Personalized teaching method and system based on generative artificial intelligence

Abstract

The invention provides a personalized teaching method and system based on generative artificial intelligence. The method comprises the following steps: constructing a knowledge graph and a teaching strategy template library of a target course; constructing a feature enhancement matrix; calculating node activation intensity, generating a plurality of candidate learning paths through a beam search algorithm, and screening an optimal learning path in combination with a preset teaching strategy template; using the generative model to generate standardized teaching content adaptive to the teaching strategy; dynamically optimizing a teaching scheme through a neuroplasticity model; and carrying out teaching adjustment based on the optimized teaching scheme. Standardized teaching content is generated by utilizing the generative model, the teaching content quality is improved, the teaching content is closely matched with a learning path and a teaching strategy of a learner, meanwhile, a teaching scheme is optimized through the neuroplasticity model, teaching parameters and cognitive resource allocation are adjusted in time according to real-time feedback and cognitive states of the learner, and the teaching efficiency is improved. Limited cognitive resources are reasonably distributed to key parts of teaching contents, and the teaching effect is improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent teaching technology, and in particular to a personalized teaching method and system based on generative artificial intelligence. Background Art

[0002] Traditional teaching methods often rely on fixed lesson plans and textbooks, making it difficult to personalize instruction based on individual learners' differences. This one-size-fits-all approach ignores learners' cognitive states and learning needs, leading to poor teaching outcomes and low learning efficiency. Furthermore, traditional teaching methods also have limitations in the selection of teaching content and strategies. Teachers often choose these based on their own experience and judgment, lacking a scientific and systematic approach. Summary of the Invention

[0003] Based on this, the purpose of the present invention is to propose a personalized teaching method and system based on generative artificial intelligence to solve the above-mentioned problems.

[0004] According to the present invention, a personalized teaching method based on generative artificial intelligence is proposed, the method comprising:

[0005] Build a knowledge map and teaching strategy template library for the target course, including a series of preset teaching strategy templates;

[0006] Based on the learner's individual learning characteristics and combined with the group learning behavior pattern, feature enhancement is performed to construct a feature enhancement matrix;

[0007] Calculating the node activation strength, and generating multiple candidate learning paths through a beam search algorithm, and screening out the optimal learning path in combination with the preset teaching strategy template;

[0008] Based on the preferred learning path, using a generative model to generate standardized teaching content adapted to the teaching strategy;

[0009] Dynamically optimizing the teaching plan through a neuroplasticity model by combining the standardized teaching content and learner characteristics reflected in the feature enhancement matrix;

[0010] Make teaching adjustments based on the optimized teaching plan.

[0011] Furthermore, the step of constructing the feature enhancement matrix includes:

[0012] Collect learning characteristic data common to individuals and groups, including learning behavior data, cognitive performance data, and learning context data;

[0013] Fusing individual learning feature data with group common feature data to obtain fused feature data;

[0014] Extract common learning patterns and explore common characteristics among learners;

[0015] Based on the fusion feature data and common learning model, a feature enhancement matrix F_enhanced is constructed, and the formula is:

[0016] F_enhanced=a·F_individual+b·F_group+c·ΔF_temporal,

[0017] Among them, a, b, and c are the weight coefficients used to adjust the individual learning feature F_individual, the group learning feature F_group, and the influence of the feature evolution over time, respectively. ΔF_temporal reflects the evolution trend of the learning feature over time.

[0018] Furthermore, the step of screening out the preferred learning path includes:

[0019] Mapping the feature enhancement matrix to a set of nodes in the knowledge graph and determining the learner's initial position in the knowledge graph to clarify the starting point of learning;

[0020] Calculate the node activation strength using the formula:

[0021] h_v (l+1) =σ(∑{u∈N(v)}α_{vu}Wh_ul),

[0022] Among them, h_v (l+1) is the activation strength of node v in the l+1 layer, N(v) is the set of neighbor nodes of node v, α_{vu} is the attention coefficient, which is used to measure the importance of node u to node v, W is the weight matrix, h_ul is the eigenvector of node u in the l layer, and σ is the activation function;

[0023] Based on the node activation strength, K candidate paths are generated by a beam search algorithm;

[0024] The candidate paths are scored in combination with the preset teaching strategy template, and the path with the highest score is selected as the preferred learning path.

[0025] Furthermore, the step of using the generative model to generate standardized teaching content adapted to the teaching strategy includes:

[0026] Build a generative model that includes semantic parsing, multimodal generation, and cognitive load assessment modules;

[0027] Understanding the teaching requirements in the preferred learning path through the BERT architecture in the semantic parsing module and generating structured data of the teaching requirements;

[0028] Based on the teaching demand structured data, initial teaching content is generated through the collaborative work of CLIP and DALL·E in the multimodal generation module;

[0029] Performing cognitive load evaluation on the initial teaching content through the LSTM network in the cognitive load evaluation module to obtain a cognitive load evaluation result;

[0030] According to the cognitive load assessment result, the initial teaching content is adjusted by the multimodal generation module to obtain adjusted teaching content.

[0031] Furthermore, the step of using the generative model to generate standardized teaching content adapted to the teaching strategy also includes:

[0032] Design a dual discriminator architecture, including a content quality discriminator and a teaching strategy discriminator;

[0033] The adjusted teaching content is evaluated for quality by the content quality discriminator. If the quality evaluation result does not meet expectations, it is fed back to the multimodal generation module to further optimize the teaching content;

[0034] The teaching strategy discriminator is used to judge whether the optimized teaching content meets the teaching strategy requirements in the preset teaching strategy template. If not, the judgment result is fed back to the multimodal generation module for targeted adjustment.

[0035] Furthermore, the step of using the generative model to generate standardized teaching content adapted to the teaching strategy also includes:

[0036] In the process of teaching content generation, according to the course difficulty coefficient δ, the generation constraints of the large language model are adjusted through the total loss function L_total. The total loss function L_total is:

[0037] L_total=L_content+δ·L_strategy+L_regularization,

[0038] Among them, L_content is the teaching content quality loss, L_strategy is the teaching strategy loss, and L_regularization is the regularization term used to prevent model overfitting;

[0039] By continuously optimizing the total loss function, standardized teaching content that matches the optimal learning path and teaching strategy is eventually generated.

[0040] Furthermore, the step of dynamically optimizing the teaching plan through the neuroplasticity model includes:

[0041] Build a neural plasticity-driven optimizer, including a synaptic weight adjustment module and a cognitive resource allocation module;

[0042] Preliminarily adjusting the teaching parameters according to the feedback from the use of the standardized teaching content through the synaptic weight adjustment module based on the STDP rule;

[0043] Through the attention mechanism in the cognitive resource allocation module, cognitive resources are allocated according to the preliminarily adjusted teaching parameters and the learner's cognitive state reflected by the feature enhancement matrix.

[0044] Furthermore, the step of dynamically optimizing the teaching plan through the neuroplasticity model also includes:

[0045] Design a hierarchical reward function, the formula is:

[0046] R_total=ω_1·Knowledge acquisition rate+ω_2·Cognitive efficiency+ω_3·Learning persistence,

[0047] Among them, ω_1, ω_2, and ω_3 are weight coefficients used to balance the learning objectives of different courses;

[0048] According to the learner's performance under the current teaching plan, the initial reward value R_total is calculated through the hierarchical reward function to reflect the comprehensive effect of the current teaching plan;

[0049] The preliminary reward value R_total is fed back to the synaptic weight adjustment module and the cognitive resource allocation module so that the teaching parameters can be further optimized by the synaptic weight adjustment module, and the cognitive resource allocation module can reallocate cognitive resources based on the further optimized teaching parameters and the latest cognitive state of the learner reflected by the feature enhancement matrix.

[0050] Furthermore, the step of performing teaching adjustment includes:

[0051] The step-by-step control of the difficulty of practice is performed according to the order of knowledge points in the preferred learning path and the knowledge mastery β reflected in the feature enhancement matrix. The adjustment formula is:

[0052] D_{new}=D_{current}×(1+κ(β-θ)),

[0053] Among them, D_{new} is the practice difficulty after adjustment, D_{current} is the practice difficulty before adjustment, κ is the adjustment coefficient, and θ is the target knowledge mastery threshold.

[0054] The present invention also proposes a personalized teaching system based on generative artificial intelligence, which is used to implement the above-mentioned personalized teaching method based on generative artificial intelligence. The system includes:

[0055] Graph construction module: used to construct the knowledge graph and teaching strategy template library of the target course, including a series of preset teaching strategy templates;

[0056] Feature enhancement module: used to enhance features based on learners’ individual learning features and group learning behavior patterns, and construct a feature enhancement matrix;

[0057] Learning path screening module: used to calculate the node activation strength and generate multiple candidate learning paths through the beam search algorithm, and screen out the optimal learning path in combination with the preset teaching strategy template;

[0058] Teaching content generation module: used to generate standardized teaching content adapted to the teaching strategy based on the preferred learning path using a generative model;

[0059] Teaching plan optimization module: used to dynamically optimize the teaching plan through the neuroplasticity model by combining the standardized teaching content and the learner characteristics reflected in the feature enhancement matrix;

[0060] Teaching adjustment module: used to adjust teaching based on the optimized teaching plan.

[0061] In summary, the personalized teaching method based on generative artificial intelligence of the present invention first constructs the knowledge map of the target course and the teaching strategy template library; then based on the individual learning characteristics of the learners, combined with the group learning behavior pattern, the feature enhancement is performed to construct a feature enhancement matrix. The feature enhancement matrix, as a quantitative representation of the learner's characteristics, can comprehensively and accurately reflect the learner's cognitive state and learning needs; then, the node activation strength is calculated, and the priority and direction of the learning path are determined according to the activation strength of each node, and multiple candidate learning paths are generated by the beam search algorithm. Combined with the preset teaching strategy template for scoring, the preferred learning path that meets the learner's needs can be efficiently screened out; then, based on the preferred learning path, the learning path is selected. Path, using generative models to generate standardized teaching content that adapts to teaching strategies. This generation method not only improves the quality and consistency of teaching content, but also enables the teaching content to closely match the learner's learning path and teaching strategy; combined with the standardized teaching content and learner characteristics reflected in the feature enhancement matrix, the teaching plan is dynamically optimized through the neuroplasticity model. This optimization mechanism can adjust teaching parameters and cognitive resource allocation in a timely manner according to the learner's real-time feedback and cognitive state, thereby reasonably allocating limited cognitive resources to the key parts of the teaching content and improving teaching effectiveness; finally, based on the optimized teaching plan, teaching adjustments are made to significantly improve teaching effectiveness and learners' learning efficiency.

[0062] Additional aspects and advantages of the present invention will be set forth in part in the following description and, in part, will be obvious from the following description, or may be learned through embodiments of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments with reference to the accompanying drawings, in which:

[0064] Figure 1 This is a flowchart of a personalized teaching method based on generative artificial intelligence according to the first embodiment of the present invention;

[0065] Figure 2 This is a system block diagram of a personalized teaching system based on generative artificial intelligence according to the second embodiment of the present invention. DETAILED DESCRIPTION

[0066] To facilitate understanding of the present invention, the present invention will be described more fully below with reference to the accompanying drawings. The drawings illustrate several embodiments of the present invention. However, the present invention may be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and comprehensive understanding of the present invention.

[0067] It should be noted that when an element is referred to as being "fixed to" another element, it may be directly on the other element or there may be an intermediate element. When an element is referred to as being "connected to" another element, it may be directly connected to the other element or there may be an intermediate element. The terms "vertical," "horizontal," "left," "right," and similar expressions used herein are for illustrative purposes only.

[0068] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this invention pertains. The terms used in this specification of the present invention are for the purpose of describing specific embodiments only and are not intended to limit the present invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0069] Example 1

[0070] See also Figure 1 The present invention proposes a personalized teaching method based on generative artificial intelligence, which includes steps S100 to S600:

[0071] S100, builds the knowledge map and teaching strategy template library of the target course, including a series of preset teaching strategy templates.

[0072] Specifically, the knowledge graph of the target course uses various knowledge points as nodes and the relationships between knowledge points as edges, thus forming a complex network structure of the course knowledge point system, which is used to clearly display the relationship and hierarchical structure of various knowledge points in the course.

[0073] When building a knowledge graph for a target course, first collect all relevant knowledge points from the course's syllabus, textbooks, and teaching materials. For example, in a mathematics course, these knowledge points might include the concept of function, the properties of linear functions, and the graph of quadratic functions. The logical relationships between these knowledge points are analyzed. For example, before learning quadratic functions, one must first master linear functions. Function is a broad concept, and linear and quadratic functions are subconcepts of functions. Then, use the NetworkX library in Python to construct a knowledge graph, treating knowledge points as nodes and relationships between them as edges to build the knowledge graph for the target course.

[0074] At the same time, a teaching strategy template library was constructed. First, various common teaching strategies were collected, such as the "step-by-step" strategy (suitable for gradually deepening knowledge from the basics), the "focus" strategy (emphasizing the teaching of key knowledge points), and the "focus on practice" strategy (deepening understanding through actual practice).

[0075] Design a detailed template for each teaching strategy, including the name of the teaching strategy, applicable knowledge point types (e.g., conceptual knowledge, procedural knowledge), applicable learner characteristics (e.g., learning style, knowledge base), teaching steps (e.g., introduction, explanation, practice, summary), and teaching resource requirements (e.g., textbooks, teaching aids, multimedia resources). Store the designed teaching strategy templates in a database for easy subsequent query and access.

[0076] S200, based on the learner's individual learning characteristics and combined with the group learning behavior pattern, performs feature enhancement and constructs a feature enhancement matrix.

[0077] Further optionally, in step S200, the step of constructing the feature enhancement matrix includes steps S210 to S240:

[0078] S210, collecting common learning characteristic data of individuals and groups, including learning behavior data, cognitive performance data and learning situation data.

[0079] Specifically, we collect learning behavior data such as learners' answer records, learning time, interaction frequency, and operation behavior, as well as cognitive performance data such as cognitive level test results and knowledge mastery assessment, and learning situation data such as learning time and learning environment.

[0080] The collected data are preprocessed by cleaning, normalizing, and other operations to remove noise and outliers, and convert the data into a format suitable for subsequent analysis.

[0081] S220, fusing the individual learning features with the group common feature data to obtain fused feature data.

[0082] Specifically, the similarity between individual learning features and group average features can be calculated, and the individual learning features can be mapped in the group learning feature space to achieve the fusion of individual learning features and group common feature data to obtain fused feature data.

[0083] S230, extract common learning patterns and explore common features among learners.

[0084] Specifically, clustering algorithms can be used to analyze the learning behaviors and cognitive performances of different learners on the same knowledge points, find out common learning rules and patterns, and thus explore the common characteristics among learners.

[0085] S240: Based on the fused feature data and the common learning model, construct a feature enhancement matrix F_enhanced, which is:

[0086] F_enhanced=a·F_individual+b·F_group+c·ΔF_temporal,

[0087] Among them, a, b, and c are the weight coefficients used to adjust the individual learning feature F_individual, the group learning feature F_group, and the influence of the feature evolution over time, respectively. ΔF_temporal reflects the evolution trend of the learning feature over time.

[0088] Specifically, by constructing a feature enhancement matrix, a more accurate representation of learner characteristics can be obtained. This feature enhancement matrix comprehensively considers individual learning characteristics, group learning characteristics, and the trend of characteristic evolution over time. Specifically, it includes information on multiple dimensions, including learning behavior characteristics (such as answering, learning progress, and interactive behavior), cognitive state characteristics (such as knowledge mastery and cognitive ability level), and learning context characteristics (such as learning time and learning environment). Cognitive ability level can be assessed based on dimensions such as attention concentration, memory retention, and comprehension speed.

[0089] S300, mapping the feature enhancement matrix to the node set of the knowledge graph, calculating the node activation strength, and generating multiple candidate learning paths through a beam search algorithm, combining the preset teaching strategy template to score the paths and screen out the preferred learning paths.

[0090] Further optionally, in step S300, the step of screening out the preferred learning path includes steps S310 to S340:

[0091] S310, mapping the feature enhancement matrix to the node set of the knowledge graph, and determining the initial position of the learner in the knowledge graph to clarify the starting point of learning.

[0092] Specifically, the learner's feature information in the feature enhancement matrix is ​​mapped to a set of nodes in the knowledge graph. That is, the learner's feature information is projected onto the course's knowledge structure diagram to determine the learner's position in the current knowledge structure. First, the initial position is determined. Based on the learner's feature information and the structure of the knowledge graph, the learner's initial position in the knowledge graph is determined. This position represents the learner's current knowledge level or the starting point of learning. For example, if the learner has mastered the knowledge point "the concept of function," then their initial position may be on the "function" node, or more specifically, on a subnode representing "mastery of the function concept." Once the learner's initial position in the knowledge graph is determined, their learning starting point is also clarified. This starting point will serve as the starting point for the subsequent generation of the learning path.

[0093] S320, calculate the node activation strength, the formula is:

[0094] h_v (l+1) =σ(∑{u∈N(v)}α_{vu}Wh_ul),

[0095] Among them, h_v (l+1) is the activation strength of node v in the l+1 layer, N(v) is the set of neighbor nodes of node v, α_{vu} is the attention coefficient, which is used to measure the importance of node u to node v, W is the weight matrix, h_ul is the eigenvector of node u in the l layer, and σ is the activation function.

[0096] Specifically, in a graph neural network, each node has an associated feature vector that describes the node’s attributes or state. A graph neural network consists of multiple layers, each of which updates and transforms the node’s features.

[0097] The node activation intensity reflects the importance of each node in the current learning task. In the knowledge graph, knowledge points (i.e., nodes) with high activation intensity may be more relevant to the learner's current learning needs.

[0098] In the process of calculating the node activation strength through the graph neural network, for the target node v, its activation strength h_v at the l+1 layer is calculated. (l+1)First, traverse all neighbor nodes u of node v (u∈N(v), where N(v) represents the set of neighbor nodes of node v); for each neighbor node u, calculate its attention coefficient α_{vu}, which is used to measure the importance of node u to node v; then, multiply the feature vector h_ul of neighbor node u in the lth layer by the weight matrix W to obtain the transformed feature vector; then, multiply the transformed feature vector by the attention coefficient α_{vu} and sum the results of all neighbor nodes; finally, perform a nonlinear transformation on the sum result through the activation function σ to obtain the activation strength h_v of node v in the l+1 layer (l+1) .

[0099] S330 : Based on the node activation strength, generate K candidate paths using a beam search algorithm.

[0100] Specifically, based on the node activation strength, a beam search algorithm generates K candidate paths. This algorithm searches for possible paths from the starting node (i.e., initial position) to the target node in the knowledge graph. The beam search algorithm retains only the K most promising candidate paths at each level, reducing the search space and improving search efficiency. A candidate path represents the sequence in which a learner progresses from one knowledge point to another.

[0101] When screening candidate paths, first, determine the starting node, which is the initial position of the learner in the knowledge graph determined above, that is, the starting point of learning; initialize an empty candidate path set to store the candidate paths of the current layer; then, add the starting node as a separate path to the candidate path set; for each candidate path of the current layer, find the last node corresponding to the path; then, according to the node activation strength, select K neighboring nodes connected to the node and with higher activation strength, and sort them according to the node activation strength to select the top K nodes; for each selected neighboring node, add it to the end of the current candidate path to form a new path; and add all newly generated paths to the candidate path set of the next layer; if the number of candidate paths in the next layer exceeds K, the candidate paths are sorted according to the total activation strength of the path, and the top K paths after sorting are retained, and the remaining paths are discarded; repeat the above iterative expansion and path screening process until the iteration termination condition is met (such as finding the target node); finally, select K paths from the candidate path set of the last layer as the final candidate path output.

[0102] S340 , scoring the candidate paths in combination with the preset teaching strategy template, and selecting the path with the highest score as the preferred learning path.

[0103] Specifically, the screened candidate paths are scored in combination with the preset teaching strategy template, and the path with the highest score is selected as the preferred learning path. At the same time, the preferred learning path information including the knowledge point topological sequence, cognitive load index and media adaptation suggestions is output to select the learning path that best meets the teaching requirements and learner characteristics from multiple candidate paths, providing accurate guidance for learning.

[0104] After using the beam search algorithm to generate K possible candidate learning paths, each path consists of a series of knowledge points, reflecting different learning sequences and content combinations. Based on the rules in the pre-set teaching strategy template, the candidate paths are quantitatively evaluated for various indicators to determine their quality. The path with the highest score is then selected as the preferred learning path.

[0105] Specifically, first, according to the preset teaching strategy template, determine the key indicators for scoring, such as knowledge point coverage, rationality of learning sequence, suitability of difficulty gradient, and consistency with teaching objectives. And assign corresponding weights to each indicator. For example, if the teaching objective emphasizes the systematic nature of knowledge, then the weights of knowledge point coverage and rationality of learning sequence are set higher. Then, for each candidate path, calculate its score on each scoring indicator separately. For example, when calculating the knowledge point coverage score, the ratio of the number of key knowledge points contained in the statistical path to the total number of key knowledge points is calculated; when calculating the rationality of learning sequence score, the logical relationship and degree of dependence between knowledge points are evaluated. Then, multiply the score of each indicator by the corresponding weight and sum it up to obtain the total score of each candidate path, and select the path with the highest score as the preferred learning path.

[0106] S400: Based on the preferred learning path, using a generative model to generate standardized teaching content adapted to the teaching strategy.

[0107] Further optionally, in step S400, the step of using the generative model to generate standardized teaching content adapted to the teaching strategy includes steps S410 to S450:

[0108] S410, constructing a generative model including a semantic parsing module, a multimodal generation module, and a cognitive load assessment module.

[0109] Specifically, BERT (Bidirectional Encoder Representations from Transformers) was chosen as the core of the semantic parsing module. BERT is a pre-trained language model that performs deep semantic analysis on input text and extracts key teaching requirements information. Integrating BERT into the generative model enables it to process text input from the preferred learning path and generate structured data on teaching requirements.

[0110] The collaborative mechanism of CLIP (Contrastive Language-Image Pre-training) and DALL·E is integrated. CLIP is used to understand the relationship between images and text, while DALL·E generates corresponding image content based on text descriptions. A multimodal generation module is built into the generative model, enabling it to generate multimodal teaching content, such as illustrated teaching materials or video scripts, based on the structured teaching data output by the semantic parsing module.

[0111] We built a cognitive load assessment module based on an LSTM network. LSTM networks excel at processing sequential data and can assess the cognitive load of generated instructional content. By integrating this module into the generative model, we enabled it to evaluate the generated instructional content in real time and adjust the difficulty and complexity of the instructional content based on the assessment results.

[0112] The semantic parsing module, multimodal generation module, and cognitive load assessment module are seamlessly integrated into the generative model to form a complete modular generator. The model parameters are then optimized to ensure that the various modules work together to generate high-quality teaching content that meets teaching needs.

[0113] S420: Understanding the teaching requirements in the preferred learning path through the BERT architecture in the semantic parsing module, and generating teaching requirement structured data.

[0114] Specifically, the selected optimal learning path is passed as input to the semantic parsing module. The optimal learning path contains a series of knowledge points and their topological sequence, reflecting the order and logical relationships of the content that the learner needs to learn. Next, the text information in the optimal learning path is preprocessed, including removing irrelevant characters, word segmentation, and part-of-speech tagging.

[0115] The preprocessed text is input into the BERT model, a pretrained language model based on the Transformer architecture with powerful semantic understanding capabilities. Leveraging its deep semantic understanding capabilities, the BERT model performs in-depth semantic analysis of the text, capturing contextual information and inter-lexical dependencies within the text. It then extracts key teaching requirements information from the optimal learning path and generates structured teaching requirements data. This includes structured information such as a list of knowledge points, explanation methods, and media formats. Structured data consists of clear fields and values. For example, a structured data table is generated containing fields such as "Knowledge Point ID," "Explanation Method," and "Media Format," with each field corresponding to a specific value.

[0116] S430 : Based on the teaching requirement structured data, initial teaching content is generated through the collaborative work of CLIP and DALL·E in the multimodal generation module.

[0117] Specifically, CLIP (Contrastive Language-Image Pre-training) is a multimodal model used to learn the associations between images and text. Through contrastive learning, it maps images and text into the same high-dimensional space, placing similar images and text close together and dissimilar ones further apart. The CLIP model can retrieve related images based on text descriptions or generate corresponding text descriptions based on images.

[0118] DALL·E is a generative model based on the Transformer architecture that can generate high-quality images that closely match text descriptions. It consists of an encoder and a decoder, where the encoder processes the text description and converts it into a representation in the latent space; the decoder generates the corresponding image based on this representation.

[0119] The system receives structured instructional requirements data from the semantic parsing module and extracts key instructional requirements information, such as a detailed description of each knowledge point, the desired media format (e.g., image, video, audio), and specific instructional requirements (e.g., emphasizing a specific concept or demonstrating a specific process). It then loads a pre-trained CLIP model and uses CLIP's image encoder to encode a large number of images related to function images and map them into a high-dimensional feature space, where similar images are clustered together and dissimilar images are separated. Based on the knowledge point descriptions in the instructional requirements structured data, it generates corresponding text descriptions and uses CLIP's text encoder to map these descriptions into the same high-dimensional feature space as the images. In the CLIP feature space, it calculates the similarity between the feature vectors of the text descriptions and the feature vectors of all function images, and selects the function image with the highest similarity to the text description feature vector as the candidate image. If the CLIP model fails to find a function image that fully meets the instructional requirements, or if a function image with a specific style, angle, or detail is required, the DALL·E model is invoked and the text description is input into the DALL·E model to generate a series of candidate function images. The CLIP model then filters the candidate function images generated by the DALL·E model, selecting the image that best matches the text description and teaching requirements. Finally, the selected function images are integrated with the text description to form multimodal teaching content, which serves as the initial teaching content. If the teaching requirements include other media formats (such as video demonstrations of function transformations or audio explanations of function properties), these should be processed and integrated accordingly.

[0120] The following example is used to illustrate. It is assumed that the teaching requirement structured data contains a function "f(x)=x^2", and the expected media form is parsed as the function graph. The specific teaching requirement is to emphasize the monotonically increasing property and minimum point of the function.

[0121] Load the pre-trained CLIP model and encode thousands of images showing different functions (including quadratic functions) to form a feature space. In this feature space, similar function images (such as images showing quadratic function images) will be clustered together, while dissimilar images (such as images showing other types of function images) will be separated.

[0122] Then, based on the teaching requirements, the function description in the structured data is generated to describe the corresponding text: "The graph of the function f(x) = x^2, showing its monotonically increasing property from left to right, with the minimum point marked." In the feature space, the cosine similarity between the feature vector of the text description and the feature vectors of all images in the function image feature space is calculated. The quadratic function image with the highest similarity is selected as the candidate image. This image should clearly demonstrate the monotonically increasing property and the minimum point of the function.

[0123] Next, the text description "Graph of the function f(x) = x^2, showing its monotonically increasing property from left to right, and marking the minimum point" was input into the DALL·E model. This generated a series of candidate function graphs, and the CLIP model was used to select the one that best met the requirements.

[0124] Finally, combine the selected quadratic function graph with the text description: "The graph of the function f(x) = x^2 shows its monotonically increasing property from left to right, with the minimum point marked." If desired, you can also add a video demonstration (e.g., showing how the function graph changes with the value of x) or an audio commentary (e.g., a detailed explanation of the function's properties).

[0125] S440: Perform cognitive load evaluation on the initial teaching content through the LSTM network in the cognitive load evaluation module to obtain a cognitive load evaluation result.

[0126] Specifically, the LSTM network is a special recurrent neural network (RNN) that can process long-term dependencies in sequence data. In this embodiment, the LSTM network is used to analyze factors such as the structure, difficulty distribution, and information presentation method of the initial teaching content.

[0127] The initial teaching content text is input into the trained LSTM network model. The model calculates its cognitive load assessment based on the input text. The cognitive load assessment result can be a numerical value (indicating the degree of cognitive load) or a classification label (such as low load, medium load, and high load). Cognitive load refers to the amount of mental resources required by learners to process information. Excessive cognitive load may cause learner fatigue and reduced learning outcomes.

[0128] When training the LSTM network model, a large amount of labeled cognitive load data is collected as a training set. This data includes the initial instructional content and its corresponding cognitive load assessment results. This training data is preprocessed and then fed into the LSTM network model. The model is trained using a loss function and optimization algorithm, and the network structure is adjusted during training based on performance feedback.

[0129] S450: According to the cognitive load assessment result, the initial teaching content is adjusted by the multimodal generation module to obtain adjusted teaching content.

[0130] Specifically, based on the cognitive load assessment results, the multimodal generation module adjusts the initial teaching content. If the cognitive load is too high, the content can be simplified, examples can be added, or a more intuitive presentation can be adopted. If the cognitive load is too low, the depth and breadth of the content can be increased. The adjusted teaching content serves as input to the content quality discriminator.

[0131] Further optionally, in step S400, the step of using the generative model to generate standardized teaching content adapted to the teaching strategy includes steps S460 to S480:

[0132] S460, design a dual discriminator architecture, including content quality discriminator and teaching strategy discriminator.

[0133] A dual-discriminator architecture is constructed to evaluate the quality of teaching content and the conformance of teaching strategies. Specifically, for content quality assessment, features related to content quality, such as grammatical correctness, logical coherence, and knowledge accuracy, are extracted, and corresponding evaluation metrics are defined. For teaching strategy conformance assessment, features related to teaching strategies, such as content organization, emphasis on key knowledge points, and practical sessions, are extracted, and corresponding judgment criteria are defined. Then, based on the results of feature extraction and metric definition, the model structures of the content quality discriminator and the teaching strategy discriminator are constructed. Machine learning algorithms, deep learning algorithms, etc. can be selected to implement the discriminator functions. Next, the discriminator model is trained using a labeled dataset, and model parameters are adjusted to optimize performance. The trained model is then evaluated to ensure that it can accurately assess the quality of teaching content and conformance to teaching strategies.

[0134] Finally, the trained content quality discriminator and teaching strategy discriminator are integrated into the framework of the generative model, and the interface and process are designed to ensure that the two discriminators can work together and seamlessly connect with other modules of the generative model.

[0135] S470: Performing a quality assessment on the adjusted teaching content using the content quality discriminator to obtain a quality assessment result. If the quality assessment result does not meet expectations, feeding the quality assessment result back to the multimodal generation module to further optimize the adjusted teaching content.

[0136] Specifically, the teaching content adjusted by the multimodal generation module is input into the content quality discriminator, which then evaluates the quality of the teaching content, such as grammatical correctness, logical coherence, and knowledge accuracy. The quality assessment results can be output as a score or evaluation report. If the quality assessment results do not meet expectations, such as high-quality teaching standards, a feedback mechanism is triggered, and the evaluation results are fed back to the multimodal generation module to further optimize the adjusted teaching content. Problems identified in the evaluation report, such as incorrect knowledge points, unclear logic, and unsmooth presentation, are then modified and improved accordingly.

[0137] S480: The teaching strategy discriminator determines whether the optimized teaching content meets the teaching strategy requirements in the preset teaching strategy template. If not, the judgment result is fed back to the multimodal generation module for targeted adjustment.

[0138] Specifically, the teaching content after content quality optimization is input into the teaching strategy discriminator, and based on the preset teaching strategy template, it is judged whether the optimized teaching content meets the preset teaching strategy requirements. If the optimized teaching content does not match the preset teaching strategy requirements, the feedback mechanism is triggered, and the judgment result is fed back to the multimodal generation module for targeted adjustment. The present invention ensures that the teaching content meets high standards in terms of quality and strategy matching through a dual discriminator architecture, and generates higher-quality teaching content through continuous feedback and optimization.

[0139] Further optionally, in step S400, the step of using the generative model to generate standardized teaching content adapted to the teaching strategy further includes steps S401 to S402:

[0140] S401, during the teaching content generation process, according to the course difficulty coefficient δ, the generation constraints of the large language model are adjusted through the total loss function L_total. The total loss function L_total is:

[0141] L_total=L_content+δ·L_strategy+L_regularization,

[0142] Among them, L_content is the teaching content quality loss, which is determined based on the quality evaluation result of the content quality discriminator; L_strategy is the teaching strategy loss, which is determined based on the judgment result of the teaching strategy discriminator; L_regularization is the regularization term used to prevent the model from overfitting.

[0143] S402, by continuously optimizing the total loss function, the quality of teaching content, strategy matching and difficulty moderation are optimally balanced, and finally standardized teaching content that matches the preferred learning path and teaching strategy is generated.

[0144] Specifically, calculate each loss, calculate the content quality loss L_content based on the quality evaluation result of the content quality discriminator; calculate the teaching strategy loss L_strategy based on the judgment result of the teaching strategy discriminator; determine the regularization term L_regularization to prevent the model from overfitting. Then calculate the total loss function value according to the total loss function formula. By continuously optimizing the total loss function and adjusting the generation constraints of the large language model, the quality of the teaching content, the strategy matching degree and the difficulty moderation are optimally balanced. The total loss function adjustment mechanism of this embodiment can ensure that the generated teaching content can meet the requirements in multiple dimensions, and ultimately generate standardized teaching content that matches the preferred learning path and teaching strategy.

[0145] S500, dynamically optimizing the teaching plan through a neuroplasticity model based on the standardized teaching content and learner characteristics reflected in the feature enhancement matrix;

[0146] Further optionally, in step S500, the step of dynamically optimizing the teaching plan through the neuroplasticity model includes steps S510 to S530:

[0147] S510, constructing a neural plasticity-driven optimizer, including a synaptic weight adjustment module and a cognitive resource allocation module.

[0148] Specifically, we designed a neuroplasticity-driven optimizer architecture consisting of two core modules: a synaptic weight adjustment module and a cognitive resource allocation module. We developed the synaptic weight adjustment module, which adjusts instructional parameters (such as teaching speed and difficulty) based on the principles of neuroplasticity. We also developed a cognitive resource allocation module, which allocates instructional resources (such as attention and time) based on the learner's cognitive state. We then integrated the synaptic weight adjustment module and cognitive resource allocation module into the optimizer.

[0149] The neuroplasticity-driven optimizer can be used to analogize the teaching system to the nervous system, mapping various elements of the teaching process (such as teaching content, teaching methods, and learner characteristics) to the components of the nervous system. Based on feedback from the teaching process, the optimizer dynamically adjusts the parameters and structure of the teaching system to improve teaching effectiveness and learner experience, just as the nervous system adapts to environmental changes through plasticity.

[0150] S520 , preliminarily adjusting the teaching parameters according to the usage feedback of the standardized teaching content through the synaptic weight adjustment module based on the STDP rule.

[0151] Specifically, standardized teaching content generates various user feedback during the actual teaching process, such as learners' accuracy rate, learning progress, and interaction frequency. The system extracts the characteristics of this feedback information and defines a mapping rule that maps the different levels or categories of feedback characteristics to the time difference ranges in the STDP rule. Based on the mapping results, the corresponding weight changes are calculated to determine the direction and magnitude of adjustment of teaching parameters. Based on the calculated weight changes, teaching parameters are adjusted, including teaching speed, difficulty level of teaching content, and the order of explanation of knowledge points. By adjusting these parameters, the teaching plan can be adaptively adjusted based on learner feedback.

[0152] S530 , allocating cognitive resources according to the initially adjusted teaching parameters and the learner's cognitive state reflected by the feature enhancement matrix through the attention mechanism in the cognitive resource allocation module.

[0153] Specifically, learners' cognitive state characteristics are extracted from the feature enhancement matrix. These characteristics may include learners' concentration, memory retention, comprehension speed, etc.; the preliminarily adjusted teaching parameters and learners' cognitive state characteristics are input into the attention mechanism model, and the attention weight of each teaching resource (such as knowledge points, exercises, explanation videos, etc.) is calculated through the model; according to the calculated attention weight, cognitive resources (such as time, attention, teaching resources, etc.) are allocated to different teaching contents or teaching links, and an effective match between learners' cognitive state characteristics and teaching resources is ensured to improve learners' learning efficiency and effectiveness; according to learners' real-time feedback (such as academic performance, learning speed, satisfaction, etc.) and changes in cognitive state, the allocation of cognitive resources is dynamically adjusted, so that limited cognitive resources are reasonably allocated to the key parts of the teaching content.

[0154] Further optionally, in step S500, the step of dynamically optimizing the teaching plan through the neuroplasticity model further includes steps S501 to S503:

[0155] S501, design a hierarchical reward function, the formula is:

[0156] R_total=ω_1·Knowledge acquisition rate+ω_2·Cognitive efficiency+ω_3·Learning persistence,

[0157] Among them, the knowledge acquisition rate is used to reflect the learner's mastery of knowledge, the cognitive efficiency is used to reflect the learner's resource utilization efficiency in the learning process, and the learning continuity is used to measure the learner's ability to maintain the learning state. ω_1, ω_2, and ω_3 are weight coefficients used to balance the learning objectives of different courses.

[0158] S502 , according to the learner's performance under the current teaching plan, a preliminary reward value R_total is calculated by the hierarchical reward function to reflect the comprehensive effect of the current teaching plan.

[0159] Specifically, identify indicators for evaluating the overall effectiveness of the current teaching plan, including knowledge acquisition rate, cognitive efficiency, and learning persistence. Based on the learning objectives of different courses, determine the weight coefficients ω_1, ω_2, and ω_3 for knowledge acquisition rate, cognitive efficiency, and learning persistence to balance the importance of different indicators. Construct a hierarchical reward function based on knowledge acquisition rate, cognitive efficiency, and learning persistence. Based on learner performance under the current teaching plan, calculate a preliminary reward value, R_total, to reflect the overall effectiveness of the current teaching plan.

[0160] S503, feeding back the preliminary reward value R_total to the synaptic weight adjustment module and the cognitive resource allocation module, so as to further optimize the teaching parameters through the synaptic weight adjustment module, and reallocate cognitive resources through the cognitive resource allocation module according to the further optimized teaching parameters and the latest cognitive state of the learner reflected by the feature enhancement matrix.

[0161] Specifically, if the initial reward value is low, it indicates that the current teaching plan has some problems. The synaptic weight adjustment module will analyze whether the shortcomings are in knowledge acquisition rate, cognitive efficiency, or learning continuity. It will then adjust the teaching parameters accordingly, such as adding practice sessions to improve knowledge acquisition rate, optimizing the teaching content presentation to improve cognitive efficiency, etc. The optimized teaching parameters will serve as input to the cognitive resource allocation module again.

[0162] The cognitive resource allocation module reallocates cognitive resources based on the teaching parameters optimized by the synaptic weight adjustment module and the learner's latest cognitive state as reflected by the feature enhancement matrix. This allocation more accurately meets the learner's needs and further improves learning efficiency. The teaching plan after reallocation of cognitive resources serves as input for the hierarchical reward function evaluation.

[0163] The hierarchical reward function plays a role in global evaluation and guidance in the optimization process of the entire teaching plan. It adaptively adjusts the weight coefficient according to the teaching objectives of the course type and motivates the teaching plan to develop in a better direction by calculating the reward value.

[0164] S600, performs teaching adjustments based on the optimized teaching plan.

[0165] Further optionally, in step S600, the step of dynamically switching the media type includes:

[0166] When it is detected based on the feature enhancement matrix that the attention entropy value exceeds the target entropy value threshold, it indicates that the learner's attention is distracted, and a media conversion strategy is triggered according to the media adaptation suggestion in the standardized teaching content.

[0167] Specifically, attention entropy is used to measure a learner's level of concentration. Using the feature enhancement matrix, the system can monitor the learner's attention distribution in real time during the learning process. When the attention entropy exceeds the target entropy threshold, it means that the learner's attention is no longer focused on the current learning content and is becoming dispersed. This detection mechanism helps the system promptly detect changes in the learner's attention state, providing a basis for teaching adjustments.

[0168] Standardized teaching content includes media adaptation suggestions. For example, for content with a strong theoretical nature, it may be recommended to use a document format with both text and pictures; for content that requires intuitive presentation, it may be recommended to use a video format.

[0169] When the attention entropy value exceeds a threshold, the system triggers a corresponding media conversion strategy based on media adaptation recommendations. For example, if the current instructional content is pure text, it may be converted to multimedia content containing images, charts, or animations. By dynamically switching media types, learners' attention can be attracted, refocused on the learning content, and improved learning outcomes.

[0170] Further optionally, in step S600, the step of performing teaching adjustment includes:

[0171] The step-by-step control of the difficulty of practice is performed according to the order of knowledge points in the preferred learning path and the knowledge mastery β reflected in the feature enhancement matrix. The adjustment formula is:

[0172] D_{new}=D_{current}×(1+κ(β-θ)),

[0173] Among them, D_{new} is the practice difficulty after adjustment, D_{current} is the practice difficulty before adjustment, κ is the adjustment coefficient, and θ is the target knowledge mastery threshold.

[0174] Specifically, the preferred learning path clearly defines the order in which knowledge points should be learned. Different knowledge points have different levels of difficulty and importance. Practicing in a reasonable learning order helps learners gradually master the knowledge.

[0175] The feature enhancement matrix can reflect the learner's mastery of each knowledge point, which is represented by β. The larger the β value, the better the learner's mastery of the knowledge point.

[0176] By adjusting the formula D_{new}=D_{current}×(1+κ(β-θ)), and controlling the difficulty of practice in steps according to the order of knowledge points in the optimal learning path and the knowledge mastery degree β reflected in the feature enhancement matrix,

[0177] In the adjustment formula, D_{new} is the adjusted exercise difficulty, calculated based on the learner's knowledge mastery and the learning path requirements. D_{current} is the pre-adjusted exercise difficulty, that is, the difficulty of the current exercise the learner is currently performing. κ is the adjustment coefficient, which controls the magnitude of the exercise difficulty adjustment. A larger κ value results in a larger adjustment, while a smaller κ value results in a smaller adjustment. θ is the target knowledge mastery threshold, a pre-set standard value used to determine whether the learner has mastered the knowledge point to the expected level.

[0178] During the adjustment process, when β>θ, that is, the learner's mastery of the knowledge point exceeds the target threshold, and at this time (β-θ)>0, the exercise difficulty is adjusted according to the adjustment formula. The exercise difficulty after adjustment will be greater than the exercise difficulty before adjustment, making the exercise difficulty increased and providing learners with more challenging exercises.

[0179] When β<θ, meaning the learner's mastery of the knowledge point has not reached the target threshold, and (β-θ)<0, the exercise difficulty is adjusted according to the adjustment formula. The adjusted exercise difficulty will be lower than the original difficulty, thus reducing the learner's sense of frustration due to excessive difficulty.

[0180] This embodiment ensures that the difficulty of practice matches the ability of the learner through this step-by-step control of the practice difficulty adjustment method, so that the learner can get appropriate training at different learning stages and promote the effective mastery of knowledge.

[0181] In summary, the personalized teaching method based on generative artificial intelligence of the present invention first constructs the knowledge map of the target course and the teaching strategy template library; then based on the individual learning characteristics of the learners, combined with the group learning behavior pattern, the feature enhancement is performed to construct a feature enhancement matrix. The feature enhancement matrix, as a quantitative representation of the learner's characteristics, can comprehensively and accurately reflect the learner's cognitive state and learning needs; then, the node activation strength is calculated, and the priority and direction of the learning path are determined according to the activation strength of each node, and multiple candidate learning paths are generated by the beam search algorithm. Combined with the preset teaching strategy template for scoring, the preferred learning path that meets the learner's needs can be efficiently screened out; then, based on the preferred learning path, the learning path is selected. Path, using generative models to generate standardized teaching content that adapts to teaching strategies. This generation method not only improves the quality and consistency of teaching content, but also enables the teaching content to closely match the learner's learning path and teaching strategy; combined with the standardized teaching content and learner characteristics reflected in the feature enhancement matrix, the teaching plan is dynamically optimized through the neuroplasticity model. This optimization mechanism can adjust teaching parameters and cognitive resource allocation in a timely manner according to the learner's real-time feedback and cognitive state, thereby reasonably allocating limited cognitive resources to the key parts of the teaching content and improving teaching effectiveness; finally, based on the optimized teaching plan, teaching adjustments are made to significantly improve teaching effectiveness and learners' learning efficiency.

[0182] Example 2

[0183] See also Figure 2 The present invention proposes a personalized teaching system based on generative artificial intelligence, which includes:

[0184] Graph construction module: used to construct the knowledge graph and teaching strategy template library of the target course, including a series of preset teaching strategy templates;

[0185] Feature enhancement module: used to enhance features based on learners’ individual learning features and group learning behavior patterns, and construct a feature enhancement matrix;

[0186] Learning path screening module: used to calculate the node activation strength and generate multiple candidate learning paths through the beam search algorithm, and screen out the optimal learning path in combination with the preset teaching strategy template;

[0187] Teaching content generation module: used to generate standardized teaching content adapted to the teaching strategy based on the preferred learning path using a generative model;

[0188] Teaching plan optimization module: used to dynamically optimize the teaching plan through the neuroplasticity model by combining the standardized teaching content and the learner characteristics reflected in the feature enhancement matrix;

[0189] Teaching adjustment module: used to adjust teaching based on the optimized teaching plan.

[0190] Further optionally, the feature enhancement module is further configured to:

[0191] Collect learning characteristic data common to individuals and groups, including learning behavior data, cognitive performance data, and learning context data;

[0192] Fusing individual learning feature data with group common feature data to obtain fused feature data;

[0193] Extract common learning patterns and explore common characteristics among learners;

[0194] Based on the fusion feature data and common learning model, a feature enhancement matrix F_enhanced is constructed, and the formula is:

[0195] F_enhanced=a·F_individual+b·F_group+c·ΔF_temporal,

[0196] Among them, a, b, and c are the weight coefficients used to adjust the individual learning feature F_individual, the group learning feature F_group, and the influence of the feature evolution over time, respectively. ΔF_temporal reflects the evolution trend of the learning feature over time.

[0197] Further optionally, the learning path screening module is further configured to:

[0198] Mapping the feature enhancement matrix to a set of nodes in the knowledge graph and determining the learner's initial position in the knowledge graph to clarify the starting point of learning;

[0199] Calculate the node activation strength using the formula:

[0200] h_v (l+1) =σ(∑{u∈N(v)}α_{vu}Wh_ul),

[0201] Among them, h_v (l+1) is the activation strength of node v in the l+1 layer, N(v) is the set of neighbor nodes of node v, α_{vu} is the attention coefficient, which is used to measure the importance of node u to node v, W is the weight matrix, h_ul is the eigenvector of node u in the l layer, and σ is the activation function;

[0202] Based on the node activation strength, K candidate paths are generated by a beam search algorithm;

[0203] The candidate paths are scored in combination with the preset teaching strategy template, and the path with the highest score is selected as the preferred learning path.

[0204] Further optionally, the teaching content generation module is further used to:

[0205] Build a generative model that includes semantic parsing, multimodal generation, and cognitive load assessment modules;

[0206] Understanding the teaching requirements in the preferred learning path through the BERT architecture in the semantic parsing module and generating structured data of the teaching requirements;

[0207] Based on the teaching demand structured data, initial teaching content is generated through the collaborative work of CLIP and DALL·E in the multimodal generation module;

[0208] Performing cognitive load evaluation on the initial teaching content through the LSTM network in the cognitive load evaluation module to obtain a cognitive load evaluation result;

[0209] According to the cognitive load assessment result, the initial teaching content is adjusted by the multimodal generation module to obtain adjusted teaching content.

[0210] Further optionally, the teaching content generation module is further used to:

[0211] Design a dual discriminator architecture, including a content quality discriminator and a teaching strategy discriminator;

[0212] The adjusted teaching content is evaluated for quality by the content quality discriminator. If the quality evaluation result does not meet expectations, it is fed back to the multimodal generation module to further optimize the teaching content;

[0213] The teaching strategy discriminator is used to judge whether the optimized teaching content meets the teaching strategy requirements in the preset teaching strategy template. If not, the judgment result is fed back to the multimodal generation module for targeted adjustment.

[0214] Further optionally, the teaching content generation module is further used to:

[0215] In the process of teaching content generation, according to the course difficulty coefficient δ, the generation constraints of the large language model are adjusted through the total loss function L_total. The total loss function L_total is:

[0216] L_total=L_content+δ·L_strategy+L_regularization,

[0217] Among them, L_content is the teaching content quality loss, L_strategy is the teaching strategy loss, and L_regularization is the regularization term used to prevent model overfitting;

[0218] By continuously optimizing the total loss function, standardized teaching content that matches the optimal learning path and teaching strategy is eventually generated.

[0219] Further optionally, the teaching plan optimization module is further used to:

[0220] Build a neural plasticity-driven optimizer, including a synaptic weight adjustment module and a cognitive resource allocation module;

[0221] Preliminarily adjusting the teaching parameters according to the feedback from the use of the standardized teaching content through the synaptic weight adjustment module based on the STDP rule;

[0222] Through the attention mechanism in the cognitive resource allocation module, cognitive resources are allocated according to the preliminarily adjusted teaching parameters and the learner's cognitive state reflected by the feature enhancement matrix.

[0223] Further optionally, the teaching plan optimization module is further used to:

[0224] Design a hierarchical reward function, the formula is:

[0225] R_total=ω_1·Knowledge acquisition rate+ω_2·Cognitive efficiency+ω_3·Learning persistence,

[0226] Among them, ω_1, ω_2, and ω_3 are weight coefficients used to balance the learning objectives of different courses;

[0227] According to the learner's performance under the current teaching plan, the initial reward value R_total is calculated through the hierarchical reward function to reflect the comprehensive effect of the current teaching plan;

[0228] The preliminary reward value R_total is fed back to the synaptic weight adjustment module and the cognitive resource allocation module so that the teaching parameters can be further optimized by the synaptic weight adjustment module, and the cognitive resource allocation module can reallocate cognitive resources based on the further optimized teaching parameters and the latest cognitive state of the learner reflected by the feature enhancement matrix.

[0229] Further optionally, the teaching adjustment module is further used to:

[0230] The step-by-step control of the difficulty of practice is performed according to the order of knowledge points in the preferred learning path and the knowledge mastery β reflected in the feature enhancement matrix. The adjustment formula is:

[0231] D_{new}=D_{current}×(1+κ(β-θ)),

[0232] Among them, D_{new} is the practice difficulty after adjustment, D_{current} is the practice difficulty before adjustment, κ is the adjustment coefficient, and θ is the target knowledge mastery threshold.

[0233] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A personalized teaching method based on generative artificial intelligence, characterized by: The method comprises: Build a knowledge map and teaching strategy template library for the target course, including a series of preset teaching strategy templates; Based on the learner's individual learning characteristics and combined with the group learning behavior pattern, feature enhancement is performed to construct a feature enhancement matrix; Calculating the node activation strength, and generating multiple candidate learning paths through a beam search algorithm, and screening out the optimal learning path by combining the preset teaching strategy template; Based on the preferred learning path, using a generative model to generate standardized teaching content adapted to the teaching strategy; Dynamically optimizing the teaching plan through a neuroplasticity model by combining the standardized teaching content and learner characteristics reflected in the feature enhancement matrix; Make teaching adjustments based on the optimized teaching plan.

2. The personalized teaching method based on generative artificial intelligence according to claim 1 is characterized in that: The step of constructing the feature enhancement matrix comprises: Collect learning characteristic data common to individuals and groups, including learning behavior data, cognitive performance data, and learning context data; Fusing individual learning feature data with group common feature data to obtain fused feature data; Extract common learning patterns and explore common characteristics among learners; Based on the fusion feature data and common learning model, a feature enhancement matrix F_enhanced is constructed, and the formula is: F_enhanced=a·F_individual+b·F_group+c·ΔF_temporal, Among them, a, b, and c are the weight coefficients used to adjust the individual learning feature F_individual, the group learning feature F_group, and the influence of the feature evolution over time, respectively. ΔF_temporal reflects the evolution trend of the learning feature over time.

3. The personalized teaching method based on generative artificial intelligence according to claim 1 is characterized in that: The step of screening out the preferred learning path includes: Mapping the feature enhancement matrix to a set of nodes in the knowledge graph and determining the learner's initial position in the knowledge graph to clarify the starting point of learning; Calculate the node activation strength using the formula: h_v (l+1) =σ(∑{u∈N(v)}α_{vu}Wh_ul), Among them, h_v (l+1) is the activation strength of node v in the l+1 layer, N(v) is the set of neighbor nodes of node v, α_{vu} is the attention coefficient, which is used to measure the importance of node u to node v, W is the weight matrix, h_ul is the eigenvector of node u in the l layer, and σ is the activation function; Based on the node activation strength, K candidate paths are generated by a beam search algorithm; The candidate paths are scored in combination with the preset teaching strategy template, and the path with the highest score is selected as the preferred learning path.

4. The personalized teaching method based on generative artificial intelligence according to claim 1 is characterized in that: The step of using the generative model to generate standardized teaching content adapted to the teaching strategy includes: Build a generative model that includes semantic parsing, multimodal generation, and cognitive load assessment modules; Understanding the teaching requirements in the preferred learning path through the BERT architecture in the semantic parsing module and generating structured data of the teaching requirements; Based on the teaching demand structured data, initial teaching content is generated through the collaborative work of CLIP and DALL·E in the multimodal generation module; Performing cognitive load evaluation on the initial teaching content through the LSTM network in the cognitive load evaluation module to obtain a cognitive load evaluation result; According to the cognitive load assessment result, the initial teaching content is adjusted by the multimodal generation module to obtain adjusted teaching content.

5. The personalized teaching method based on generative artificial intelligence according to claim 4 is characterized in that: The step of using the generative model to generate standardized teaching content adapted to the teaching strategy also includes: Design a dual discriminator architecture, including a content quality discriminator and a teaching strategy discriminator; The adjusted teaching content is evaluated for quality by the content quality discriminator. If the quality evaluation result does not meet expectations, it is fed back to the multimodal generation module to further optimize the teaching content; The teaching strategy discriminator is used to judge whether the optimized teaching content meets the teaching strategy requirements in the preset teaching strategy template. If not, the judgment result is fed back to the multimodal generation module for targeted adjustment.

6. The personalized teaching method based on generative artificial intelligence according to claim 5 is characterized in that: The step of using the generative model to generate standardized teaching content adapted to the teaching strategy also includes: In the process of teaching content generation, according to the course difficulty coefficient δ, the generation constraints of the large language model are adjusted through the total loss function L_total. The total loss function L_total is: L_total=L_content+δ·L_strategy+L_regularization, Among them, L_content is the teaching content quality loss, L_strategy is the teaching strategy loss, and L_regularization is the regularization term used to prevent model overfitting; By continuously optimizing the total loss function, standardized teaching content that matches the optimal learning path and teaching strategy is eventually generated.

7. The personalized teaching method based on generative artificial intelligence according to claim 1 is characterized in that: The steps of dynamically optimizing the teaching plan through the neuroplasticity model include: Build a neural plasticity-driven optimizer, including a synaptic weight adjustment module and a cognitive resource allocation module; Preliminarily adjusting the teaching parameters according to the feedback from the use of the standardized teaching content through the synaptic weight adjustment module based on the STDP rule; Through the attention mechanism in the cognitive resource allocation module, cognitive resources are allocated according to the preliminarily adjusted teaching parameters and the learner's cognitive state reflected by the feature enhancement matrix.

8. The personalized teaching method based on generative artificial intelligence according to claim 7 is characterized in that: The step of dynamically optimizing the teaching plan through the neuroplasticity model further includes: Design a hierarchical reward function, the formula is: R_total=ω_1·Knowledge acquisition rate+ω_2·Cognitive efficiency+ω_3·Learning persistence, Among them, ω_1, ω_2, and ω_3 are weight coefficients used to balance the learning objectives of different courses; According to the learner's performance under the current teaching plan, the initial reward value R_total is calculated through the hierarchical reward function to reflect the comprehensive effect of the current teaching plan; The preliminary reward value R_total is fed back to the synaptic weight adjustment module and the cognitive resource allocation module so that the teaching parameters can be further optimized by the synaptic weight adjustment module, and the cognitive resource allocation module can reallocate cognitive resources based on the further optimized teaching parameters and the latest cognitive state of the learner reflected by the feature enhancement matrix.

9. The personalized teaching method based on generative artificial intelligence according to claim 1, characterized in that: The steps of performing teaching adjustment include: The step-by-step control of the difficulty of practice is performed according to the order of knowledge points in the preferred learning path and the knowledge mastery β reflected in the feature enhancement matrix. The adjustment formula is: D_{new}=D_{current}×(1+κ(β-θ)), Among them, D_{new} is the practice difficulty after adjustment, D_{current} is the practice difficulty before adjustment, κ is the adjustment coefficient, and θ is the target knowledge mastery threshold.

10. A personalized teaching system based on generative artificial intelligence, used to implement the personalized teaching method based on generative artificial intelligence according to any one of claims 1 to 9, characterized in that: The system comprises: Graph construction module: used to construct the knowledge graph and teaching strategy template library of the target course, including a series of preset teaching strategy templates; Feature enhancement module: used to enhance features based on learners’ individual learning features and group learning behavior patterns, and construct a feature enhancement matrix; Learning path screening module: used to calculate the node activation strength and generate multiple candidate learning paths through the beam search algorithm, and screen out the optimal learning path in combination with the preset teaching strategy template; Teaching content generation module: used to generate standardized teaching content adapted to the teaching strategy based on the preferred learning path using a generative model; Teaching plan optimization module: used to dynamically optimize the teaching plan through the neuroplasticity model by combining the standardized teaching content and the learner characteristics reflected in the feature enhancement matrix; Teaching adjustment module: used to adjust teaching based on the optimized teaching plan.

Images