Artificial intelligence-based teaching interactive content intelligent generation method

Abstract

This invention relates to the field of artificial intelligence education and teaching technology, specifically to an intelligent method for generating interactive teaching content based on artificial intelligence. The method includes: receiving a personalized learning profile of a target learner containing a sequence of knowledge mastery, learning behavior patterns, and cognitive ability tags; generating a dynamic learner state vector through multimodal learning feature extraction and fusion; and constructing a dynamic knowledge subgraph by selecting related knowledge nodes from a pre-defined knowledge graph based on this vector. The dynamic learner state vector and the dynamic knowledge subgraph are input into a content generation engine, which simultaneously generates interconnected explanatory texts, interactive questions, and contextualized cases. The internal logical consistency of these three types of content is checked, and the difficulty is adjusted. After integration, a complete interactive teaching content unit is formed. This method generates related teaching content based on the learner's dynamic state and dynamic knowledge structure, optimizing the content logic and difficulty suitability.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence education and teaching technology, and in particular to a method for intelligent generation of interactive teaching content based on artificial intelligence. Background Technology

[0002] Current interactive teaching content generation largely relies on pre-set question banks and fixed text templates. It only performs simple matching and filtering based on learners' basic learning information, and the processing of learners' learning characteristics is limited to single-dimensional data extraction. Knowledge systems are formed by manual organization into static knowledge modules, without adjusting the knowledge structure in light of learners' real-time status. Conventional content generation methods can only produce explanatory texts or scattered interactive questions. Case materials are mostly retrieved independently, without a unified content generation logic, and the selection of knowledge nodes always relies on a fixed knowledge graph framework.

[0003] The teaching content generated by this technology has many shortcomings. Explanatory texts, interactive questions, and case materials are disconnected, failing to form an internal knowledge connection. Furthermore, the content is not logically validated after generation, making it prone to logical gaps during knowledge transfer. Additionally, it fails to adjust the difficulty level according to learners' cognitive abilities and knowledge mastery, resulting in a discrepancy between the content's difficulty and learners' actual learning status. The selection of knowledge nodes does not align with learners' dynamic learning characteristics, leading to a low degree of personalization and an inability to match learners' real-time learning progress.

[0004] The teaching content generation process cannot combine dynamic learner state vectors and dynamic knowledge subgraphs to synchronously produce interconnected explanatory texts, interactive questions, and contextualized cases. The generated three types of teaching content cannot be internally checked for logical consistency or adjusted for difficulty balance. The relevance, logical coherence, and difficulty suitability of the teaching content cannot meet the actual requirements of personalized teaching. The degree of fit between the teaching content and learners is difficult to meet the actual teaching needs. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing an intelligent method for generating interactive teaching content based on artificial intelligence.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: an intelligent generation method for interactive teaching content based on artificial intelligence, comprising:

[0007] Receive the personalized learning profile of the target learner, which includes a knowledge mastery sequence, learning behavior patterns and cognitive ability tags;

[0008] Perform multimodal learning feature extraction and fusion operations on the personalized learning profile to generate a dynamic learner state vector;

[0009] Based on the dynamic learner state vector, related knowledge nodes are retrieved and filtered from the preset knowledge graph to form a dynamic knowledge subgraph to be taught.

[0010] The dynamic learner state vector and the dynamic knowledge subgraph are input together into a content generation engine;

[0011] Using the content generation engine, based on the structure of the dynamic knowledge subgraph and the learner characteristics represented by the dynamic learner state vector, interconnected explanatory texts, interactive questions, and contextualized cases are generated synchronously.

[0012] The internal logical consistency of the synchronously generated explanatory text, interactive questions, and contextualized cases is checked and the difficulty balance is adjusted.

[0013] The verified and adjusted explanatory text, interactive questions, and contextualized cases are integrated to form a complete interactive teaching content unit.

[0014] As a further aspect of the present invention, multimodal learning feature extraction and fusion operations are performed on the personalized learning profile to generate a dynamic learner state vector, including:

[0015] The historical accuracy, forgetting curve features, and most recent practice time points of different knowledge points are extracted from the knowledge mastery sequence, and the information is encoded into a knowledge point mastery feature vector.

[0016] The dwell time distribution, question frequency, relearning markers, and interaction preferences are extracted from the learning behavior patterns, and the information is encoded into learning behavior feature vectors.

[0017] The cognitive ability labels are mapped into multi-dimensional numerical cognitive ability assessment vectors;

[0018] The knowledge point mastery feature vector, the learning behavior feature vector, and the cognitive ability assessment vector are concatenated to form a high-dimensional preliminary fusion feature vector.

[0019] The preliminary fused feature vector is input into a feature self-attention network for weighted fusion. The feature self-attention network automatically calculates the importance weights of different feature dimensions for modeling the current learner state.

[0020] The importance weights are used to reweight each dimension of the preliminary fused feature vector, and the weighted results are then passed through a fully connected layer for dimensionality reduction and integration to output the dynamic learner state vector.

[0021] As a further aspect of the present invention, based on the dynamic learner state vector, related knowledge nodes are retrieved and filtered from a preset knowledge graph to form a dynamic knowledge subgraph to be taught, including:

[0022] Using the dynamic learner state vector as the query vector, semantic similarity matching is performed in the preset knowledge graph to initially recall a set of related knowledge nodes;

[0023] From the initially recalled set of knowledge nodes, select the preceding and subsequent knowledge points that are directly connected to the weak knowledge points in the knowledge mastery sequence;

[0024] Calculate the correlation score between the dynamic learner's state vector and the embedded representation of each associated knowledge node in the knowledge graph;

[0025] The related knowledge nodes are sorted according to the relevance scores, and a preset number of the top-ranked knowledge nodes are selected as core teaching nodes.

[0026] Centered on the core teaching node, extract its first-degree related edges and adjacent nodes from the preset knowledge graph to form the dynamic knowledge subgraph;

[0027] In the dynamic knowledge subgraph, each node is labeled with its relevance level to the current learning objective.

[0028] As a further aspect of the present invention, the content generation engine is used to synchronously generate interconnected explanatory texts, interactive questions, and contextualized cases based on the structure of the dynamic knowledge subgraph and the learner characteristics represented by the dynamic learner state vector, including:

[0029] The structural information of the dynamic knowledge subgraph, including node type, node attributes and edge relationships, is transformed into a serialized representation of the graph structure.

[0030] The serialized representation of the graph structure is concatenated with the dynamic learner state vector and used together as the conditional input of the content generation engine;

[0031] The content generation engine includes parallel text generation channels, question generation channels, and case generation channels;

[0032] In the text generation channel, a pre-trained language model generates the explanatory narrative text that conforms to the logic of the dynamic knowledge subgraph based on the conditional input;

[0033] In the question generation channel, a dedicated question generation model analyzes the core concepts and relationships in the dynamic knowledge subgraph and combines them with the cognitive abilities reflected in the dynamic learner state vector to generate interactive questions with different cognitive levels.

[0034] In the case generation channel, a case generation model constructs a contextualized case containing specific plots and data based on the application scenario of the dynamic knowledge subgraph and the interaction preferences reflected in the learning behavior pattern.

[0035] As a further aspect of the present invention, the internal logical consistency verification and difficulty balance adjustment of the synchronously generated explanatory narrative text, the interactive questions, and the contextualized cases include:

[0036] Establish a shared fact checklist, which extracts all key facts, concept definitions, and relational assertions from the dynamic knowledge subgraph;

[0037] The explanatory text, the stems and options of the interactive questions, and the descriptions of the contextualized cases are scanned sequentially to check whether they contain statements that contradict the fact-checking list.

[0038] If a contradictory statement is found, the generation channel from which it originates is located, and the local regeneration of the generation channel is triggered to correct the contradiction.

[0039] Assess the cognitive difficulty level of the interactive questions and the comprehension complexity of the contextualized cases, respectively.

[0040] The assessed cognitive difficulty level and comprehension complexity are compared with the cognitive ability assessment vector in the dynamic learner state vector;

[0041] If the level of cognitive difficulty or the complexity of understanding exceeds the learner's adaptability, the interactive questions or contextualized cases will be simplified by replacing complex terms or breaking down complex situations.

[0042] As a further aspect of the present invention, the integration of the verified and adjusted explanatory text, the interactive questions, and the contextualized cases forms a complete interactive teaching content unit, including:

[0043] Design a standard content assembly template that defines the location and connection logic of the explanation area, question insertion point, and case display area;

[0044] Fill the explanation text into the explanation area of ​​the content assembly template;

[0045] At key concepts or reasoning steps in the explanatory text, the corresponding interactive questions are embedded according to the question insertion point markers of the content assembly template.

[0046] Fill the case display area of ​​the content assembly template with the contextualized case that is most relevant to the current topic of explanation;

[0047] Transitional sentences are automatically generated between the explanatory narrative text, the embedded interactive questions, and the filled-in contextualized cases to ensure a smooth overall narrative.

[0048] Metadata tags are added to the final assembled content. These metadata tags include the corresponding knowledge node identifier, the expected learning duration, and the interaction type.

[0049] As a further aspect of the present invention, the method further includes: sending the teaching interactive content unit to the learner's interactive terminal for presentation, and collecting the learner's feedback response data during the interaction process in real time;

[0050] Using the collected feedback response data, the knowledge mastery sequence and learning behavior pattern in the personalized learning profile are incrementally updated;

[0051] Based on the updated personalized learning profiles, a new round of intelligent generation process for interactive teaching content will be launched.

[0052] The real-time collection of learners' feedback and response data during the interaction process includes:

[0053] Capture learners’ real-time behavior flow while learning the instructional interactive content unit, the real-time behavior flow including scrolling speed, dwell position and highlighting in the explanatory text area;

[0054] Record the learner's answer selection, number of modifications, and final submitted answer for each embedded interactive question;

[0055] Collect learners' interaction traces in the contextualized case demonstration area, including the click order of case steps, the reaction results to the simulated operation, and the text input in the case discussion area;

[0056] Record the total time spent by learners to complete the entire interactive teaching content unit, as well as the number and timing of their interruptions.

[0057] The captured real-time behavior stream, recorded answers and interaction traces, as well as the total duration and interruption information, are packaged together into structured feedback response data.

[0058] As a further aspect of the present invention, the incremental update of the knowledge mastery sequence and learning behavior pattern in the personalized learning profile using the collected feedback response data includes:

[0059] The answers to the interactive questions in the feedback response data are analyzed to determine their correctness, and different mastery score changes are assigned according to the cognitive difficulty of the questions.

[0060] Based on the change in the mastery score, update the mastery score of the corresponding knowledge point in the knowledge mastery sequence and the last practice timestamp;

[0061] Analyze the scrolling speed and dwell position patterns in the real-time behavior stream to identify learners’ reading focus intervals and suspected points of confusion, and update the content preferences and difficulty distribution in the learning behavior patterns.

[0062] Analyze the total duration and interruption information to assess learners' learning endurance and time management habits, and update the learning rhythm characteristics in the learning behavior pattern;

[0063] The updated knowledge mastery sequence and the learning behavior pattern are smoothly integrated with the unupdated historical records to form an incrementally updated personalized learning profile.

[0064] As a further aspect of the present invention, the step of initiating a new round of intelligent generation process for interactive teaching content based on the updated personalized learning profile includes:

[0065] The incrementally updated personalized learning profile is used as new input to trigger the multimodal learning feature extraction and fusion operation;

[0066] When generating a new dynamic learner state vector, the feature self-attention network adjusts the importance weights of different feature dimensions according to the new learning behavior pattern.

[0067] When the new dynamic learner state vector is used to retrieve the preset knowledge graph, the weak knowledge points have changed because the knowledge mastery sequence has been updated, resulting in different core teaching nodes being retrieved compared to before.

[0068] The content generation engine generates a new round of explanatory texts, interactive questions, and contextualized cases that adapt to the learner's latest state, based on the new dynamic knowledge subgraph and the new dynamic learner state vector.

[0069] As a further aspect of the present invention, calculating the relevance score between the dynamic learner state vector and the embedding representation of each associated knowledge node in the knowledge graph includes:

[0070] The dynamic learner state vector is extracted, which represents the learner's current comprehensive knowledge state and learning characteristics;

[0071] From the preset knowledge graph, obtain the pre-trained vectorized embedding representation of each associated knowledge node;

[0072] Perform a vector dot product operation on the dynamic learner state vector and the embedded representation of an associated knowledge node to obtain an initial relevance value;

[0073] The initial correlation values ​​are normalized and mapped to a standardized score range between zero and one.

[0074] By combining the global importance weights of the associated knowledge nodes recorded in the knowledge graph with the current knowledge learning objective, the standardized score is weighted and corrected to obtain the corrected relevance score;

[0075] For all associated knowledge nodes initially recalled through semantic similarity matching in the knowledge graph, the vector dot product operation, normalization process and weighted correction steps are repeatedly executed to obtain the relevance score corresponding to each associated knowledge node.

[0076] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0077] The dynamic learner state vector and dynamic knowledge subgraph are input into the content generation engine. Based on the structure of the dynamic knowledge subgraph and the learner characteristics represented by the dynamic learner state vector, interconnected explanatory texts, interactive questions, and contextualized cases are generated simultaneously. The three types of teaching content establish a knowledge-level correspondence during the generation stage. The knowledge context of the content matches the node connections of the dynamic knowledge subgraph, aligning with the learner's knowledge mastery sequence, learning behavior patterns, and cognitive ability labels. The teaching content and the learner's real-time learning status are precisely correlated, the knowledge connection between the content is natural and fits the teaching logic, and the knowledge relationship between the content closely fits the structural features of the dynamic knowledge subgraph.

[0078] The generated explanatory texts, interactive questions, and contextualized cases undergo internal logical consistency checks and difficulty balance adjustments to eliminate knowledge logic conflicts among the three types of content, ensuring that the knowledge transmission logic remains coherent and unified. The difficulty level of the content is adjusted based on learners' cognitive ability tags and knowledge mastery sequences, so that the difficulty gradient of the teaching content matches the learners' learning characteristics. The integrated interactive teaching content units maintain logical coherence, and the difficulty settings are tailored to the learners' actual learning levels. The logical state and difficulty level of the teaching content are aligned with the content presentation needs of personalized teaching, and the usage of the content is adapted to the learners' learning characteristics. Attached Figure Description

[0079] Figure 1 This is a state diagram of the AI-based intelligent generation method for interactive teaching content as described in this invention.

[0080] Figure 2 A flowchart for generating dynamic learner state vectors;

[0081] Figure 3 This is a flowchart for constructing a dynamic knowledge subgraph. Detailed Implementation

[0082] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0083] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0084] See Figure 1This invention provides an intelligent method for generating interactive teaching content based on artificial intelligence. The implementation method will be described in detail below, focusing on the specific steps of the technical solution of this invention. The system receives a personalized learning profile of the target learner, which integrates the learner's historical and real-time data. After receiving the personalized learning profile, which includes a knowledge mastery sequence, learning behavior patterns, and cognitive ability tags, the system first performs in-depth analysis and fusion, extracting multimodal information into a dynamic learner state vector representing the learner's current state. This vector is used for precise navigation within a pre-set knowledge graph, retrieving and filtering the most relevant knowledge nodes that best match the learner's current state, thereby constructing a focused, structured dynamic knowledge subgraph. This dynamic knowledge subgraph and the dynamic learner state vector are jointly fed into a content generation engine. Based on the internal logic of the knowledge subgraph and the learner's individual characteristics, the engine generates in parallel closely related explanatory texts, interactive questions, and contextualized cases. The generated content elements undergo internal logical consistency verification and difficulty balance adjustment to eliminate contradictions and ensure appropriate difficulty. Finally, these verified and adjusted elements are integrated according to the preset template and logic, and embedded with connecting statements to form a complete, fluent, and immediately usable interactive teaching content unit.

[0085] In one embodiment of the present invention, see [reference] Figure 2 The process of generating dynamic learner state vectors from personalized learning profiles involves the extraction and fusion of multimodal features. The operation first extracts historical accuracy rates, forgetting curve characteristics, and recent practice times for different knowledge points from the knowledge mastery sequence. This temporal and statistical information is encoded into a dense numerical vector, called the knowledge point mastery feature vector. Simultaneously, multi-dimensional behavioral indicators such as dwell time distribution, questioning frequency, relearning markers, and interaction preferences are extracted from learning behavior patterns. These indicators are encoded into learning behavior feature vectors. Cognitive ability labels, such as logical reasoning and spatial imagination, are mapped into a multi-dimensional numerical cognitive ability assessment vector. Subsequently, the knowledge point mastery feature vector, learning behavior feature vector, and cognitive ability assessment vector are concatenated along their feature dimensions to form a higher-dimensional preliminary fused feature vector. This preliminary fused feature vector is input into a feature self-attention network for deep processing. This network automatically calculates and assigns importance weights to different feature dimensions in characterizing the current learner's state. The network uses these importance weights to reweight each dimension of the preliminary fused feature vector, highlighting features that contribute significantly to the current state judgment and suppressing features that contribute less. The weighted features are then combined with dimensionality reduction and nonlinearity through a fully connected layer, ultimately outputting a compact and information-rich dynamic learner state vector.

[0086] In practical implementation, the process of generating dynamic learner state vectors from personalized learning profiles involves multimodal learning feature extraction and fusion. Personalized learning profiles include knowledge mastery sequences, learning behavior patterns, and cognitive ability labels. From the knowledge mastery sequence, historical accuracy rates, forgetting curve features, and recent practice timestamps for different knowledge points are extracted. Historical accuracy rates are obtained by calculating the proportion of times a learner answers correctly on a specific knowledge point throughout history. Forgetting curve features are obtained by fitting the learner's historical answer records into an exponential decay model to obtain parameters representing memory retention rates. Recent practice timestamps are converted into time differences relative to the current time. This extracted information is encoded as a knowledge point mastery feature vector. From the learning behavior patterns, the process extracts dwell time distribution, question frequency, relearning markers, and interaction preferences. Dwell time distribution is a histogram of the statistical distribution of the time a learner spends on different learning content blocks. Question frequency is the average number of questions a learner asks per unit of time. Relearning markers are the set of identifiers that learners actively mark as content items that need to be relearned. Interaction preferences are a normalized vector of the frequency at which learners choose among various interaction methods. This extracted information is encoded as a learning behavior feature vector. Cognitive ability labels are mapped to multi-dimensional numerical cognitive ability assessment vectors. These labels include logical reasoning, working memory, and spatial ability. The mapping process is completed by querying a predefined mapping table between labels and numerical vectors, with each label corresponding to an assessment score for one dimension. The knowledge point mastery feature vector, learning behavior feature vector, and cognitive ability assessment vector are concatenated along their feature dimensions to form a high-dimensional preliminary fused feature vector. The dimension of this preliminary fused feature vector is the sum of the dimensions of the knowledge point mastery feature vector, the learning behavior feature vector, and the cognitive ability assessment vector.

[0087] In some embodiments, the initially fused feature vector is input into a feature self-attention network for weighted fusion. The feature self-attention network automatically calculates the importance weights of different feature dimensions for modeling the current learner state. The feature self-attention network includes a trainable query weight matrix, key weight matrix, and value weight matrix. The network first multiplies the initially fused feature vector with the query weight matrix, key weight matrix, and value weight matrix respectively to obtain the query vector, key vector, and value vector. An attention score is calculated by the dot product operation of the query vector and key vector. The attention score is processed by a normalization function to obtain the importance weight. The importance weight is used to perform a weighted summation of the value vectors to obtain the weighted fused feature representation. The feature self-attention network can employ a multi-head attention mechanism, projecting the initially fused feature vector onto multiple subspaces to compute attention in parallel. The calculation process of the importance weight is expressed using the scaled dot product attention formula:

[0088]

[0089] in: This represents the query vector matrix obtained by linear transformation of the initial fused feature vectors. Let T represent the key vector matrix obtained by the linear transformation of the initial fused feature vectors. (in Chinese) represents the matrix transpose operation. This represents the value vector matrix obtained by the linear transformation of the initial fused eigenvectors. Indicates the dimension of the key vector. Scaling factor. Used to adjust the size of the dot product result. It can be understood that the softmax function converts the attention score into importance weights in the form of a probability distribution.

[0090] The initial fused feature vector is reweighted using importance weights across its dimensions. This reweighting process multiplies the importance weights by the corresponding dimensions of the initial fused feature vector, resulting in a weighted set of feature values. This weighted set of feature values ​​is then dimensionality-reduced and integrated through a fully connected layer. This fully connected layer contains trainable weight and bias parameters and introduces a non-linear transformation using a non-linear activation function, outputting a dynamic learner state vector. The dimension of this dynamic learner state vector is a pre-defined fixed value. Optionally, the feature self-attention network can contain multiple stacked attention layers, with the output of the previous layer serving as the input to the next, to capture more complex feature interactions. After the fully connected layer, layer normalization can be applied to stabilize the distribution of the output vector. In essence, the dynamic learner state vector integrates multimodal information about knowledge mastery, learning behavior, and cognitive ability, which is used for subsequent knowledge retrieval and content generation.

[0091] In one embodiment of the present invention, see [reference] Figure 3The process of constructing a dynamic knowledge subgraph is a graph retrieval and subgraph extraction process based on learner states. This process uses the dynamic learner state vector as the query vector, performing semantic similarity matching with the embedded representations of all knowledge nodes in the pre-defined knowledge graph to initially recall a set of relevant knowledge nodes. From this initially recalled set, it further filters out predecessor and successor knowledge points directly connected to the weak knowledge points identified in the knowledge mastery sequence to strengthen knowledge coherence. To quantify the degree of association, the system calculates the relevance score between the dynamic learner state vector and the embedded representation of each related knowledge node in the knowledge graph. The calculation of the relevance score first extracts the dynamic learner state vector, which encodes the learner's overall state, and obtains the pre-trained embedded representation of each related knowledge node from the knowledge graph. A vector dot product operation is performed on the dynamic learner state vector and the embedded representation of a target knowledge node to obtain an initial relevance value. This initial value is then normalized to map it to a standard score range. Subsequently, the standardized score is weighted and adjusted based on the predefined importance weights of the knowledge node relative to the global learning objective within the knowledge graph, yielding the final relevance score. This calculation process is performed on all related nodes. Based on the calculated relevance scores, all related knowledge nodes are sorted, and a predetermined number of nodes with the highest rankings are selected as the core teaching nodes for this lesson. Using these core teaching nodes as centers, the first-degree edges directly connected to them and their adjacent nodes are extracted from the predefined knowledge graph. These nodes and edges together constitute the dynamic knowledge subgraph used in this lesson. In this dynamic knowledge subgraph, each node is labeled with its relevance level to the current learning objective.

[0092] In practical implementation, the process of constructing a dynamic knowledge subgraph based on the dynamic learner state vector includes retrieval, filtering, and subgraph extraction operations. The dynamic learner state vector represents the learner's current comprehensive knowledge state and learning characteristics. The system uses the dynamic learner state vector as the query vector and performs semantic similarity matching in a preset knowledge graph. The preset knowledge graph contains a large number of knowledge nodes represented in embedded form. Semantic similarity matching is achieved by calculating the cosine similarity between the dynamic learner state vector and the embedded representation of each knowledge node, initially recalling a set of relevant knowledge points with similarity scores exceeding a preset threshold. From the initially recalled set of knowledge nodes, the system further filters out predecessor and successor knowledge points directly connected to weak knowledge points in the knowledge mastery sequence. Weak knowledge points are identifiers of knowledge points in the knowledge mastery sequence whose mastery scores are lower than a preset standard. The filtering process is completed by querying the predecessor-successor relationship edges between knowledge points recorded in the knowledge graph. The system calculates the relevance score between the dynamic learner's state vector and the embedded representation of each associated knowledge node in the knowledge graph. The dynamic learner's state vector, with a fixed dimension, is extracted. Pre-trained vectorized embedded representations of each associated knowledge node are obtained from the pre-defined knowledge graph; these representations reside in the same vector space and have the same dimension as the dynamic learner's state vector. The relevance score is calculated by performing a vector dot product operation between the dynamic learner's state vector and the embedded representation of an associated knowledge node. The result of this dot product is a scalar, denoted as the initial relevance value. This initial relevance value is then normalized using the Sigmoid function to map it to a standardized score range between zero and one. Finally, the standardized score is weighted and corrected by combining the global importance weights of the associated knowledge nodes recorded in the knowledge graph with the current knowledge learning objective. These global importance weights are pre-assigned weights to each knowledge node based on expert knowledge or statistical frequency during knowledge graph construction. This weighting correction is performed through multiplication, yielding the final corrected relevance score. For all associated knowledge nodes initially recalled through semantic similarity matching in the knowledge graph, the vector dot product operation, normalization process, and weighted correction steps are repeatedly performed to obtain the final relevance score for each associated knowledge node. The calculation process of the final relevance score is expressed by the following formula:

[0093]

[0094] in: This represents the final relevance score between the state vector of the i-th dynamic learner and the j-th associated knowledge node. This represents the state vector of the i-th dynamic learner. This represents the embedding representation of the j-th associated knowledge node in a knowledge graph, using the symbol... This represents the vector dot product operation. This represents the Sigmoid normalization function. This represents the global importance weight of the j-th knowledge node recorded in the knowledge graph.

[0095] The related knowledge nodes are sorted according to their final relevance scores, from highest to lowest. A predetermined number of these nodes are selected as core teaching nodes, the number determined by the capacity of the teaching content. Centered on the core teaching nodes, one-degree-of-connection edges and adjacent nodes are extracted from a predetermined knowledge graph. A one-degree connection refers to nodes directly connected by a directed or undirected edge in the knowledge graph. The extracted core teaching nodes, edges, and adjacent nodes constitute a dynamic knowledge subgraph. In this dynamic knowledge subgraph, each node is labeled with its relevance level to the current learning objective. The relevance level is determined by whether the node is a core teaching node or an adjacent node, and the interval in which its final relevance score falls. In some embodiments, semantic similarity matching can combine multi-view comparisons of knowledge node embeddings and dynamic learner state vectors, such as simultaneously calculating cosine similarity and the reciprocal of Euclidean distance. The initial recalled knowledge node set can have a fixed capacity, for example, recalling the top 100 nodes with the highest similarity. Optionally, when filtering for predecessor and successor knowledge points directly connected to weak knowledge points, the type of connecting edges can be limited. For example, only edges with strong logical relationships such as "prerequisite" or "partial of" can be considered. Global importance weighting. The parameters can be designed to be dynamically adjustable, reassigned according to updates to the syllabus. In some embodiments, the construction of the dynamic knowledge subgraph can include not only first-degree connections but also, in domains with simple knowledge structures, second-degree connections to provide richer context. Optionally, relevance level labels can be discrete, such as "high relevance," "medium relevance," and "low relevance," which are used to differentiate the level of detail in subsequent content generation. It can be understood that a dynamic knowledge subgraph is the most relevant connected substructure in the knowledge graph to the learner's current state. It can also be understood that by calculating and ranking the final relevance scores, the system can focus on the most suitable cluster of knowledge nodes for teaching from the vast knowledge graph.

[0096] In one embodiment of the present invention, after receiving a dynamic knowledge subgraph and a dynamic learner state vector, the content generation engine simultaneously generates multiple teaching elements. The engine first transforms the structural information of the dynamic knowledge subgraph, including node types, node attributes, and edge relationships, into a machine-readable graph structure serialization representation. This serialization representation is concatenated with the dynamic learner state vector, serving as the conditional input for the content generation engine to generate all content. The content generation engine internally includes three parallel generation channels: a text generation channel, a question generation channel, and a case generation channel. In the text generation channel, a pre-trained language model generates explanatory text that conforms to the logical relationships of the dynamic knowledge subgraph based on the aforementioned conditional input in an autoregressive manner. In the question generation channel, a dedicated question generation model analyzes the core concepts and relationships in the dynamic knowledge subgraph and associates them with the cognitive ability dimensions reflected in the dynamic learner state vector, generating interactive questions with different cognitive levels. In the case generation channel, a case generation model constructs contextualized cases with strong realism, containing specific plots and data, based on the knowledge application scenarios pointed to by the dynamic knowledge subgraph and combined with the interaction preferences reflected in the learning behavior patterns. After the three types of content are generated, internal logical consistency checks and difficulty balance adjustments are required. The system establishes a fact-checking list that extracts all key facts, concept definitions, and relational assertions from a dynamic knowledge subgraph. The system sequentially scans the explanatory text, the stems and options of interactive questions, and the descriptions of contextualized cases to check for statements contradicting the fact-checking list. If a contradiction is found, the system locates the generation channel from which it originated and triggers a partial regeneration of the contradictory portion to correct the error. The system evaluates the cognitive difficulty level of the generated interactive questions and the comprehension complexity of the contextualized cases, comparing the evaluation results with the cognitive ability evaluation vector in the dynamic learner state vector. When the cognitive difficulty level of the interactive questions or the comprehension complexity of the contextualized cases exceeds the learner's adaptation range, the system simplifies the interactive questions or contextualized cases, such as replacing complex terminology or decomposing complex contexts.

[0097] In its implementation, the content generation engine receives a dynamic knowledge subgraph and a dynamic learner state vector, and simultaneously generates various teaching elements. The engine first transforms the structural information of the dynamic knowledge subgraph into a serialized graph representation. This structural information includes node types, node attributes, and edge relationships. The transformation process uses a depth-first search to traverse all nodes in the dynamic knowledge subgraph, concatenating the type, attributes, and relationships with adjacent nodes of each node in the order of access using a specific delimiter to form a text sequence, thus creating the serialized graph representation. This serialized graph representation is then concatenated with the dynamic learner state vector. This concatenation operation converts the text sequence into a vector through an embedding layer, which is then connected to the dynamic learner state vector along the feature dimension, serving as the conditional input for the content generation engine. Internally, the content generation engine includes parallel text generation, question generation, and case generation channels. These three channels share conditional inputs but have independent neural network model parameters. In the text generation channel, a pre-trained language model autoregressively generates explanatory text that conforms to the logic of a dynamic knowledge subgraph based on conditional input. The pre-trained language model uses the conditional input as a prefix to predict and generate subsequent explanatory text sequences word by word, ensuring that the generated narrative covers the core nodes and relationships in the dynamic knowledge subgraph. In the question generation channel, a dedicated question generation model analyzes the core concepts and relationships in the dynamic knowledge subgraph and generates interactive questions based on the cognitive abilities reflected in the dynamic learner's state vector. The question generation model identifies key knowledge entities and relationship pairs from the conditional input and determines the cognitive level of the generated question based on the scores of the corresponding dimensions in the cognitive ability evaluation vector. Cognitive levels correspond to different levels such as memory, understanding, application, and analysis. In the case generation channel, a case generation model constructs contextualized cases based on the application scenarios of the dynamic knowledge subgraph and the interactive preferences reflected in the learning behavior patterns. The case generation model parses the application scenario descriptions of knowledge points from the conditional input and extracts interactive preference information from the learning behavior feature vector, such as a preference for visual or story-based cases, thereby generating contextualized case descriptions containing specific plots and data.

[0098] After generating the explanatory text, interactive questions, and contextualized cases, the system performs internal logical consistency checks and difficulty balancing adjustments. A shared fact-checking list is established, extracting all key facts, concept definitions, and relational assertions from a dynamic knowledge subgraph. The extraction process traverses each node and edge of the dynamic knowledge subgraph, recording the standardized knowledge representations it carries in the list. The explanatory text, the stems and options of the interactive questions, and the descriptions of the contextualized cases are scanned sequentially to check for statements contradicting the fact-checking list. This check is performed by segmenting the text into sentences or phrases and performing semantic or rule matching with the entries in the fact-checking list. If a contradictory statement is found, the generation channel from which it originates is located, triggering a local regeneration of that channel to correct the contradiction. This local regeneration re-feeds the contradictory text fragment, the correct statement from the fact-checking list, and the original input conditions back into the corresponding generation model, requesting it to generate the corrected content. The cognitive difficulty level of interactive questions and the comprehension complexity of contextualized cases are assessed separately. The cognitive difficulty level of interactive questions is scored based on the type of cognitive operations involved, the abstractness of the knowledge points, and the deceptiveness of the options. The comprehension complexity of contextualized cases is scored based on the number of knowledge points involved, the unfamiliarity of the context, and the length of the reasoning chain. The assessed cognitive difficulty level and comprehension complexity are compared with the cognitive ability assessment vector in the dynamic learner's state vector. During the comparison, the quantified scores of cognitive difficulty level and comprehension complexity are compared with the score thresholds of the corresponding dimensions in the cognitive ability assessment vector. If the cognitive difficulty level or comprehension complexity exceeds the learner's adaptability, the interactive questions or contextualized cases are simplified. Simplification includes replacing complex terminology or breaking down complex contexts, such as replacing technical terms with synonymous common words, or breaking down a multi-step complex case into several consecutive simple cases. For the assessment of the cognitive difficulty level of interactive questions, please refer to Table 1.

[0099] Table 1: Difficulty Balance Adjustment Rules

[0100] Cognitive ability dimension Ability score range The highest permissible level of cognition Acceptable terminology complexity Acceptable number of context-related knowledge points Logical reasoning [0,0.3) memory Low ≤2 Logical reasoning [0.3,0.7) application middle ≤4 Logical reasoning [0.7,1.0] analyze high ≤6 Working memory [0,0.3) understand Low ≤1 Working memory [0.3,0.7) application middle ≤3 Working memory [0.7,1.0] analyze high ≤5

[0101] In some embodiments, the construction of the fact-checking list can incorporate constraints from node attributes in a knowledge graph as logical verification rules. The question generation channel can generate interactive questions in various formats, including single-choice, multiple-choice, and fill-in-the-blank questions. The case generation channel can select to generate plain text cases, mixed text and image cases, or interactive simulation cases based on interaction preference information. Optionally, logical consistency verification can utilize a pre-trained natural language reasoning model to determine whether the generated sentences have an implicit, contradictory, or neutral relationship with the statements in the fact list. Difficulty balancing adjusts the complexity of understanding contextualized cases. Quantization can be performed using a linear weighted formula, expressed as follows:

[0102]

[0103] in: This indicates the number of knowledge points explicitly involved in the case study. The rating of unfamiliarity of the context with respect to the learner's historical record. This represents the minimum number of reasoning steps required to arrive at the answer to the question from the case description. These are preset weighting coefficients. It's understandable that internal logical consistency checks ensure the accuracy of the generated teaching content in terms of knowledge. It's also understandable that difficulty balancing adjustments match the generated teaching content with learners' cognitive abilities, thus maintaining an appropriate learning challenge.

[0104] In one embodiment of the invention, the generated and verified teaching elements are integrated into a complete interactive teaching content unit. The system uses a predefined standard content assembly template, which clearly defines the relative positions and connection logic of the explanation area, multiple question insertion points, and case display area within the content unit. The generated explanation text is filled into the explanation area of ​​the content assembly template. At the locations of key concepts or important reasoning steps identified in the explanation text, corresponding interactive questions are embedded at the question insertion points marked in the content assembly template. From the generated contextualized cases, one or more cases most relevant to the current explanation topic are selected and filled into the case display area of ​​the content assembly template. Between the explanation text, the embedded interactive questions, and the filled contextualized cases, the system automatically generates transitional statements to ensure a natural and smooth transition between different parts and maintain the overall coherence of the narrative. Metadata tags are added to the finally assembled content unit, which include the identifiers of the core knowledge nodes taught in the unit, the estimated learning time required for the learner to complete the task, and the types of interactions included. After integration, the interactive teaching content unit is sent to the learner's interactive terminal for presentation. During the presentation, the system collects learners' feedback response data in real time during the interaction. The collection process captures learners' real-time behavioral flow while learning interactive content units, including scrolling speed, pause positions, and highlighting actions in the explanatory text area. The system records learners' answer selections, modification counts, and final submissions for each embedded interactive question. The system collects learners' interaction traces in the contextualized case study display area, including the click order of case study steps, reactions to simulated operations, and text input in the case study discussion area. The system records the total time learners spend learning the entire interactive content unit, as well as the number and specific times of interruptions. The captured real-time behavioral flow, recorded answers and interaction traces, total duration, and interruption information are packaged together into structured feedback response data.

[0105] In practical implementation, the system integrates validated and adjusted explanatory text, interactive questions, and contextualized cases to form complete interactive teaching content units. A standard content assembly template is designed, defining the positions and connection logic of the explanation area, question insertion points, and case display area. The content assembly template is a structured document framework. The explanation area holds continuous narrative text, question insertion points are anchor points scattered throughout the explanation area, identified by specific markers, and the case display area is an independent area located at the end of the content unit or in the sidebar. The explanation text is filled into the explanation area of ​​the content assembly template by copying the entire generated explanation text string to the placeholder in the explanation area. At key concepts or reasoning steps in the explanation text, the system embeds corresponding interactive questions based on the question insertion point markers in the content assembly template. Key concepts or reasoning steps are located through text analysis, identifying terms or conclusive statements in the explanation text that are strongly related to the core nodes of the dynamic knowledge subgraph. Question insertion point markers are pre-set after the positions corresponding to these terms or statements. The embedding operation inserts the question stem, options, and interactive logic controls of the interactive question into the marked position. The most relevant contextualized cases to the current topic are filled into the case display area of ​​the content assembly template. Relevance is determined by calculating the semantic similarity between the text description of the contextualized case and the explanatory text; the case with the highest similarity is selected. Between the explanatory text, embedded interactive questions, and the filled contextualized cases, the system automatically generates transitional sentences to ensure smooth narration. These transitional sentences utilize a lightweight text generation model, using context as a condition, to produce connecting text such as "Based on the above concepts, please consider the following questions" or "To deepen understanding, we can examine the following practical scenarios." Metadata tags are added to the final assembled content. These tags include the corresponding knowledge node identifier, estimated learning time, and interaction type. The knowledge node identifier comes from the unique code of the core teaching node in the dynamic knowledge subgraph. The estimated learning time is estimated by multiplying the number of words in the explanatory text, the number of questions, and the case complexity by a preset unit time coefficient. The interaction type records the question formats and case interaction formats included in the content unit.

[0106] After integration, the interactive teaching content units are sent to learners' interactive terminals for presentation, and learners' feedback response data is collected in real time during the interaction process. The sending operation transmits the data packets of the content units to the learners' terminal devices through the application programming interface. Real-time collection of learners' feedback response data during the interaction process includes capturing the learner's real-time behavioral flow while learning the interactive teaching content units. This real-time behavioral flow includes scrolling speed, dwell position, and highlighting in the explanatory text area. Scrolling speed is the number of pixels scrolled per unit time; dwell position is the set of coordinates where the learner lingers in a specific coordinate area on the page for more than a threshold time; and highlighting is the record of the learner selecting and marking text using the mouse or touchscreen. The system also records the learner's answer selections, modification counts, and final submitted answer for each embedded interactive question. Answer selection is the first option number clicked by the learner in the multiple-choice question options; modification count is the total number of times the learner changes the options before submitting the final answer; and the final submitted answer is the option number selected or the text entered by the learner when confirming submission. The system collects learners' interaction traces in the contextualized case demonstration area. These traces include the click sequence of case steps, the response results to simulated operations, and text input in the case discussion area. The click sequence of case steps is the time sequence of learners clicking different step buttons in the case flow. The response results to simulated operations are the system feedback caused by the parameters input or choices made by learners in the simulated operation interface. The text input in the case discussion area is the text content of questions or comments entered by learners in the discussion area. The system records the total time spent by learners to complete the entire teaching interaction content unit, as well as the number and time points of interruptions. The total time is the time from the start of the content unit presentation to the learner's final mark of completion. The number of interruptions is the number of times learners actively switch to other applications or tabs within the total time. The interruption time point is the specific moment each interruption occurs. The captured real-time behavior flow, recorded answers and interaction traces, and total time and interruption information are packaged together into structured feedback response data. During the packaging process, all data is organized according to a preset JSON schema and attached with a timestamp and a unique identifier for the content unit (see Table 2).

[0107] Table 2: Content Assembly Template Structure Table

[0108] Template area Location and signage Source of populated content Connection Logic Explanation area Main content area, ID: "main_content" The generated explanatory text Nothing, as the subject of the narrative Question insertion point 1 Located after the first paragraph of the explanation section, marked with: "{Q1}" The first question in the interactive question list Immediately after explaining the core concepts, embed questions. Question insertion point 2 Located after the description of the key reasoning steps, marked: "{Q2}" The second question in the interactive question list Embed questions after demonstrating the reasoning to test comprehension. Case Study Area At the end of the main content area, the ID is "case_area". The most relevant contextualized cases generated Presented after all explanations and questions, for comprehensive application.

[0109] In some embodiments, the lightweight text generation model that automatically generates transition sentences can be based on rule templates, selecting appropriate transition sentences from a pre-defined sentence library according to the type of the preceding and following content. The content assembly template can be configurable, allowing different template layouts to be selected based on subject differences, such as placing the case study area in the sidebar. Optionally, the formula for estimating the expected learning time can be expressed as:

[0110]

[0111] Where: D represents the estimated learning time of the final assembled interactive teaching content unit. This indicates the total number of words in the explanatory and narrative text. This indicates the number of embedded interactive questions. Indicates the complexity level of contextualized cases. These are preset time weighting coefficients for different content types. This is a preset benchmark value for average reading and interaction speed. Dwell positions in the real-time behavior stream can be aggregated into multiple attention hotspots. These hotspots are calculated using a clustering algorithm to analyze the coordinates of the dwell positions. It can be understood that standardized content assembly templates ensure consistent structure and user experience across different batches of generated interactive teaching content units. It can also be understood that real-time collection of multi-dimensional feedback response data provides comprehensive input for subsequent incremental updates to personalized learning profiles.

[0112] In one embodiment of the present invention, the collected feedback response data is used to iteratively update the personalized learning profile. The system analyzes the answers to interactive questions in the feedback response data, determines their correctness, and assigns a mastery score change value to the corresponding knowledge point based on the preset cognitive difficulty level of the question. Based on the mastery score change value, the system updates the current mastery score and last practice timestamp of the corresponding knowledge point in the knowledge mastery sequence. The system analyzes the scrolling speed and dwell position patterns in the real-time behavior flow to identify the learner's reading focus intervals and suspected points of confusion within content units, thereby updating the content preferences and difficulty distribution information in the learning behavior pattern. The system analyzes the total learning time and interruption information to assess the learner's learning endurance and time management habits, and updates the learning rhythm characteristics in the learning behavior pattern. The updated knowledge mastery sequence and learning behavior pattern are weighted and smoothly fused with the unupdated historical records in the profile to form an incrementally updated personalized learning profile reflecting the learner's latest state. Based on the updated personalized learning profile, the system automatically initiates a new round of intelligent generation of interactive teaching content. The incrementally updated personalized learning profile is used as input to trigger multimodal learning feature extraction and fusion operations. When generating new dynamic learner state vectors, the feature self-attention network adaptively adjusts the importance weights of different feature dimensions based on the new learning behavior pattern data. When retrieving the preset knowledge graph using the new dynamic learner state vectors, the system's assessment of weak knowledge points may change due to the updated knowledge mastery sequence, resulting in different core teaching nodes being retrieved and selected compared to the previous round. The content generation engine, based on the new dynamic knowledge subgraph and the new dynamic learner state vectors, generates a new round of explanatory texts, interactive questions, and contextualized cases, adapting this content to the learner's latest knowledge state and behavioral characteristics.

[0113] In practice, the collected feedback response data is used to incrementally update the knowledge mastery sequence and learning behavior patterns in personalized learning portfolios. The answers to interactive questions in the feedback response data are analyzed and their correctness is determined. This determination process compares the learner's submitted answers with the standard answers to the questions, assigning different mastery score changes based on the cognitive difficulty of the questions. Cognitive difficulty is divided into three levels: high, medium, and low, each corresponding to a different score change coefficient. Based on the mastery score change value, the mastery score and last practice timestamp of the corresponding knowledge point in the knowledge mastery sequence are updated. The update operation locates the record corresponding to the knowledge point in the knowledge mastery sequence, adds the original mastery score to the mastery score change value to obtain the new mastery score, and updates the last practice timestamp to the time recorded in the current feedback response data. This study analyzes scrolling speed and dwell patterns in real-time behavior streams to identify learners' reading focus intervals and potential points of confusion. Reading focus intervals are continuous text areas where scrolling speed consistently falls below a slow threshold and dwell positions are densely packed. Potential points of confusion are page locations with abnormally concentrated dwell positions accompanied by repeated back-and-forth scrolling behavior. Based on the identified reading focus intervals and potential points of confusion, the study updates content preferences and difficulty distribution in the learning behavior patterns. Content preferences are patterns where learners exhibit longer dwell times on specific types of content, and difficulty distribution is the set of knowledge point identifiers associated with potential points of confusion. The study also analyzes total duration and interruption information to assess learners' learning endurance and time management habits. Learning endurance is assessed by combining the ratio of total duration to the expected learning time per content unit and the frequency of interruptions. Time management habits are inferred by analyzing whether interruptions occur after specific content segments, and the study update the learning rhythm characteristics in the learning behavior patterns. These characteristics include average continuous learning duration, typical interruption intervals, and learning efficiency labels for different time periods. The updated knowledge mastery sequence and learning behavior patterns are smoothly integrated with the unupdated historical records. This smooth integration is achieved through an exponentially weighted moving average algorithm, forming an incrementally updated personalized learning profile that reflects the learner's latest knowledge status and behavioral characteristics.

[0114] Based on the updated personalized learning profiles, the system initiates a new round of intelligent generation of interactive teaching content. The incrementally updated personalized learning profiles serve as new input, triggering multimodal learning feature extraction and fusion operations. When generating new dynamic learner state vectors, the feature self-attention network adjusts the importance weights of different feature dimensions based on new learning behavior patterns. For example, when new learning behavior patterns show a significant increase in learners' dwell time on chart content, the feature self-attention network may increase the weights of feature dimensions related to visual learning preferences. When retrieving the preset knowledge graph using the new dynamic learner state vectors, because the knowledge mastery sequence has been updated, weak knowledge points have changed. These changes stem from the increase or decrease in mastery scores for certain knowledge points during this update, resulting in different retrieved core teaching nodes. The system recalculates the correlation between the dynamic learner state vectors and knowledge graph nodes and selects a new set of core teaching nodes. Based on the new dynamic knowledge subgraph and the new dynamic learner state vectors, the content generation engine generates a new round of explanatory texts, interactive questions, and contextualized cases adapted to the learners' latest state. The new dynamic knowledge subgraph consists of new core teaching nodes and their associated edges. Knowledge mastery score during the smooth integration and update process The update can be expressed by the following formula:

[0115]

[0116] in: This represents the score indicating the new mastery of knowledge points after smooth integration. This represents the updated mastery score calculated from the feedback response data. This indicates the historical smoothness score of mastery of this knowledge point before this update. This is a smoothing factor between 0 and 1, used to control the fusion ratio of new and historical data. In some embodiments, different mastery score changes are assigned based on the cognitive difficulty of the question. A difficulty weight mapping table can be defined, where higher cognitive difficulty questions receive higher positive score changes for correct answers and lower negative scores for incorrect answers. When identifying suspected points of confusion, cross-validation can be performed by combining learners' performance on corresponding interactive questions. If an answer is incorrect and the dwell time at that point is unusually long, it is confirmed as a difficult point. Optionally, the learning efficiency labels for different time periods in the learning rhythm characteristics can be determined by analyzing the knowledge mastery gain per unit time in different time periods within historical learning records. Smoothing factor It can be designed as a dynamic value, adjusted according to the early or late stage of the learning process, with higher weights given to new data in the early stages. In some embodiments, the feature-based self-attention network adjusting weights according to new learning behavior patterns is an online learning process, and some parameters of the network can be fine-tuned based on the distribution differences between new and historical data. It can be understood that incremental updates based on feedback response data enable personalized learning profiles to reflect learners' progress and changes in real time. It can also be understood that initiating a new round of generation ensures that the interactive teaching content can dynamically adapt to the learner's latest learning status, achieving the continuous evolution of personalized teaching paths.

[0117] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An artificial intelligence-based teaching interactive content intelligent generation method, characterized in that, include: Receive the personalized learning profile of the target learner, which includes a knowledge mastery sequence, learning behavior patterns and cognitive ability tags; Perform multimodal learning feature extraction and fusion operations on the personalized learning profile to generate a dynamic learner state vector; Based on the dynamic learner state vector, related knowledge nodes are retrieved and filtered from the preset knowledge graph to form a dynamic knowledge subgraph to be taught. The dynamic learner state vector and the dynamic knowledge subgraph are input together into a content generation engine; Using the content generation engine, based on the structure of the dynamic knowledge subgraph and the learner characteristics represented by the dynamic learner state vector, interconnected explanatory texts, interactive questions, and contextualized cases are generated synchronously. The internal logical consistency of the synchronously generated explanatory text, interactive questions, and contextualized cases is checked and the difficulty balance is adjusted. The verified and adjusted explanatory text, interactive questions, and contextualized cases are integrated to form a complete interactive teaching content unit. 2.The method of claim 1, wherein the method further comprises: Perform multimodal learning feature extraction and fusion operations on the personalized learning profile to generate a dynamic learner state vector, including: The historical accuracy, forgetting curve features, and most recent practice time points of different knowledge points are extracted from the knowledge mastery sequence, and the information is encoded into a knowledge point mastery feature vector. The dwell time distribution, question frequency, relearning markers, and interaction preferences are extracted from the learning behavior patterns, and the information is encoded into learning behavior feature vectors. The cognitive ability labels are mapped into multi-dimensional numerical cognitive ability assessment vectors; The knowledge point mastery feature vector, the learning behavior feature vector, and the cognitive ability assessment vector are concatenated to form a high-dimensional preliminary fusion feature vector. The preliminary fused feature vector is input into a feature self-attention network for weighted fusion. The feature self-attention network automatically calculates the importance weights of different feature dimensions for modeling the current learner state. The importance weights are used to reweight each dimension of the preliminary fused feature vector, and the weighted results are then passed through a fully connected layer for dimensionality reduction and integration to output the dynamic learner state vector.

3. The method for intelligent generation of interactive teaching content based on artificial intelligence as described in claim 1, characterized in that, Based on the dynamic learner state vector, related knowledge nodes are retrieved and filtered from the preset knowledge graph to form a dynamic knowledge subgraph to be taught, including: Using the dynamic learner state vector as the query vector, semantic similarity matching is performed in the preset knowledge graph to initially recall a set of related knowledge nodes; From the initially recalled set of knowledge nodes, select the preceding and subsequent knowledge points that are directly connected to the weak knowledge points in the knowledge mastery sequence; Calculate the correlation score between the dynamic learner's state vector and the embedded representation of each associated knowledge node in the knowledge graph; The related knowledge nodes are sorted according to the relevance scores, and a preset number of the top-ranked knowledge nodes are selected as core teaching nodes. Centered on the core teaching node, extract its first-degree related edges and adjacent nodes from the preset knowledge graph to form the dynamic knowledge subgraph; In the dynamic knowledge subgraph, each node is labeled with its relevance level to the current learning objective.

4. The method for intelligent generation of interactive teaching content based on artificial intelligence as described in claim 3, characterized in that, Using the content generation engine, based on the structure of the dynamic knowledge subgraph and the learner characteristics represented by the dynamic learner state vector, interconnected explanatory texts, interactive questions, and contextualized cases are generated synchronously, including: The structural information of the dynamic knowledge subgraph, including node type, node attributes and edge relationships, is transformed into a serialized representation of the graph structure. The serialized representation of the graph structure is concatenated with the dynamic learner state vector and used together as the conditional input of the content generation engine; The content generation engine includes parallel text generation channels, question generation channels, and case generation channels; In the text generation channel, a pre-trained language model generates the explanatory narrative text that conforms to the logic of the dynamic knowledge subgraph based on the conditional input; In the question generation channel, a dedicated question generation model analyzes the core concepts and relationships in the dynamic knowledge subgraph and combines them with the cognitive abilities reflected in the dynamic learner state vector to generate interactive questions with different cognitive levels. In the case generation channel, a case generation model constructs a contextualized case containing specific plots and data based on the application scenario of the dynamic knowledge subgraph and the interaction preferences reflected in the learning behavior pattern.

5. The method for intelligent generation of interactive teaching content based on artificial intelligence as described in claim 4, characterized in that, The process of performing internal logical consistency checks and difficulty balance adjustments on the synchronously generated explanatory text, interactive questions, and contextualized cases includes: Establish a shared fact checklist, which extracts all key facts, concept definitions, and relational assertions from the dynamic knowledge subgraph; The explanatory text, the stems and options of the interactive questions, and the descriptions of the contextualized cases are scanned sequentially to check whether they contain statements that contradict the fact-checking list. If a contradictory statement is found, the generation channel from which it originates is located, and the local regeneration of the generation channel is triggered to correct the contradiction. Assess the cognitive difficulty level of the interactive questions and the comprehension complexity of the contextualized cases, respectively. The assessed cognitive difficulty level and comprehension complexity are compared with the cognitive ability assessment vector in the dynamic learner state vector; If the level of cognitive difficulty or the complexity of understanding exceeds the learner's adaptability, the interactive questions or contextualized cases will be simplified by replacing complex terms or breaking down complex situations.

6. The method for intelligent generation of interactive teaching content based on artificial intelligence as described in claim 5, characterized in that, The integrated, verified and adjusted explanatory text, interactive questions, and contextualized cases form a complete interactive teaching content unit, including: Design a standard content assembly template that defines the location and connection logic of the explanation area, question insertion point, and case display area; Fill the explanation text into the explanation area of ​​the content assembly template; At key concepts or reasoning steps in the explanatory text, the corresponding interactive questions are embedded according to the question insertion point markers of the content assembly template. Fill the case display area of ​​the content assembly template with the contextualized case that is most relevant to the current topic of explanation; Transitional sentences are automatically generated between the explanatory narrative text, the embedded interactive questions, and the filled-in contextualized cases to ensure a smooth overall narrative. Metadata tags are added to the final assembled content. These metadata tags include the corresponding knowledge node identifier, the expected learning duration, and the interaction type.

7. The method for intelligent generation of interactive teaching content based on artificial intelligence as described in claim 1, characterized in that, The method further includes: sending the teaching interactive content unit to the learner's interactive terminal for presentation, and collecting the learner's feedback response data in real time during the interaction process; Using the collected feedback response data, the knowledge mastery sequence and learning behavior pattern in the personalized learning profile are incrementally updated; Based on the updated personalized learning profiles, a new round of intelligent generation process for interactive teaching content will be launched. The real-time collection of learners' feedback and response data during the interaction process includes: Capture learners’ real-time behavior flow while learning the instructional interactive content unit, the real-time behavior flow including scrolling speed, dwell position and highlighting in the explanatory text area; Record the learner's answer selection, number of modifications, and final submitted answer for each embedded interactive question; Collect learners' interaction traces in the contextualized case demonstration area, including the click order of case steps, the reaction results to the simulated operation, and the text input in the case discussion area; Record the total time spent by learners to complete the entire interactive teaching content unit, as well as the number and timing of their interruptions. The captured real-time behavior stream, recorded answers and interaction traces, as well as the total duration and interruption information, are packaged together into structured feedback response data.

8. The method for intelligent generation of interactive teaching content based on artificial intelligence as described in claim 7, characterized in that, The step of incrementally updating the knowledge mastery sequence and learning behavior pattern in the personalized learning profile using the collected feedback response data includes: The answers to the interactive questions in the feedback response data are analyzed to determine their correctness, and different mastery score changes are assigned according to the cognitive difficulty of the questions. Based on the change in the mastery score, update the mastery score of the corresponding knowledge point in the knowledge mastery sequence and the last practice timestamp; Analyze the scrolling speed and dwell position patterns in the real-time behavior stream to identify learners’ reading focus intervals and suspected points of confusion, and update the content preferences and difficulty distribution in the learning behavior patterns. Analyze the total duration and interruption information to assess learners' learning endurance and time management habits, and update the learning rhythm characteristics in the learning behavior pattern; The updated knowledge mastery sequence and the learning behavior pattern are smoothly integrated with the unupdated historical records to form an incrementally updated personalized learning profile.

9. The method for intelligent generation of interactive teaching content based on artificial intelligence as described in claim 8, characterized in that, Based on the updated personalized learning profile, a new round of intelligent generation process for interactive teaching content is initiated, including: The incrementally updated personalized learning profile is used as new input to trigger the multimodal learning feature extraction and fusion operation; When generating a new dynamic learner state vector, the feature self-attention network adjusts the importance weights of different feature dimensions according to the new learning behavior pattern. When the new dynamic learner state vector is used to retrieve the preset knowledge graph, the weak knowledge points have changed because the knowledge mastery sequence has been updated, resulting in different core teaching nodes being retrieved compared to before. The content generation engine generates a new round of explanatory texts, interactive questions, and contextualized cases that adapt to the learner's latest state, based on the new dynamic knowledge subgraph and the new dynamic learner state vector.

10. The method for intelligent generation of interactive teaching content based on artificial intelligence as described in claim 3, characterized in that, Calculating the relevance score between the dynamic learner's state vector and the embedded representation of each associated knowledge node in the knowledge graph includes: The dynamic learner state vector is extracted, which represents the learner's current comprehensive knowledge state and learning characteristics; From the preset knowledge graph, obtain the pre-trained vectorized embedding representation of each associated knowledge node; Perform a vector dot product operation on the dynamic learner state vector and the embedded representation of an associated knowledge node to obtain an initial relevance value; The initial correlation values ​​are normalized and mapped to a standardized score range between zero and one. By combining the global importance weights of the associated knowledge nodes recorded in the knowledge graph with the current knowledge learning objective, the standardized score is weighted and corrected to obtain the corrected relevance score; For all associated knowledge nodes initially recalled through semantic similarity matching in the knowledge graph, the vector dot product operation, normalization process and weighted correction steps are repeatedly executed to obtain the relevance score corresponding to each associated knowledge node.

Images