A large model-based education question and answer method and system

Abstract

The application discloses an education question and answer method and system based on a large model, relates to the technical field of artificial intelligence education question and answer, and comprises the following steps: receiving and analyzing an education question and answer request input by a student, generating a structured question object, identifying a student error mode in combination with a student historical learning record, and generating a hierarchical prompt strategy; retrieving relevant content from an education knowledge base according to the structured question object, forming a hierarchical evidence package according to evidence sources and course adaptation marks; performing evidence unit splitting on the hierarchical evidence package, extracting an education proposition to be verified, and calculating an education evidence conflict degree through relationship annotation; resolving conflicting evidence according to the education evidence conflict degree and an evidence level, and generating a controlled evidence set; and constructing an education question and answer prompt instance, and calling a large model to generate a candidate answer. The application can improve the output reliability of an education question and answer result in a multi-source evidence conflict scene, and improve the pertinence, guidance and learning assistance effect of the education question and answer.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence educational question-answering technology, and in particular to an educational question-answering method and system based on a large model. Background Technology

[0002] In recent years, the development of artificial intelligence and educational informatization technologies has driven the application of intelligent education systems. Large-scale model-based educational question-answering methods, due to their capabilities in natural language understanding, knowledge generation, and multi-turn interaction, are increasingly being applied to scenarios such as intelligent teaching assistants, question analysis, knowledge explanation, and learning tutoring. Existing technical solutions typically receive text, voice, question images, or learning files input by students. They then identify student questions through processing methods such as speech-to-text, image recognition, text parsing, and knowledge point matching. These questions are then retrieved by combining textbook resources, question bank analysis, teacher courseware, or historical question-answering content. The retrieval results are then input into the large-scale model as prompts to generate corresponding problem-solving processes, concept explanations, or learning suggestions. These methods, to a certain extent, improve the automation and interaction efficiency of educational question-answering, providing technical support for student self-directed learning and teacher-assisted question answering.

[0003] However, existing methods still have the following shortcomings: On the one hand, existing methods tend to focus on recalling similar educational materials and directly using them for large-scale model generation, lacking an assessment of the supporting, refuting, and conflicting relationships between educational evidence from different sources. When there are differences in expression or applicable conditions between textbook content, teacher handouts, question bank analysis, or historical Q&A, it is easy to generate definitive answers lacking controlled evidence. On the other hand, existing methods usually generate uniform answers around the question itself, making it difficult to identify learning obstacles such as unclear concepts, misuse of formulas, omission of conditions, calculation errors, or misreading of questions by combining students' historical learning records. This results in a mismatch between the prompting order, analysis depth, and openness of the answer and the students' actual error patterns, affecting the relevance and reliability of educational Q&A. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an educational question-answering method and system based on a large model, which solves the problems of uncontrollable handling of conflicts in multi-source educational evidence and insufficient adaptation of question-answering prompts to student error patterns.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides an educational question-answering method based on a large model, comprising:

[0008] It receives educational question and answer requests from students, parses the requests, generates structured question objects, identifies student error patterns based on the structured question objects and students' historical learning records, and generates a hierarchical prompting strategy based on the student error patterns.

[0009] Relevant content is retrieved from the educational knowledge base based on the structured question object, and a graded evidence package is generated according to the evidence source and curriculum adaptation marker; the graded evidence package includes the evidence content and the corresponding evidence level;

[0010] The graded evidence package is split into evidence units, and educational propositions to be verified are extracted based on structured question objects. The relationship between the evidence units and the educational propositions to be verified is labeled, and the degree of conflict of educational evidence is calculated based on the labeling results.

[0011] The evidence units are adjudicated based on the degree of conflict and the level of evidence in the education process, and a controlled evidence set is generated.

[0012] Based on the structured question object, hierarchical prompting strategy, and controlled evidence set, construct educational question-and-answer prompt instances, and call the large model to generate candidate answers based on the educational question-and-answer prompt instances;

[0013] The candidate answers are validated for consistency of evidence to generate the final educational question and answer results.

[0014] As a preferred embodiment of the large-model-based educational question-answering method of the present invention, the generation of structured question objects includes: receiving educational question-answering requests including text questions, voice questions, question images, or uploaded files; performing basic verification based on student identifiers, course identifiers, grade information, and textbook version information; determining the main input content and auxiliary input content based on the input modality; and generating a question-answering processing task; calling the corresponding parsing component according to the input modality in the question-answering processing task; processing and fusing the text, speech-to-text, question image recognition content, or uploaded file parsing content; extracting the standard question text, course subject, grade level, textbook version, question type, candidate knowledge points, known conditions, and target to be obtained; and generating a structured question object.

[0015] As a preferred embodiment of the large-model-based educational question-answering method of the present invention, the step of identifying student error patterns based on structured question objects and students' historical learning records, and generating a hierarchical prompting strategy based on student error patterns, includes: reading students' historical learning records within the candidate knowledge points and associated knowledge point ranges based on student identifiers associated with structured question objects, matching the historical learning records with the question type, candidate knowledge points, known conditions, and target in the structured question object to identify the student error pattern corresponding to the current question; determining the primary error pattern and auxiliary error patterns based on the student error patterns, and configuring the prompting start method, answer openness level, and prompting order according to the primary error pattern to generate a hierarchical prompting strategy.

[0016] As a preferred embodiment of the large-model-based educational question-answering method of the present invention, the step of retrieving relevant content from the educational knowledge base based on the structured question object and generating a hierarchical evidence package according to the evidence source and course adaptation marker includes: generating evidence retrieval conditions based on the structured question object and retrieving content related to the current question from the educational knowledge base; performing adaptation screening on the retrieved content to generate initial evidence records including evidence content, evidence source, textbook version matching marker, knowledge point matching marker, target relevance marker, and knowledge point tag; determining the evidence level according to the evidence source; and sorting the initial evidence records according to the evidence level, textbook version matching marker, knowledge point matching marker, and target relevance marker to generate a hierarchical evidence package.

[0017] As a preferred embodiment of the large-model-based educational question-answering method of the present invention, the step of splitting the hierarchical evidence package into evidence units and extracting educational propositions to be verified based on structured question objects includes: reading the evidence records in the hierarchical evidence package, splitting them into evidence units according to the knowledge expression granularity of definitions, formulas, applicable conditions, problem-solving rules, steps and conclusions, and configuring evidence unit types and evidence citation identifiers for the evidence units; generating educational propositions to be verified based on the question type, candidate knowledge points, known conditions and objectives to be sought in the structured question objects, combined with the course knowledge point catalog.

[0018] As a preferred embodiment of the large-model-based educational question-answering method of the present invention, the step of labeling the evidence units with the educational propositions to be verified and calculating the degree of educational evidence conflict based on the labeling results includes: matching the evidence units with the educational propositions to be verified according to the correspondence between knowledge point tags and types, and labeling the evidence units as supporting, refuting, or irrelevant relationships according to the consistency between the evidence content and the proposition content and applicable conditions; for each educational proposition to be verified, statistically analyzing the supporting and refuting evidence units, and calculating the degree of educational evidence conflict based on the evidence level weights.

[0019] As a preferred embodiment of the large-model-based educational question-answering method of the present invention, the generation of the controlled evidence set includes: adjudicating evidence units with supporting and refuting relationships based on the degree of conflict of educational evidence, the cumulative amount of supporting evidence, the cumulative amount of rebuttal evidence, and the evidence level corresponding to each educational proposition to be verified, and determining the definitive evidence or propositions to be confirmed that can be used as the basis for a definitive answer; generating the controlled evidence set based on the adjudication results, and writing a prohibition on definitive output flag for irrelevant evidence, propositions without valid evidence, evidence corresponding to propositions to be confirmed, and conflicting evidence that has not been adopted.

[0020] As a preferred embodiment of the large-model-based educational question-answering method of the present invention, the step of constructing educational question-answering prompt instances based on structured question objects, hierarchical prompting strategies, and controlled evidence sets, and then calling the large model to generate candidate answers based on the educational question-answering prompt instances, includes: generating a question description area based on structured question objects, generating an answer strategy area based on hierarchical prompting strategies, and generating an evidence constraint area based on determinable evidence, adjudication status, and prohibition of deterministic output markers in the controlled evidence set, thus constructing educational question-answering prompt instances; and calling a text large model or a multimodal large model based on the educational question-answering prompt instances and the input modality, so that the large model determines the answer initiation method, answer opening level, and prompting order according to the answer strategy area, and generates candidate answers containing proposition numbers or evidence citation identifiers within the scope limited by the evidence constraint area.

[0021] As a preferred embodiment of the large-model-based educational question-and-answer method of the present invention, the generation of the final educational question-and-answer result includes: based on the proposition number or evidence citation identifier carried in the candidate answer, comparing the candidate answer fragments with the determinable evidence, adjudication status, and prohibition of deterministic output markers in the controlled evidence set, eliminating candidate answer fragments without evidence support or with conflicting evidence, and generating a set of outputtable answer fragments; and organizing the output content according to the set of outputtable answer fragments and the hierarchical prompting strategy, in accordance with the prompting start method, prompting order, and answer openness level, to generate the final educational question-and-answer result.

[0022] Secondly, the present invention provides an educational question-answering system based on a large model, comprising:

[0023] The educational question-and-answer parsing module is used to receive educational question-and-answer requests from students, parse and merge input content from different modalities, and generate structured question objects.

[0024] The tiered prompting module is used to identify student error patterns based on structured question objects and students' historical learning records, and to generate tiered prompting strategies.

[0025] The evidence retrieval and grading module is used to retrieve relevant content from the educational knowledge base based on structured question objects, and generate graded evidence packages according to evidence sources and course adaptation tags.

[0026] The evidence conflict adjudication module is used to split the graded evidence package into evidence units, extract the educational propositions to be verified, label the relationships and calculate the degree of conflict, and generate a controlled evidence set.

[0027] The controlled question-answer generation module is used to construct educational question-answer prompt instances based on structured question objects, hierarchical prompting strategies, and controlled evidence sets, call the large model to generate candidate answers, and generate the final educational question-answer results after evidence consistency verification.

[0028] The beneficial effects of this invention are as follows: By performing hierarchical retrieval, evidence unit decomposition, and relationship annotation on multi-source content in the educational knowledge base, it can identify conflicting relationships between supporting and rebuttal evidence under the same educational proposition. Furthermore, it generates a controlled evidence set based on evidence level and conflict degree, thereby avoiding large models directly outputting definitive answers based on inconsistent data, and improving the reliability and standardization of educational question-and-answer results. Simultaneously, by combining structured question objects and students' historical learning records to identify student error patterns, and generating a hierarchical prompting strategy accordingly, the answers can adjust the starting point, analysis depth, and openness of the prompts based on the student's actual learning obstacles, enhancing the relevance, guidance, and learning assistance effects of educational question-and-answer. Attached Figure Description

[0029] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 This is a flowchart of the educational question-answering method based on a large model in Example 1.

[0031] Figure 2 This is a structural diagram of the educational question-answering system based on a large model in Example 1.

[0032] Figure 3 This is a structural diagram of the multi-source educational evidence processing in Example 1.

[0033] Figure 4 A framework diagram is constructed for the educational question-and-answer prompt example in Example 1. Detailed Implementation

[0034] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0035] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0036] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0037] Example 1, referring to Figures 1-4 This is the first embodiment of the present invention, which provides an educational question-answering method based on a large model, including the following steps:

[0038] S1. Receive educational question and answer requests input by students, parse the educational question and answer requests, generate structured question objects, identify student error patterns based on structured question objects and students' historical learning records, and generate hierarchical prompt strategies based on student error patterns.

[0039] Receive educational Q&A requests and generate Q&A processing tasks.

[0040] Specifically, students input educational Q&A requests through the client. First, basic verification is performed based on student ID, course ID, and textbook version information to confirm that the request belongs to the scope of the course that can be processed. After the verification is passed, the type of input content in the educational Q&A request is identified, and input modality tags are configured for the input content.

[0041] When a single educational question-and-answer request contains multiple input options, the primary and secondary input options are determined based on the completeness of the question: If the request includes a question image or uploaded file, the question content identified from the image or file is used as the primary input, while the student's text or voice follow-up questions are used as secondary input; if the request only contains text questions, the text questions are used as the primary input; if the request only contains voice questions, the transcribed text is used as the primary input. This primary-secondary input division method aims to use content containing the question stem, formulas, options, or context as the core basis for subsequent analysis, avoiding the mistaken treatment of student follow-up questions as complete questions.

[0042] The student identifier, course identifier, grade information, textbook version information, main input content, auxiliary input content, input modality marker, and request time are encapsulated into a question-and-answer processing task.

[0043] It should be noted that educational Q&A requests include at least one of the following: text questions, voice questions, question images, or uploaded files, and carry student identifiers, course identifiers, grade information, textbook version information, and request time; input modal tags include text modality, voice modality, image modality, and document modality.

[0044] Multimodal parsing is performed on the question-and-answer processing task to generate structured question objects.

[0045] Specifically, the corresponding parsing component is invoked based on the input modality marker. For text modality, the text content is standardized by removing invalid characters, unifying punctuation and mathematical symbols, and extracting the question text, keywords, numerical conditions, unit information, and interrogative expressions. For speech modality, the speech content is first transcribed into text, and then the subject terms, formula readings, and unit readings in the transcribed text are corrected by combining course identifiers and educational terminology glossaries to obtain the speech parsing text. For image modality, the question image is preprocessed and OCR is performed to extract the question stem, options, formulas, chart descriptions, and annotations from the image, and the question reading order is restored according to the page layout. For document modality, the text, tables, and formula content in the uploaded file are extracted, and the question segment corresponding to the current request is located based on the question number, heading level, or student auxiliary input.

[0046] After completing the analysis of each modality, the analysis results of the main input content are used as the main body of the question, and the analysis results of the auxiliary input content are used as follow-up questions or supplementary conditions for students. The results are then integrated in the order of question stem, conditions, options, chart descriptions, and student follow-up questions to generate a standard question text. If the auxiliary input content belongs to the same knowledge point or the same question object as the main input content, the auxiliary input content is written into the supplementary requirements. If the auxiliary input content is unrelated to the main input content, it is not included as the core content of the question in the subsequent analysis.

[0047] Based on the standard question text, further extract question elements to identify the course subject, question type, candidate knowledge points, known conditions, and the goal to be achieved. Encapsulate the standard question text, course subject, grade level, textbook version, question type, candidate knowledge points, known conditions, and the goal to be achieved into a structured question object.

[0048] It should be noted that the known conditions include numerical values, units, object relationships, graphical relationships, experimental conditions, or grammatical conditions given in the problem; the target to be found includes the target content that needs to be calculated, proved, selected, explained, corrected, or generated; candidate knowledge points are determined based on the knowledge point index range limited by the course identifier, grade information, and textbook version information, as well as keywords, formula symbols, numerical units, question type characteristics, and subject terminology extracted from the standard problem text. For example, in one implementation, for the eighth-grade math problem "Given the linear function y=2x+3, find the value of y when x=4, and determine whether the point (4, 11) lies on the graph of the function", the scope of the linear function chapter is limited according to the textbook version, and the candidate knowledge points are determined based on "linear function", "y=2x+3", "function graph", etc., as linear function expression, function value calculation, and the relationship between the point and the function graph. This example is only used to illustrate the process of determining candidate knowledge points. For other subjects, grades, or question types, matching can be performed in the same way based on the corresponding textbook version and knowledge point index.

[0049] Structured problem objects are used to represent the content of the student's current problem, the course boundaries, the scope of knowledge points, the known conditions, and the solution objective, and serve as the basic data for subsequent student error pattern recognition, educational knowledge retrieval, and large-scale model question answering generation.

[0050] Based on the student identifier associated with the structured question object, the student's historical learning records within the candidate knowledge points and associated knowledge points are read, and the historical learning records are matched with the question type, candidate knowledge points, known conditions and the target to be found in the structured question object to identify the student's error pattern corresponding to the current question;

[0051] Specifically, the system reads structured question objects and obtains student identifiers through question-and-answer processing tasks associated with the structured question objects; it extracts question types, candidate knowledge points, known conditions, and objectives from the structured question objects, and reads the student's historical learning records within the candidate knowledge points and associated knowledge points based on the student identifiers.

[0052] It should be noted that related knowledge points are those in the course knowledge graph that have a preceding, following, parent-child, or child relationship with the candidate knowledge points; historical learning records include historical wrong question records, historical follow-up question records, prompt review records, and knowledge point mastery records.

[0053] Furthermore, historical learning records are matched with structured question objects to identify student error patterns corresponding to the current question. This includes: when historical error records or follow-up questions indicate that a student has misunderstood the definition, principle, or conceptual meaning of a corresponding knowledge point, it is identified as a lack of understanding of concepts; when the standard question text contains formula symbols, formula names, or calculation trigger words, the corresponding candidate formulas and their applicable conditions are determined based on the candidate knowledge points, and the known conditions and the desired objective in the structured question object are matched with the applicable conditions of the formulas; if the known conditions or the desired objective do not meet the applicable conditions of the formulas, or if there are errors in formula selection or application in the historical error records, then the student error pattern corresponding to the current question is identified as... Formula misuse; when the known conditions cannot meet the key conditions required for the desired outcome, or when there are omissions in the historical error records, such as missing conditions, ignoring units, or ignoring the scope of limitations; when the candidate knowledge points and solution path can be determined, but the historical error records focus on calculation, substitution, unit conversion, or symbol processing, it is identified as a calculation error; when the desired outcome is inconsistent with the direction of follow-up questions in the student's auxiliary input, or when there are irrelevant answers or misinterpretations of the question in the historical error records, it is identified as a question reading error; when the question type is an explanation question, proof question, essay correction question, or open-ended question, and the historical records show that the student has problems with incomplete step expression, non-standard conclusion expression, or unclear argumentation level, it is identified as an expression error.

[0054] It should be noted that whether the known conditions meet the key conditions required for the desired objective is determined based on the question type rules, formula application conditions, or textbook analysis templates corresponding to the candidate knowledge points; when the known conditions lack the necessary input items required by the question type rules, formula application conditions, or textbook analysis templates, it is determined that the key conditions are incomplete.

[0055] When students' historical learning records are insufficient, the initial error pattern is directly determined based on the structured question object: if the question type is knowledge explanation, it is determined to be unclear concept; if the standard question text contains formula symbols or calculation relationships, it is determined to be formula misuse or calculation error; if the known conditions lack the key conditions required for the desired objective, it is determined to be condition omission; if the desired objective is unclear or inconsistent with the question stem, it is determined to be question misreading; if the question type is explanation, proof, or open expression, it is determined to be expression error.

[0056] Based on the students' error patterns, the primary error pattern and secondary error pattern are determined, and the prompting start method, answer access level and prompting order are configured according to the primary error pattern to generate a hierarchical prompting strategy.

[0057] Specifically, when the same question meets the judgment conditions for multiple error modes, the error mode directly triggered by the current structured question object is first determined as the primary error mode; if no directly triggered error mode exists, the error mode with the highest frequency in the historical wrong question record is determined as the primary error mode; if the historical wrong question record is insufficient to distinguish, the error mode corresponding to the candidate knowledge point with the lowest knowledge point mastery is determined as the primary error mode; in addition to the primary error mode, the other error modes that meet the judgment conditions are used as auxiliary error modes; a hierarchical prompting strategy is generated based on the primary error mode, which includes the prompting start method, the answer open level, and the prompting order.

[0058] It should be noted that the prompt initiation method is used to determine the first content to be output in the answer, the answer open level is used to determine whether to output only prompts, whether to output key steps, or whether to output the complete analysis, and the prompt order is used to determine the order of the various prompt contents; the answer open level includes prompt level, step level, and complete analysis level; different main error modes are only used to determine which level to use first and the order of openness.

[0059] Specifically, when the main error pattern is unclear concepts, the prompting method is set to conceptual explanation, the answer reveal level is set to first prompt the concept and example, then reveal the problem analysis based on the student's follow-up questions, and the prompting order is set to concept definition, applicable scenarios, and problem analysis; when the main error pattern is misuse of formulas, the prompting method is set to prompt the formula's applicable conditions, the answer reveal level is set to first prompt the basis for formula selection, then reveal the substitution steps, and the prompting order is set to formula conditions, variable meanings, formula selection, and substitution calculation; when the main error pattern is missing conditions, the prompting method is set to prompt key conditions, the answer reveal level is set to first prompt or ask about the missing conditions, then reveal the solution steps, and the prompting order is set to locate the conditions in the question stem, prompt the missing conditions, and then reveal the solution. The system guides responses to various errors. When the primary error pattern is a calculation error, the prompting starts with checking intermediate steps, reveals the answer level by first showing key intermediate steps and checking methods, and does not immediately output the complete answer. The prompting order is formula substitution, key calculations, and result verification. When the primary error pattern is a misinterpretation of the question, the prompting starts with restating the question, reveals the answer level by first showing the question requirements and the desired outcome, and then revealing the solution process. The prompting order is question requirements, known conditions, desired outcome, and solution direction. When the primary error pattern is an expression error, the prompting starts with providing the answer structure, reveals the answer level by first revealing the expression framework, and then revealing the reference statement. The prompting order is answer structure, step expression, and conclusion statement. These primary error patterns, secondary error patterns, prompting start methods, answer reveal levels, and prompting order are encapsulated into a layered prompting strategy.

[0060] S2. Retrieve relevant content from the educational knowledge base based on the structured question object, and generate a graded evidence package according to the evidence source and course adaptation mark; the graded evidence package includes the evidence content and the corresponding evidence level;

[0061] Specifically, the system reads structured question objects and generates evidence retrieval conditions. These conditions include: using course subject, grade level, and textbook version as the search scope; using candidate knowledge points and related knowledge points as knowledge point tags; using question type and target as evidence type; and using keywords, formula symbols, and subject-specific terms in the standard question text, known conditions, and target as content matching conditions. Based on these evidence retrieval conditions, relevant content is retrieved from the educational knowledge base. During the retrieval, records matching the course subject, grade level, and textbook version are first filtered from the original knowledge records according to the search scope conditions. Then, records matching candidate knowledge points or related knowledge points are filtered according to the knowledge point tag conditions to obtain candidate knowledge records. Next, definitions, formulas, applicable conditions, problem-solving rules, example analyses, or answer requirements to be recalled are determined according to the evidence type conditions. Finally, the content matching conditions are matched with the titles, text, tags, or analysis fields of the candidate knowledge records to obtain retrieval content relevant to the current question.

[0062] It should be noted that the educational knowledge base includes a teacher explicit requirements database, a current textbook knowledge base, a teacher lecture notes database, a question bank analysis database, and a historical question and answer database. The original knowledge records in the educational knowledge base are pre-configured with course subject, grade level, textbook version, knowledge point tags, evidence source, evidence type, title, and main text content. The teacher explicit requirements database provides teachers with guidelines for explaining knowledge points and specifying answer requirements. The current textbook knowledge base provides definitions, formulas, examples, and explanations from the corresponding textbook version. The teacher lecture notes database provides supplementary classroom content. The question bank analysis database provides standard explanations for reviewed questions. The historical question and answer database provides confirmed historical question and answer content.

[0063] Furthermore, the retrieved content undergoes adaptability screening, generating initial evidence records. These initial evidence records include evidence content, evidence source, textbook version matching marker, knowledge point matching marker, target relevance marker, and knowledge point tag. Evidence levels are determined based on the evidence source in the initial evidence records, forming evidence records in total. Evidence records are grouped according to their evidence levels, and within the same evidence level, they are sorted according to textbook version matching marker, knowledge point matching marker, and target relevance marker, generating hierarchical evidence packages. Each hierarchical evidence package includes evidence content, evidence source, evidence level, textbook version matching marker, knowledge point matching marker, target relevance marker, and knowledge point tag.

[0064] It should be noted that the textbook version matching marker is determined based on whether the textbook version in the structured question object is consistent with the textbook version in the candidate knowledge record; the knowledge point matching marker is determined based on whether the candidate knowledge point in the structured question object is the same as or belongs to the same knowledge point directory node as the knowledge point tag in the candidate knowledge record; the target relevance marker is determined based on whether the evidence type, title or text content of the target to be sought corresponds to the candidate knowledge record; for content that does not match the textbook version, is irrelevant to the knowledge point, or only has keywords that happen to match but do not correspond to the target to be sought, no initial evidence record is generated.

[0065] It should also be noted that the determination of evidence level based on the source of evidence includes: evidence from a teacher's explicit request database is classified as first-level evidence; evidence from the current textbook knowledge base is classified as second-level evidence; evidence from the teacher's lecture notes database is classified as third-level evidence; evidence from the question bank analysis database is classified as fourth-level evidence; and evidence from the historical question and answer database is classified as fifth-level evidence. The priority of evidence levels one through five decreases sequentially.

[0066] S3. The graded evidence package is split into evidence units, and the educational propositions to be verified are extracted based on the structured question objects. The relationship between the evidence units and the educational propositions to be verified is labeled, and the degree of conflict of educational evidence is calculated based on the labeling results.

[0067] Specifically, the process involves reading a tiered evidence package and processing the evidence records within it. The evidence content is then broken down according to the granularity of knowledge expression, dividing segments that can independently express definitions, formulas, applicable conditions, problem-solving rules, key steps, or conclusions into evidence units. Each evidence unit is assigned an evidence unit number and type. Evidence unit types include definition, formula, applicable condition, rule, step, and conclusion types. Each evidence unit inherits the evidence source, evidence level, textbook version matching marker, knowledge point matching marker, target relevance marker, and knowledge point tag from its corresponding evidence record. An evidence citation identifier is generated based on the evidence source, evidence unit number, and the segment position of the evidence unit within its corresponding evidence content. The process also involves reading a structured problem object and extracting the standard problem text, question type, candidate knowledge points, known conditions, and target from the structured problem object. Based on the candidate knowledge points, the process determines the definitions, formulas, formula applicable conditions, problem-solving rules, and standard conclusion types that the current problem may involve from the course knowledge point catalog. The process also determines the type of proposition to be verified based on the question type. Finally, the process generates an educational proposition to be verified corresponding to the current problem based on the known conditions and target.

[0068] It should be noted that when the question type is knowledge explanation, the conceptual meaning, definition boundary, or applicable scenario corresponding to the candidate knowledge point will be generated as a definition proposition to be verified; when the question type is calculation or derivation, the candidate formula, formula application conditions, variable correspondence, and problem-solving rules corresponding to the candidate knowledge point will be generated as a formula proposition to be verified, an application condition proposition to be verified, and a problem-solving rule proposition to be verified; when the question type is proof, judgment, or question-and-answer, the object to be proved, the object to be judged, or the object to be answered in the standard question text will be generated as a conclusion proposition to be verified; when the standard question text contains multiple consecutive problem-solving steps, the key intermediate steps will be generated as step propositions to be verified. Each educational proposition to be verified includes a proposition number, proposition content, proposition type, corresponding knowledge point, and applicable conditions. The proposition content represents the specific judgment that needs to be verified by evidence. The proposition type represents which category the proposition belongs to: definition, formula, applicable conditions, problem-solving rules, steps, or conclusion. The corresponding knowledge point comes from the candidate knowledge points in the structured problem object. The applicable conditions come from the known conditions and the target to be found in the structured problem object.

[0069] For example, if the structured problem object shows the current problem as a calculation problem of "calculating the area of ​​a circle given its radius," the following educational propositions to be verified can be generated: the area of ​​a circle should be calculated using the formula for the area of ​​a circle; the formula for the area of ​​a circle is applicable when the radius or diameter is known; the known conditions in the problem correspond to the radius variable in the formula; and the final objective is the area result. Subsequent evidence units then determine whether these propositions are supported by textbooks, handouts, or teacher requirements.

[0070] Furthermore, first determine whether the knowledge point tags inherited by the evidence unit are the same as or related to the corresponding knowledge points of the educational proposition to be verified, and then determine whether the evidence unit type corresponds to the proposition type of the educational proposition to be verified; if the knowledge point tags are different and unrelated, or the evidence unit type does not correspond to the proposition type, then the evidence unit is marked as irrelevant; if the knowledge point tags are the same or related, and the evidence unit type corresponds to the proposition type, then the evidence unit is determined as the evidence unit to be marked.

[0071] The evidence content of the unit to be labeled is compared with the content and applicable conditions of the educational proposition to be verified. The comparison dimensions include at least one of the following: definition, formula name, applicable conditions of the formula, meaning of variables, unit scope, problem-solving rules, order of steps, and direction of conclusion.

[0072] When the evidence unit to be labeled is consistent with the educational proposition to be verified in the corresponding comparison dimension, and can serve as the basis for the definition, formula, condition, rule, step, or conclusion of the educational proposition to be verified, the evidence unit to be labeled is marked as a supporting relationship; for example, if the educational proposition to be verified is "this problem should use the formula for the area of ​​a circle", and the evidence unit states "the area of ​​a circle is equal to π multiplied by the square of the radius, which is applicable to problems where the area is known by the radius", then the evidence unit constitutes a supporting relationship for the proposition.

[0073] When the evidence unit to be labeled and the educational proposition to be verified belong to the same or related knowledge point range, but there are contents that cannot be simultaneously true in the corresponding comparison dimension, the evidence unit to be labeled is marked as a refutation relationship; for example, the educational proposition to be verified is "this question can be directly applied to the formula of uniform linear motion", while the evidence unit states "this formula is only applicable to motion processes with constant velocity", and the current known conditions show that the velocity changes, then the evidence unit constitutes a refutation relationship for this proposition;

[0074] When an evidence unit to be labeled overlaps with the educational proposition to be verified in terms of keywords, but cannot provide valid evidence for the educational proposition in terms of definition, formula, condition, rule, step or conclusion, the evidence unit to be labeled is marked as irrelevant; for example, if the educational proposition to be verified involves "circle area calculation", but the evidence unit only records the "circumference formula" and does not involve area calculation, then the evidence unit is marked as irrelevant to the proposition.

[0075] The relationship labeling results are bound to the evidence unit number and the educational proposition number to be verified to generate a relationship labeling record. For each educational proposition to be verified, the evidence units labeled as supporting relationships are included in the supporting evidence unit set, the evidence units labeled as refuting relationships are included in the refuting evidence unit set, and the evidence units labeled as irrelevant relationships are excluded from the conflict degree calculation.

[0076] Based on the sets of supporting and rebuttal evidence units, calculate the degree of educational evidence conflict for each educational proposition to be verified:

[0077]

[0078]

[0079]

[0080] In the formula, Indicates the first The degree of conflict of educational evidence for an unverified educational proposition. Expressing support for the first The cumulative amount of evidence for an unverified educational proposition Indicating a rebuttal to the first The cumulative amount of evidence for an unverified educational proposition To prevent extremely small constants with a denominator of zero, Indicates the first The weight of the evidentiary level for each evidentiary unit Expressing support for the first A set of evidence units for an educational proposition to be verified. Indicating a rebuttal to the first A set of evidence units for an educational proposition to be verified.

[0081] It should be noted that, The closer a number is to 0, the stronger the supporting or rebuttal evidence is, indicating a weaker conflict. The closer a number is to 1, the closer the cumulative amount of supporting and rebuttal evidence is, and the stronger the conflict.

[0082] The levels of evidence in the graded evidence package are determined by the following criteria: Level 1, Level 2, Level 3, Level 4, and Level 5 evidence correspond to 1.00, 0.85, 0.70, 0.55, and 0.40, respectively. The values ​​are chosen to give higher weight to evidence with more authoritative sources and stronger course relevance in the total amount of supporting or rebuttal evidence.

[0083] When a certain educational proposition to be verified has no supporting evidence unit and no rebuttal evidence unit, the educational proposition to be verified is marked as a proposition without valid evidence and the degree of educational evidence conflict is not calculated; when only supporting evidence unit or only rebuttal evidence unit exists, the cumulative quantity of evidence on the side that does not exist is 0, and the corresponding degree of educational evidence conflict is calculated.

[0084] S4. Adjudicate the evidence units based on the degree of conflict and the level of evidence in the education evidence, and generate a controlled evidence set.

[0085] Specifically, the ruling is determined based on the set of supporting evidence units, the set of rebuttal evidence units, the cumulative amount of supporting evidence, the cumulative amount of rebuttal evidence, the degree of conflict of educational evidence, and the evidence level of each evidence unit for each educational proposition to be verified. For educational propositions to be verified that have been marked as having no valid evidence, they remain in the state of having no valid evidence and are prohibited from being used as the basis for a definitive answer.

[0086] When the educational proposition to be verified has only supporting evidence units, the evidence unit with the highest evidence level in the supporting evidence unit set is identified as definable evidence, and the ruling status is marked as established. When the educational proposition to be verified has only rebuttal evidence units, the evidence unit with the highest evidence level in the rebuttal evidence unit set is identified as definable evidence, and the ruling status is marked as invalid, prohibiting the output of the educational proposition to be verified as a positive definitive conclusion. When there are multiple evidence units with the same highest evidence level, priority is given to retaining evidence units whose textbook version matching mark, knowledge point matching mark, and target relevance mark all meet the matching requirements.

[0087] When the same educational proposition to be verified has both supporting and rebuttal evidence units, if the degree of conflict of educational evidence is lower than a preset conflict threshold, the side with the greater cumulative amount of supporting and rebuttal evidence is determined as the side to be adopted in the ruling, and the evidence unit with the highest evidence level from the side to be adopted in the ruling is determined as the definable evidence; if the degree of conflict of educational evidence is not lower than the preset conflict threshold, the highest evidence level among the supporting and rebuttal evidence units is compared first, and the side with the higher highest evidence level is determined as the side to be adopted in the ruling; if the highest evidence levels of both sides are the same, the difference between the cumulative amount of supporting and rebuttal evidence is compared. When the difference reaches the preset ruling boundary, the side with the higher cumulative amount is determined as the side to be adopted in the ruling. When the difference does not reach the preset ruling boundary, the corresponding educational proposition to be verified is marked as a proposition to be confirmed and is prohibited from being used as the basis for a definitive answer.

[0088] Based on the ruling, a controlled evidence set is generated. The controlled evidence set includes identifiable evidence, corresponding proposition number, evidence content, evidence source, evidence level, knowledge point label, relationship annotation result, educational evidence conflict degree, ruling status, and evidence citation identifier. For irrelevant evidence units, evidence units without valid evidence corresponding to propositions, conflicting evidence units corresponding to propositions to be confirmed, and conflicting evidence units that have not been adopted as identifiable evidence, a prohibition on deterministic output is written.

[0089] It should be noted that the preset conflict threshold is exemplarily set to 0.60 to 0.80, preferably 0.70. The basis for this value is that when the degree of conflict of educational evidence reaches this range, the cumulative amount of supporting evidence and the cumulative amount of rebuttal evidence are relatively close, indicating that there is a substantial disagreement under the same educational proposition to be verified, and further adjudication by the level of evidence is required. When it is below this range, it can usually be considered that one side of the evidence has formed a clear advantage. The preset adjudication boundary is exemplarily set to 10% to 20% of the maximum value of the cumulative amount of supporting evidence and the cumulative amount of rebuttal evidence, preferably 15%. The basis for this value is that when the difference between the cumulative amounts of the two sides reaches this proportion, it can be considered that one side of the evidence has a sufficient advantage in terms of quantity or level weighting to support the adjudication. When it does not reach this proportion, it indicates that the difference between the two sides is insufficient to form a definite conclusion and should be marked as a proposition to be confirmed.

[0090] S5. Construct educational question-and-answer prompt instances based on structured question objects, hierarchical prompting strategies, and controlled evidence sets, and call the large model to generate candidate answers based on the educational question-and-answer prompt instances;

[0091] Specifically, the structured question object, hierarchical prompting strategy, and controlled evidence set are obtained. The standard question text, question type, known conditions, and target to be sought in the structured question object are written into the question description area. The prompting start method, answer open level, and prompting order in the hierarchical prompting strategy are written into the answering strategy area. The determinable evidence, adjudication status, evidence citation identifier, and prohibition of deterministic output marker in the controlled evidence set are written into the evidence constraint area, thereby constructing an educational question-and-answer prompt instance.

[0092] It should be noted that the educational question-and-answer prompt examples are used to limit the scope and manner of the large model's responses: when the decision status is "true," the large model is allowed to generate positive explanations based on corresponding verifiable evidence; when the decision status is "false," the large model is allowed to generate explanations indicating that the conditions are not met or the proposition is false; when the decision status is "pending confirmation" or "no valid evidence," the large model is only allowed to generate cautious prompts and is not allowed to output definitive conclusions; for evidence units marked with a prohibition on definitive output, they are not included in the scope of verifiable evidence. The educational question-and-answer prompt examples also limit the large model to associate corresponding evidence citation identifiers when generating definitions, formulas, problem-solving steps, or conclusions, and to determine the starting content, depth of development, and order of prompts according to a hierarchical prompting strategy.

[0093] Furthermore, the generated educational question-and-answer prompt instances are input into the target big model, enabling the target big model to generate candidate answers under the joint constraints of the question description area, the answer strategy area, and the evidence constraint area. When the structured question object already contains the complete standard question text, known conditions, and the target to be sought, the text big model is invoked. When the current question still needs to be understood in conjunction with the question image, chart, or layout content, the multimodal big model is invoked.

[0094] The target model generates candidate answer fragments based on the answering strategies in the educational question-and-answer prompt examples. When the prompt starts with a concept explanation, a concept explanation fragment is generated first. When the prompt starts with a formula application condition prompt, a formula selection explanation fragment is generated first. When the prompt starts with a key condition prompt, a condition location or rhetorical question prompt fragment is generated first. When the answer open level is limited to intermediate steps, only key steps and checking methods are generated, and the complete conclusion is not directly output.

[0095] The system receives the output content generated by the target large model based on educational question-and-answer prompt instances, and organizes the output content into fragmented candidate answers according to the output format requirements in the educational question-and-answer prompt instances. The fragmented candidate answers include at least one of the following: prompt guidance fragments, knowledge explanation fragments, problem-solving step fragments, conclusion explanation fragments, and cautionary explanation fragments. Each fragment involving certain knowledge content retains the corresponding proposition number or evidence citation identifier.

[0096] It should be noted that the target large model is a pre-deployed natural language processing (NLP) large model or multimodal large model, used to generate candidate answers based on educational question-and-answer prompts. The NLP large model handles question-and-answer tasks where the structured question object already contains the complete standard question text, known conditions, and the desired objective. The multimodal large model handles question-and-answer tasks that still require understanding the question image, chart, or layout content. The educational question-and-answer prompts are used as input to the target large model, which outputs candidate answers. The NLP large model and multimodal large model can be selected from any existing model with corresponding processing capabilities. In one optional implementation, the NLP large model is a pre-trained language model based on the Transformer architecture, and the multimodal large model is a visual language model with image-text joint understanding capabilities.

[0097] S6. Perform evidence consistency verification on candidate answers and generate the final educational question and answer results.

[0098] Specifically, based on candidate response fragments, for fragments involving definitions, formulas, applicable conditions, problem-solving steps, or conclusions, the corresponding definable evidence is matched in the controlled evidence set according to the proposition number or evidence citation identifier they carry. The content of the candidate response fragment is then compared for consistency with the content and ruling status of the definable evidence. When a candidate response fragment can be matched with definable evidence, and its content is consistent with the definable evidence and its ruling status, the candidate response fragment is marked as supported by definable evidence. When a candidate response fragment corresponds to a proposition to be confirmed, a proposition without valid evidence, or a prohibition on definitive output, and only expresses a non-deterministic hint, it is marked as a cautionary statement. When a candidate response fragment is only used for learning guidance and does not contain definitive knowledge content, it is marked as strategy guidance. When a candidate response fragment cannot be matched with definable evidence, it is marked as unsupported by evidence. When a candidate response fragment contradicts definable evidence or a ruling status, or outputs the content corresponding to a prohibition on definitive output as a definitive conclusion, it is marked as conflicting evidence.

[0099] Candidate response segments marked as supported by definite evidence, cautious explanations, or strategy-guided segments are added to the set of output response segments, and candidate response segments marked as unsupported by evidence or conflicting evidence are removed, resulting in the set of output response segments.

[0100] Based on the set of outputtable answer fragments, the starting content of the final educational question and answer result is determined according to the prompting start method in the hierarchical prompting strategy. Outputtable answer fragments are organized according to the prompting order, and the completeness of the final educational question and answer result is controlled according to the answer openness level.

[0101] When the answer is open at the hint level, only hints, explanations of concepts, the basis for formula selection, or key conditions are output. When the answer is open at the step level, necessary problem-solving steps and checking methods are output, but the complete answer is not directly output. When the answer is open at the complete analysis level, the complete analysis, necessary steps, and conclusions are output within the allowed range of the set of outputtable answer fragments. For candidate answer fragments marked as "cautious explanations," only non-deterministic hints are output, not as definitive conclusions.

[0102] The sorted outputtable answer segments are combined into the final educational question and answer results; among them, candidate answer segments marked as unsupported by evidence or with conflicting evidence are not included in the final educational question and answer results.

[0103] This embodiment also provides an educational question-answering system based on a large model, including:

[0104] The educational question-and-answer parsing module is used to receive educational question-and-answer requests from students, parse and merge input content from different modalities, and generate structured question objects.

[0105] The tiered prompting module is used to identify student error patterns based on structured question objects and students' historical learning records, and to generate tiered prompting strategies.

[0106] The evidence retrieval and grading module is used to retrieve relevant content from the educational knowledge base based on structured question objects, and generate graded evidence packages according to evidence sources and course adaptation tags.

[0107] The evidence conflict adjudication module is used to split the graded evidence package into evidence units, extract the educational propositions to be verified, label the relationships and calculate the degree of conflict, and generate a controlled evidence set.

[0108] The controlled question-answer generation module is used to construct educational question-answer prompt instances based on structured question objects, hierarchical prompting strategies, and controlled evidence sets, call the large model to generate candidate answers, and generate the final educational question-answer results after evidence consistency verification.

[0109] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An educational question-answering method based on a large model, characterized in that, include: It receives educational question and answer requests from students, parses the requests, generates structured question objects, identifies student error patterns based on the structured question objects and students' historical learning records, and generates a hierarchical prompting strategy based on the student error patterns. Relevant content is retrieved from the educational knowledge base based on the structured question object, and a graded evidence package is generated according to the evidence source and curriculum adaptation marker; the graded evidence package includes the evidence content and the corresponding evidence level; The graded evidence package is split into evidence units, and educational propositions to be verified are extracted based on structured question objects. The relationship between the evidence units and the educational propositions to be verified is labeled, and the degree of conflict of educational evidence is calculated based on the labeling results. The evidence units are adjudicated based on the degree of conflict and the level of evidence in the education process, and a controlled evidence set is generated. Based on the structured question object, hierarchical prompting strategy, and controlled evidence set, construct educational question-and-answer prompt instances, and call the large model to generate candidate answers based on the educational question-and-answer prompt instances; The candidate answers are validated for consistency of evidence to generate the final educational question and answer results.

2. The educational question-answering method based on a large model as described in claim 1, characterized in that: The process of generating structured question objects includes: receiving educational Q&A requests including text questions, voice questions, question images, or uploaded files; performing basic verification based on student identifiers, course identifiers, grade information, and textbook version information; determining the main input content and auxiliary input content based on the input modality; and generating a Q&A processing task. Based on the input modality in the Q&A processing task, the corresponding parsing component is invoked to process and fuse the text, speech-to-text, question image recognition content, or uploaded file parsing content, extracting the standard question text, course subject, grade level, textbook version, question type, candidate knowledge points, known conditions, and target to be obtained, thereby generating a structured question object.

3. The educational question-answering method based on a large model as described in claim 2, characterized in that: The step of identifying student error patterns based on structured question objects and students' historical learning records, and generating a hierarchical prompting strategy based on student error patterns, includes: reading students' historical learning records within the candidate knowledge points and related knowledge point ranges based on the student identifier associated with the structured question object, matching the historical learning records with the question type, candidate knowledge points, known conditions, and target in the structured question object to identify the student error pattern corresponding to the current question; determining the primary error pattern and auxiliary error patterns based on the student error patterns, and configuring the prompting start method, answer open level, and prompting order according to the primary error pattern to generate a hierarchical prompting strategy.

4. The educational question-answering method based on a large model as described in claim 3, characterized in that: The step of retrieving relevant content from the educational knowledge base based on the structured question object and generating a graded evidence package according to the evidence source and course fit marker includes: generating evidence retrieval conditions based on the structured question object and retrieving content related to the current question from the educational knowledge base; performing fit screening on the retrieved content to generate initial evidence records including evidence content, evidence source, textbook version matching marker, knowledge point matching marker, target relevance marker, and knowledge point tag; determining the evidence level according to the evidence source; and sorting the initial evidence records according to the evidence level, textbook version matching marker, knowledge point matching marker, and target relevance marker to generate a graded evidence package.

5. The educational question-answering method based on a large model as described in claim 4, characterized in that: The step of splitting the hierarchical evidence package into evidence units and extracting educational propositions to be verified based on structured problem objects includes: reading the evidence records in the hierarchical evidence package, splitting them into evidence units according to the knowledge expression granularity of definitions, formulas, applicable conditions, problem-solving rules, steps and conclusions, and configuring evidence unit types and evidence citation identifiers for the evidence units; and generating educational propositions to be verified based on the question type, candidate knowledge points, known conditions and objectives to be sought in the structured problem objects, combined with the course knowledge point catalog.

6. The educational question-answering method based on a large model as described in claim 5, characterized in that: The step of labeling the evidence units with the educational propositions to be verified and calculating the degree of conflict of educational evidence based on the labeling results includes: matching the evidence units with the educational propositions to be verified according to the correspondence between knowledge point tags and types, and labeling the evidence units as supporting, refuting, or irrelevant relationships according to the consistency between the evidence content and the proposition content and applicable conditions; for each educational proposition to be verified, statistically analyzing the supporting and refuting evidence units, and calculating the degree of conflict of educational evidence based on the evidence level weight.

7. The educational question-answering method based on a large model as described in claim 6, characterized in that: The process of generating a controlled evidence set includes: adjudicating evidence units with supporting and rebuttal relationships based on the degree of conflict of educational evidence, the cumulative amount of supporting evidence, the cumulative amount of rebuttal evidence, and the evidence level corresponding to each educational proposition to be verified, and determining the definitive evidence or propositions to be confirmed that can be used as the basis for a definitive answer; generating a controlled evidence set based on the adjudication results, and writing a prohibition on definitive output flag for irrelevant evidence, propositions without valid evidence, evidence corresponding to propositions to be confirmed, and conflicting evidence that was not adopted.

8. The educational question-answering method based on a large model as described in claim 7, characterized in that: The process of constructing educational question-and-answer prompt instances based on structured question objects, hierarchical prompting strategies, and controlled evidence sets, and then calling a large model to generate candidate answers based on these prompt instances, includes: generating a question description area based on the structured question object, generating an answer strategy area based on the hierarchical prompting strategy, and generating an evidence constraint area based on determinable evidence, adjudication status, and prohibition of deterministic output markers in the controlled evidence set; and calling a text-based large model or a multimodal large model based on the educational question-and-answer prompt instances and the input modality, so that the large model determines the answer initiation method, answer openness level, and prompting order according to the answer strategy area, and generates candidate answers containing proposition numbers or evidence citation identifiers within the scope defined by the evidence constraint area.

9. The educational question-answering method based on a large model as described in claim 8, characterized in that: The process of generating the final educational question-and-answer result includes: based on the proposition number or evidence citation identifier carried in the candidate answer, comparing the candidate answer fragments with the determinable evidence, adjudication status, and prohibition of deterministic output markers in the controlled evidence set, eliminating candidate answer fragments without evidence support or with conflicting evidence, and generating a set of outputtable answer fragments; and organizing the output content according to the outputtable answer fragment set and the hierarchical prompting strategy, based on the prompting start method, prompting order, and answer openness level, to generate the final educational question-and-answer result.

10. A large-model-based educational question-answering system, based on the large-model-based educational question-answering method according to any one of claims 1 to 9, characterized in that, include: The educational question-and-answer parsing module is used to receive educational question-and-answer requests from students, parse and merge input content from different modalities, and generate structured question objects. The tiered prompting module is used to identify student error patterns based on structured question objects and students' historical learning records, and to generate tiered prompting strategies. The evidence retrieval and grading module is used to retrieve relevant content from the educational knowledge base based on structured question objects, and generate graded evidence packages according to evidence sources and course adaptation tags. The evidence conflict adjudication module is used to split the graded evidence package into evidence units, extract the educational propositions to be verified, label the relationships and calculate the degree of conflict, and generate a controlled evidence set. The controlled question-answer generation module is used to construct educational question-answer prompt instances based on structured question objects, hierarchical prompting strategies, and controlled evidence sets, call the large model to generate candidate answers, and generate the final educational question-answer results after evidence consistency verification.

Images