Intelligent auxiliary teacher lesson preparation system and method based on large education model

Abstract

The invention discloses an intelligent auxiliary teacher lesson preparation system and method based on an education large model. The method comprises the following steps: S1, generating a standardized teacher lesson preparation data set; s2, constructing a teaching knowledge graph; s3, generating preliminary intelligent lesson preparation content; s4, presenting the preliminarily generated intelligent lesson preparation content, and providing an interactive interface combining a natural language and voice recognition; s5, generating the adjusted intelligent lesson preparation content; and S6, according to historical lesson preparation data of the teacher and the interactive feedback data, performing adaptive learning and optimization on the intelligent lesson preparation content, adjusting weights of key knowledge points in the teaching knowledge graph, updating a content generation strategy of the large education model, and generating a final intelligent lesson preparation scheme. According to the method, the response capability of the intelligently generated content to the teaching intention of a teacher is remarkably improved, the teaching content and the test question style are supported to dynamically cooperate with classroom feedback along with the teaching rhythm, and an intelligent decision support basis is provided for large-scale personalized teaching.

Description

Technical Field

[0001] The present invention relates to the field of educational technology, and in particular, to an intelligent auxiliary teacher lesson preparation system and method based on an educational large model. Background Art

[0002] With the rapid development of artificial intelligence and educational information technology, intelligent question generation has gradually become an important direction for teaching assistance in the field of basic education. In primary and secondary schools, it shows broad application potential in reducing the teacher's lesson preparation burden, improving the consistency of question quality, and enhancing personalized teaching support. Currently, the mainstream methods for generating questions in primary and secondary schools mostly rely on manual writing or the call of question banks based on fixed templates. Usually, teachers compile questions according to the teaching syllabus, teaching materials, and teaching experience. Although this traditional mode has a certain teaching pertinence, it faces obvious bottlenecks when dealing with the needs of large-scale, interdisciplinary, and personalized question generation.

[0003] In the prior art, some systems introduce natural language processing technology to try to achieve automatic question generation. However, most of them are limited to the extraction of keywords from teaching materials and template matching, lacking in-depth understanding of the teaching context, curriculum objectives, and knowledge point structure. The generated questions are often shallow in content and single in structure, unable to meet the high standards of logical rigor, semantic accuracy, and knowledge coverage required by teaching evaluation. At the same time, the prior art generally lacks an effective teacher interaction feedback mechanism, and cannot dynamically feedback the teacher's modification opinions on the question content, structure, and difficulty in actual teaching to the question generation model, resulting in the inability to continuously optimize the generation quality.

[0004] In terms of teaching content assistance, although some platforms have the ability to intelligently push courseware or teaching resources, their underlying logic is mostly content recommendation based on statistics. Lacking the support of a structured teaching knowledge graph, it is difficult to establish a semantic closed-loop between teaching objectives - knowledge points - questions, resulting in the disconnection between content push and classroom objectives. Furthermore, in the current system, the expression of the teacher's lesson preparation intention mostly relies on static parameter selection, lacking the semantic understanding ability of natural language or voice input intention, which limits the interaction efficiency and intelligence level between teachers and intelligent systems.

[0005] Therefore, there is an urgent need to propose an intelligent question generation method for primary and secondary schools that can integrate deep semantic understanding, structured teaching knowledge modeling, and a human-machine feedback closed-loop mechanism to effectively make up for the deficiencies of the prior art. Summary of the Invention

[0006] An object of the present invention is to propose an intelligent auxiliary teacher lesson preparation system and method based on an educational large model. The present invention significantly improves the response ability of the intelligently generated content to the teacher's teaching intention, supports the dynamic coordination of teaching content and question styles with the teaching rhythm and classroom feedback, and provides an intelligent decision-making support basis for large-scale personalized teaching.

[0007] An intelligent auxiliary teacher lesson preparation method based on an educational large model according to an embodiment of the present invention includes the following steps:

[0008] S1. Collect and integrate multi-source teacher lesson preparation data related to teacher lesson preparation, and preprocess the teacher lesson preparation data to generate a standardized teacher lesson preparation data set;

[0009] S2. Use semantic parsing of the standardized teacher lesson preparation data set to construct a teaching knowledge graph;

[0010] S3. Based on the educational large model, perform semantic understanding and context analysis on the lesson preparation requirement data input by the teacher, and match the lesson preparation requirement data with the teaching knowledge graph to generate preliminary intelligent lesson preparation content;

[0011] S4. Present the preliminarily generated intelligent lesson preparation content, and provide an interactive interface that combines natural language and speech recognition, allowing the teacher to perform real-time preview, editing, and feedback on the intelligent lesson preparation content to generate interactive feedback data;

[0012] S5. Analyze the interactive feedback data, extract content modifications, key adjustments, and personalized requirements involved in the feedback, and dynamically optimize the preliminarily generated intelligent lesson preparation content based on the feedback information to generate adjusted intelligent lesson preparation content;

[0013] S6. According to the teacher's historical lesson preparation data and the interactive feedback data, perform adaptive learning and optimization on the intelligent lesson preparation content, adjust the weights of key knowledge points in the teaching knowledge graph, and update the content generation strategy of the educational large model to generate a final intelligent lesson preparation plan.

[0014] Optionally, S1 includes the following steps:

[0015] S11. Obtain the original educational data set D raw The original educational data set includes the primary and secondary school textbook content set T m , the teaching syllabus content set T s , the set of examination questions over the years T e and the primary and secondary school question bank content set T q ;

[0016] S12. Perform text cleaning on the original educational data set D raw Remove and standardize the garbled information, redundant symbols, non-teaching content, and format error content included in the original educational data set to obtain the cleaned educational text set D clean , where the cleaning process uses unified specification rules to ensure the consistency of the expression structures of the content from each source;

[0017] S13. For the cleaned educational text set D clean perform format standardization processing, perform standardized coding on the paragraph structure, question type identification, knowledge point description, and teaching language in it, and obtain the standardized educational data set D norm , where the format standardization processing maintains the consistent expression form of all samples in the semantic structure and information level;

[0018] S14. Based on the standardized educational data set D norm perform key knowledge point annotation, match and annotate each educational data sample with a certain knowledge point in its corresponding teaching knowledge point set K = {k1, k2,..., k n}, and form the annotated educational data set D tag , where k i represents the i-th knowledge point, and the annotation result is embedded into each sample data in the form of a structured label;

[0019] S15. Re-index and group the annotated educational data set D tag according to the subject, grade, course content, and knowledge point category, organize each sample data into a structured form, and construct the structured educational data set D struct , where each structured educational data includes: educational text sample x i , corresponding knowledge point label k i and its classification information y i , and the classification information includes subject attributes, grade dimensions, and question type categories.

[0020] Optionally, S2 includes the following steps:

[0021] S21. Based on the structured educational data set D struct perform teaching semantic unit segmentation on each educational text sample x i to construct a teaching semantic block set B i ={b i1 , b i2 ,..., b im}, and the teaching semantic block represents the smallest information unit with consistent teaching logic, and is used to capture the semantic dependency structure of teaching objectives and knowledge points;

[0022] S22. Apply the context embedding representation function to each teaching semantic block b i in the teaching semantic block set B ij to generate a semantic embedding vector E ij , and map the semantic embedding vector to the semantic compression space S sem , and retain the semantic factors related to question generation;

[0023] S23. Introduce the semantic activation function A(k t ,E ij ) = cosine(φ k (k t ),E ij ) to calculate the teaching semantic block b ij for the semantic support strength of the knowledge point k t , where φ k (k t ) represents the embedding representation of the knowledge point, which is used to guide the construction of edges between the knowledge point and the context semantics;

[0024] S24. Based on all knowledge point label sets K, teaching objective sets G, and teaching semantic block sets B, construct the teaching knowledge graph G edu = (V, E, W), where the node set V includes knowledge point nodes v k , teaching objective nodes v g and teaching semantic block nodes v b , the edge set E includes semantic support edges, target dependency edges, and chapter structure edges, and the weight set W is jointly composed of semantic activation function values and context embedding representation functions;

[0025] S25. Based on the teaching knowledge graph G edu execute the context-aware path reasoning algorithm for the given teaching objective node v g and knowledge point node v k to calculate the test question generation teaching path P inf :

[0026]

[0027] where P inf is the final teaching path for generating test questions, is the reachable path set from all teaching objective nodes v g to knowledge point nodes v k , Redundancy(p) is the path redundancy penalty function, which is used to avoid repeated reference to the same semantic block, and λ div is the balance coefficient for controlling the balance between diversity rewards and structural overlap. The obtained teaching path is used to guide the education large model to generate target-oriented test questions;

[0028] S26. Encode the teaching path P inf into a semantic path vector π, which is used as the input semantic control factor of the education large model generation module to guide the model to generate a set of generated test questions with rigorous logic, consistent semantics, and clear knowledge points

[0029] S27. Analyze the knowledge point calls for generating the test question set Extract the set of knowledge points called in the actual generation Compare with the set of teaching objectives And compare with the target sets K, G in the path reasoning stage, and calculate the semantic coverage deviation:

[0030]

[0031] Among them, Δ sem is the semantic coverage deviation of the current generated test question semantic path for the teaching objective, and β is the weight factor of the semantic coverage deviation;

[0032] S28. Based on the semantic coverage deviation feedback result, adjust the edge weight set W in the teaching knowledge graph, increase the edge weights of the semantic nodes that appear in the path but are not called, optimize the edge strength of the mis-activated path, and form a closed-loop optimization of the semantic graph driven by test question generation:

[0033]

[0034] Among them, W ′ (u, v) is the edge weight set after feedback adjustment, and η is the structure update learning rate, indicates whether the edge participates in the current round of reasoning path.

[0035] Optionally, the S3 includes the following steps:

[0036] S31. Receive the set of lesson preparation requirement data D input by the teacher plan , perform semantic encoding processing on all text information in the set of lesson preparation requirement data D plan input by the teacher, and generate the teacher's lesson preparation vector representation Ψ plan through the context understanding module in the education large model. The teacher's lesson preparation vector representation integrates information on teaching objective intentions, course context, learning group characteristics, and teaching method requirements;

[0037] S32. Call the teaching knowledge graph G edu Match the teacher's lesson preparation vector representation Ψ plan with all the teaching semantic path vector sets Π in the teaching knowledge graph one by one, calculate the degree of semantic relevance, and obtain the semantic relevance score between the teacher's lesson preparation vector representation and each semantic path;

[0038] S33. Based on the calculated semantic relevance scores, screen out the path set that satisfies the preset semantic relevance threshold with the teacher's lesson preparation vector representation Ψ plan as the semantic support path set Π for generating intelligent lesson preparation content * ;

[0039] S34. Based on the set of semantic support paths Π obtained by screening * and the teacher's lesson preparation vector representation Ψ plan , the educational large model generates a preliminary set of intelligent lesson preparation content C plan ={c1, c2,..., c k}, where the preliminary set of intelligent lesson preparation content includes a draft teaching syllabus, a preliminary courseware, lecture notes, teaching cases, and personalized exercise suggestions. Each piece of content is generated by the semantic generation function of the educational large model under the dual constraints of path semantics and teacher intentions;

[0040] S35. Perform teaching consistency detection and content structure integrity assessment on each piece of content in the preliminary set of intelligent lesson preparation content C plan . Through the built-in evaluation module of the educational large model, an adaptation score vector Ω = {ω1, ω2,..., ω k} is output. The score vector is used to represent the structural adaptation and alignment with teaching objectives of each piece of content relative to the teacher's input intention;

[0041] S36. Return the intelligent lesson preparation content with a score lower than the preset tolerance threshold τ in the adaptation score vector Ω to the educational large model for regeneration and correction, and retain the revision record in the form of a structured label in the regenerated version, finally generating a preliminary set of intelligent lesson preparation content The preliminary set of intelligent lesson preparation content meets the requirements of teaching objectives, context logic, and individual characteristics.

[0042] Optionally, the set of lesson preparation requirement data includes teaching objective descriptions, a set of course theme keywords, teaching stage information, teaching method preferences, and a set of student portrait parameters.

[0043] Optionally, S4 includes the following steps:

[0044] S41. Input the preliminary set of intelligent lesson preparation content into the intelligent interaction presentation module for structured display. The intelligent interaction presentation module uses a document view, a slide view, and an interactive component view to present the draft teaching syllabus, the preliminary courseware, lecture notes, teaching cases, and personalized exercise suggestions respectively according to the content type;

[0045] S42. Based on the structured display interface, construct a fusion interaction interface for the natural language interaction sub-module and the speech recognition interaction sub-module. The natural language interaction sub-module receives the set of text instructions Γ text input by the teacher, and the speech recognition interaction sub-module receives the set of teacher speech inputs Γ voice , where the set of text instructions and the set of speech inputs are uniformly mapped to the teaching feedback semantic space;

[0046] S43. Co - encode the natural language input set Γ text and the speech input set Γ voice into a teacher interaction feedback representation set Γ feedback . The teacher interaction feedback representation set is used to express the teacher's intention of editing and modifying the preliminary intelligent lesson preparation content, restructuring suggestions, content preferences, or teaching style adjustment requirements;

[0047] S44. Match each item in the teacher interaction feedback representation set Γ feedback with the intelligent lesson preparation content set to identify the content index set I of the feedback target corresponding content match , and extract the original semantic path vector set Π match of the corresponding content segment;

[0048] S45. For each teacher interaction feedback representation and its mapped content segment, execute the semantic alignment function to extract the offset features and adjustment directions of the teacher's intention in the original path semantics, and form the feedback offset vector set Δ edit ;

[0049] S46. Generate an interaction feedback data set where each feedback data record includes the original semantic path vector π l , the feedback offset vector δ l , and the content index i of the feedback target corresponding content l .

[0050] Optionally, the S5 includes the following steps:

[0051] S51. Receive and parse the interaction feedback data set F interact , and extract the content modification information, teaching key point adjustment requirements, and personalized teaching preferences involved in the feedback;

[0052] S52. According to the semantic features of the teacher feedback offset vector δ l , perform semantic reconstruction on the corresponding original semantic path vector π l to generate an adjusted path semantic vector π' l . During the semantic reconstruction process, the original structure logic of the path is maintained, and only the node content, edge weight distribution, or knowledge point organization method is optimized with minimal perturbation;

[0053] S53. Use all the reconstructed path semantic vector set Π' = {π'1, π'2,..., π' s} as the semantic driving factor and input it into the content regeneration module of the education large - model, and jointly generate an adjusted intelligent lesson preparation content set plan ​

[0054] S54. On the basis of the interactive feedback dataset F interact , taking the teacher feedback offset vector delta l and the original semantic path vector pi l as inputs, semantically guided reconstruction is performed on the intelligent lesson preparation content set to generate an adjusted intelligent lesson preparation content set with fine-tuned content. On the basis of retaining the original semantic path logical structure, the adjusted intelligent lesson preparation content set integrates content modifications, teaching focus adjustments, and personalized needs expressed in teacher feedback, specifically including a draft teaching outline, a first draft of courseware, class handouts, teaching cases, and personalized exercise suggestions;

[0055] S55. Finally, the adjusted intelligent lesson preparation content set is output The adjusted intelligent lesson preparation content set is used to support teachers' teaching implementation in real time and dynamically support personalized teaching design.

[0056] Optionally, the draft teaching outline: Based on the teaching objective nodes covered by the original semantic path vector π l , if the teacher feedback offset vector contains information on teaching objective decomposition and knowledge point reordering, the priorities of the teaching objective nodes in the path are rearranged, and the entry order and objective expression level in the draft teaching outline are updated accordingly;

[0057] The first draft of courseware: Based on the graphic and text nodes and interactive logic in the original semantic path vector π l , if the teacher feedback offset vector contains preferences for slide layout, suggestions for illustration methods, or content simplification requirements, the weights of the nodes related to visual expression in the path are adjusted, and the courseware page structure and information distribution are reconstructed;

[0058] The class handouts: Based on the knowledge point coverage sections in the original semantic path vector π l , if the teacher feedback offset vector contains requirements for example replacement, expansion of the handout structure, or adjustment of the graphic-to-text ratio, the corresponding knowledge point segments are locally updated, and the exemplary content nodes in the path are replaced;

[0059] The teaching cases: Based on the teaching scenario nodes and guiding semantic chains bound in the original semantic path vector π l , if the teacher feedback offset vector contains requirements for teaching scenario reconstruction, background replacement, or style correction, the dialogue framework and scenario setting for this path segment are regenerated, and the start and end points of the path are kept unchanged;

[0060] The personalized exercise suggestions: Based on the original semantic path vector π lThe associated knowledge point nodes and their graph edge weight information. If the teacher feedback offset vector contains requirements for ability adaptation, difficulty control, or examination point strengthening, then fine-tune the node weights and edge structures in this path segment, reconstruct the matching degree between question types and difficulty, and maintain consistency with the original path logic.

[0061] An intelligent auxiliary teacher lesson preparation system based on an educational large model, which is applied to an intelligent auxiliary teacher lesson preparation method based on an educational large model, and includes the following modules:

[0062] A data processing module, which is used to collect primary and secondary school textbooks, teaching syllabuses, past examination questions, and exercise question banks, and perform text cleaning, format standardization, and key knowledge point annotation on the collected educational data to generate a structured educational data set;

[0063] An educational large model construction module, which is used to pre-train and fine-tune the structured educational data set to construct an educational large model for primary and secondary school scenarios;

[0064] A teaching knowledge graph construction module, which is used to construct a teaching knowledge graph based on teaching semantic blocks, teaching objectives, and knowledge point relationships to form a semantic path reasoning mechanism;

[0065] An intelligent lesson preparation content generation module, which is used to receive the lesson preparation requirement data input by the teacher, and combine the educational large model and the teaching knowledge graph to generate preliminary intelligent lesson preparation content such as a draft teaching outline, a first draft of courseware, classroom lecture notes, teaching cases, and personalized exercise question suggestions;

[0066] An interactive feedback processing module, which is used to collect the teacher's editing and modification, style preferences, and teaching adjustment requirements for the intelligent lesson preparation content through a natural language and speech recognition interactive interface to generate a semantic offset representation and an interactive feedback data set;

[0067] A dynamic optimization module, which is used to reconstruct the original semantic path and update the graph path based on the teacher feedback content to generate dynamically adjusted intelligent lesson preparation content, and is used to drive the subsequent question generation process.

[0068] The beneficial effects of the present invention are:

[0069] (1) By constructing a multi-level graph structure including teaching objective nodes, knowledge point nodes, and teaching semantic block nodes, introducing a semantic activation function and a path redundancy penalty function, using a path reasoning algorithm to obtain an optimal teaching path vector, and dynamically adjusting the graph edge weights based on the semantic coverage deviation between the generated questions and the target knowledge points after question generation, the present invention realizes a generation-driven self-learning update mechanism for the graph structure, breaks through the limitation of the traditional static use of the knowledge graph, makes the semantic path of question generation highly consistent with the teaching structure, and significantly improves the path generation accuracy and semantic coverage rate.

[0070] (2) The present invention uniformly maps the natural language input of teachers and voice commands into interactive feedback representations, generates a feedback offset vector through semantic alignment analysis, and acts on the original path semantic vector for minimum perturbation reconstruction to achieve structured modification of multiple types of content such as teaching outlines, lecture notes, cases, and test question suggestions. On this basis, the feedback data synchronously updates the teaching knowledge graph and the generation model to form a closed-loop content optimization mechanism, significantly improving the response ability of the intelligently generated content to the teaching intentions of teachers, supporting the dynamic coordination of teaching content and test question styles with the teaching rhythm and classroom feedback, and providing an intelligent decision-making support basis for large-scale personalized teaching. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] The accompanying drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention, and do not constitute a limitation to the present invention. In the drawings:

[0072] Figure 1 is a flowchart of an intelligent auxiliary teacher lesson preparation method based on an educational large model proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0073] Now, the present invention will be further described in detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only showing the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0074] Refer to Figure 1 , an intelligent auxiliary teacher lesson preparation method based on an educational large model, includes the following steps:

[0075] S1. Collect and integrate multi-source teacher lesson preparation data related to teacher lesson preparation, and preprocess the teacher lesson preparation data to generate a standardized teacher lesson preparation data set;

[0076] S2. Use semantic parsing on the standardized teacher lesson preparation data set to construct a teaching knowledge graph;

[0077] S3. Based on the educational large model, perform semantic understanding and context analysis on the lesson preparation requirement data input by the teacher, and match the lesson preparation requirement data with the teaching knowledge graph to generate preliminary intelligent lesson preparation content;

[0078] S4. Present the preliminarily generated intelligent lesson preparation content, and provide an interactive interface that combines natural language and speech recognition, allowing the teacher to perform real-time preview, editing, and feedback on the intelligent lesson preparation content to generate interactive feedback data;

[0079] S5. Parse the interactive feedback data, extract the content modifications, key adjustments, and personalized requirements involved in the feedback, and dynamically optimize the initially generated intelligent lesson preparation content based on the feedback information to generate the adjusted intelligent lesson preparation content;

[0080] S6. According to the teacher's historical lesson preparation data and interactive feedback data, conduct adaptive learning and optimization on the intelligent lesson preparation content, adjust the weights of key knowledge points in the teaching knowledge graph, and update the content generation strategy of the educational large model to generate the final intelligent lesson preparation plan.

[0081] In this implementation manner, S1 includes the following steps:

[0082] S11. Obtain the original educational data set D raw The original educational data set includes the primary and secondary school textbook content set T m , the teaching syllabus content set T s , the collection of examination questions over the years T e and the primary and secondary school learning question bank content set T q ;

[0083] S12. Perform text cleaning on the original educational data set D raw , remove and standardize the garbled information, redundant symbols, non-teaching content, and format error content included in the original educational data set to obtain the cleaned educational text set D clean , where the cleaning process adopts a unified standard rule to ensure the consistency of the expression structure of each source content;

[0084] S13. Perform format standardization on the cleaned educational text set D clean , perform standardized coding on the paragraph structure, question type identification, knowledge point description, and teaching terms therein to obtain the format-standardized educational data set D norm , where the format standardization process maintains the consistent expression form of all samples in the semantic structure and information level;

[0085] S14. Based on the format-standardized educational data set D norm , perform key knowledge point annotation, match and annotate each educational data sample with a certain knowledge point in its corresponding teaching knowledge point set K = {k1, k2,..., k n}, to form the annotated educational data set D tag , where k i represents the i-th knowledge point, and the annotation result is embedded into each sample data in the form of a structured label;

[0086] S15. On the annotated educational data set D tagRe - index and group according to the subject, grade, course content, and knowledge - point category, organize each sample data into a structured form, and construct a structured education dataset D struct , where each piece of structured education data includes: an educational text sample x i , the corresponding knowledge - point label k i and its classification information y i . The classification information includes subject attributes, grade dimensions, and question - type categories.

[0087] In this embodiment, S2 includes the following steps:

[0088] S21. Based on the structured education dataset D struct Perform teaching semantic unit segmentation on each educational text sample x i to construct a set of teaching semantic blocks B i ={b i1 ,b i2 ,...,b im}, where a teaching semantic block represents the smallest information unit with consistent teaching logic and is used to capture the semantic dependency structure of teaching objectives and knowledge points;

[0089] S22. Apply a context embedding representation function to each teaching semantic block b i in the set of teaching semantic blocks B ij to generate a semantic embedding vector E ij , and map the semantic embedding vector to a semantic compression space S sem , retaining the semantic factors related to question generation;

[0090] S23. Introduce a semantic activation function A(k t ,E ij ) = cosine(φ k [[ID=4:6]](k t ),E ij ) to calculate the semantic support strength of the teaching semantic block b ij for the knowledge point k t , where φ k (k t ) represents the embedding representation of the knowledge point and is used to guide the establishment of edges between the knowledge point and the context semantics;

[0091] S24. Based on the set of all knowledge - point labels K, the set of teaching objectives G, and the set of teaching semantic blocks B, construct a teaching knowledge graph G edu =(V, E, W), where the node set V includes knowledge - point nodes v k , teaching - objective nodes v g and teaching - semantic - block nodes v b, the edge set E includes semantic support edges, target dependency edges, and chapter structure edges, and the weight set W is jointly composed of semantic activation function values and context embedding representation functions;

[0092] S25. Based on the teaching knowledge graph G edu For a given teaching objective node v g and knowledge point node v k Execute the context-aware path reasoning algorithm to calculate the teaching path P for question generation inf :

[0093]

[0094] where P inf is the teaching path finally used for generating questions, is the set of reachable paths from all teaching objective nodes v g to knowledge point nodes v k , Redundancy(p) is the path redundancy penalty function used to avoid repeated reference to the same semantic block, and λ div is the balance coefficient for controlling the balance between diversity reward and structural overlap. The obtained teaching path is used to guide the educational large model to generate goal-oriented questions;

[0095] S26. Encode the teaching path P inf into a semantic path vector π, which is used as the input semantic control factor of the educational large model generation module to guide the model to generate a set of generated questions with rigorous logic, consistent semantics, and clear knowledge points

[0096] S27. Conduct knowledge point call analysis on the set of generated questions to extract the set of knowledge points actually called during generation and compare it with the teaching objective set and calculate the semantic coverage deviation compared with the target sets K, G in the path reasoning stage:

[0097]

[0098] where Δ sem is the semantic coverage deviation of the current generated question semantic path from the teaching objective, and β is the weight factor of the semantic coverage deviation;

[0099] S28. Based on the semantic coverage deviation feedback result, adjust the edge weight set W in the teaching knowledge graph, increase the edge weights of semantic nodes that appear in the path but are not called, optimize the edge strength of mis-activated paths, and form a closed-loop optimization of the semantic graph driven by question generation:

[0100]

[0101] Among them, \(W′(u, v)\) is the set of edge weights after feedback adjustment, and \(\eta\) is the structure update learning rate, which indicates whether the edge participated in the current round of reasoning path.

[0102] In this embodiment, S3 includes the following steps:

[0103] S31. Receive the set of lesson preparation requirement data \(D\) input by the teacher plan , and perform semantic encoding processing on all the text information in the set of lesson preparation requirement data \(D\) plan input by the teacher, and generate the teacher's lesson preparation vector representation \(\varPsi\) through the context understanding module in the educational large model plan , where the teacher's lesson preparation vector representation integrates information on teaching objective intentions, course context, learning group characteristics, and teaching method requirements;

[0104] S32. Invoke the teaching knowledge graph \(G\) edu Match the teacher's lesson preparation vector representation \(\varPsi\) plan one by one with all the teaching semantic path vector sets \(\varPi\) in the teaching knowledge graph, calculate the degree of semantic relevance, and obtain the semantic relevance scores between the teacher's lesson preparation vector representation and each semantic path;

[0105] S33. Based on the calculated semantic relevance scores, filter out the path set that satisfies the preset semantic relevance threshold with the teacher's lesson preparation vector representation \(\varPsi\) plan as the semantic support path set \(\varPi\) for generating intelligent lesson preparation content * ;

[0106] S34. Based on the filtered semantic support path set \(\varPi\) * and the teacher's lesson preparation vector representation \(\varPsi\) plan , generate the preliminary intelligent lesson preparation content set \(C\) plan =\(\{c_1, c_2,..., c\) k \}\) by the educational large model. The preliminary intelligent lesson preparation content set includes a draft teaching outline, a preliminary courseware, class notes, teaching cases, and personalized exercise suggestions, and each content is generated by the semantic generation function of the educational large model under the dual constraints of path semantics and teacher intentions;

[0107] S35. Perform teaching consistency detection and content structure integrity assessment on each content in the preliminary intelligent lesson preparation content set \(C\) plan , and through the evaluation module built into the educational large model, output the content fitness score vector \(\varOmega=\{\omega_1, \omega_2,..., \omega\) k \}\). The score vector is used to represent the structural fitness and teaching objective alignment degree of each content with respect to the teacher's input intention;

[0108] S36. Return the intelligent lesson preparation content in the fitness score vector Ω with scores lower than the preset tolerance threshold τ to the educational large model for re-generation and correction, and retain the revision records in the form of structured tags in the re-generated version, and finally generate a preliminary set of intelligent lesson preparation content The preliminary set of intelligent lesson preparation content meets the requirements of teaching objectives, contextual logic and individual characteristics.

[0109] In this embodiment, the lesson preparation requirement data set includes teaching objective descriptions, a set of course theme keywords, teaching stage information, teaching method preferences, and a set of student portrait parameters.

[0110] In this embodiment, S4 includes the following steps:

[0111] S41. The preliminary set of intelligent lesson preparation content is input into the intelligent interactive presentation module for structured display. The intelligent interactive presentation module uses document views, slide views and interactive component views to present teaching outline drafts, initial courseware, classroom handouts, teaching cases and personalized exercise suggestions respectively according to the content type;

[0112] S42. Based on the structured display interface, a fusion interaction interface of the natural language interaction sub-module and the speech recognition interaction sub-module is constructed. The natural language interaction sub-module receives the set of text instructions Γ text input by the teacher, and the speech recognition interaction sub-module receives the set of teacher speech inputs Γ voice , where the set of text instructions and the set of speech inputs are uniformly mapped to the teaching feedback semantic space;

[0113] S43. The set of natural language inputs Γ text and the set of speech inputs Γ voice are jointly encoded into the set of teacher interaction feedback representations Γ feedback , and the set of teacher interaction feedback representations is used to express the teacher's intention to edit and modify the preliminary intelligent lesson preparation content, reconstruction suggestions, content preferences or teaching style adjustment requirements;

[0114] S44. Match each item in the set of teacher interaction feedback representations Γ feedback with the intelligent lesson preparation content set to identify the set of content indexes I match corresponding to the feedback target, and extract the set of original semantic path vectors Π match of the corresponding content segments;

[0115] S45. For each teacher interaction feedback representation and the content segment it maps to, execute the semantic alignment function, extract the offset features and adjustment directions of the teacher's intention in the original path semantics, and form the set of feedback offset vectors Δ edit ;

[0116] S46. Generate an interactive feedback data set Each feedback data record includes the original semantic path vector π l , the feedback offset vector δ l and the content index i corresponding to the feedback target l .

[0117] In this embodiment, S5 includes the following steps:

[0118] S51. Receive and parse the interactive feedback data set F interact , and extract the content modification information, teaching focus adjustment requirements, and personalized teaching preferences involved in the feedback;

[0119] S52. According to the semantic features of the teacher feedback offset vector δ l , perform semantic reconstruction on the corresponding original semantic path vector π l to generate an adjusted path semantic vector π′ l . During the semantic reconstruction process, the original structure logic of the path is maintained, and only minimal perturbation optimization is performed on the node content, edge weight distribution, or knowledge point organization method;

[0120] S53. Use the set of all reconstructed path semantic vectors Π′ = {π′1, π′2,..., π′ s} as semantic driving factors and input them into the content regeneration module of the education large model, and link with the original teacher lesson preparation vector representation Ψ plan to jointly generate an adjusted intelligent lesson preparation content set

[0121] S54. Based on the interactive feedback data set F interact , use the teacher feedback offset vector delta l and the original semantic path vector pi l as inputs to perform semantic-guided reconstruction on the intelligent lesson preparation content set to generate an adjusted intelligent lesson preparation content set with fine-tuned content On the basis of retaining the original semantic path logical structure, the adjusted intelligent lesson preparation content set integrates the content modification, teaching focus adjustment, and personalized needs expressed in the teacher feedback, and specifically includes a draft teaching outline, a first draft of the courseware, class notes, teaching cases, and personalized exercise suggestions;

[0122] S55. Finally, output the adjusted intelligent lesson preparation content set The adjusted intelligent lesson preparation content set is used to support teachers' teaching implementation in real time and dynamically support personalized teaching design.

[0123] In this embodiment, the draft teaching syllabus: based on the original semantic path vector π l For the teaching objective nodes covered, if the teacher feedback offset vector contains information on the decomposition of teaching objectives and the reordering of knowledge points, the priorities of the teaching objective nodes in the path are rearranged, and the entry order and target expression level in the draft teaching syllabus are updated accordingly;

[0124] The first draft of the courseware: based on the graphic and text nodes and interaction logic in the original semantic path vector π l If the teacher feedback offset vector contains preferences for slide layout, suggestions for illustration methods, or requirements for content simplification, the weights of the nodes related to visual expression in the path are adjusted, and the page structure and information distribution of the courseware are reconstructed;

[0125] The class handout: based on the knowledge point coverage section in the original semantic path vector π l If the teacher feedback offset vector contains requirements for replacing examples, expanding the handout structure, or adjusting the graphic - text ratio, the corresponding knowledge point segments are updated locally, and the exemplary content nodes in the path are replaced;

[0126] The teaching case: based on the teaching situation nodes and guiding semantic chains bound in the original semantic path vector π l If the teacher feedback offset vector contains requirements for reconstructing the teaching situation, replacing the background, or correcting the style, the dialogue framework and situation setting of this path segment are regenerated, while keeping the start and end points of the path unchanged;

[0127] Personalized exercise suggestions: based on the knowledge point nodes and their graph edge weight information associated with the original semantic path vector π l If the teacher feedback offset vector contains requirements for ability adaptation, difficulty control, or strengthening of examination points, the weights of the nodes and the edge structure in this path segment are slightly adjusted, the matching degree between the question type and difficulty is reconstructed, and the logic is kept consistent with the original path.

[0128] An intelligent teacher lesson - preparation assistance system based on an educational large - model, which is applied to an intelligent teacher lesson - preparation method based on an educational large - model, includes the following modules:

[0129] The data processing module is used to collect primary and secondary school textbooks, teaching syllabuses, past examination questions, and exercise question banks, and perform text cleaning, format standardization, and key knowledge point annotation on the collected educational data to generate a structured educational data set;

[0130] The educational large - model construction module is used to pre - train and fine - tune the structured educational data set to construct an educational large - model for primary and secondary school scenarios;

[0131] The teaching knowledge graph construction module is used to construct a teaching knowledge graph based on teaching semantic blocks, teaching objectives, and knowledge point relationships to form a semantic path reasoning mechanism;

[0132] An intelligent lesson preparation content generation module, which is used to receive the lesson preparation requirement data input by teachers, and combine the education large model and the teaching knowledge graph to generate preliminary intelligent lesson preparation content such as a draft teaching outline, a preliminary courseware, class handouts, teaching cases, and personalized exercise suggestions;

[0133] An interactive feedback processing module, which is used to collect teachers' editing modifications, style preferences, and teaching adjustment requirements for the intelligent lesson preparation content through a natural language and speech recognition interactive interface, and generate a semantic deviation representation and an interactive feedback data set;

[0134] A dynamic optimization module, which is used to reconstruct the original semantic path and update the graph path based on the teachers' feedback content, generate dynamically adjusted intelligent lesson preparation content, and is used to drive the subsequent test question generation process.

[0135] Example 1:

[0136] On October 12, 2024, the ninth-grade mathematics group of A Experimental Middle School first introduced this invention to conduct an intelligent proposition assistance experiment in the teaching topic of "The Image and Properties of Quadratic Functions".

[0137] At 08:16 in the morning, mathematics teacher Li logged in to the intelligent teaching assistance system through the school's teaching platform and entered the following lesson preparation requirements:

[0138] Grade: Ninth grade; Subject: Mathematics; Teaching objectives: Master the mutual conversion relationship between the vertex form and the general form of quadratic functions; Be able to judge the function properties according to the image; Current progress: The third class of the teaching unit, and the basic image changes of "y = ax 2 + bx + c" have been covered; Teachers' personalized preferences: Tend to structured image questions, multiple questions for one topic, and life situation guiding questions; Teachers' input text instructions: Please generate a set of exercises containing moderately difficult quadratic function image questions, including multiple-choice, fill-in-the-blank, and answer questions, covering three core knowledge points, and it is recommended that the total number of questions be controlled within 10 questions.

[0139] After receiving the input, the system starts the semantic analysis module at 08:17:41, converts the teacher's input into a semantic embedding representation Ψ plan , calls the structured education data set for semantic path matching, and constructs a knowledge graph subgraph related to the teaching objectives at 08:18:07, which contains 24 nodes and a total of 51 semantic path candidate sets.

[0140] At 08:18:09, the system executes context-aware path reasoning, selects the path sets covering "the axisymmetry of the quadratic function image", "the relationship between the function opening direction and the coefficient positive and negative", and "the vertex coordinate change", and generates path vectors π1, π2, π3, and the path structure depths are 5, 7, and 6 respectively.

[0141] At 08:18:14, the education large model generated 13 preliminary candidate questions according to the path semantic vector drive, including:

[0142] Multiple-choice questions: 5; Fill-in-the-blank questions: 3; Short-answer questions: 5; The stem of the scenario questions involves three types of life context, namely "traffic light change cycle", "parabolic basketball projection", and "water flow fountain design".

[0143] At 08:18:16, the system started the mutant fish swarm algorithm to optimize the global fitness of the 13 questions. The set evaluation indicators include:

[0144] Knowledge point coverage C k : It must cover 3 input target knowledge points; Difficulty distribution D s : Controlled to be moderately difficult to difficult (system-defined score between 4.0 - 4.5 / 5); Language normality score L f : It is required that the sentence pattern is smooth and the stem is complete; The top 9 questions with the comprehensive score of the fitness function are selected. The optimization took 0.91 seconds, and finally the preliminary test question set was generated

[0145] At 08:19, Teacher Li started previewing the test questions and gave the following modification instructions through voice interaction:

[0146] "The image coordinates of question 4 are not clear. It is recommended to add a graph"; "The question of question 6 is too broad. Can it be split into two questions?"; "The context of question 8 can be changed to a campus problem scenario, such as the design of a parabolic school gate";

[0147] The system real-time recognizes the voice content and generates feedback offset vectors δ4, δ6, δ8, which act on the relevant nodes in the path vectors π2, π3 to perform path-level adjustment.

[0148] At 08:21:08, the system automatically reconstructed the test question generation path and updated the test question content as follows:

[0149] Question 4 adds dynamic image auxiliary information; Question 6 is split from 1 question into 2 sub-questions, and the second sub-question focuses on "the influence of the sign of coefficient a on the opening direction"; The scenario of question 8 is switched to "the arched design of the campus main entrance", and the image uses a parabola to express the gate structure.

[0150] At 08:22, the teacher clicked "Save and Generate PDF", and the system automatically generated a structured practice paper, including questions, answers, test point annotations, and preset scoring criteria, and synchronously uploaded it to the school platform and the class assignment system.

[0151] The teacher feedback collection system counted the data on the same day as follows:

[0152] Time taken for the system to generate test questions: 4 minutes and 26 seconds; Number of times teachers manually modified: 3 times (all through voice interaction); Student completion rate: 98.1% (47 out of 48 students in the class completed); Average completion time: 21 minutes and 43 seconds; Accuracy of knowledge points covered in the test paper (AI statistics): 95.8%; Teacher satisfaction score: 4.82 / 5.

[0153] As a comparison, the time taken for the practice paper of the previous unit "Pythagorean Theorem and Triangles" prepared by Li using the traditional method was 47 minutes, the teachers manually modified it 7 times, the student completion rate was only 92.4%, the accuracy of knowledge points covered was 82.7%, and the satisfaction score was 3.94 / 5.

[0154] In terms of test question training data, the total corpus of function-type test questions used in this round of training is:

[0155] High-quality historical quadratic function questions: 4,130; Structured annotation data matching textbook semantic nodes: 3,120; Teacher manual correction feedback records: 984 (used to fine-tune path semantic deviation); Total number of path nodes called during system operation: 1,724.

[0156] As can be seen from Example 1, the present invention can not only automatically generate a set of test questions that meet the needs of teachers, but also dynamically adapt to teacher feedback through means such as knowledge graph path reasoning, semantic embedding, and graph structure optimization, realizing the output of test questions with semantic consistency, reasonable difficulty, and high consistency with teaching objectives, greatly improving the efficiency of teacher question preparation, the fit between students' answering and the overall teaching evaluation accuracy, and solving the problems of low efficiency in test question generation, difficulty in controlling quality, and lack of personalized interactive feedback in existing methods.

[0157] The present invention constructs a multi-level graph structure including teaching objective nodes, knowledge point nodes, and teaching semantic block nodes, introduces a semantic activation function and a path redundancy penalty function, uses a path reasoning algorithm to obtain the optimal teaching path vector, and dynamically adjusts the edge weights of the graph based on the semantic coverage deviation between the generated test questions and the target knowledge points after test question generation, realizing a generation-driven self-learning update mechanism for the graph structure, breaking through the limitation of the static use of traditional knowledge graphs, making the semantic path of test question generation highly consistent with the teaching structure, and significantly improving the path generation accuracy and semantic coverage rate.

[0158] The present invention uniformly maps the natural language input of teachers and voice commands into interactive feedback representations, generates a feedback offset vector through semantic alignment analysis, and acts on the original path semantic vector for minimum perturbation reconstruction, realizing the structured modification of multiple types of content such as teaching outlines, lecture notes, cases, and test question suggestions. On this basis, the feedback data synchronously updates the teaching knowledge graph and the generation model to form a closed-loop content optimization mechanism, significantly improving the response ability of the intelligently generated content to the teaching intentions of teachers, supporting the dynamic coordination of teaching content and test question styles with the teaching rhythm and classroom feedback, and providing an intelligent decision-making support basis for large-scale personalized teaching.

[0159] As described above, the above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. An intelligent auxiliary teacher lesson preparation method based on an educational large model, characterized in that, It includes the following steps: S1. Collect and integrate multi-source teacher lesson preparation data related to teacher lesson preparation, and preprocess the teacher lesson preparation data to generate a standardized teacher lesson preparation dataset; S2. Use semantic parsing on the standardized teacher lesson preparation dataset to construct a teaching knowledge graph; S3. Based on the education large model, perform semantic understanding and context analysis on the lesson preparation requirement data input by the teacher, and match the lesson preparation requirement data with the teaching knowledge graph to generate preliminary intelligent lesson preparation content; S4. Present the preliminarily generated intelligent lesson preparation content, and provide an interactive interface that combines natural language and speech recognition, allowing teachers to preview, edit, and provide feedback on the intelligent lesson preparation content in real time to generate interactive feedback data; S5. Parse the interactive feedback data, extract content modifications, key adjustments, and personalized requirements involved in the feedback, and dynamically optimize the preliminarily generated intelligent lesson preparation content based on the feedback information to generate adjusted intelligent lesson preparation content; S6. According to the teacher's historical lesson preparation data and the interactive feedback data, perform adaptive learning and optimization on the intelligent lesson preparation content, adjust the weights of key knowledge points in the teaching knowledge graph, and update the content generation strategy of the education large model to generate the final intelligent lesson preparation plan.

2. The intelligent auxiliary teacher lesson preparation method based on an educational large model according to claim 1, wherein The S1 includes the following steps: S11. Obtain the original education data set D raw The original education data set includes the primary and secondary school textbook content set T m , the syllabus content set T s , the collection of examination questions over the years T e and the primary and secondary school learning question bank content set T q ; S12. For the original educational data set D raw Perform text cleaning processing, remove and standardize the garbled information, redundant symbols, non-teaching content, and formatting errors contained in the original educational data set, and obtain the cleaned educational text set D clean ; S13. For the cleaned educational text set D clean Perform format standardization processing, perform standardized coding on the paragraph structure, question type identification, knowledge point description, and teaching language in it, and obtain the standardized educational data set D norm ; S14. Based on the standardized education data set D norm Perform key knowledge point annotation, match and annotate each education data sample with a certain knowledge point in its corresponding teaching knowledge point set K = {k1, k2,..., k n}, and form the annotated education data set D tag , where k i represents the i-th knowledge point, and the annotation result is embedded into each sample data in the form of a structured label; S15. For the labeled educational data set D tag Re-index and group it according to the subject, grade, course content, and knowledge point category it belongs to, organize each sample data into a structured form, and construct a structured educational data set D struct , where each piece of structured educational data includes: an educational text sample x i , the corresponding knowledge point label k i and its classification information y i , and the classification information includes subject attributes, grade dimensions, and question type categories.

3. The intelligent auxiliary teacher lesson preparation method based on an educational large model according to claim 2, wherein The S2 includes the following steps: S21. Based on the structured educational dataset D struct For each educational text sample x i Perform teaching semantic unit segmentation to construct a set B of teaching semantic blocks i ={b i1 , b i2 ,..., b im}, where a teaching semantic block represents the smallest information unit with consistent teaching logic and is used to capture the semantic dependency structure between teaching objectives and knowledge points; S22. For each teaching semantic block b in the teaching semantic block set B i apply the context embedding representation function to generate a semantic embedding vector E ij and map the semantic embedding vector to the semantic compression space S ij while retaining the semantic factors related to test question generation; sem ​ S23. Introduce the semantic activation function A(k t , E ij ) = cosine(φ k (k t ), E ij ) to calculate the semantic support strength of the teaching semantic block b ij for the knowledge point k t . Among them, φ k (k t ) represents the embedding representation of the knowledge point, which is used to guide the establishment of edges between the knowledge point and the context semantics; S24. Construct a teaching knowledge graph \(G\) based on the set \(K\) of all knowledge point tags, the set \(G\) of teaching objectives, and the set \(B\) of teaching semantic chunks edu =(V, E, W), where the node set \(V\) includes knowledge point nodes \(v\) k , teaching objective nodes \(v\) g and teaching semantic chunk nodes \(v\) b , the edge set \(E\) includes semantic support edges, target dependency edges, and chapter structure edges, and the weight set \(W\) is jointly composed of semantic activation function values and context embedding representation functions; S25. Based on the teaching knowledge graph G edu For the given teaching objective node v g and the knowledge point node v k Execute the context-aware path reasoning algorithm to calculate the test question generation teaching path P inf : Among them, P inf is the teaching path finally used to generate test questions, is the set of reachable paths from all teaching objective nodes v g to knowledge point nodes v k , Redundancy(p) is the path redundancy penalty function used to avoid repeated reference to the same semantic block, and λ div is the balance coefficient for controlling the balance between diversity rewards and structural overlap. The obtained teaching path is used to guide the educational large model to generate goal-oriented test questions; S26. Encode the teaching path P inf into a semantic path vector π, which serves as the input semantic control factor for the educational large model generation module, guiding the model to generate a set of generated test questions with rigorous logic, consistent semantics, and clear knowledge points S27. Generate a test question set Conduct a knowledge point call analysis, and extract the set of knowledge points called in the actual generation And compare it with the teaching objective set And compare it with the target sets K and G in the path reasoning stage, and calculate the semantic coverage deviation: Among them, Δ sem is the semantic coverage deviation of the current generated test question semantic path from the teaching objective, and β is the weight factor of the semantic coverage deviation; S28. Based on the semantic coverage deviation feedback result, adjust the edge weight set W in the teaching knowledge graph, increase the edge weights of semantic nodes that appear in the path but are not called, and optimize the edge strength of the misactivated path to form a semantic graph optimization closed loop driven by question generation: Among them, \(W'(u, v)\) is the set of edge weights after feedback adjustment, and \(\eta\) is the structure update learning rate, which indicates whether the edge participated in the current round of inference path.

4. The intelligent auxiliary teacher lesson preparation method based on an educational large model according to claim 3, wherein, The S3 includes the following steps: S31. Receive the set of lesson preparation requirement data D input by the teacher plan , and perform semantic encoding processing on all the text information in the set of lesson preparation requirement data D plan input by the teacher to generate a teacher lesson preparation vector representation Ψ through the context understanding module in the education large model plan , where the teacher lesson preparation vector representation integrates information on teaching objective intentions, course context, learning group characteristics, and teaching method requirements; S32. Invoke the teaching knowledge graph G edu The teacher's lesson preparation vector representation Ψ plan is matched one by one with all the teaching semantic path vector sets Π in the teaching knowledge graph, the semantic correlation degree is calculated, and the semantic correlation score between the teacher's lesson preparation vector representation and each semantic path is obtained; S33. Based on the calculated semantic relevance scores, filter out the paths that satisfy the preset semantic relevance threshold with the teacher's lesson preparation vector representation Ψ plan as the semantic support path set Π for generating intelligent lesson preparation content * ; S34. Based on the set of semantic support paths Π obtained by screening * and the teacher's lesson preparation vector representation Ψ plan , the educational large model generates a preliminary set of intelligent lesson preparation content C plan = {c1, c2,..., c k}, where the preliminary set of intelligent lesson preparation content includes a draft teaching outline, a preliminary courseware, class notes, teaching cases, and personalized exercise suggestions, and each piece of content is generated by the semantic generation function of the educational large model under the dual constraints of path semantics and teacher intentions; S35. Perform teaching consistency detection and content structure integrity assessment on each item in the preliminary intelligent lesson preparation content set C plan and output a content fitness score vector Ω = {ω1, ω2,..., ω k} through the evaluation module built into the education large model. The score vector is used to represent the structural fitness and teaching objective alignment degree of each item relative to the teacher's input intention; S36. Return the intelligent lesson preparation content with scores lower than the preset tolerance threshold τ in the fitness score vector Ω to the educational large model for re-generation and correction, and retain the revision record in the form of structured tags in the re-generated version, and finally generate a preliminary set of intelligent lesson preparation content The preliminary set of intelligent lesson preparation content meets the requirements of teaching objectives, context logic, and individual characteristics.

5. The intelligent auxiliary teacher lesson preparation method based on an educational large model according to claim 4, characterized in that, The lesson preparation requirement data set includes teaching objective descriptions, a set of course theme keywords, teaching stage information, teaching method preferences, and a set of student portrait parameters.

6. The intelligent auxiliary teacher lesson preparation method based on an educational large model according to claim 4, wherein, The S4 includes the following steps: S41. Input the set of preliminary intelligent lesson preparation content into the intelligent interaction presentation module for structured display. The intelligent interaction presentation module uses the document view, slide view, and interactive component view to present the draft teaching outline, initial courseware, class handouts, teaching cases, and personalized exercise suggestions diversely according to the content type; S42. Build a fusion interaction interface for the natural language interaction sub-module and the speech recognition interaction sub-module based on the structured display interface. The natural language interaction sub-module receives the text instruction set Γ input by the teacher text , and the speech recognition interaction sub-module receives the teacher speech input set Γ voice , where the text instruction set and the speech input set are uniformly mapped to the teaching feedback semantic space; S43. Co-code the natural language input set Γ text with the speech input set Γ voice into a set of teacher interaction feedback representations Γ feedback , where the set of teacher interaction feedback representations is used to express the teacher's intention to edit and modify the preliminary intelligent lesson preparation content, reconstruction suggestions, content preferences, or requirements for teaching style adjustment; S44. For the set Γ of teacher interaction feedback expressions feedback match each content in the set of intelligent lesson preparation content one by one to identify the content index set I corresponding to the feedback target match and extract the set Π of original semantic path vectors of the corresponding content segments match ; S45. For each teacher interaction feedback representation and its mapped content segment, execute a semantic alignment function to extract the offset features and adjustment directions of the teacher's intention in the original path semantics, and form a feedback offset vector set Δ edit ; S46. Generate an interactive feedback dataset Each feedback data record includes an original semantic path vector π l , a feedback offset vector δ l , and a content index i corresponding to the feedback target l .

7. A smart assisted teacher lesson preparation method based on an education large model according to claim 1, characterized in that, The S5 includes the following steps: S51. Receive and parse the interactive feedback data set F interact , and extract the content modification information, teaching key adjustment requirements, and personalized teaching preferences involved in the feedback from it; S52. According to the semantic features of the teacher feedback offset vector δ l perform semantic reconstruction on the corresponding original semantic path vector π l to generate an adjusted path semantic vector π′ l , and the semantic reconstruction process keeps the original structural logic of the path unchanged, and only performs minimum perturbation optimization on the node content, edge weight distribution, or knowledge point organization method; S53. Set all the reconstructed path semantic vectors Π′={π′1,π′2,...,π s ′} is input into the content regeneration module of the education model as a semantic driving factor, and the original teacher preparation vector representation Ψ plan Jointly generate an adjusted set of intelligent lesson preparation content S54. On the basis of the interactive feedback dataset F interact and using the teacher feedback offset vector delta l and the original semantic path vector pi l as inputs, perform semantic-guided reconstruction on the intelligent lesson preparation content set to generate an adjusted intelligent lesson preparation content set with fine-tuned content. On the basis of retaining the original semantic path logical structure, the adjusted intelligent lesson preparation content set integrates content modifications, teaching focus adjustments, and personalized needs expressed in the teacher feedback, specifically including a draft teaching outline, a first draft of the courseware, class lecture notes, teaching cases, and personalized exercise suggestions; S55. Final output: adjusted set of intelligent lesson preparation content The adjusted set of intelligent lesson preparation content is used to support teachers' teaching implementation in real time and dynamically support personalized teaching design.

8. An intelligent auxiliary teacher lesson preparation system and method based on an education large model according to claim 7, characterized in that The draft teaching syllabus: Based on the original semantic path vector π l For the teaching objective nodes covered, if the teacher feedback offset vector contains information on the decomposition of teaching objectives and the reordering of knowledge points, the priorities of the teaching objective nodes in the path are rearranged, and the entry order and objective expression level in the draft teaching syllabus are updated accordingly; The initial draft of the courseware: Based on the original semantic path vector π l For the graphic and text nodes and interaction logic in it, if the teacher's feedback offset vector contains preferences for slide layout, suggestions for illustration methods, or requirements for content simplification, then adjust the node weights related to visual expression in the path, and reconstruct the courseware page structure and information distribution; The classroom lecture notes: Based on the original semantic path vector π l In the knowledge point coverage section, if the teacher feedback offset vector contains requirements for example replacement, lecture note structure expansion, or graphic - text ratio adjustment, then locally update the corresponding knowledge point fragment and replace the exemplary content nodes in the path; The teaching case: Based on the original semantic path vector π l For the teaching situation nodes and guiding semantic chains bound in it, if the teacher feedback offset vector contains requirements for teaching situation reconstruction, background replacement, or style correction, regenerate the dialogue framework and situation setting for this path segment while keeping the start and end points of the path unchanged; The personalized exercise suggestions: Based on the original semantic path vector π l For the associated knowledge point nodes and their graph edge weight information, if the teacher feedback offset vector contains requirements for ability adaptation, difficulty control, or test point strengthening, then fine-tune the node weights and edge structures in this path segment, reconstruct the matching degree between the question type and difficulty, and maintain consistency with the original path logic.

9. An intelligent auxiliary teacher lesson preparation system based on an educational large model, which is applied to an intelligent auxiliary teacher lesson preparation method according to any one of claims 1-8, and is characterized in that, It includes the following modules: The data processing module is used to collect primary and secondary school textbooks, teaching syllabuses, past exam questions, and exercise question banks, and perform text cleaning, format standardization, and key knowledge point annotation on the collected education data to generate a structured education dataset; The education large model construction module is used to pre-train and fine-tune the structured education dataset to construct an education large model for primary and secondary school scenarios; The teaching knowledge graph construction module is used to construct a teaching knowledge graph based on teaching semantic blocks, teaching objectives, and knowledge point relationships to form a semantic path reasoning mechanism; The intelligent lesson preparation content generation module is used to receive the lesson preparation requirement data input by the teacher, and combine the education large model with the teaching knowledge graph to generate preliminary intelligent lesson preparation content such as a draft teaching outline, a first draft of courseware, class notes, teaching cases, and personalized exercise suggestions; The interactive feedback processing module is used to collect the editing modifications, style preferences, and teaching adjustment requirements of teachers for the intelligent lesson preparation content through a natural language and speech recognition interactive interface to generate a semantic deviation representation and an interactive feedback dataset; The dynamic optimization module is used to reconstruct the original semantic path and update the graph path based on the teacher feedback content to generate dynamically adjusted intelligent lesson preparation content, and is used to drive the subsequent question generation process.

Images