to be answered. The evaluation value of each agent's viewpoint is calculated in combination with the adjusted credibility weight. The viewpoint with the highest evaluation value is selected as the negotiation conclusion.

[0027] The steps for cross-validation analysis of the reasoning basis of each agent include:

[0028] Extract the knowledge sources and reasoning step sequences from the reasoning basis provided by each intelligent agent, and construct the reasoning path graph of each intelligent agent. The reasoning steps in the reasoning path graph are used as nodes, and the logical deduction relationship between the steps is used as directed edges.

[0029] Node matching is performed on the reasoning path graphs of different agents. Node pairs that are the same in terms of reasoning steps and knowledge sources are identified as mutual corroboration nodes. The proportion of mutual corroboration nodes in the reasoning path graphs of each agent is counted. The higher the proportion, the greater the increase in the credibility weight of the agent.

[0030] Extract the questioned reasoning jump positions or missing premise positions from the questioning content provided by each agent, mark the questioned nodes in the reasoning path graph of the corresponding questioned agent, calculate the criticality of the questioned node in the reasoning path graph, the criticality is calculated by counting the number of subsequent reasoning steps that depend on the node, and reduce the credibility weight of the agent based on the number of questioned nodes and criticality.

[0031] The steps of integrating the uncontroversial parts of each candidate answer with the negotiated conclusion to generate a consistent answer result and present it to the learner include:

[0032] A knowledge node matching analysis is performed on the undisputed part and the negotiation conclusion to identify knowledge node pairs with semantic overlap. For each knowledge node pair, its knowledge completeness in the undisputed part and the negotiation conclusion is calculated. The knowledge completeness is calculated by statistically associating the number of preconditions and the number of reasoning steps. The expression of the party with higher knowledge completeness is retained and the expression of the other party is deleted.

[0033] Logical breakpoint detection is performed on the connection position between the non-disputed part and the negotiation conclusion to identify the missing intermediate reasoning steps or preconditions between the knowledge nodes before and after the connection position, and the reasoning path connecting the knowledge nodes before and after the connection position is extracted from the content of each candidate solution as a logical connection segment.

[0034] The Q&A results are constructed in the order of non-controversial parts and negotiated conclusions. The logical connection segments are inserted at the connecting positions to generate a consistent Q&A result to be presented to the learner.

[0035] A second aspect of this invention provides a multi-agent collaborative Q&A interactive system for educational scenarios, comprising:

[0036] The question retrieval module is used to retrieve unanswered questions submitted by learners.

[0037] The question distribution module is used to distribute the question to be answered to multiple agents, and each agent independently generates candidate answers for the question to be answered.

[0038] The semantic alignment analysis module is used to perform semantic alignment analysis on the content of each candidate solution, identify disputed segments with logical conflicts or knowledge contradictions, and extract the focus of the dispute corresponding to the disputed segments.

[0039] The negotiation and interaction module is used to construct a negotiation and interaction process between intelligent agents for the disputed issue. In the negotiation and interaction process, intelligent agents with different viewpoints provide reasoning evidence to support their viewpoints in turn, and evaluate each viewpoint based on the logical completeness and knowledge consistency of the reasoning evidence. The viewpoint with the highest evaluation value is selected as the negotiation conclusion for the disputed issue.

[0040] The result fusion module is used to merge the uncontroversial parts of each candidate answer with the negotiated conclusion to generate a consistent answer result for the learner.

[0041] A third aspect of the present invention provides an electronic device, comprising:

[0042] processor;

[0043] Memory used to store processor-executable instructions;

[0044] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0045] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0046] This invention distributes the question to be answered to multiple agents and generates candidate solutions independently, enabling responses from different knowledge perspectives and reasoning paths, thus avoiding the knowledge blind spots and cognitive limitations of a single agent. By constructing a negotiation and interaction process, agents with different viewpoints provide reasoning evidence and challenge each other. Evaluation is based on logical completeness and knowledge consistency, allowing the most reliable viewpoints to be selected through debate and verification among agents, thereby improving the accuracy and credibility of the answers. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating the multi-agent collaborative Q&A interaction method in an educational setting according to an embodiment of the present invention.

[0048] Figure 2 A flowchart for extracting points of contention. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0051] Figure 1 This is a flowchart illustrating the multi-agent collaborative Q&A interaction method in an educational scenario according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0052] Obtain the unanswered questions raised by learners; distribute the unanswered questions to multiple agents, and each agent independently generates candidate solutions to the unanswered questions;

[0053] Semantic alignment analysis is performed on each candidate solution to identify disputed segments with logical conflicts or knowledge contradictions, and the focus of the dispute corresponding to the disputed segments is extracted.

[0054] For the aforementioned point of contention, a negotiation and interaction process between intelligent agents is constructed. In this process, intelligent agents with different viewpoints provide reasoning evidence to support their viewpoints in turn. The viewpoints are evaluated based on the logical completeness and knowledge consistency of the reasoning evidence, and the viewpoint with the highest evaluation value is selected as the negotiation conclusion for the point of contention.

[0055] The uncontroversial parts of each candidate answer are integrated with the negotiated conclusion to generate a consistent answer result, which is then presented to the learner.

[0056] In one optional implementation, the step of distributing the question to be answered to multiple agents, each agent independently generating candidate solutions for the question to be answered, includes:

[0057] The process involves: acquiring unanswered questions raised by learners; decomposing the unanswered questions into knowledge points to obtain multiple knowledge point units constituting the unanswered questions and the dependencies between these knowledge point units; selecting multiple agents from the agent resource pool to participate in collaborative question answering based on the knowledge point units and the dependencies, wherein each agent in the agent resource pool has a capability feature vector that characterizes its knowledge coverage and historical question answering quality.

[0058] Based on the capability feature vector, each agent is assigned a scope of responsibility for answering questions, and the scope of responsibility for answering questions limits the subset of knowledge point units that each agent should cover.

[0059] A question distribution instruction containing the question to be answered and the scope of responsibility for answering the question is sent to the corresponding intelligent agent, and each intelligent agent independently generates candidate answer content based on the scope of responsibility for answering the question.

[0060] For example, the agent is a question-answering unit with independent knowledge retrieval and reasoning generation capabilities. It can be implemented based on existing pre-trained language model technology, such as using a language model fine-tuned with data from the education field as the reasoning engine, and configuring a knowledge base specific to a particular subject as the retrieval data source. Each agent comprises three functional components: a knowledge retrieval unit, a reasoning generation unit, and an interaction response unit. The knowledge retrieval unit retrieves knowledge fragments related to the input question from the configured knowledge base. The knowledge base can use a vector database to store textbook texts, question bank resources, and academic literature, etc. During retrieval, it returns relevant content by calculating the semantic similarity between the question text and the knowledge fragments. The reasoning generation unit concatenates the retrieved knowledge fragments with the input question and inputs them into the language model to generate an answer text in natural language. The interaction response unit receives external instructions via a RESTful API or message queue, parses the question text and task parameters in the instructions, schedules the knowledge retrieval unit and the reasoning generation unit to execute tasks, and returns the answer result.

[0061] The agent resource pool maintains the configuration information and running status of multiple agent instances. Each agent instance is configured with a different knowledge base scope and model parameters. The resource pool records the knowledge coverage, historical Q&A quality scores, and current load status of each agent instance. It supports dynamically selecting the appropriate agent instance to execute the Q&A task based on the problem requirements, and supports multiple agent instances to process complex problems in parallel or collaboratively.

[0062] The process of distributing unanswered questions to multiple agents begins with obtaining the unanswered questions submitted by learners. These questions are input in the form of natural language text. The question text is preprocessed to remove irrelevant punctuation and redundant spaces while preserving the complete semantic structure.

[0063] The educational knowledge graph is a pre-constructed knowledge structure containing concept nodes within a subject area and edges representing relationships between concepts. Node attributes include concept name, subject, difficulty level, and prerequisite knowledge identifiers. Edge attributes indicate dependencies, inclusion relationships, or comparison relationships between concepts. Graph construction is achieved by parsing the textbook's table of contents and knowledge point system. Each textbook chapter corresponds to a set of associated concept nodes, and the logical order between chapters is translated into dependency edges between nodes. Mapping question text to graph nodes uses a keyword matching and context analysis method. Terms in the question text are matched against graph node names, and successfully matched terms are mapped to corresponding knowledge point units. Dependencies between knowledge points are determined by identifying connectors indicating logical relationships in the question text. When a question contains words such as "based on," "utilizing," or "through," the knowledge points before and after these words are extracted, and dependency edges are established.

[0064] When decomposing a problem into knowledge points, the problem text is mapped to a predefined set of knowledge point units, each corresponding to an independent concept node in the educational knowledge graph. The dependencies between knowledge point units are determined by analyzing the conceptual reasoning paths involved in the problem. If understanding knowledge point A requires prior knowledge point B, a directed edge from B to A is created in the dependency graph. The dependency graph is stored using an adjacency list structure, supporting quick lookup of all prerequisite dependencies and subsequent extensions of a given knowledge point.

[0065] The agent resource pool maintains capability feature vectors for all available agents. Each agent's capability feature vector consists of a knowledge coverage vector and historical Q&A quality scores. The dimension of the knowledge coverage vector is the same as the total number of knowledge point units in the educational knowledge graph. Each element in the vector ranges from 0 to 1, representing the agent's mastery of the corresponding knowledge point. The mastery level is calculated by the frequency and accuracy of the agent's answers to the knowledge point in historical Q&A sessions. The calculation method is: frequency × accuracy ÷ normalization coefficient, where the normalization coefficient is the maximum value of the product of the frequency and accuracy of all knowledge points for that agent. The historical Q&A quality score is a weighted calculation based on three indicators: learner satisfaction feedback, logical completeness of the answer, and knowledge accuracy of the answers previously generated by the agent. Learner satisfaction feedback is obtained by collecting learner ratings for the answer content, ranging from 1 to 5 points, and normalized by dividing the rating by 5. Logical completeness of the answer is calculated by statistically analyzing the completeness of the reasoning steps in the answer. The completeness is the number of reasoning steps in the answer divided by the number of reasoning steps required for the standard answer to the question, with an upper limit of 1. Knowledge accuracy is calculated by comparing the answer content with the correct answers in the standard knowledge base. Accuracy is the number of correctly stated knowledge points in the answer divided by the total number of knowledge points involved in the answer. Satisfaction feedback is weighted at 0.3, logical completeness at 0.4, and knowledge accuracy at 0.3. The three indicators are normalized to the range of 0 to 1 and then weighted and summed to obtain the quality score.

[0066] When selecting multiple agents from the agent resource pool to participate in collaborative question answering, the knowledge coverage matching degree between each agent and the question to be answered is calculated. The matching degree is obtained by extracting the element values ​​of the knowledge point units required for the question to be answered from the agent's capability feature vector and calculating the average of these element values. The top N agents with the highest matching degree are selected to participate in collaborative question answering. The value of N ranges from 3 to 8, with a default value of 5. It can be adjusted according to the number of knowledge point units involved in the question to be answered and the complexity of the dependency relationship. When the number of knowledge point units exceeds 10 or the longest path length in the dependency relationship graph exceeds 5, N is set to an upper limit of 8.

[0067] When assigning answering responsibilities to each agent, a knowledge point allocation graph is constructed based on the knowledge point units involved in the question to be answered and their dependencies. The knowledge point allocation graph uses knowledge point units as nodes and dependencies as edges. Node attributes include knowledge point complexity, which is calculated based on the depth of the knowledge point in the educational knowledge graph and the number of connected edges. The calculation method is depth × number of connected edges ÷ 10, with a value ranging from 0.1 to 5.0. The knowledge point allocation graph is divided into multiple subgraphs, and each subgraph is assigned to an agent as its answering responsibility scope. The partitioning strategy prioritizes assigning clusters of knowledge points with strong dependencies and high complexity to the agent with the highest corresponding element value in the capability feature vector, while ensuring that the difference in the number of knowledge point units assigned to each agent does not exceed 20% of the total to achieve load balancing. The answering responsibility scope is stored in the form of a set of knowledge point unit identifiers. Each agent's responsibility scope limits the subset of knowledge point units it should focus on covering when generating candidate answers.

[0068] A question distribution instruction is constructed, containing the question to be answered and the scope of responsibility for answering it. The instruction is encapsulated in JSON format and includes fields for the question text, the scope of responsibility for answering it, recommended knowledge sources, generation time limit, and format specifications. The scope of responsibility field stores a list of knowledge point unit identifiers. The recommended knowledge sources field provides an index of textbook chapters or reference materials related to the question. The generation time limit field is set to a value between 30 and 120 seconds, with a default of 60 seconds. The format specifications field specifies the structural requirements of the candidate solution content, such as paragraph division and citation methods. The question distribution instruction is sent asynchronously to the corresponding agent's interface endpoint via a message queue. After receiving the instruction, the agent calls its internal knowledge retrieval module and reasoning generation module based on the scope of responsibility for answering it. The knowledge retrieval module extracts relevant content of knowledge points within the scope of responsibility from the knowledge base. The reasoning generation module combines the retrieved content with the question to be answered to generate candidate solution content. The candidate solution content is returned in structured text format, including the solution text, the identifiers of the cited knowledge sources, and a confidence score. The confidence score ranges from 0 to 1 and is calculated by the agent's internal evaluation mechanism based on the completeness of the reasoning steps and the knowledge coverage.

[0069] This implementation method accurately matches the capabilities of agents with the needs of problems by using capability feature vectors, and reasonably allocates the scope of answering responsibilities based on knowledge point dependencies, so that each agent can focus on its area of ​​expertise to generate high-quality candidate answers.

[0070] In one optional implementation, the steps of performing semantic alignment analysis on each candidate answer to identify disputed segments with logical conflicts or knowledge contradictions, and extracting the points of contention corresponding to the disputed segments, include:

[0071] The content of each candidate answer is segmented into semantic units, and each candidate answer is decomposed into multiple answer fragments carrying independent semantics. Each answer fragment is labeled with the source identifier of the candidate answer content to which it belongs.

[0072] For each knowledge dimension in the question to be answered, based on the source identifier of the candidate answer content, extract the expression fragments for that knowledge dimension from the answer fragments of different candidate answer content, and form a fragment comparison group by combining the expression fragments from different candidate answer content source identifiers for the same knowledge dimension.

[0073] Logical consistency and knowledge contradiction detection are performed on the statement fragments in each comparison group. If there are mutually exclusive reasoning premises or mutually contradictory conclusions between the statement fragments in the comparison group, the statement fragment is marked as a disputed fragment, and the divergence points that cause the mutually exclusive reasoning premises or mutually contradictory conclusions are extracted as the focus of the dispute.

[0074] Combination Figure 2 The flowchart for extracting points of contention is used to illustrate this. For example, a segmentation algorithm based on syntactic analysis and semantic boundary recognition is used to segment the candidate answer content into semantic units, decomposing each candidate answer content into multiple answer fragments carrying independent semantics.

[0075] Semantic boundary recognition is achieved by detecting logical connectors, punctuation marks, and semantic integrity markers in sentences. When connectors indicating causality such as "therefore" or "so," connectors indicating contrast such as "but" or "however," or strong separators such as periods are detected, a semantic boundary is set at that location to ensure that each solution fragment contains a complete subject-verb-object structure and independent semantic expression. The data structure of a solution fragment includes a fragment text field, a fragment identifier field, a candidate solution content source identifier field, and a semantic type field. The candidate solution content source identifier is used to trace the agent to which the solution fragment belongs, and the semantic type field indicates whether the fragment belongs to the definition, reasoning, example, or conclusion category.

[0076] When extracting descriptive fragments from each knowledge dimension of a problem to be solved, the problem is pre-divided into knowledge dimensions. Each knowledge dimension corresponds to a different knowledge aspect involved in the problem. For example, if the problem involves determining the monotonicity of a function, the knowledge dimensions include the definition of the derivative, the calculation of the derivative, and the rules for determining monotonicity. This knowledge dimension division is achieved by mapping the problem to be solved to multiple knowledge nodes in an educational knowledge graph. Each knowledge node corresponds to a knowledge dimension, and the node attributes include the dimension name and a set of keywords.

[0077] Based on the source identifier of the candidate answer content, expression fragments for specific knowledge dimensions are extracted from answer fragments of different candidate answer content. The extraction strategy is to calculate the matching degree between the answer fragment text and the keyword set of the knowledge dimension. The matching degree is calculated by dividing the number of keywords appearing in the answer fragment by the total number of keywords in the keyword set. When the ratio is greater than or equal to 0.3, the answer fragment is considered to be for that knowledge dimension.

[0078] Description fragments from different candidate answer content source identifiers but targeting the same knowledge dimension are grouped into fragment comparison groups. These groups are stored in a list structure, where each element contains the complete data structure of a description fragment. For a given knowledge dimension, if only one candidate answer content source identifier provides a description fragment, no comparison group is formed. A comparison group is only constructed when description fragments from at least two different source identifiers are present.

[0079] When performing logical consistency checks and knowledge contradiction detection on the statement fragments in each comparison group, propositional expressions are extracted from each statement fragment. Each propositional expression contains a subject, predicate, object, and logical relation identifiers, which indicate the implication, parallelism, or negation relationship between propositions. Pairwise comparisons are performed on the propositional expressions of different statement fragments within the same comparison group. The comparison process includes checking whether the propositional subject points to the same knowledge object, whether the operations or attributes expressed by the propositional predicate are consistent, and whether there are any conflicts in the values ​​of the propositional object.

[0080] When two propositions have the same subject but mutually exclusive object values, they are considered to have mutually exclusive premises. For example, one proposition states "the derivative of a function is greater than 0 in the interval [0, 1]" while another proposition states "the derivative of a function is less than 0 in the interval [0, 1]". The object values ​​"greater than 0" and "less than 0" are mutually exclusive. When two propositions have the same premises but opposite conclusions, they are considered to have mutually negating conclusions. For example, one proposition states "if the derivative is greater than 0, the function is monotonically increasing" while another proposition states "if the derivative is greater than 0, the function is monotonically decreasing". The two propositions have opposite conclusions.

[0081] If, within a comparison group of excerpts, the reasoning premises of the excerpts are mutually exclusive or their conclusions are mutually contradictory, the excerpts involving conflict are marked as disputed excerpts. The marking information of the disputed excerpts is stored in the disputed marker field of the excerpt data structure, with the field value set to a Boolean type truth value. The points of divergence that lead to mutually exclusive reasoning premises or mutually contradictory conclusions are extracted as the focus of the dispute. These divergences are determined by comparing the differences between the conflicting propositions. Specifically, this involves extracting the object parts with different values ​​or the predicate parts with different conclusion statements from the two conflicting propositions, and combining the differences and their corresponding knowledge dimensions into a descriptive text for the focus of the dispute.

[0082] The data structure of the dispute focus includes a dispute focus identifier field, a dispute focus description text field, a list of associated dispute fragment identifiers field, and a list of candidate solution content source identifiers field. The dispute focus identifier is generated using a unique hash value. The list of associated dispute fragment identifiers records all fragment identifiers that are marked as dispute fragments and are related to the dispute focus. The list of candidate solution content source identifiers records the identifiers of agents holding different viewpoints.

[0083] The system outputs a set of disputed points and a set of undisputed fragments. The set of disputed points contains all identified disputed points and their associated information, while the set of undisputed fragments contains all solution fragments not marked as disputed. The data structure of the set of disputed points is stored in dictionary form, with dictionary keys representing disputed point identifiers and dictionary values ​​representing the complete data structure of the disputed points. This allows for quick retrieval of dispute details using the disputed point identifier. The set of undisputed fragments is stored in groups according to the source identifiers of the candidate solution content, with each group containing all undisputed fragments from the same agent.

[0084] This implementation method accurately identifies the differences in the answers provided by different intelligent agents through semantic unit segmentation and knowledge dimension matching. It also quantitatively detects reasoning conflicts and knowledge contradictions based on propositional logic comparison and extracts the points of contention for subsequent negotiation processes.

[0085] In one optional implementation, the steps of performing logical consistency checks and knowledge contradiction detection on the expression fragments in each fragment comparison group include: extracting the proposition set of each expression fragment in the fragment comparison group; performing logical relationship deduction on the proposition sets of different expression fragments within the same fragment comparison group, identifying proposition pairs with direct negation relationships and marking them as logical conflicts, wherein a direct negation relationship means that if one proposition is true, the other proposition must be false; for propositions that do not have direct negation relationships but involve the same knowledge object, constructing a set of preconditions and a set of conclusions for each proposition, identifying proposition pairs with attribute value conflicts between the precondition sets or logically incompatible conclusion sets and marking them as knowledge contradictions, wherein an attribute value conflict means that the same knowledge attribute is assigned different attribute values, and logically incompatible means that two conclusions cannot be true simultaneously under the same premises.

[0086] For example, a proposition extraction algorithm based on dependency parsing is used to decompose each expression fragment into a propositional expression containing a subject, predicate, object, and logical relations.

[0087] During proposition extraction, the core verb in the expression fragment is identified as the predicate. The subject and object components associated with this verb are identified through dependency relationships. The subject component corresponds to the agent or described object, and the object component corresponds to the patient or attribute value. The data structure of the propositional expression includes a proposition identifier field, a subject field, a predicate field, an object field, a logical relation field, and an identifier field for the expression fragment to which it belongs. The proposition identifier is generated using a unique hash value. The logical relation field indicates the implication, parallel, or negation relationships between propositions, determined by analyzing logical connectors in the expression fragment. The proposition set is stored in list form, where each element is a complete data structure for a propositional expression.

[0088] When deducing logical relationships among proposition sets of different statement fragments within the same comparison group, select any two statement fragments from the comparison group, denoted as statement fragment A and statement fragment B respectively. Extract the proposition sets of statement fragment A and statement fragment B. Iterate through each proposition in the proposition set of statement fragment A, denoted as proposition A1, and for each proposition in the proposition set of statement fragment B, denoted as proposition B1, perform pairwise comparisons to identify direct negation relationships.

[0089] The identification of direct negation is achieved by examining the consistency and opposition of the subject, predicate, and object of a proposition pair. When proposition A1 and proposition B1 have the same subject and predicate but their objects express opposite semantics, they are determined to have a direct negation relationship. The determination of opposite semantics in the objects uses a predefined antonym list and logical opposition rules. The antonym list includes common pairs of opposites such as "increase" and "decrease," "greater than" and "less than," and "correct" and "incorrect." The logical opposition rules identify object expressions containing negation words; for example, "monotonically increasing" and "not monotonically increasing" constitute logical opposition. If proposition A1 is true and proposition B1 is necessarily false, or proposition B1 is true and proposition A1 is necessarily false, then the two propositions are considered to have a direct negation relationship, and the proposition pair is marked as logically conflicting. The logical conflict marking information is stored in the conflict record data structure, which includes a conflict type field, a proposition identifier field, and a conflict description field. The conflict type field is set to "direct negation", and the proposition identifier field records the identifiers of proposition A1 and proposition B1.

[0090] For propositions that do not have a direct negation relationship but involve the same knowledge object, the determination of whether they involve the same knowledge object is made by checking whether the subject of the proposition points to the same entity or concept. The determination of whether the subject points to the same knowledge object adopts entity disambiguation and synonym matching technology. When two proposition subjects are mapped to the same knowledge node in the educational knowledge graph or have a synonym relationship, they are identified as involving the same knowledge object.

[0091] When constructing the premise set and conclusion set for each proposition, clauses expressing conditional relationships in the propositional expression are identified as premises, and main clauses are identified as conclusions. The premise set contains all the precondition expressions upon which the proposition's reasoning depends, and the conclusion set contains the judgment or result expression derived from the proposition. The premise set is structured as a list, with each element containing a condition description text field, a knowledge attribute field related to the condition, and an attribute value field. The conclusion set is also structured as a list, with each element containing a conclusion description text field and a conclusion truth value field.

[0092] The precondition sets of two propositions involving the same knowledge object are compared in pairs. Each precondition in the precondition set of proposition A1 is denoted as Precondition A1, and each precondition in the precondition set of proposition B1 is denoted as Precondition B1. Precondition A1 and Precondition B1 are checked to see if they involve the same knowledge attribute. The determination of identical knowledge attributes is achieved by comparing the attribute fields in the preconditions. When the attribute field values ​​are the same, it is determined that they involve the same knowledge attribute. If they involve the same knowledge attribute but the attribute values ​​of Precondition A1 and Precondition B1 are different and mutually exclusive, it is determined that there is an attribute value conflict. For example, Precondition A1 states "the function is continuous within the interval" while Precondition B1 states "the function is discontinuous within the interval," and the attribute values ​​"continuous" and "discontinuous" of the two preconditions are mutually exclusive.

[0093] The conclusion sets of two propositions involving the same knowledge object and having the same set of premises are compared. The determination of identical premise sets is achieved by comparing each premise in the two sets one by one; the premise sets are considered identical when the descriptive text and attribute values ​​of all premises are completely consistent. Under the same premises, if the conclusion sets of proposition A1 and proposition B1 contain mutually exclusive conclusion expressions, they are deemed logically incompatible. The determination of mutually exclusive conclusions employs antonyms and logical opposition rules similar to direct negation; the conclusions are considered mutually exclusive when the truth values ​​of the two conclusions are opposite or when the descriptive text of the conclusions expresses contradictory semantics.

[0094] Proposition pairs whose premise sets conflict in terms of attribute values ​​or whose conclusion sets cannot be simultaneously true are marked as knowledge contradictions. The marking information of knowledge contradictions is stored in a contradiction record data structure, which includes a contradiction type field, a proposition identifier field, and a contradiction description field. The contradiction type field is set to "attribute value conflict" or "conclusions cannot be simultaneously true" based on the detection result. The proposition identifier field records the proposition identifiers involved in the contradiction, and the contradiction description field generates text explaining the specific content of the contradiction.

[0095] This implementation method accurately identifies direct negation relationships and knowledge contradictions between statement fragments through propositional logical deduction and attribute value comparison, providing a quantitative basis for marking disputed fragments and extracting the focus of disputes.

[0096] In one optional implementation, for the point of contention, a negotiation and interaction process between agents is constructed. In this process, agents holding different viewpoints sequentially provide reasoning to support their viewpoints, and each viewpoint is evaluated based on the logical completeness and knowledge consistency of the reasoning. The step of selecting the viewpoint with the highest evaluation value as the negotiation conclusion for the point of contention includes:

[0097] Extract the current point of contention and its associated contention segments, and determine the agents holding different viewpoints and their speaking order based on the source identifiers of the candidate solutions to the contention segments.

[0098] According to the speaking order, the reasoning basis request instruction is sent to each agent in sequence. The reasoning basis request instruction contains the current point of contention and the reasoning basis of the speaking agents. Each agent generates reasoning basis, which includes a reasoning step sequence, knowledge source and question content. The question content points to the reasoning jump position or the missing premise position in the reasoning step sequence of the speaking agents.

[0099] Cross-validation analysis is performed on the reasoning basis of each agent to identify the overlapping parts of knowledge sources and extract mutually corroborating knowledge nodes to improve the credibility weight of the corresponding agent; the number of questioned reasoning jump positions and missing premise positions in the reasoning basis of each agent is counted to reduce the credibility weight of the corresponding agent.

[0100] The logical completeness score is calculated based on the number of reasoning jumps and missing premises. The knowledge consistency score is calculated based on the overlap between the knowledge source and the knowledge domain of the question to be answered. The evaluation value of each agent's viewpoint is calculated in combination with the adjusted credibility weight. The viewpoint with the highest evaluation value is selected as the negotiation conclusion.

[0101] For example, a data structure for reading disputed points from a set of disputed points is constructed to build a negotiation and interaction process between intelligent agents. This data structure includes a disputed point identifier, a disputed point description text, a list of associated disputed fragment identifiers, and a list of candidate solution content source identifiers.

[0102] The source identifiers of candidate solutions for each disputed segment are extracted from the associated list of disputed segment identifiers. After deduplicating the source identifiers, a list of agent identifiers holding different viewpoints is obtained. The speaking order is determined by a sorting strategy based on the historical Q&A quality scores in the agent's ability feature vector. The historical Q&A quality scores of each agent are extracted, and the agent identifiers are arranged in descending order of scores. Agents with higher scores speak first, and the speaking order is stored as an ordered list structure.

[0103] The data structure for the inference basis request command is encapsulated in JSON format, including the current dispute focus field, the inference basis field of the speaking agents, the response time limit field, and the inference basis format specification field. The current dispute focus field stores a description of the dispute focus. The inference basis field of the speaking agents is a list structure, where each element contains the identifier of a speaking agent and the complete content of their provided inference basis. Initially, this field is empty when the speaking agent receives the command; subsequent speaking agents will receive a field containing the inference basis of all previously speaking agents. The response time limit field is set to a value between 60 and 180 seconds, with a default of 120 seconds. Commands are sent sequentially to the corresponding agent's interface endpoint via a message queue in the order of speaking.

[0104] The data structure upon which the reasoning is based includes a reasoning step sequence field, a knowledge source field, and a questioning content field. The reasoning step sequence is stored as an ordered list, where each element represents a reasoning step. The reasoning step data structure includes a step identifier, a step description, the preconditions the step depends on, and the conclusion reached by the step. Step identifiers are integers starting from 1 and incrementing. The knowledge source field stores identification information for references, textbook chapters, or authoritative materials supporting the reasoning steps. The knowledge source identifier includes the source type, source name, and specific chapter or page number.

[0105] The "Question Content" field is generated based on the sequence of reasoning steps of the speaking agents. Reasoning jumps are identified by checking the logical continuity between adjacent steps. If there is a lack of intermediate derivation between the conclusion of step N and the premise of step N+1, a reasoning jump is marked between step N and step N+1. Premise missing locations are identified by checking the completeness of the premises of the reasoning steps. If the conclusion of a reasoning step depends on an assumption not listed in the premises, that reasoning step is marked as having a premise missing. The "Question Content" data structure includes the identifier of the agent being questioned, the identifier of the step being questioned, the "Question Type" field, and the text of the reason for questioning. The "Question Type" field takes the value "Reasoning Jump" or "Premise Missing".

[0106] Cross-validation analysis extracts knowledge source fields from the reasoning provided by each agent, constructing a knowledge source association matrix. Matrix rows represent agent identifiers, columns represent knowledge source identifiers, and matrix elements are Boolean values ​​indicating whether the corresponding agent references the corresponding knowledge source. Overlapping knowledge sources are identified by traversing the matrix columns; if multiple elements in a column are true, the knowledge source corresponding to that column is considered overlapping. Mutually corroborating knowledge nodes are extracted by extracting the knowledge point units involved in the reasoning steps of each agent that references that knowledge source for each overlapping knowledge source. If multiple agents consistently express the same knowledge point in their reasoning steps, that knowledge point is marked as a mutually corroborating knowledge node.

[0107] The initial credibility weight for each agent is set to 1.0, with the increase calculated as the number of mutually corroborating knowledge nodes × 0.1, and the upper limit of the weight set to 2.0. The number of questions raised is accumulated by traversing all questioning content fields provided by all agents, and the decrease is calculated as the number of questions raised × 0.15, with the lower limit of the weight set to 0.2. The adjusted credibility weight is the initial value + increase - decrease.

[0108] Logical completeness score = 1.0 - number of reasoning jumps × 0.2 - number of missing premises × 0.2, with a lower limit of 0. Knowledge consistency score is calculated based on the overlap between the knowledge source and the knowledge domain of the question to be solved. The knowledge domain is determined by the subject area to which the knowledge points involved in the question belong. The domain attribute of the knowledge source is extracted from the metadata of the knowledge source identifier. The overlap is calculated as the number of knowledge sources referenced by the agent that belong to the knowledge domain of the question to be solved ÷ the total number of knowledge sources referenced by the agent. The knowledge consistency score equals the overlap value, ranging from 0 to 1.

[0109] The evaluation value is calculated as follows: Logical completeness score × 0.4 + Knowledge consistency score × 0.3 + Adjusted credibility weight × 0.3, with a range of 0 to 2.6. The evaluation values ​​of all agents are compared, and the viewpoint of the agent with the highest evaluation value is taken as the negotiation conclusion for that point of contention. The negotiation conclusion data structure includes the conclusion text, the identifiers of the agents supporting the conclusion, the evaluation value, and the complete content of the reasoning basis.

[0110] This invention quantitatively evaluates the logical completeness and knowledge consistency of each agent's viewpoint through multi-round negotiation and cross-validation based on reasoning, and selects the optimal viewpoint as the negotiation conclusion based on a dynamic adjustment mechanism of credibility weight.

[0111] In one optional implementation, the step of performing cross-validation analysis on the reasoning basis of each agent includes:

[0112] Extract the knowledge sources and reasoning step sequences from the reasoning basis provided by each intelligent agent, and construct the reasoning path graph of each intelligent agent. The reasoning steps in the reasoning path graph are used as nodes, and the logical deduction relationship between the steps is used as directed edges.

[0113] Node matching is performed on the reasoning path graphs of different agents. Node pairs that are the same in terms of reasoning steps and knowledge sources are identified as mutual corroboration nodes. The proportion of mutual corroboration nodes in the reasoning path graphs of each agent is counted. The higher the proportion, the greater the increase in the credibility weight of the agent.

[0114] Extract the questioned reasoning jump positions or missing premise positions from the questioning content provided by each agent, mark the questioned nodes in the reasoning path graph of the corresponding questioned agent, calculate the criticality of the questioned node in the reasoning path graph, the criticality is calculated by counting the number of subsequent reasoning steps that depend on the node, and reduce the credibility weight of the agent based on the number of questioned nodes and criticality.

[0115] For example, cross-validation analysis extracts knowledge sources and reasoning step sequences from the reasoning basis provided by each agent. Knowledge sources include reference identifiers, textbook chapter identifiers, or authoritative material identifiers. The reasoning step sequence is an ordered list, and each element in the list includes a step identifier, step description text, preconditions, and conclusion.

[0116] The reasoning path graph is constructed using a directed graph data structure. Reasoning steps are stored as nodes in a node set. The node data structure includes a node identifier field, a step description text field, a knowledge source field, a prerequisite field, and a conclusion field. The node identifier field generates a unique identifier by combining the agent identifier and the step identifier, ensuring that nodes from different agents are distinguishable. The logical deduction relationships between steps are stored as directed edges in an edge set. The edge data structure includes a start node identifier field, an end node identifier field, and a deduction type field. The deduction type field indicates whether the edge is a causal deduction, a conditional deduction, or a conclusion deduction. The identification of directed edges is achieved by analyzing the logical relationships between adjacent steps in the reasoning step sequence. When the conclusion of step M serves as a prerequisite for step N, a directed edge is established from node M to node N.

[0117] Node matching in the inference path graphs of different agents is achieved by comparing the step description text and knowledge source of each node pair. Node pairs are selected from the node sets of the inference path graphs of two different agents, and all possible combinations are iterated. Identification of identical inference steps is performed using semantic similarity calculation. The step description texts of the two nodes are extracted, and the cosine similarity of the text vectors is calculated. A similarity threshold of 0.85 is set; if the similarity is greater than or equal to this threshold, the inference steps are considered identical. Identification of identical knowledge sources is achieved by comparing the knowledge source fields of the two nodes; if the knowledge source identifiers are completely identical, the knowledge sources are considered identical.

[0118] Node pairs with identical reasoning steps and knowledge sources are marked as mutually corroborating nodes. The data structure for mutually corroborating nodes includes a node pair identifier, a list of involved agent identifiers, and a corroboration type field. The corroboration type field indicates whether the node pair belongs to definition corroboration, reasoning corroboration, or conclusion corroboration. The percentage of mutually corroborating nodes in each agent's reasoning path graph is calculated by dividing the number of mutually corroborating nodes the agent participates in by the total number of nodes in the agent's reasoning path graph; the percentage ranges from 0 to 1.

[0119] The increase in credibility weight is positively correlated with the proportion. The increase is calculated as proportion × weight adjustment coefficient. With the weight adjustment coefficient set to 0.5, the increase is 0.1 when the proportion is 0.2, 0.3 when the proportion is 0.6, and 0.5 when the proportion is 1.0. Based on the initial adjustment of credibility weights made in the previous cross-validation analysis, the weights are further increased here according to the proportion of mutually corroborating nodes. The final credibility weight is the weight adjusted in the previous analysis plus the increase.

[0120] The content of the challenge is extracted from the reasoning provided by each agent. The data structure of the challenge content includes the identifier of the agent being challenged, the identifier of the step being challenged, the challenge type field, and the text of the reason for the challenge. The challenged reasoning jump position corresponds to the edge between two nodes in the reasoning path graph that lacks an intermediate derivation step, and the challenged premise missing position corresponds to the incomplete set of premises of a node in the reasoning path graph.

[0121] In the reasoning path graph of the agent being questioned, the nodes being questioned are marked by modifying the annotation field of the node data structure. The annotation field is a Boolean value, where a true value indicates that the node is questioned, and a false value indicates that the node is not questioned. For questions about reasoning jump positions, the starting and ending nodes involved are marked. For questions about missing premise positions, the node corresponding to that reasoning step is marked.

[0122] The criticality of a questioned node in the inference path graph is calculated by counting the number of subsequent inference steps that depend on that node. Dependencies are determined by traversing the directed edges of the inference path graph. When there is a node reachable from the questioned node, that node is a subsequent inference step that depends on the questioned node. The criticality is calculated using a depth-first search algorithm that traverses the inference path graph, visiting all reachable nodes along the directed edges starting from the questioned node, and counting the number of reachable nodes as the criticality value. The criticality value ranges from 0 to the total number of nodes in the inference path graph minus 1. A criticality of 0 indicates that the node is a terminal node with no subsequent dependencies, while a higher criticality indicates that the node is earlier in the inference chain and has a larger influence range.

[0123] The reduction in credibility weight is related to the number and criticality of the nodes being questioned. The reduction is calculated as: number of questioned nodes × average criticality × weight reduction coefficient, with the weight reduction coefficient set to 0.1. The average criticality is the sum of the criticalities of all questioned nodes ÷ the number of questioned nodes. Based on the initial adjustment of credibility weights made in the previous cross-validation analysis, the weights are further reduced here according to the number and criticality of the nodes being questioned. The final credibility weight is the previously adjusted weight minus the reduction. The lower limit of the weight is set to 0.2; if the calculated result is less than the lower limit, the lower limit value is used.

[0124] Taking the reasoning path graphs of two agents as an example, agent A's reasoning path graph contains 5 nodes, A1 to A5 representing reasoning steps respectively. Agent B's reasoning path graph contains 4 nodes, B1 to B4. Node matching reveals that the semantic similarity of the step description texts of nodes A2 and B2 is 0.92, and both have the knowledge source identifier of Chapter 3, Section 2 of the textbook, thus marking them as mutually corroborating nodes. The proportion of mutually corroborating nodes for agent A is 1 ÷ 5 = 0.2, with an improvement of 0.2 × 0.5 = 0.1. Agent B's challenge points out that node A3 of agent A has a missing premise, marking node A3 as the challenged node. The criticality of node A3 is calculated to be 2, because nodes A4 and A5 depend on node A3. The number of challenged nodes for agent A is 1, the average criticality is 2, and the reduction is 1 × 2 × 0.1 = 0.2.

[0125] This invention uses a reasoning path graph to structurally represent the reasoning process. Based on node matching and key metric evaluation, it assesses the degree of mutual corroboration and the impact of being questioned in the reasoning of different agents, providing a precise basis for adjusting credibility weights.

[0126] In one optional implementation, the step of fusing the uncontroversial portions of each candidate answer with the negotiated conclusion to generate a consistent answer result and presenting it to the learner includes:

[0127] A knowledge node matching analysis is performed on the undisputed part and the negotiation conclusion to identify knowledge node pairs with semantic overlap. For each knowledge node pair, its knowledge completeness in the undisputed part and the negotiation conclusion is calculated. The knowledge completeness is calculated by statistically associating the number of preconditions and the number of reasoning steps. The expression of the party with higher knowledge completeness is retained and the expression of the other party is deleted.

[0128] Logical breakpoint detection is performed on the connection position between the non-disputed part and the negotiation conclusion to identify the missing intermediate reasoning steps or preconditions between the knowledge nodes before and after the connection position, and the reasoning path connecting the knowledge nodes before and after the connection position is extracted from the content of each candidate solution as a logical connection segment.

[0129] The Q&A results are constructed in the order of non-controversial parts and negotiated conclusions. The logical connection segments are inserted at the connecting positions to generate a consistent Q&A result to be presented to the learner.

[0130] For example, the uncontroversial portion is extracted from the set of uncontroversial fragments obtained from the preceding semantic alignment analysis, and the negotiation conclusion is obtained from the output of the negotiation interaction process. The extraction of knowledge nodes is achieved by mapping the uncontroversial portion and the negotiation conclusion to concept nodes in the educational knowledge graph. Each knowledge node corresponds to an independent knowledge concept. The node data structure includes a node identifier, a concept name, an associated text fragment identifier, and a source identifier field. The source identifier field indicates that the knowledge node comes from the uncontroversial portion or the negotiation conclusion.

[0131] Semantically overlapping knowledge node pairs are identified by calculating the semantic similarity between knowledge nodes from different sources. A node is selected from the set of knowledge nodes in the non-disputed portion, denoted as node A, and a node is selected from the set of knowledge nodes in the negotiated conclusion, denoted as node B. The text fragments associated with the two nodes are extracted, and the cosine similarity of the text vectors is calculated. A similarity threshold of 0.8 is set. When the similarity is greater than or equal to this threshold, node A and node B are considered to constitute a semantically overlapping knowledge node pair. The knowledge node pair data structure includes a node pair identifier, a node A identifier, a node B identifier, and a similarity value field.

[0132] Knowledge completeness is calculated based on the preconditions and reasoning steps associated with knowledge nodes. Preconditions are extracted from the text fragments associated with the knowledge nodes; they are the underlying assumptions or known facts required to support the expression of the knowledge node. Extraction is achieved by identifying conditional clauses and pre-explanatory statements in the text fragments. Reasoning steps are also extracted from the text fragments associated with the knowledge nodes; these are the logical steps that derive from the preconditions to the expression of the knowledge node. Extraction is achieved by analyzing the reasoning structure of the text fragments and dividing them into reasoning stages. Knowledge completeness is calculated as the number of preconditions + the number of reasoning steps × 2. The weighting coefficient for reasoning steps is set to 2 to emphasize the importance of the reasoning process. The value of knowledge completeness ranges from 0 to an unlimited integer.

[0133] For each knowledge node pair, the knowledge completeness of node A in the non-disputed portion and the knowledge completeness of node B in the negotiation conclusion are calculated separately. The two knowledge completeness values ​​are compared, and the statement with the higher knowledge completeness is retained while the other is deleted. Retention is achieved by keeping the associated text fragments of the corresponding knowledge node in the fusion result, while deletion is achieved by removing the associated text fragments of the corresponding knowledge node from the fusion result. When the two knowledge completeness values ​​are equal, the statement in the negotiation conclusion is retained first, because the negotiation conclusion has been verified through multi-agent negotiation.

[0134] The connection point between the uncontroversial part and the negotiated conclusion is determined by analyzing the knowledge node sequences of the two parts. The knowledge nodes of the uncontroversial part are arranged in a sequence according to their order of appearance in the candidate solutions, and the knowledge nodes of the negotiated conclusion are also arranged in a sequence. The connection point is the link between the last node of the uncontroversial part sequence and the first node of the negotiated conclusion sequence. The connection point data structure includes a previous node identifier, a next node identifier, and a connection type field.

[0135] Logical breakpoint detection is achieved by examining the logical continuity between knowledge nodes before and after a connecting point. It extracts the conclusion statement of the preceding node and the precondition statement of the following node, determining whether the conclusion of the preceding node satisfies the requirements of the preconditions of the following node. If there is a lack of logical connection between the conclusion of the preceding node and the preconditions of the following node, a logical breakpoint is marked at that connecting point. Missing intermediate reasoning steps are identified by analyzing the semantic gap between the conclusion of the preceding node and the preconditions of the following node. The semantic gap represents the reasoning bridge required to derive the preconditions from the conclusion of the preceding node. Missing preconditions are identified by analyzing the completeness of the statement in the following node. When the statement in the following node relies on an assumption not stated in the preceding node, that assumption is marked as a missing precondition.

[0136] Logical connection segments are extracted from each candidate answer. The extraction strategy involves retrieving reasoning paths connecting knowledge nodes before and after the connection point within the candidate answer content. Reasoning path identification employs a reasoning path graph matching algorithm, constructing the target connection path between the node before and after the connection point. Subgraphs containing the target connection path are then searched within the reasoning path graph corresponding to each candidate answer. The node sequence of the subgraph serves as the logical connection segment. The logical connection segment data structure includes a segment identifier, node sequence, source candidate answer content identifier, and segment text field. When multiple candidate answers contain reasoning paths that meet the criteria, the path with the fewest reasoning steps is selected as the logical connection segment to maintain the simplicity of the answer result.

[0137] The Q&A results are constructed using a text concatenation and insertion strategy. The text content is arranged in the order of non-disputed sections and negotiated conclusions. The text content of the non-disputed sections is derived from the text fragments associated with the retained knowledge nodes, combined in their original order. The text content of the negotiated conclusions is extracted from the conclusion text field of the negotiated conclusions data structure. Logical connecting fragments are inserted at the connection points. The insertion operation is achieved by adding a fragment of text that connects the non-disputed sections between the end of the non-disputed sections and the beginning of the negotiated conclusions. When adding the logical connecting fragment, a newline character is inserted before and after the logical connecting fragment to maintain the clarity of the paragraph structure.

[0138] The data structure for the consistency Q&A results includes a text field for the Q&A result, a list of cited knowledge sources, and a generation timestamp field. The text field stores the complete, merged Q&A content, while the list of cited knowledge sources summarizes all knowledge source identifiers involved in the undisputed parts, the negotiated conclusions, and logically connected segments. The Q&A results are output to the learner's terminal through a user interface presentation module, supporting segmented display and knowledge source annotation.

[0139] Taking a fusion scenario containing 3 uncontroversial segments and 1 negotiated conclusion as an example, 5 knowledge nodes are extracted from the uncontroversial part and 3 knowledge nodes are extracted from the negotiated conclusion. Matching analysis reveals that node N2 in the uncontroversial part and node C1 in the negotiated conclusion form a semantically overlapping node pair with a similarity of 0.87. The knowledge completeness of node N2 is 2 premises + 3 reasoning steps × 2 = 8, and the knowledge completeness of node C1 is 3 premises + 2 reasoning steps × 2 = 7. The statement of node N2 is retained, and the statement of node C1 is deleted. The connection position is between node N5 at the end of the uncontroversial part and node C2 at the beginning of the negotiated conclusion. Logical breakpoint detection reveals that the conclusion statement "the derivative of the function is greater than 0" in node N5 lacks a reasoning step to determine monotonicity from the sign of the derivative between the premise "the function is monotonically increasing" in node C2. The segment containing this reasoning step is extracted from the candidate solution content as a logical connection segment, inserted into the connection position, and a complete answer is generated.

[0140] This invention removes redundant statements through knowledge integrity measurement and comparison, and generates logically coherent and complete consistent Q&A results by detecting logical breakpoints and extracting reasoning paths to complete the reasoning chain at the connection points.

[0141] A second aspect of this invention provides a multi-agent collaborative Q&A interactive system for educational scenarios, comprising:

[0142] The question retrieval module is used to retrieve unanswered questions submitted by learners.

[0143] The question distribution module is used to distribute the question to be answered to multiple agents, and each agent independently generates candidate answers for the question to be answered.

[0144] The semantic alignment analysis module is used to perform semantic alignment analysis on the content of each candidate solution, identify disputed segments with logical conflicts or knowledge contradictions, and extract the focus of the dispute corresponding to the disputed segments.

[0145] The negotiation and interaction module is used to construct a negotiation and interaction process between intelligent agents for the disputed issue. In the negotiation and interaction process, intelligent agents with different viewpoints provide reasoning evidence to support their viewpoints in turn, and evaluate each viewpoint based on the logical completeness and knowledge consistency of the reasoning evidence. The viewpoint with the highest evaluation value is selected as the negotiation conclusion for the disputed issue.

[0146] The result fusion module is used to merge the uncontroversial parts of each candidate answer with the negotiated conclusion to generate a consistent answer result for the learner.

[0147] A third aspect of the present invention provides an electronic device, comprising:

[0148] processor;

[0149] Memory used to store processor-executable instructions;

[0150] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0151] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0152] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0153] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-agent collaborative Q&A interaction method in educational scenarios, characterized in that: include: Obtain the unanswered questions raised by learners; distribute the unanswered questions to multiple agents, and each agent independently generates candidate solutions to the unanswered questions; Semantic alignment analysis is performed on each candidate solution to identify disputed segments with logical conflicts or knowledge contradictions, and the focus of the dispute corresponding to the disputed segments is extracted. For the aforementioned point of contention, a negotiation and interaction process between intelligent agents is constructed. In this process, intelligent agents with different viewpoints provide reasoning evidence to support their viewpoints in turn. The viewpoints are evaluated based on the logical completeness and knowledge consistency of the reasoning evidence, and the viewpoint with the highest evaluation value is selected as the negotiation conclusion for the point of contention. The uncontroversial parts of each candidate answer are integrated with the negotiated conclusion to generate a consistent answer result, which is then presented to the learner.

2. The method according to claim 1, characterized in that, The step of distributing the unsolved question to multiple agents, with each agent independently generating candidate solutions for the unsolved question, includes: The problem to be solved is decomposed into knowledge points and analyzed in terms of cognitive hierarchy to obtain multiple knowledge point units that constitute the problem to be solved and the dependencies between each knowledge point unit. Multiple agents are selected from the agent resource pool to participate in collaborative Q&A based on the knowledge point units and the dependencies. Each agent in the agent resource pool has a feature vector that represents its knowledge coverage and historical Q&A quality. Based on the capability feature vector, each agent is assigned a scope of responsibility for answering questions, and the scope of responsibility for answering questions limits the subset of knowledge point units that each agent should cover. The question distribution instruction, which includes the question to be answered and the scope of responsibility for answering the question, is sent to the corresponding intelligent agent, and each intelligent agent independently generates candidate answer content.

3. The method according to claim 1, characterized in that, The steps of performing semantic alignment analysis on each candidate solution, identifying disputed segments with logical conflicts or knowledge contradictions, and extracting the points of contention corresponding to the disputed segments include: Each candidate solution is decomposed into multiple solution fragments carrying independent semantics, and each solution fragment is labeled with the source identifier of the candidate solution content to which it belongs; For each knowledge dimension in the question to be answered, based on the source identifier of the candidate answer content, extract the expression fragments for that knowledge dimension from the answer fragments of different candidate answer content, and form a fragment comparison group by combining the expression fragments from different candidate answer content source identifiers for the same knowledge dimension. Logical consistency and knowledge contradiction detection are performed on the statement fragments in each comparison group. If there are mutually exclusive reasoning premises or mutually contradictory conclusions between the statement fragments in the comparison group, the statement fragment is marked as a disputed fragment, and the divergence points that cause the mutually exclusive reasoning premises or mutually contradictory conclusions are extracted as the focus of the dispute.

4. The method according to claim 3, characterized in that, The steps for performing logical consistency checks and knowledge contradiction detection on the statement fragments in each comparison group include: Extract the set of propositions for each statement segment in the segment comparison group; Logical relationship deduction is performed on the set of propositions with different statements within the same segment comparison group. Proposition pairs with direct negation relationship are identified and marked as logical conflicts. The direct negation relationship means that if one proposition is true, the other proposition must be false. For propositions that do not have a direct negation relationship but involve the same knowledge object, a set of premises and a set of conclusions for each proposition are constructed. Proposition pairs that have conflicting attribute values ​​between their premise sets or whose logical conclusions cannot be simultaneously true are identified and marked as knowledge contradictions. The conflicting attribute values ​​refer to the same knowledge attribute being assigned different attribute values. The logical incompatibility refers to the two conclusions not being true at the same time under the same premises.

5. The method according to claim 1, characterized in that, To address the points of contention, a negotiation and interaction process is constructed between intelligent agents. In this process, agents holding different viewpoints sequentially provide reasoning supporting their arguments. Each viewpoint is evaluated based on the logical completeness and knowledge consistency of its reasoning, and the viewpoint with the highest evaluation value is selected as the negotiation conclusion for that point of contention. The steps include: Extract the current point of contention and its associated contention segments, and determine the agents holding different viewpoints and their speaking order based on the source identifiers of the candidate solutions to the contention segments. According to the speaking order, the reasoning basis request instruction is sent to each agent in sequence. The reasoning basis request instruction contains the current point of contention and the reasoning basis of the speaking agents. Each agent generates reasoning basis, which includes a reasoning step sequence, knowledge source and question content. The question content points to the reasoning jump position or the missing premise position in the reasoning step sequence of the speaking agents. Cross-validation analysis is performed on the reasoning basis of each agent to identify the overlapping parts of knowledge sources and extract mutually corroborating knowledge nodes to improve the credibility weight of the corresponding agent; the number of questioned reasoning jump positions and missing premise positions in the reasoning basis of each agent is counted to reduce the credibility weight of the corresponding agent. The logical completeness score is calculated based on the number of reasoning jumps and missing premises. The knowledge consistency score is calculated based on the overlap between the knowledge source and the knowledge domain of the question to be answered. The evaluation value of each agent's viewpoint is calculated in combination with the adjusted credibility weight. The viewpoint with the highest evaluation value is selected as the negotiation conclusion.

6. The method according to claim 5, characterized in that, The steps for cross-validation analysis of the reasoning basis of each agent include: Extract the knowledge sources and reasoning step sequences from the reasoning basis provided by each intelligent agent, and construct the reasoning path graph of each intelligent agent. The reasoning steps in the reasoning path graph are used as nodes, and the logical deduction relationship between the steps is used as directed edges. Node matching is performed on the reasoning path graphs of different agents. Node pairs that are the same in terms of reasoning steps and knowledge sources are identified as mutual corroboration nodes. The proportion of mutual corroboration nodes in the reasoning path graphs of each agent is counted. The higher the proportion, the greater the increase in the credibility weight of the agent. Extract the questioned reasoning jump positions or missing premise positions from the questioning content provided by each agent, mark the questioned nodes in the reasoning path graph of the corresponding questioned agent, calculate the criticality of the questioned node in the reasoning path graph, the criticality is calculated by counting the number of subsequent reasoning steps that depend on the node, and reduce the credibility weight of the agent based on the number of questioned nodes and criticality.

7. The method according to claim 1, characterized in that, The steps of integrating the uncontroversial parts of each candidate answer with the negotiated conclusion to generate a consistent answer result and present it to the learner include: A knowledge node matching analysis is performed on the undisputed part and the negotiation conclusion to identify knowledge node pairs with semantic overlap. For each knowledge node pair, its knowledge completeness in the undisputed part and the negotiation conclusion is calculated. The knowledge completeness is calculated by statistically associating the number of preconditions and the number of reasoning steps. The expression of the party with higher knowledge completeness is retained and the expression of the other party is deleted. Logical breakpoint detection is performed on the connection position between the non-disputed part and the negotiation conclusion to identify the missing intermediate reasoning steps or preconditions between the knowledge nodes before and after the connection position, and the reasoning path connecting the knowledge nodes before and after the connection position is extracted from the content of each candidate solution as a logical connection segment. The Q&A results are constructed in the order of non-controversial parts and negotiated conclusions. The logical connection segments are inserted at the connecting positions to generate a consistent Q&A result to be presented to the learner.

8. A multi-agent collaborative Q&A interactive system for educational scenarios, used to implement the method of any one of claims 1-7, characterized in that, include: The question retrieval module is used to retrieve unanswered questions submitted by learners. The question distribution module is used to distribute the question to be answered to multiple agents, and each agent independently generates candidate answers for the question to be answered. The semantic alignment analysis module is used to perform semantic alignment analysis on the content of each candidate solution, identify disputed segments with logical conflicts or knowledge contradictions, and extract the focus of the dispute corresponding to the disputed segments. The negotiation and interaction module is used to construct a negotiation and interaction process between intelligent agents for the disputed issue. In the negotiation and interaction process, intelligent agents with different viewpoints provide reasoning evidence to support their viewpoints in turn, and evaluate each viewpoint based on the logical completeness and knowledge consistency of the reasoning evidence. The viewpoint with the highest evaluation value is selected as the negotiation conclusion for the disputed issue. The result fusion module is used to merge the uncontroversial parts of each candidate answer with the negotiated conclusion to generate a consistent answer result for the learner.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images