Method for dynamically generating computer professional classroom teaching resources empowered by generative AI

Abstract

This invention relates to the field of educational informatization and intelligent teaching technology, specifically to a generative AI-enabled method for dynamically generating computer science classroom teaching resources. This method acquires teaching objectives and content to construct a knowledge graph; collects classroom data, generates learning gains and emotional representations within a time window, forms a sequence through confidence assessment and denoising, calculates the emotional core coupling index, and gates the core weights; updates the core and emotional weights based on the learning gains and emotional sequence, and applies stability constraints; generates teaching resources and adjusts their order according to emotional weights; combines learning and resource usage to generate evaluations, updates objectives or parameters, and achieves iterative optimization; updates the graph weights through a dual-channel approach of learning gains / emotions, coupling gating, and stability constraints, suppressing the "emotional injection" pollution of core edge weights by interaction heat; and generates resources based on the graph and adjusts their order according to emotional weights, achieving immediate response without deviating from the main course line.

Description

Technical Field

[0001] This invention relates to the field of educational informatization and intelligent teaching technology, and more specifically, to a method for dynamically generating computer science classroom teaching resources empowered by generative AI. Background Technology

[0002] In computer science classes, knowledge points exhibit strong prerequisite dependencies and reasoning chains. For example, topics such as data structure invariants, algorithm complexity analysis, and boundary condition proofs often determine the upper limit of subsequent comprehensive abilities. Current teaching practices employ two main approaches: one organizes resources based on the course syllabus, textbook chapters, or teacher experience, ensuring stable knowledge coverage and teaching pace; another approach introduces classroom process data (such as in-class quizzes, code submission records, classroom interactions, and attention shifts) for post-class evaluation, personalized recommendations, or fine-tuning of the class pace; simultaneously, knowledge networks or knowledge graphs are used to express knowledge points and their relationships, supporting structured organization and resource association; and generative artificial intelligence is increasingly being used to automatically generate explanatory texts, code examples, and visualization materials, thereby reducing resource production costs and improving response efficiency.

[0003] However, when classroom process data is used for real-time strategy adjustments, especially when behavioral signals such as classroom interaction and attention are directly mapped to the edge weights of the knowledge graph and used to drive the order of resource presentation, a more subtle risk of bias may emerge: highly interactive or more engaging content is more likely to receive higher weights in behavioral signals, thus being gradually "pushed to the center" during iterative updates; while core theoretical chains that require continuous reasoning input and have insignificant short-term interaction may be relatively "squeezed out." This phenomenon is not necessarily present in traditional static resource organization, but is more likely to accumulate under a dynamic closed-loop mechanism that simultaneously collects behavioral data and updates weights in real time. This manifests as classroom resources becoming increasingly biased towards surface-level content that is highly discussion-oriented and demonstrative. Students become more proficient in demonstration and reproduction, but their grasp of key invariants, proof chains, and complexity constraints is not solid enough, thus exposing their ability gaps in final comprehensive questions, competition training, or engineering interview transfer tasks.

[0004] Therefore, the industry needs a more robust approach to dynamic classroom management: one that can utilize classroom process data to improve classroom responsiveness and resource organization efficiency, distinguish the boundaries of learning gains and emotional activation during weight updates, suppress the long-term contamination of "core judgments" by short-term popularity through confidence assessment, temporal denoising, coupling gating, and stability constraints, and reverse-correct teaching objectives or update parameters based on classroom evaluation results to ensure that iterative optimization always converges around the core competency objectives of the course. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method for dynamically generating computer science classroom teaching resources empowered by generative AI.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A generative AI-enabled method for dynamically generating computer science classroom teaching resources includes at least the following steps: Step 1: Obtain teaching objectives and content information for computer science classes; Step 2: Construct a classroom knowledge network based on teaching objective information and teaching content information, and then construct a classroom knowledge graph based on the classroom knowledge network; Step 3: During classroom teaching, collect classroom process data through at least one smart terminal; Step 4: Within a preset time window, update and dynamically adjust the core weight and sentiment weight of each edge in the classroom knowledge graph through two channels, specifically including: Constructing learning gain representations and emotion representations based on classroom process data; The confidence scores of the learning gain representation and the emotion representation are evaluated and normalized to obtain the learning confidence score and the emotion confidence score. Then, exponential smoothing or recursive estimation is used to perform temporal denoising on the learning gain representation and the emotion representation to obtain the learning gain sequence and the emotion activation sequence. Calculate the core coupling index of emotion and update the gating of the core weights based on the core coupling index of emotion; The core weights are updated based on the learning gain sequence, and the emotion weights are updated based on the emotion activation sequence, with stability constraints introduced in the update process. Step 5: Generate a set of teaching resources based on the nodes, edges, and core weights of the classroom knowledge graph, and adjust the presentation order of the teaching resource set based on the emotion weight; Step 6: Generate classroom evaluation results based on the learning process data and the usage of teaching resources. The classroom evaluation results should include at least the knowledge point mastery index and the error distribution index. Update the teaching objective information or the update parameters of the classroom knowledge graph according to the classroom evaluation results to drive the cyclical execution of steps 3 to 5 in the next time window, thereby realizing the dynamic generation and dynamic control of classroom teaching resources.

[0007] Furthermore, the teaching objectives information should at least include the knowledge points, ability objectives, and prerequisite constraints of this lesson, and the teaching content information should at least include the key points of the lesson plan, classroom tasks, and sample code provided by the teacher.

[0008] Furthermore, step 2 involves constructing a classroom knowledge network by: using generative artificial intelligence to perform semantic analysis on the teaching content information, extracting concept nodes, grammar nodes, and algorithm nodes, and constructing a classroom knowledge network based on the semantic relationships between the nodes.

[0009] Furthermore, the classroom knowledge graph includes nodes and edges between nodes, and initializes core weights and emotion weights for each edge.

[0010] Furthermore, classroom process data includes at least classroom interaction data, classroom attention data, and learning process data; among which, classroom interaction data includes at least the number of questions asked, the number of discussions, and the number of collaborations; classroom attention data includes at least the duration of gaze or level of focus; and learning process data includes at least the code writing log, execution results, and error type distribution.

[0011] Furthermore, the learning gain representation should at least represent at least one of the following: accuracy change, error convergence speed, and completion rate of key concept-related tasks corresponding to the learning process data within the time window; the emotion representation should at least represent at least one of the following: participation intensity, attention stability, and volatility corresponding to the classroom interaction data and classroom attention data within the time window.

[0012] Furthermore, the emotion core coupling index is used to characterize the strength and consistency of the influence of the emotion activation sequence on the learning gain sequence. The emotion core coupling index is obtained by at least a combination of correlation measures, mutual information measures, or coupling coefficients based on the state space model, and is weighted and corrected by the learning confidence and emotion confidence.

[0013] Furthermore, when the emotional core coupling index exceeds the adaptive gating threshold and the emotional confidence level is higher than the preset confidence threshold, decoupling control is performed. Decoupling control includes at least: pausing or reducing the contribution of classroom interaction data and classroom attention data to core weight updates, and limiting core weight updates to be driven by the learning gain sequence. When the emotional core coupling index does not exceed the adaptive gating threshold or the emotional confidence level is not higher than the preset confidence threshold, collaborative control is performed, allowing emotional weights to participate in the optimization of the teaching resource presentation order while maintaining core weight updates according to the learning gain sequence.

[0014] Furthermore, stability constraints include at least one of the following: upper limit of weight change, entropy constraint of weight change, or drift penalty, in order to suppress the amplification of short-term fluctuations on weight updates.

[0015] Furthermore, the teaching resource set includes at least code examples, layered explanatory texts, and visualization resources; when decoupling control is performed in step 4, the generative artificial intelligence is further controlled to generate bridging teaching resources. The bridging teaching resources are used to associate high-popularity nodes ranked by emotion weights with core nodes determined by core weights. The bridging teaching resources include at least a derivation explanation from phenomenon to principle, a minimum runnable code experiment containing core node knowledge points, or an algorithm logic flowchart.

[0016] Compared with the prior art, the present invention has the following beneficial effects: In step 4 of this invention, the classroom process data is split into two channels: learning gain representation and emotional representation. Confidence assessment, normalization, and exponential smoothing / recursive estimation are then performed to denoise the data, forming a learning gain sequence and an emotional activation sequence. Subsequently, the emotional core coupling index is calculated, and gating is applied to the core weight update. This ensures that the core weight is robustly driven by the learning gain sequence, while the emotional weight is only used to reflect classroom engagement and attention. This mechanism reduces the erroneous update of "high popularity ≠ high core value," and avoids the knowledge structure drift dominated by entertainment content after multiple time window iterations. This invention generates a set of teaching resources using nodes, edges, and their core weights in a classroom knowledge graph. This ensures that resource generation aligns with knowledge dependencies, guaranteeing that key theoretical chains (such as invariants, complexity, and boundary conditions) remain central to the teaching process. Simultaneously, it utilizes emotional weights to adjust the order of resource presentation, enabling the system to organize and explain current classroom focus points as "entry paths" without directly altering the core judgments of knowledge relationships. Combined with the stability constraints in step 4, this reduces frequent jumps and strategy oscillations caused by short-term interaction fluctuations, thereby simultaneously achieving "explainable consistency of the main line" and "agile adaptation to classroom conditions." This invention incorporates learning process data and teaching resource usage into classroom evaluation, outputting evaluation results that include at least knowledge point mastery indicators and error distribution indicators. Based on these results, it updates teaching objective information or classroom knowledge graph update parameters, driving the cyclical execution of steps 3 through 5 in the next time window. This closed loop shifts the system optimization focus from "more interactive / interesting" to "improved mastery and error convergence." When error clusters or insufficient mastery of key edges occur, it can make core edge weight updates more sensitive and resource generation more focused. When mastery meets the standard but fluctuates significantly, it can suppress the impact of short-term noise, avoiding the precipitous drop of "being able to talk and demonstrate but having a broken theoretical chain," thus improving transferable ability performance in final comprehensive questions, competition tasks, and engineering interview scenarios. Attached Figure Description

[0017] Figure 1 A flowchart illustrating the method for dynamically generating classroom teaching resources for computer science that empowers generative AI; Figure 2 A schematic diagram for constructing a classroom knowledge graph and initializing dual weights; Figure 3 This is a schematic diagram of dual-channel update and dynamic control within a preset time window; Figure 4 This diagram illustrates the closed-loop iteration of teaching resource generation, emotion regulation, and evaluation. Detailed Implementation

[0018] Reference Figures 1 to 4 A generative AI-enabled method for dynamically generating computer science classroom teaching resources includes at least the following steps: Step 1: Obtain teaching objectives and content information for computer science classes; this provides input boundaries and constraints for subsequent knowledge structure modeling and teaching resource generation. By obtaining teaching objectives, the scope of knowledge points to be covered, the target skills, and prerequisites for the lesson can be clearly defined. By obtaining teaching content information, the main teaching line, classroom tasks, and example code can be determined, ensuring that the subsequent analysis, generation, and control of generative AI revolve around the established teaching objectives and content information, preventing resource generation from deviating from the intended teaching objectives. In one specific implementation, the teaching objective and content information for a computer science class can be configured by the instructor before class and confirmed at the start of the class. The teaching objective information includes at least the knowledge points, skill objectives, and prerequisite constraints for the lesson. The teaching content information includes at least the key points of the lesson plan, classroom tasks, and sample code provided by the instructor. For example, for the topic of binary heaps and priority queues in a data structures course, the teaching objective information could be set as follows: the knowledge points for the lesson include the definition of a heap, up-filtering and down-filtering, the heapification process, and the priority queue interface; the skill objectives include being able to analyze the time complexity of a heap and independently implement min-heap insertion and deletion operations; and the prerequisite constraints include array indexing rules, binary tree basics, and comparison operations. The teaching content information could further provide key points such as introducing priority handling from a task scheduling scenario, transitioning to heap properties using a complete binary tree array representation, and then implementing the details using a heapification algorithm. The classroom task could be set as completing a priority queue that supports push, pop, and top operations and using it to solve the problem of finding the top K smallest elements. The sample code could provide a min-heap skeleton stored in an array, a down-filtering loop with key boundary conditions, and exception handling. In this way, the teaching objectives and teaching content are clearly defined and structured within the same semantic framework, providing sufficient and traceable input for subsequent steps to construct a classroom knowledge network based on the teaching objectives and teaching content, and to construct a classroom knowledge graph based on the classroom knowledge network. At the same time, it ensures that classroom tasks and example code strictly correspond to the knowledge points and ability objectives of this lesson.

[0019] Step 2: Construct a classroom knowledge network based on teaching objective information and teaching content information, and then construct a classroom knowledge graph based on this network. This transforms the previously scattered teaching objective and content information into a computable, associative, and updatable structured knowledge representation, providing a unified knowledge carrier for mechanisms such as weight updates, resource generation, and presentation order adjustments. Constructing the classroom knowledge network makes the relationships between knowledge points explicit. Further constructing the classroom knowledge graph allows for the carrying of core weights and emotional weights at the node and edge levels, providing a recordable, comparable, and iteratively updatable object foundation for dynamic characteristics such as "core focus" and "popularity / attention" in classroom teaching. In one specific implementation, after acquiring teaching objective information and teaching content information, the next step is to construct a classroom knowledge network based on these two pieces of information. This involves using generative artificial intelligence (AI) to perform semantic analysis on the teaching content information, extracting concept nodes, grammar nodes, and algorithm nodes, and constructing the classroom knowledge network based on the semantic relationships between these nodes. Specifically, generative AI performs multi-level semantic decomposition on the key points of the teacher's lesson plan, classroom tasks, and example code. It identifies abstract theoretical terms in the text descriptions as concept nodes; for example, when teaching quicksort, it identifies concept nodes such as divide-and-conquer, pivot element, and recursive structure. Simultaneously, it performs grammatical hierarchical analysis on the program structure in the example code, extracting loop structures, conditional statements, recursive calls, and array access as grammar nodes. Furthermore, it combines the core solution processes involved in the classroom tasks, extracting processes such as interval partitioning, comparison and swapping, and recursion termination checks as algorithm nodes. Based on contextual dependencies, causal relationships, and implementation dependencies, generative AI establishes theoretical support relationships between concept nodes and algorithm nodes, establishes implementation mapping relationships between algorithm nodes and grammar nodes, and forms directional semantic relationships, thereby constructing the classroom knowledge network. For example, in the teaching scenario of quicksort, the divide-and-conquer approach can be used to support recursive calls, which in turn rely on a function stack structure. The division of intervals depends on multiple layers of relationships, including loops and conditional statements. Through this approach, the classroom knowledge network not only reflects the superficial parallel relationships between knowledge points but also embodies the deep semantic coupling between theory, implementation, and structure. This lays a structured foundation for subsequently constructing a classroom knowledge graph based on the classroom knowledge network and for dual-channel updating and dynamic control of core weights and sentiment weights. In one specific implementation, after constructing the classroom knowledge network, a classroom knowledge graph is further constructed based on the classroom knowledge network. The classroom knowledge graph includes nodes and edges between nodes, and core weights and sentiment weights are initialized for each edge. Specifically, the concept nodes, syntax nodes, and algorithm nodes in the classroom knowledge network are uniformly mapped to the node set of the classroom knowledge graph, and the semantic relationships between nodes are mapped to the edge set of the classroom knowledge graph. The direction of the edges is used to represent the sequential relationship of dependencies or derivations, and the type of the edges is used to distinguish different semantic relationships such as theoretical support, implementation mapping, and step dependencies. To ensure a comparable benchmark for the dual-channel updating and dynamic adjustment of the core weight and emotional weight of each edge in the classroom knowledge graph within the preset time window, the initial core weight can be set based on the knowledge points and ability objectives of the lesson in the teaching objective information. For example, edges that are directly related to the ability objectives and appear more frequently in classroom tasks can be given higher initial core weights. The initial emotional weight can be set based on the difficulty prompts of classroom tasks, teacher emphasis marks, and common error annotations in the example code in the teaching content information. For example, edges that correspond to common error annotations can be given higher initial emotional weights to reflect potential attention. For example, in the topic of binary heaps and priority queues, heap properties, upper filtering, lower filtering, heapification process, and array representation can be mapped as nodes. Heap properties to lower filtering, array representation to upper filtering, and heapification process to priority queue interface can be mapped as edges. The edges from heapification process to lower filtering are given a higher initial core weight to reflect their importance to the capability goal. At the same time, the edges from upper filtering to boundary condition processing are given a higher initial sentiment weight to reflect their characteristics of easily attracting attention and discussion in the example code. Thus, the classroom knowledge graph has a structured semantic relationship and an updatable dual weight foundation at the beginning of the class.

[0020] Step 3: During classroom teaching, collect classroom process data through at least one smart terminal; this provides a real, continuous, and quantifiable source of classroom evidence for the dynamic updating of the classroom knowledge graph, enabling subsequent weight updates and resource generation to reflect the actual classroom situation rather than relying solely on static lesson plans. Collecting classroom process data through at least one smart terminal can cover multiple dimensions such as classroom interaction, classroom attention, and the learning process, allowing the system to obtain a process-oriented description of phenomena such as how well students are learning, where they are struggling, and their level of classroom participation. This lays the data foundation for the next step of constructing learning gain representations and emotional representations. In one specific implementation, during classroom teaching, classroom process data is collected through at least one smart terminal to ensure that the core weight and emotional weight of each edge in the classroom knowledge graph are updated and dynamically adjusted through dual channels within a preset time window, providing a continuous and quantifiable basis. The classroom process data includes at least classroom interaction data, classroom attention data, and learning process data. The classroom interaction data includes at least the number of questions asked, the number of discussions, and the number of collaborations. The classroom attention data includes at least the gaze duration or concentration level. The learning process data includes at least the code writing log, running results, and error type distribution. Specifically, instructors can organize students to complete classroom tasks using their personal computing terminals and record code writing logs and execution results within the same time window. The pass / fail status, execution time, test case coverage, and error type distribution (compilation errors, runtime errors, logical errors, etc.) of each run can be used as learning process data. Simultaneously, combining classroom activities such as raising hands to ask questions, oral discussions, and group paired programming, the number of questions, discussions, and collaborations can be statistically analyzed within a time window to create classroom interaction data. Furthermore, students' attention span and switching behavior regarding teaching content can be recorded on their personal computing terminals or classroom interactive screens, forming classroom attention data in the form of gaze duration or focus level. For example, when teaching the topics of binary heaps and priority queues, if students raise multiple questions and discuss the lower filter boundary condition within a certain time window, resulting in an increase in the number of questions and discussions in the classroom interaction data, and a concentrated distribution of error types such as array out-of-bounds errors and loop termination condition errors in the learning process data, then the classroom process data within that time window can be considered to have fully revealed the learning status and classroom attention status of the relevant implementation links of this knowledge point. This provides traceable and fine-grained data support for subsequently constructing learning gain representations and emotional representations.

[0021] Step 4: Within a preset time window, the core weight and emotional weight of each edge in the classroom knowledge graph are updated and dynamically controlled through dual channels. Specifically, this includes transforming classroom process data into a dynamic depiction of knowledge relationships at the time window granularity, enabling parallel advancement of "learning-driven core updates" and "classroom-state-driven emotional updates," and controlling the coupling strength between the two through a gating mechanism to ensure the interpretability and controllability of the updates. This step ensures that the classroom knowledge graph is no longer a static structure built only once, but continuously evolves during the classroom process according to changes in learning gains and emotional states, thus providing a real-time basis for the dynamic generation and presentation order adjustment of subsequent teaching resources. In one specific embodiment, to achieve traceable attribution of "classroom process data → each edge in the classroom knowledge graph", this embodiment establishes an edge-evidence mapping table M before the start of class, and performs edge-level aggregation of classroom process data according to M within each preset time window. The mapping table M includes at least: edge identifier e(i,j), corresponding knowledge point / concept node set, corresponding classroom task step set, corresponding unit test set, and corresponding typical error label set.

[0022] The unit test set can be obtained by teachers breaking down classroom tasks before class and setting minimum verification test cases for each key concept; the typical error label set can be obtained by statistical analysis of historical classroom error distribution or by generative artificial intelligence performing semantic clustering of example code and common student error descriptions, and then solidified after teacher confirmation.

[0023] At the end of each time window t, the system aggregates the running results, error type distribution, and key step completion marks in the learning process data according to the mapping table M, into side-level learning evidence E_learn(e,t); and aggregates classroom interaction data and attention data into side-level emotion evidence. This ensures that subsequent learning gain representations, emotion representations, coupling indexes, and weight updates can all be implemented and reproduced at the level of "every edge".

[0024] In one specific implementation, constructing learning gain representations and emotion representations based on classroom process data can be performed as follows: Within a preset time window, learning process data is aligned and aggregated. Code writing logs, running results, and error type distributions are merged according to individual students and classroom tasks. A sequence is established for multiple runs of the same task to characterize changes in accuracy and error convergence speed. For example, in classroom tasks on binary heaps and priority queues, the unit test pass rate, number of compilation errors, number of runtime errors, and number of logical errors for each submission are sorted by time to obtain a comparable trajectory over time. Construct a learning gain representation that represents at least one of the following: accuracy change, error convergence speed, and completion of key concept-related tasks within the time window: for example, set checkpoints for key steps related to the down-filtering operation, and count whether boundary condition processing, minimum child node selection, and loop termination determination are completed, thereby forming the completion of key concept-related tasks. Within the same preset time window, classroom interaction data and classroom attention data are simultaneously calibrated. The number of questions, discussions, and collaborations are categorized by teaching segments, and the mean, variance, and continuity indicators of gaze duration or focus are calculated by time window to characterize participation intensity, attention stability, and volatility. For example, when explaining the stacking process, if the number of discussions and collaborations increases significantly and the gaze duration remains high and has low variance within the time window, then participation intensity is considered high and attention stability is good. Conversely, if the gaze duration drops sharply multiple times and is accompanied by a concentrated burst of questions, then volatility is considered high. Constructing an emotional representation requires at least one of the following: participation intensity, attention stability, and volatility, corresponding to classroom interaction data and classroom attention data within a time window. For example, during the explanation of the priority queue interface, if students raise concentrated questions about the boundary conditions in the example code and the attention curve shows high-frequency fluctuations, the emotional representation can output higher participation intensity and higher volatility to support subsequent confidence assessment and normalization of the learning gain representation and emotional representation, and to enter the temporal denoising process.

[0025] In one specific embodiment, for either side In the time window Built-in learning gain representation With emotional representation The following definition can be used.

[0026] 1) Learning gain representation :make This represents the pass rate of the unit test corresponding to the edge at the end of the window. If the initial pass rate is given, then the accuracy change term is: ;make The error convergence term is the number of errors occurring in the typical error label set corresponding to the edge within the window (which can be counted separately for compilation errors, runtime errors, and logical errors, and then weighted and summarized). ;make If the proportion of key steps completed on the corresponding edge within the window is given, then the learning gain representation can be defined as follows: ;in and Example of possible values: , , Because of: changes in accuracy. The most direct reflection of mastery improvement has the highest weight; error convergence (S) reflects the efficiency of error correction in the learning process, and is second; the completion rate of key steps (c) is easily affected by the granularity of task breakdown and fluctuates greatly, so its weight is relatively low.

[0027] 2) Emotional representation :make This represents the number of questions / discussions / collaboration events corresponding to that side within the window (weighted by event type). The average level of focus. For the variance of attention, To determine the number of times focus "drops" sharply, the participation intensity, attention stability, and fluctuation items can be defined as follows: ; ; And take emotional representation as ; acceptable , , , , , , Because of: the number of classroom interaction events. It can quickly reflect "participation intensity," but it is easily affected by topical noise, therefore... Medium; average focus level is more critical for consistent engagement, assigning a higher value. Variance and sudden drops characterize distraction and volatility; to suppress situations where there is "excitement but deviation from the main trend," a penalty term should be retained. , However, it should not be too large. Composite weights At the level of participation, both intensity and stability jointly determine the usable "classroom state," and both should be given equal importance. The fluctuation term is used more for fine-tuning the rhythm, therefore Slightly lower.

[0028] In one specific implementation, confidence assessment and normalization are performed on the learning gain representation and the sentiment representation, and temporal denoising is completed using exponential smoothing or recursive estimation. This can be done in the following steps: Within a preset time window, interpretable confidence assessment criteria are established. Learning confidence is calculated for learning gain representation from two aspects: data sufficiency and consistency. For example, the completeness of code writing logs, the number of valid execution results, and the coverage of error type distribution are used as sufficiency criteria, while the consistency criteria are whether the accuracy changes of the same classroom task within adjacent windows show reasonable continuity. For emotion representation, emotion confidence is calculated from two aspects: sampling reliability and group representativeness. For example, the missing proportion of gaze duration or focus and the proportion of abnormal jumps are used as sampling reliability criteria, while the dispersion of classroom interaction data among different groups is used as group representativeness criteria. Thus, learning confidence and emotion confidence are obtained within the same semantic framework. Normalization is applied to learning gain representations and emotion representations to make values ​​comparable across different classes, task difficulties, and terminal sampling scales. Learning gain representations can be normalized using minimum-maximum boundary normalization based on historical windows or normalization based on robust quantiles to compress accuracy variations, error convergence speed, and key concept-related task completion rates into a uniform numerical range. Emotion representations can be normalized in segments based on classroom time, including participation intensity, attention stability, and volatility, allowing for relative comparisons between discussion-intensive and lecture-intensive segments. Exponential smoothing or recursive estimation is used to perform temporal denoising on the normalized learning gain representation and emotion representation. Exponential smoothing can adaptively set the smoothing coefficient according to the learning confidence and emotion confidence, so that more new information is retained when the confidence is high and the smoothing strength is increased to suppress noise when the confidence is low. Recursive estimation can regard the representation of adjacent time windows as a state transition and introduce the observation error term, so that short-term abnormal peaks are gradually absorbed by the recursive constraint, thereby obtaining a more stable learning gain sequence and emotion activation sequence. Output and verify the rationality of the learning gain sequence and the emotion activation sequence. Compare the sequence with the key nodes of the classroom task. If there is a sudden increase in the learning gain sequence but the learning confidence is consistently low, backtrack and check whether there is batch copying or abnormal recording of the running results in the code writing log. If there is a sudden drop in the emotion activation sequence but the emotion confidence is high, combine classroom attention data to confirm whether the reduced interaction is due to entering the independent coding stage.

[0029] In one specific embodiment, confidence is learned. It can be composed of the sufficiency and consistency of learning data.

[0030] 1) Learning confidence Sufficiency is determined by the number of valid runs within the window. Minimum requirements The ratio is used to characterize consistency; consistency is represented by the change in accuracy between adjacent windows. and The differences between them can be characterized as follows. ;in This is the consistency attenuation coefficient.

[0031] 2) Emotional confidence Emotion confidence can be determined by the sampling missing rate. With abnormal jump rate Composition, for example ;in This represents the abnormality suppression coefficient.

[0032] 3) Robust normalization treatment: for and respectively adopt based on the nearest A robust quantile normalization method for each time window, with quantiles and quantiles Linear scaling is applied to the boundary, and the result is truncated to the interval. .

[0033] 4) Adaptive Exponential Smoothing Denoising: Adaptive exponential smoothing is applied to the learning gain and the emotional representation, defined as follows: ; The smoothing coefficient is adjusted by the corresponding confidence level. ; ;and Based on the smoothing parameter. (Preferred) , Because: learning evidence (unit testing, error convergence) is relatively stable and more critical to core updates, the smoothing coefficient should be larger to follow the real improvement more quickly; sentiment data is more susceptible to occasional interactions, noise and sampling fluctuations, so the basic smoothing coefficient should be smaller to enhance noise filtering and stability, and avoid short-term "excitement" driving strategy oscillations.

[0034] In one specific implementation, to characterize the strength and consistency of the influence of the emotion activation sequence on the learning gain sequence, the emotion core coupling index can be constructed by the following steps: Within a preset time window, the learning gain sequence and the emotion activation sequence are time-aligned and scale-matched. Paired sequences are established according to the same sampling granularity, and lag terms are introduced to cover the common propagation delay of emotion preceding learning performance. For example, zero lag, one time window lag, and two time window lag are set to form a candidate alignment set, so that subsequent measurements can reflect both the synchronous effect and the delay effect. Calculate at least one correlation measure to characterize linear or monotonic consistency, preferably using a robust correlation measure to reduce the perturbation of results by outlier windows, and take the maximum consistency under different lag configurations as the candidate correlation score; for example, in the topic of binary heaps and priority queues, students show an increase in emotional activation during the explanation of the lower filter boundary conditions, and the increase in learning gain is more obvious in the subsequent time window. In this case, the correlation score under one-window lag may be higher than the score under zero lag, which is more in line with the cognitive patterns in the classroom. The mutual information metric or the coupling coefficient based on the state-space model is used to characterize nonlinear effects and dynamic transmission. The mutual information metric can be calculated by discretizing the sequence by quantiles or estimating the kernel density to obtain information gain. It is used to identify nonlinear relationships where high emotional activation does not necessarily correspond to high learning gain, but it significantly improves within a specific interval. The coupling coefficient based on the state-space model can treat the learning gain sequence as a state variable and the emotional activation sequence as an exogenous driver or observation factor. The driving term coefficient and its stability index are obtained by recursion estimation. It is used to characterize the dynamic contribution and consistency of emotion to learning. For example, when there is frequent classroom interaction but large fluctuations in attention, the model may show a positive coupling coefficient but a large variance, indicating that the influence is strong but the consistency is insufficient. The correlation measure, mutual information measure, and coupling coefficient are comparablely fused and weighted by learning confidence and sentiment confidence to form the final sentiment core coupling index. The fusion can employ a weighted summation or multiplicative consistency penalty strategy to ensure that the three measures contribute collectively within a unified numerical range. Simultaneously, the geometric mean or weighted average of learning confidence and sentiment confidence is used as a correction factor to attenuate the coupling contribution of low-confidence windows. For example, in quicksort teaching, if a missing code writing log in a certain time window leads to a low learning confidence, even if a surge in discussion frequency within that window results in high sentiment activation, the corrected sentiment core coupling index will still be suppressed to avoid misjudging data gaps as strong coupling. Conversely, when both learning confidence and sentiment confidence are high, and the mutual information and state-space model coupling coefficients are simultaneously positive and stable, the sentiment core coupling index can significantly increase. This provides an interpretable and noise-resistant coupling basis for subsequent dual-channel updates and dynamic adjustments of the core weights and sentiment weights of each edge in the classroom knowledge graph within a preset time window.

[0035] In one specific embodiment, for each edge In the time window Calculate the core coupling index of emotion First, the learning gain representation after denoising is processed. With emotional representation Introducing lag (Preferred) (Each window has a candidate coupling score, and the maximum score is taken as the edge's position within that window). The coupling index is a basic term.

[0036] 1) Construction of candidate coupling score: (1) Correlation term: for sequence and Lag The Spearman rank correlation coefficient is calculated and denoted as . (2) Mutual information item: for and Perform quantile discretization (e.g., by quantiles) (discrete individual quantiles), calculation of lag. The mutual information below is denoted as (3) Dynamic coupling terms: use a first-order state-space model or recursive regression to estimate the lag. Dynamic mapping relationship ;in The coupling coefficients are obtained through recursive estimation. As residuals; and with The fluctuation of the coupling coefficient within the recent window is used to construct a stability suppression term.

[0037] 2) Coupling index It can be defined as: ;in and , and These are learning confidence and emotional confidence, respectively. The definitions above use... Used to automatically select the one that best reflects "emotion" Learning the lag window of coupling relationships. (This is a possible approach.) , , .because: It is most robust to monotonic coupling and insensitive to dimensions, making it suitable as the dominant term. It can supplement nonlinear dependencies but is sensitive to discretization and sample size, and has slightly lower weights; dynamic coupling terms It reflects the directionality and time-varying relationship of "emotion-driven learning", but it is estimated to be susceptible to noise, so a moderate weight is taken and suppressed by stability.

[0038] 3) Adaptive gating threshold and confidence threshold: Adaptive gating threshold Based on the nearest edge The coupling index distribution of each window is given, for example, using a quantile threshold: ;in Preferred selection To adapt to the fluctuations in different classroom segments; and to set preset thresholds. (For example ), used in or When the value is below the threshold, the coupling determination strength of the window is suppressed or reduced.

[0039] In one specific implementation, to perform dual-channel updates and dynamic adjustments to the core weight and emotion weight of each edge in the classroom knowledge graph within a preset time window, an adaptive gating threshold is set for the emotion core coupling index, and decoupling and collaborative adjustments are achieved by combining it with a preset information threshold. This can be done as follows: The basis for setting the adaptive gating threshold and the preset confidence threshold is determined. The adaptive gating threshold preferably refers to the statistical process control and adaptive threshold setting methods commonly used in the existing technology, and is dynamically given based on the distribution of the core emotional coupling index of several recent time windows. For example, a sliding window mean plus standard deviation or a quantile threshold is used to adapt to the fluctuation differences of different classes and different classroom segments. The preset confidence threshold preferably refers to the common practices of data quality control and educational data collection integrity requirements, and is set according to reliability indicators such as gaze duration or attention deficit ratio, verifiable interaction event ratio, and terminal record consistency. This makes it difficult for the emotional confidence level to exceed the threshold when data is missing or abnormally jumps, thereby meeting the requirements of traceability and interpretability for compliant acceptance. At the end of each preset time window, a gating decision is made. If the core emotional coupling index exceeds the adaptive gating threshold and the emotional confidence level is higher than the preset confidence threshold, decoupling control is triggered. If the core emotional coupling index does not exceed the adaptive gating threshold or the emotional confidence level is not higher than the preset confidence threshold, collaborative control is triggered. The gating decision should also consider the maximum coupling result under the lag configuration to avoid missed judgments due to the propagation delay caused by the influence of emotions. When decoupling control is triggered, the contribution of classroom interaction data and classroom attention data to the core weight update is paused or weakened, and the core weight update is limited to being driven by the learning gain sequence. Specifically, a gating factor can be introduced into the core weight update to reduce the coefficients of the interaction contribution item and the attention contribution item to zero or decrease proportionally. At the same time, the independent update of the emotion weight is retained to reflect the classroom attention state but no longer pulls the core weight in the reverse direction. For example, in the teaching of quick sorting, if the number of discussions and collaborations surges and the gaze duration fluctuates more due to the teacher raising open-ended questions in a certain time window, the emotion activation sequence rises and becomes abnormally strongly coupled with the learning gain sequence, and the emotion confidence is high, then the decoupling control will drive the core weight update only by the change in accuracy and the error convergence speed, so as to avoid misjudging the intensity of short-term lively but not necessarily improved interaction as an increase in the importance of knowledge edges, thereby preventing the core weight from being amplified by emotional noise. When triggering collaborative regulation, emotional weights are allowed to participate in optimizing the presentation order of teaching resources while maintaining the core weights updated according to the learning gain sequence. Specifically, emotional weights can be used to fine-tune the order of explanation, example comparison, and practice gradient. For example, examples or supplementary exercises corresponding to edges with higher emotional weights and greater volatility can be presented first to stabilize attention and increase participation. However, the core weights are still strictly updated according to the learning gain sequence to maintain the stability of the knowledge importance judgment. For example, in the topics of binary heaps and priority queues, if the emotional core coupling index does not exceed the adaptive gating threshold or the emotional confidence is low, collaborative regulation can first present a visual demonstration of the lower-filter boundary conditions and targeted test cases based on the emotional weights to reduce volatility. At the same time, the core weights continue to be driven by changes in the completion and accuracy of key concept-related tasks, making the presentation order of teaching resources more in line with the classroom situation without changing the learning-driven principle of the core weights.

[0040] Core weights are updated based on the learning gain sequence, while emotion weights are updated based on the emotion activation sequence, with stability constraints introduced during the updates. The driving sources and boundaries of the dual-channel updates are clearly defined, ensuring that core weights and emotion weights each correspond to different classroom evidence, and maintaining the stability and continuity of results during dynamic updates. Core weights are updated based on the learning gain sequence, ensuring that they focus on the mastery and improvement reflected in the learning process data; emotion weights are updated based on the emotion activation sequence, ensuring that they reflect the changes in intensity reflected in classroom interaction and attention. The introduction of stability constraints during updates maintains a controllable pace of change within time windows, avoiding unnecessary jumps in the overall knowledge graph evolution caused by short-term anomalies.

[0041] In one specific embodiment, a gating factor is defined to suppress core edge weight drift caused by "classroom emotion injection". It is composed of the emotional core coupling index With emotional confidence A joint decision.

[0042] 1) Hard gating (example): when and This triggers decoupling regulation (inhibiting the reverse contamination of the core by emotions), causing... Otherwise, it will trigger coordinated regulation, causing .

[0043] 2) Soft gating (example): or use soft gating. ;in For the Sigmoid function, For the slope parameter, This is an indicator function.

[0044] 3) Dual-weight maintenance and update template: Maintain core weights for each edge. With Emotion Weight These are used to represent "core knowledge structure" and "intensity of classroom attention / engagement" respectively, to ensure "high engagement". "High-core".

[0045] (1) Core weight update: Driven solely by learning gain, the basic update of core weights can be defined as And apply a single-window change cap (clip): Further apply a drift penalty (pull back to the moving average). ): ;in Get the most recent The mean of each window, For parameters.

[0046] (2) Emotion weight update: Reflects attention / investment, but does not contaminate the core. Emotion weight update can be defined as And also apply a change limit. (e.g., using) (Constraints). Optionally, introduce a "change entropy / flip count constraint": when the most recent Within each window The number of times the direction of change and flipping exceeds the threshold Automatically lower To suppress oscillations.

[0047] (3) Key restrictions during decoupling: prohibit interactive / attention items from entering core updates, when near At that time, any interaction / attention-related items should be prohibited from entering the core update. If interaction-enhancing items exist in the system implementation... Then it should be multiplied by This would cause it to fail during decoupling, for example... This avoids "Engagement-weight Poisoning" from pushing the "more fun / more exciting" side errors into the core of the knowledge structure.

[0048] In one specific implementation, to update the core weights of each edge in the classroom knowledge graph according to the learning gain sequence and the emotion weights according to the emotion activation sequence, and to introduce stability constraints in the update process to suppress the amplification of short-term fluctuations, the following steps can be performed: At the end of each preset time window, edge-level mapping and driver term generation are performed on the learning gain sequence and the emotion activation sequence. That is, learning gain indicators such as the completion degree of knowledge point tasks, the change in accuracy and the error convergence speed corresponding to a certain edge are aggregated into the core update driver, and emotion activation indicators such as participation intensity, attention stability and volatility are aggregated into the emotion update driver. The driver terms are normalized on the same scale to ensure the comparability between different edges. The core weight is updated using an incremental update strategy driven by the learning gain sequence. This strategy adjusts the core weight when the learning gain increases and adjusts or maintains the core weight when the learning gain decreases or stagnates. At the same time, at least one stability constraint is introduced into the update term to limit instantaneous jumps. For example, an upper limit is set for weight changes so that the increase or decrease of the core weight within a single time window does not exceed a preset upper limit, or a drift penalty is introduced so that a penalty term is generated when the core weight deviates from its recent moving average, thereby pulling back the trend. The emotion weight update is performed using an isomorphic update strategy driven by the emotion activation sequence. This strategy gradually increases the emotion weight as emotion activation continues to rise and gradually decreases it as emotion activation falls. Stability constraints are also introduced to prevent drastic fluctuations in emotion weight caused by discussion outbreaks or sudden drops in attention. Among these, the weight change entropy constraint can be considered as a preferred approach. This involves imposing constraints on the distribution of the direction and magnitude of weight changes within several adjacent time windows, making the update more inclined to a small, continuous adjustment rather than frequent, large-scale reversals, thereby improving interpretability and reproducibility. The stability constraints of core weights and sentiment weights are jointly verified and written back. If it is detected that a certain edge experiences a sharp increase in both core weight and sentiment weight in a short period of time, but the learning gain sequence does not show a sustained improvement, a stronger drift penalty or a stricter upper limit on weight changes is triggered to suppress the amplification of short-term popularity on weight updates. For example, in the topics of binary heaps and priority queues, if the filtering boundary conditions cause a large number of questions in a short period of time, resulting in a spike in the sentiment activation sequence, the sentiment weight should be adjusted accordingly to prioritize the presentation of boundary condition examples and test cases. However, since the learning gain sequence often shows a stable improvement only after several time windows, the core weight update should mainly follow the continuous improvement of error convergence speed and accuracy. The upper limit on weight changes and the drift penalty limit single-window jumps to avoid misjudging a peak in classroom discussion as a significant long-term increase in the coreness of the edge. This ensures that the weight evolution is smooth and robust while responding to the classroom state.

[0049] Step 5: Generate a set of teaching resources based on the nodes, edges, and core weights of the classroom knowledge graph, and adjust the presentation order of the teaching resources based on emotional weights. This transforms the structural information of the classroom knowledge graph and the "importance" of the core weights into teaching resource outputs that can directly serve classroom teaching. The emotional weights are used to adapt the presentation rhythm, thereby achieving dynamic coordination between "what to teach" and "what to teach first." Generating a set of teaching resources based on the nodes, edges, and core weights of the classroom knowledge graph allows for the organization of explanatory materials and examples around the core knowledge relationships, ensuring that the resource content aligns with the classroom knowledge structure. Simultaneously, adjusting the presentation order of the teaching resources based on emotional weights prioritizes knowledge points with high current classroom attention or significant changes in participation, ensuring that classroom progress is more aligned with the real-time classroom situation without deviating from the core theme, thus improving classroom interaction efficiency and knowledge absorption efficiency. In one specific implementation, a set of teaching resources is generated based on the nodes, edges, and core weights of the classroom knowledge graph. The presentation order of the teaching resource set is adjusted based on emotion weights. Furthermore, when decoupling control is triggered, generative artificial intelligence is used to generate bridging teaching resources. This can be executed in the following steps: Based on the semantic definitions of nodes and edges in the classroom knowledge graph and the core weight distribution, a set of teaching resources is generated. Each edge is mapped to an executable learning path unit, and each path unit is configured with code examples, hierarchical explanatory texts, and visualization resources. The code examples cover at least two versions: correct implementation and typical incorrect implementation. The hierarchical explanatory texts include at least a concept definition layer, a reasoning link layer, and a boundary condition layer. The visualization resources include at least a state change sequence or a key variable trajectory to ensure that learners can understand the same knowledge relationship from multiple perspectives. The presentation order of teaching resources is adjusted based on emotional weights, prioritizing the presentation of resources associated with high-intensity nodes that have higher emotional weights. A rhythm control rule is introduced during presentation, prioritizing short segments and strong feedback for high-intensity resources, while maintaining the conventional teaching order for low-intensity resources to reduce attention fluctuations. For example, in the topic of binary heaps and priority queues, if the emotional weight of the lower-filter boundary condition node significantly increases, comparative code examples containing array out-of-bounds errors and loop termination conditions are presented first, along with a visual trajectory of heap array index movement, before proceeding to the standard interface definition and complexity analysis text layer. When decoupling control is performed in step 4, the generative artificial intelligence generates bridging teaching resources. These bridging teaching resources are used to connect high-popularity nodes ranked by emotion weight to core nodes determined by core weight. The generation process is controlled by the core weight-driven principle, which first determines the minimum closed-loop explanation chain between the core node and the core edge, and then connects the high-popularity nodes to the explanation chain by starting with the phenomenon. The bridging teaching resources include at least the derivation from the phenomenon to the principle, the minimum runnable code experiment containing the knowledge points of the core node, or the algorithm logic flowchart. The minimum runnable code experiment emphasizes that it can be run directly on the classroom terminal and the output difference can be observed. The algorithm logic flowchart emphasizes that the branch conditions and termination conditions are explicit to avoid the emotional popularity leading the explanation to trivial details that are not related to the core mastery. The implementation and closed-loop verification of bridging teaching resources are carried out by prioritizing high-interest nodes in the presentation order to capture classroom attention. Simultaneously, bridging resources quickly guide students back to core nodes. Immediately after presentation, a minimal verification task is introduced to verify whether the learning gains have been effectively driven. For example, when students are discussing an anomaly in the priority queue dequeue result, the high-interest node might be the comparator direction or the down-filter swap condition, while the core node is the heap order property of a binary heap and the down-filter invariant. In this case, the bridging teaching resources can first explain why a single incorrect comparison would break the heap order invariant by deriving the principle from the phenomenon. Then, a minimal runnable code experiment containing the core node knowledge points is provided, allowing students to modify only one line of the comparison condition and observe the output recovery. A flowchart illustrating the algorithm logic of the down-filter process is also provided, annotating the boundary conditions of each step. Finally, a verification task is used to check whether the error convergence speed has improved. Thus, under decoupled control, resource generation and presentation optimization is achieved by using emotional appeal to guide the entry point and core weight to maintain the main thread.

[0050] In one specific embodiment, the teaching resource set is for each edge In the window Generate resource package It includes at least: (1) Code examples: correct implementation, typical error implementation and minimum runnable experiment; (2) Layered explanatory text: concept definition layer, reasoning link layer, invariant / boundary condition layer; (3) Visual resources: key variable trajectory, state change sequence and flowchart.

[0051] 1) Sorting by combination score for presentation order: The presentation order of resources can be sorted using a combination score of "maintaining the core theme + using emotion as the entry point," for example... ;in As the core weight, For sentiment weight, This represents the ranking / ranking mapping within the candidate edge set (which can be normalized to...). ),and To ensure the stability of the main storyline. This is acceptable. , .because: Make the sorting based on core weights Maintain a firm grip on the main theme, avoiding pushing "non-critical but fun" elements to the forefront due to short-term classroom enthusiasm; while also retaining a certain amount of... Prioritize high-investment / high-attention edges as entry points to increase motivation and click-through rates, then move on to subsequent clicks. The advancement completes the convergence and migration, or adopts a two-stage sequencing when triggering decoupling control: first by... Select Top-m as the entry set And then Internal Press The process is structured in a sequential manner to achieve the goal of "introducing with emotion and concluding with a core theme."

[0052] 2) Bridged teaching resources triggered by decoupling When decoupling control is triggered ( And on one side When exhibiting characteristics of "high emotion weight and low core weight", the system generates bridging teaching resources. ,in To and The core edge with the shortest path distance and the highest core weight in the graph. The bridging resource generation satisfies the following constraints: (1) Determine the core interpretation chain: Calculate from arrive The shortest dependency link is obtained by finding the set of link edges. (2) Output "phenomenon" "Principle" Derivation Explanation: According to The derivation text is given step by step from the edge sequence, with each step corresponding to an edge on the link and explicitly associated with its concept and invariant; (3) Output the minimum runnable experiment: the construction only needs to be modified You can observe the "error phenomenon" by looking at a few key statements. (4) Output flowchart: Generate a flowchart of the derivation and experimental process, and explicitly mark the branch conditions and termination conditions to avoid the discussion from staying on non-core details.

[0053] Step 6: Generate classroom evaluation results based on the learning process data and the usage of teaching resources. These results should include at least knowledge point mastery indicators and error distribution indicators. Update the teaching objective information or classroom knowledge graph update parameters based on the evaluation results to drive the iterative execution of steps 3 through 5 in the next time window, thereby achieving dynamic generation and control of classroom teaching resources. This forms a closed-loop feedback mechanism for the classroom, enabling the system not only to "generate resources" but also to "evaluate resource effectiveness and provide feedback for correction," thus continuously improving the quality of resource generation and the effectiveness of control strategies in subsequent time windows. Generating classroom evaluation results from the learning process data and the usage of teaching resources allows for the characterization of learning outcomes and areas of concentrated problems from the perspectives of knowledge point mastery indicators and error distribution indicators. Further updating the teaching objective information or classroom knowledge graph update parameters based on the evaluation results enables the process data collection, weight updates, and resource generation in the next time window to specifically strengthen weak points, optimize resource structure and presentation order, and thus drive the iterative execution of steps 3 through 5, achieving continuous iteration of dynamic generation and control of classroom teaching resources within the classroom. In one specific implementation, classroom evaluation results are generated based on the learning process data and the usage of the teaching resource set. The teaching objective information or the update parameters of the classroom knowledge graph are then updated according to the classroom evaluation results to drive the cyclical execution of steps 3 to 5 in the next time window. This can be performed as follows: Within a preset time window, learning process data and the usage of teaching resources are aggregated to form an evaluable set of evidence. The learning process data includes at least the submission sequence of code writing logs, the unit test pass rate trajectory of running results, and the number and location of compilation errors, runtime errors, and logical errors. The usage of teaching resources includes at least the number of times code examples are opened and run, the reading depth and dwell time of layered explanation texts, and the completion rate and keyframe replay count of visualization resources. The above evidence is then attributed and mapped according to the nodes and edges of the classroom knowledge graph, so that each edge obtains a corresponding evidence slice to support the interpretable output of subsequent evaluation. The attribution mapping preferably adopts the edge-evidence mapping table M of the embodiment. The classroom evaluation results are generated, which include at least a knowledge point mastery index and an error distribution index. The knowledge point mastery index can be derived from a combination of the completion rate of key concept-related tasks, the change in accuracy, and the error convergence speed. It is also corrected by combining the effectiveness signals of resource use. For example, the weight of evidence that only reads the text but does not run the code example is reduced, while the weight of evidence that repeatedly reviewing the visualization resources and the error convergence speeds up is increased. The error distribution index can be statistically analyzed in three dimensions: error type, error location, and triggering conditions. The high-frequency error clusters and their corresponding knowledge edges are output, thereby forming evaluation items that can directly guide teaching adjustments. Based on classroom evaluation results, update the teaching objective information or the update parameters of the classroom knowledge graph to drive the iterative execution of steps 3 to 5 in the next time window. Specifically, when the knowledge point mastery index is below the target threshold and the error distribution index shows a clustered pattern, update the teaching objective information to improve the achievement requirements of that knowledge edge and break it down into more granular stage goals. Simultaneously, increase the learning drive sensitivity related to that edge in the classroom knowledge graph update parameters to accelerate the convergence of core weights. When the knowledge point mastery index has met the target but the error distribution index shows a scattered and random pattern, update the teaching objective information to shift towards transfer application and comprehensive question types, and decrease the more specific learning drive sensitivity parameters. The new parameters include short-term fluctuation response to enhance stability. For example, in the priority queue topic, if the classroom evaluation results show that the mastery of heap order properties is low and the error distribution is concentrated in the comparison condition direction and the lower filter termination condition, and students frequently run the comparison code example but still trigger logic errors multiple times, then the teaching objectives for the next time window can be updated to first achieve invariant oral presentation and minimum experimental verification, and then enter the interface comprehensive application. The parameters are adjusted and updated accordingly to make the core weights respond faster to the learning gain sequence, and the emotion weights are only used for fine-tuning the presentation order, thereby realizing the dynamic generation and dynamic control of classroom teaching resources in subsequent time windows.

[0054] In one specific embodiment, the classroom assessment module targets each edge In the window Output knowledge point mastery level With error distribution Mastery can be obtained by a linear combination of factors such as learning gain, accuracy variation, and stability, for example: ;in This represents the learning gain representation after denoising. This represents the increment of accuracy / pass rate. Indicators representing stability or consistency It can also be normalized.

[0055] Error distribution You can select "Error Type" Triggering conditions The system organizes statistics based on "location (e.g., line of code / conceptual step)" to form high-frequency error clusters, and outputs the Top-k error clusters and their corresponding edge sets. It is used to locate common misconceptions and key trigger points.

[0056] Objective and parameter update rules based on evaluation results (example): If and If clustering occurs (i.e., errors are significantly concentrated in a few clusters), then increase the learning-driven sensitivity of that edge. And lower the sensitivity of the side's sentiment updates. Simultaneously, in the next window, a minimal verification task (such as a minimal runnable counterexample and assertion check) is forcibly generated for the error cluster to quickly verify invariants / boundary conditions and close the loop for error correction.

[0057] like Meets the standard but High volatility (e.g., recently) The number of times the window changes direction and flips exceeds the threshold. To reduce the impact of emotional noise on the presentation strategy, the stability constraint strength should be increased, for example, by reducing... (Upper limit of change) or increase the entropy constraint / flip penalty to suppress the sorting oscillation caused by "heat fluctuation".

[0058] like If the target is met and the error clusters are scattered (without significant clusters), the teaching objectives will be updated to transfer and application / comprehensive questions, and the resource generation will shift from "filling gaps" to "transfer," such as introducing cross-question type comparison examples, conditional transformations, and comprehensive exercises to promote the ability to apply knowledge to new situations and structured transfer.

[0059] The above formulas are all dimensionless calculations, and the preset parameters in the formulas should be set by those skilled in the art according to the actual situation.

[0060] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for dynamically generating computer science classroom teaching resources empowered by generative AI, characterized in that: It should include at least the following steps: Step 1: Obtain teaching objectives and content information for computer science classes; Step 2: Construct a classroom knowledge network based on teaching objective information and teaching content information, and then construct a classroom knowledge graph based on the classroom knowledge network; Step 3: During classroom teaching, collect classroom process data through at least one smart terminal; Step 4: Within a preset time window, update and dynamically adjust the core weight and sentiment weight of each edge in the classroom knowledge graph through two channels, specifically including: Constructing learning gain representations and emotion representations based on classroom process data; The confidence scores of the learning gain representation and the emotion representation are evaluated and normalized to obtain the learning confidence score and the emotion confidence score. Then, exponential smoothing or recursive estimation is used to perform temporal denoising on the learning gain representation and the emotion representation to obtain the learning gain sequence and the emotion activation sequence. Calculate the core coupling index of emotion and update the gating of the core weights based on the core coupling index of emotion; The core weights are updated based on the learning gain sequence, and the emotion weights are updated based on the emotion activation sequence, with stability constraints introduced in the update process. Step 5: Generate a set of teaching resources based on the nodes, edges, and core weights of the classroom knowledge graph, and adjust the presentation order of the teaching resource set based on the emotion weight; Step 6: Generate classroom evaluation results based on the learning process data and the usage of teaching resources. The classroom evaluation results should include at least the knowledge point mastery index and the error distribution index. Update the teaching objective information or the update parameters of the classroom knowledge graph according to the classroom evaluation results to drive the cyclical execution of steps 3 to 5 in the next time window, thereby realizing the dynamic generation and dynamic control of classroom teaching resources.

2. The method for dynamically generating computer science classroom teaching resources empowered by generative AI according to claim 1, characterized in that, The teaching objectives information should at least include the knowledge points, ability objectives and prerequisites for this lesson, and the teaching content information should at least include the key points of the lesson plan, classroom tasks and sample code provided by the teacher.

3. The method for dynamically generating computer science classroom teaching resources empowered by generative AI according to claim 1, characterized in that, Step 2 involves constructing a classroom knowledge network by using generative artificial intelligence to perform semantic analysis on the teaching content information, extracting concept nodes, grammar nodes, and algorithm nodes, and constructing a classroom knowledge network based on the semantic relationships between the nodes.

4. The method for dynamically generating computer science classroom teaching resources empowered by generative AI according to claim 1, characterized in that, The classroom knowledge graph contains nodes and edges between nodes, and core weights and sentiment weights are initialized for each edge.

5. The method for dynamically generating computer science classroom teaching resources empowered by generative AI according to claim 1, characterized in that, Classroom process data should include at least classroom interaction data, classroom attention data, and learning process data. Among them, classroom interaction data should include at least the number of questions asked, the number of discussions, and the number of collaborations; classroom attention data should include at least the duration of gaze or the level of focus; and learning process data should include at least the code writing log, the running results, and the distribution of error types.

6. The method for dynamically generating computer science classroom teaching resources empowered by generative AI according to claim 1, characterized in that, Learning gain representation should at least represent at least one of the following: accuracy change, error convergence speed, and completion rate of key concept-related tasks corresponding to the learning process data within the time window; emotion representation should at least represent at least one of the following: participation intensity, attention stability, and volatility corresponding to the classroom interaction data and classroom attention data within the time window.

7. The method for dynamically generating computer science classroom teaching resources empowered by generative AI according to claim 1, characterized in that, The emotion core coupling index is used to characterize the strength and consistency of the influence of the emotion activation sequence on the learning gain sequence. The emotion core coupling index is obtained by at least a combination of correlation measure, mutual information measure, or coupling coefficient based on state space model, and is weighted and corrected by learning confidence and emotion confidence.

8. The method for dynamically generating computer science classroom teaching resources empowered by generative AI according to claim 7, characterized in that, When the emotional core coupling index exceeds the adaptive gating threshold and the emotional confidence level is higher than the preset confidence threshold, decoupling control is performed. Decoupling control includes at least: pausing or reducing the contribution of classroom interaction data and classroom attention data to core weight updates, and limiting core weight updates to be driven by the learning gain sequence. When the emotional core coupling index does not exceed the adaptive gating threshold or the emotional confidence level is not higher than the preset confidence threshold, collaborative control is performed, allowing emotional weights to participate in the optimization of the teaching resource presentation order while maintaining the core weight updates according to the learning gain sequence.

9. The method for dynamically generating computer science classroom teaching resources empowered by generative AI according to claim 1, characterized in that, Stability constraints include at least one of the following: upper limit of weight change, entropy constraint of weight change, or drift penalty, in order to suppress the amplification of short-term fluctuations on weight updates.

10. The method for dynamically generating computer science classroom teaching resources empowered by generative AI according to claim 1, characterized in that, The teaching resource set includes at least code examples, layered explanatory texts, and visualization resources. When decoupling control is performed in step 4, the generative artificial intelligence is further controlled to generate bridging teaching resources. The bridging teaching resources are used to associate high-popularity nodes ranked by emotion weights with core nodes determined by core weights. The bridging teaching resources include at least the derivation explanation from phenomena to principles, minimum runnable code experiments containing core node knowledge points, or algorithm logic flowcharts.

Images