Method for automatically generating and difficulty self-adapting of multi-modal classroom exercise questions based on cognitive load management

Abstract

The application relates to the technical field of educational informatization, and discloses a multi-modal classroom exercise question automatic generation and difficulty self-adaptive adjustment method based on cognitive load management, wherein the method comprises the following steps: acquiring a multi-level dependency chain topological description of a target knowledge point; generating a diagnosis exercise pair sequence composed of control questions and detection questions according to a topological sequence; collecting answer data and calculating an effective residual cognitive occupation amount based on exponential weighted accumulation; calculating a current available cognitive capacity and marking a fatigue high-risk level; correcting a response time difference value and an attribution confidence based on a fatigue interference coefficient, and performing layer-by-layer attribution determination; and generating targeted multi-modal exercise questions according to a bottleneck level and a bottleneck type and adjusting a difficulty level.

Description

Technical Field

[0001] This invention relates to the field of educational informatization technology, specifically to a method for automatically generating and adaptively adjusting the difficulty of multimodal classroom exercises based on cognitive load management. Background Technology

[0002] In primary and secondary school classrooms, a single practice session typically lasts 15 to 25 minutes and includes multiple consecutive practice questions. The target knowledge points involved in these exercises often exhibit a multi-level, interdisciplinary, chain-dependent structure. For example, physics word problems rely on the ability to solve mathematical equations, which in turn rely on the ability to extract conditions from the text in language arts.

[0003] Existing adaptive exercise generation technologies for primary and secondary school classrooms mainly cover the following representative research and solutions:

[0004] The first category is the single-knowledge-point adaptive practice question recommendation method based on the knowledge tracing model. This type of method (see: Corbett AT, Anderson J R. Knowledge Tracing: Modeling the Acquisition of Procedural Knowledge[J]. User Modeling and User-Adapted Interaction, 1994, 4(4): 253-278; and Yudelson MV, Koedinger KR, Gordon G J. Individualized Bayesian Knowledge Tracing Models[C]. Proceedings of AIED, 2013: 171-180) uses a Bayesian knowledge tracing model (BKT) to sequentially estimate the mastery probability of a single knowledge point and selects practice questions of corresponding difficulty based on the mastery level. The core limitation of the above method is that each attribution is performed only for a single target knowledge point. When students make mistakes on comprehensive questions with multi-level interdisciplinary dependencies, the system cannot distinguish whether the mistake stems from a lack of the target knowledge point itself or from a lack of interdisciplinary connection between its preceding dependent knowledge points and the target knowledge point.

[0005] The second category is problem difficulty modeling methods based on cognitive load theory. These methods (see: Sweller J. Cognitive Load During Problem Solving: Effects on Learning[J]. CognitiveScience, 1988, 12(2): 257-285; and Paas F, Van Merriënboer JJ G. Instructional Control of Cognitive Load in the Training of Complex CognitiveTasks[J]. Educational Psychology Review, 1994, 6(4): 351-371) statically assess the cognitive load of a single problem using task complexity or element interactivity as indicators, and set the problem difficulty accordingly. However, these methods perform independent assessments for each problem, assuming that students maintain the same cognitive capacity throughout the entire practice sequence. They fail to model the working memory residue of completed problems and its exponential decay over time, thus failing to predict and manage the risk of cumulative depletion of working memory resources in scenarios with multiple consecutive problems.

[0006] The third category is exercise generation schemes based on multimodal presentation. These schemes (see: Chinese Patent Application Publication No. CN115691271A, "An Adaptive Learning Resource Recommendation Method Based on Multimodal Knowledge Graphs"; and Mayer R E. Multimedia Learning [M]. Cambridge University Press, 2001) introduce multiple modal elements such as text, images, and animations into exercise generation, enriching the knowledge presentation through multimodal combinations. However, the combination methods of multimodal elements in the above schemes are fixed or hard-coded by rules, and are not dynamically correlated with the student's real-time cognitive state: when the student's working memory resources have been significantly depleted by previous questions, the additional modal switching burden introduced by parallel multimodal presentation exacerbates the depletion of cognitive resources, but existing schemes do not have the ability to perceive this state and adjust the modal configuration in real time.

[0007] The fourth category is dependency modeling methods oriented towards knowledge graphs. These methods (see: Liu Q, et al. EKT: Exercise-Aware Knowledge Tracing for Student Performance Prediction[J]. IEEE Transactions on Knowledge and Data Engineering, 2021, 33(1): 100-115; and Chinese Patent Announcement No. CN113657557B, A Personalized Learning Path Recommendation Method Based on Knowledge Graphs) model the dependencies between knowledge points using knowledge graphs and recommend learning paths accordingly. These methods typically plan the overall path at the level of the learning path itself, without performing layer-by-layer diagnostics within the question sequence of a single classroom exercise, nor distinguishing the specific level of the dependency chain where the bottleneck is located. Therefore, they are difficult to support the application requirements of accurately locating and promptly remedying issues within a limited classroom time.

[0008] The combined effect of these two shortcomings leads to the following technical problems: First, the method of independently assessing cognitive load ignores the residual occupation of working memory resources across questions and the nonlinear cumulative decay during the progression of the practice sequence. This results in a decline in attention and a sharp drop in accuracy for students in the later stages of practice due to cumulative fatigue, and existing methods cannot identify this risk in advance. Second, single attribution cannot distinguish which level of the dependency chain the bottleneck of answer failure is located at. The decline in performance caused by cumulative fatigue in the later stages is misjudged as a knowledge deficiency at a specific level, causing the automatically generated practice questions to deviate from the students' actual needs in terms of difficulty and knowledge point positioning, wasting limited classroom practice time. Third, the introduction of multimodal elements lacks coordination with the students' real-time cognitive state. The parallel presentation of multimodal elements in the stage of cognitive resource depletion amplifies the fatigue effect and exacerbates learning obstacles.

[0009] The aforementioned shortcomings collectively make it difficult for existing methods to simultaneously meet the requirements of precise fatigue management, interdisciplinary bottleneck step-by-step location, and modal dynamic adaptation in the context of continuous multi-question practice in primary and secondary school classrooms. There is an urgent need for a method and system that can integrate sequential-level cognitive state modeling and interdisciplinary hierarchical attribution judgment into the same framework. Summary of the Invention

[0010] This invention provides a method for automatic generation and adaptive difficulty adjustment of multimodal classroom exercises based on cognitive load management, which solves the technical problem in related technologies that cannot dynamically compensate for the generation and difficulty adjustment of exercises according to the real-time cognitive load of students, resulting in inaccurate attribution judgment results due to fatigue interference.

[0011] This invention discloses a method for automatically generating and adaptively adjusting the difficulty of multimodal classroom exercises based on cognitive load management, comprising:

[0012] A reverse traversal is performed from the multidisciplinary knowledge network with the target knowledge point as the endpoint. In-degree paths with a depth of more than two levels and spanning two or more subject identifiers are filtered to obtain a set of interdisciplinary dependency chains. The knowledge point identifier, the subject identifier, and the current mastery value of each level node in each interdisciplinary dependency chain are extracted. The information of each level node is organized in topological order to generate a multi-level dependency chain topology description of the target knowledge point.

[0013] Diagnostic practice pairs are generated layer by layer from bottom to top according to the multi-level dependency chain topology. Each diagnostic practice pair consists of a control question and a probe question. The control question only examines the knowledge point at the current level itself and is presented in the typical modality of the subject at that level. The probe question examines the connection and application of the knowledge point at the current level with the node at the previous level and is presented in a cross-modal form that combines the typical modalities of two adjacent subject levels. The diagnostic practice pairs at each level are arranged in the multi-level dependency chain topology order, and the diagnostic practice pair sequence is output.

[0014] The diagnostic practice test sequence is presented to students sequentially for them to answer. The correct / incorrect markings, response times, and operation intervals for each question are collected. After each question is completed, the cognitive load values ​​of the completed questions are exponentially weighted and accumulated using the working memory half-life parameter to calculate the effective residual cognitive load at the current moment. :

[0015]

[0016] in, This represents the number of questions completed. For the first The cognitive load of the question. For the first The time interval between the completion of the problem and the current time. For working memory half-life parameters;

[0017] The initial cognitive capacity threshold of students Subtract the effective residual cognitive occupancy Obtain the current available cognitive capacity Obtain the cognitive load requirement value for each question in the next level of diagnostic practice, and then compare the cognitive load requirement value with the current available cognitive capacity and the preset safety ratio coefficient. product The comparison is performed, and when the cognitive load requirement of any question exceeds the product, the level is marked as a high-risk fatigue level. Modal lightweighting adjustment is then performed on diagnostic practice pairs marked as high-risk fatigue levels before presentation. This modal lightweighting adjustment includes: replacing modal elements with high cognitive occupancy coefficients with equivalent modal elements with low occupancy coefficients, and converting cross-modal parallel presentation information into a temporally serialized form, thereby reducing the cognitive load requirement of the diagnostic practice pairs to a lower level. The knowledge points tested and the diagnostic effectiveness remain unchanged;

[0018] For each level The effective residual cognitive occupancy before the presentation of the control question at this level. With the initial cognitive carrying threshold The ratio is used to calculate the fatigue interference coefficient. ; using the fatigue interference coefficient according to Fatigue compensation correction is performed on the response time differences at each level to obtain the corrected response time differences. ,in and These are the response times for the detection and control questions, respectively. The baseline response time parameter is used; for levels marked as high-risk fatigue levels, the response time is determined based on the fatigue interference coefficient. The fatigue discount factor is calculated, and the initial attribution confidence of incorrect answers is weighted down to obtain the weighted attribution confidence. This is the discount adjustment constant. The lower limit of the discount factor is used; based on the corrected response time difference and the weighted attribution confidence, a bottom-up attribution determination is performed to determine the bottleneck level and bottleneck type.

[0019] Based on the bottleneck level and bottleneck type, a corresponding template is matched from the question template library to generate targeted multimodal practice questions, and the difficulty level is adjusted to the level corresponding to the current mastery value of the bottleneck level.

[0020] Preferably, the multidisciplinary knowledge network is stored in a directed graph structure, where nodes represent knowledge points of each discipline, and directed edges represent prerequisite dependencies between knowledge points. The multidisciplinary knowledge network is constructed as follows: subject experts label the granularity of knowledge points and cross-disciplinary prerequisite dependencies according to the curriculum standards to form an initial directed graph structure. After the system is running, the historical answer data is statistically analyzed based on the Bayesian knowledge tracing model. When a student's answer performance on a certain target knowledge point shows a significant positive correlation with the mastery of a certain prerequisite knowledge point, the weight of the corresponding directed edge is automatically supplemented or strengthened. Each node stores the knowledge point identifier, the discipline identifier, a list of recommended modality types, and the current student's mastery value. The mastery value is updated in real time by the Bayesian knowledge tracing model based on the student's historical answer sequence.

[0021] Preferably: a text modality, presenting the question stem and conditions in natural language; a formula modality, presenting quantitative relationships using mathematical symbols and equations, suitable for mathematical and science knowledge points; a diagrammatic modality, presenting spatial relationships or physical scenarios using static graphics; a step-by-step diagrammatic modality, presenting the process deduction using multiple static frames arranged chronologically; and an interactive operation modality, requiring students to complete answers through drag-and-drop, filling, or clicking operations. The cognitive occupancy coefficient of each modality type is stored as a preset constant in the system configuration, with the cognitive occupancy coefficient of the interactive operation modality being higher than that of the text modality, formula modality, diagrammatic modality, and step-by-step diagrammatic modality.

[0022] Preferably: the effective residual cognitive occupancy The calculation method is as follows:

[0023]

[0024] in, This represents the number of questions completed at the current moment. For the first The cognitive load of the question. For the first The time interval between the completion of the problem and the current time. For working memory half-life parameters;

[0025] The cognitive load value The calculation method is as follows:

[0026]

[0027] in, For the first The number of modal elements contained in the question. For the first The modal interaction complexity score of the question. For the first The number of operation steps required to answer the question. , , For pre-calibrated constant weights and satisfying , , , Mean normalization based on the range was performed before calculation.

[0028] Preferably: the working memory half-life parameter The initial values ​​are preset according to the students' age group: 5 to 7 minutes for lower elementary grades, 7 to 9 minutes for upper elementary grades, 9 to 12 minutes for junior high school, and 12 to 15 minutes for senior high school. After each diagnostic exercise pair is completed, the actual decay trajectory is compared with the current... The predicted trajectory is compared, and the working memory half-life parameter is updated using the least squares method. The value of is determined; simultaneously, the initial cognitive capacity threshold is updated based on the maximum number of valid answers given by the students in this diagnosis. .

[0029] Preferably, after marking the high-risk fatigue level, the diagnostic practice pairs marked as high-risk fatigue levels undergo modal lightweighting adjustment before presentation. The modal lightweighting adjustment includes: replacing modal elements with high cognitive occupancy coefficients in control and probe questions with equivalent modal elements with low occupancy coefficients, including replacing animated demonstrations with step-by-step diagram modalities and replacing complex interactive operations with selection and completion operations; converting cross-modal parallel presentation of information into a temporally serialized form presented sequentially according to time steps, so that students process only a single modal information at any given time; after the modal lightweighting adjustment, the cognitive load requirement of the diagnostic practice pairs is reduced to within the safe range of available capacity, while the knowledge assessment points and diagnostic discrimination power remain unchanged.

[0030] Preferably, the step of correcting the response time difference and attribution confidence of each level using the fatigue interference coefficient includes:

[0031] For each level Calculate the fatigue interference coefficient ,in hierarchical The effective residual cognitive occupancy before the control question is presented. As the initial cognitive capacity threshold for students, The value ranges from 0 to 1;

[0032] Calculate the difference in response time after correction ,in The response time of the probe question, To control the response time of the questions, The baseline for the expected additional time consumption of the detection problem relative to the control problem under fatigue-free conditions;

[0033] For levels marked as high-risk fatigue levels, multiply the initial attribution confidence of their incorrect responses by a fatigue discount factor. The attribution confidence score after weight reduction is obtained, among which , This is a discount adjustment constant with a value ranging from 0.3 to 0.7. The discount factor is the lower limit and ranges from 0.2 to 0.4; for levels not marked as high-risk fatigue levels, the attribution confidence remains unchanged.

[0034] The initial attribution confidence is pre-configured based on the combination of correct and incorrect labels for control and probe questions: a higher preset value is used when the control question is correct but the probe question is incorrect, a lower preset value is used when the control question itself is incorrect, and the highest preset value is used when both questions are correct.

[0035] Preferably, the bottom-up, layer-by-layer attribution determination includes:

[0036] If a certain level's control question is correct but the probe question is incorrect, and the attribution confidence after weight reduction is not lower than the reliable judgment threshold, then the interdisciplinary connection between this level and the upper level is determined to be a bottleneck, and the bottleneck type is marked as interdisciplinary connection defect type.

[0037] If a certain level of control question is incorrect, and the attribution confidence after weight reduction of that level is not lower than the reliable judgment threshold, then the knowledge point ontology of that level is determined to be a bottleneck, the bottleneck type is marked as knowledge ontology defect type, and the attribution judgment to higher levels is terminated.

[0038] If both the control and detection questions at a certain level are correct, and the difference in response time after correction does not exceed the preset abnormal time consumption threshold, then it is determined that there is no bottleneck at that level, and the attribution judgment is continued for the next level.

[0039] The reliability threshold is a preset constant ranging from 0.5 to 0.8. Levels with a reduced attribution confidence level below the reliability threshold are marked as levels to be verified, and the required shortest recovery interval is calculated using the following formula. :

[0040]

[0041] in, For working memory half-life parameters, This represents the effective residual cognitive occupancy when responding at the level to be verified. As the initial cognitive capacity threshold for students, To preset the safety ratio coefficient, To verify the cognitive load requirement of the diagnostic practice pair, a cognitive recovery interval no shorter than the minimum recovery interval is inserted at a subsequent position in the diagnostic practice pair sequence. After recovery, a simplified version of the diagnostic practice pair is added to the level to be verified for secondary confirmation. The attribution results of the secondary confirmation cover the initial attribution conclusion.

[0042] Preferably, the step of generating targeted multimodal practice questions and adjusting the difficulty level based on bottleneck level and bottleneck type includes:

[0043] For levels where the bottleneck type is knowledge ontology deficiency, single-point specialized practice templates corresponding to the knowledge points of that level are matched from the question template library. Single-point specialized practice questions are generated using the typical modalities of the subject at that level, and the difficulty level is adjusted to the level corresponding to the current mastery value of that level.

[0044] For levels where the bottleneck type is interdisciplinary connection deficiency, bridging practice templates involving knowledge points of adjacent two levels are matched from the question template library to generate bridging multimodal practice questions. The bridging multimodal practice questions connect the knowledge points of the two levels through progressive contextual transitions: the first stage only involves independent application scenarios of the lower-level knowledge points, and the second stage introduces application elements of the upper-level knowledge points based on the results of the first stage, and explicitly presents the corresponding mapping relationship between the two subject concepts in the transitional context in a multimodal manner with text annotations and graphic arrows.

[0045] The difficulty levels are divided into five levels, and the correspondence between each difficulty level and the mastery value is as follows: Mastery values ​​in the range of 0 to 0.2 correspond to difficulty level 1, where questions require the identification and restatement of basic definitions of knowledge points; mastery values ​​in the range of 0.2 to 0.4 correspond to difficulty level 2, where questions require the direct application of formulas or rules in a single-step context; mastery values ​​in the range of 0.4 to 0.6 correspond to difficulty level 3, where questions involve multi-step derivation and parameters must be extracted from the context; mastery values ​​in the range of 0.6 to 0.8 correspond to difficulty level 4, where questions require the flexible application of knowledge points in new contexts and cross-step reasoning; and mastery values ​​in the range of 0.8 to 1.0 correspond to difficulty level 5, where questions involve the comprehensive application of multiple knowledge points and open-ended reasoning requirements.

[0046] The parameters for difficulty adjustment include four dimensions: numerical complexity, number of reasoning steps, contextual abstraction, and modal combination complexity. When adjusting to a lower difficulty level, all four dimensions decrease simultaneously. When adjusting to a higher difficulty level, the four dimensions gradually increase sequentially from numerical complexity, number of reasoning steps, contextual abstraction, to modal combination complexity.

[0047] The diagnostic practice pairs are arranged in the following order: within the same dependency chain, they are arranged in bottom-up order; when there are multiple independent dependency chains, the shallower chains are completed first, followed by the deeper chains; when multiple dependency chains share a bottom-level node, the diagnostic practice pair for the shared node is generated only once, and its attribution conclusion applies to all dependency chains that share that node.

[0048] The student's initial cognitive capacity threshold Based on the student's average maximum number of valid answers in the first three practice sessions Compared with the average cognitive load per item The product is obtained by calculation, that is For new users who do not yet have historical practice data, The default values ​​are set according to age group: 3.0 for lower elementary grades, 4.0 for upper elementary grades, 5.0 for junior high school, and 6.0 for senior high school.

[0049] After generating the targeted multimodal practice questions, the practice questions are assembled and output according to the presentation rhythm control parameters, which include the display duration of each modal element, the modality switching interval, and the interaction waiting time; the display duration is based on the information density of the modal elements and the current available cognitive capacity. Certainly, the higher the information density or The lower the value, the longer the display duration; the modal switching interval is determined based on the cognitive conversion requirements between adjacent modal elements.

[0050] A multimodal classroom exercise automatic generation and difficulty adaptive adjustment system based on cognitive load management, used to perform the method described in claim 1, comprising:

[0051] The dependency chain topology extraction module is used to filter cross-disciplinary dependency chains with a depth of more than two levels and spanning two or more subject identifiers from a multi-disciplinary knowledge network. It extracts the knowledge point identifier, the subject identifier, and the current mastery value of each level node and generates a multi-level dependency chain topology description of the target knowledge point.

[0052] The diagnostic practice pair generation module is used to generate diagnostic practice pairs consisting of control questions and probe questions layer by layer from bottom to top according to the multi-level dependency chain topology description, and output a diagnostic practice pair sequence;

[0053] The module for collecting answer data and calculating cognitive load is used to present the diagnostic practice pair sequence to students in sequence and collect the correct and incorrect labels, response time and operation interval time. Based on the working memory half-life parameter, the cognitive load value of the completed questions is exponentially weighted and accumulated to calculate the effective residual cognitive occupancy at the current moment.

[0054] The cognitive capacity assessment and risk labeling module is used to subtract the effective residual cognitive occupancy from the student's initial cognitive capacity threshold to obtain the current available cognitive capacity, compare the cognitive load requirement value of the subsequent diagnostic practice pair with the product of the current available cognitive capacity and the preset safety ratio coefficient, label the fatigue high-risk level, and perform modal lightweight adjustment on the diagnostic practice pair labeled as fatigue high-risk level.

[0055] The fatigue compensation attribution module is used to calculate the fatigue interference coefficient based on the ratio of the effective residual cognitive occupancy of each level to the initial cognitive carrying threshold. The fatigue interference coefficient is used to correct the response time difference and reduce the attribution confidence of the high-risk fatigue level. Layer-by-layer attribution judgment is performed to determine the bottleneck level and bottleneck type.

[0056] The exercise generation and difficulty adjustment module is used to match the corresponding template from the question template library according to the bottleneck level and bottleneck type to generate targeted multimodal exercise questions, adjust the difficulty level to the level corresponding to the current mastery value of the bottleneck level, and assemble and output according to the presentation rhythm control parameters.

[0057] The present invention has the following beneficial effects.

[0058] First, by exponentially weighting and accumulating the cognitive load values ​​of completed questions according to the working memory half-life parameter, the effective residual cognitive occupancy is continuously calculated. This allows for the identification of high-risk fatigue levels before presenting diagnostic exercises, thus preventing questions that exceed the student's current capacity from being presented directly.

[0059] Second, the fatigue interference coefficient is used to correct the response time difference, eliminating the interference of fatigue-induced time dilation on the assessment of interdisciplinary connection ability; the fatigue discount factor is used to reduce the attribution confidence of high-risk fatigue levels, so that the results of the hierarchical attribution judgment reflect the students' knowledge mastery level, rather than the performance decline caused by the depletion of cognitive resources.

[0060] Third, for the levels to be verified that have insufficient attribution confidence after weight reduction, verification is carried out by inserting cognitive recovery intervals and adding secondary confirmation, so as to avoid over-correction that would cause the real knowledge bottleneck to be missed.

[0061] Fourth, based on the bottleneck levels and types determined after fatigue compensation correction, single-point specialized practice questions or bridging multimodal practice questions are generated respectively, so that the difficulty level and knowledge point positioning of subsequent practice questions correspond to the students' real ability bottleneck nodes, thereby improving the pertinence of practice within the limited classroom practice time. Attached Figure Description

[0062] Figure 1This is a flowchart of the method for automatic generation and adaptive difficulty adjustment of multimodal classroom exercises based on cognitive load management provided in this embodiment of the invention;

[0063] Figure 2 This is a schematic diagram of the distribution of knowledge point mastery in a three-layer interdisciplinary dependency chain provided in an embodiment of the present invention;

[0064] Figure 3 This is a schematic diagram comparing the sequence cognitive load requirement value and presentation modality of diagnostic practice provided in an embodiment of the present invention;

[0065] Figure 4 This is a schematic diagram illustrating the temporal change of the effective residual cognitive occupancy after completing each question, as provided in an embodiment of the present invention.

[0066] Figure 5 This is a schematic diagram illustrating the cognitive capacity assessment provided by an embodiment of the present invention: a comparison of the initial threshold, residual occupancy, and available capacity.

[0067] Figure 6 This is a schematic diagram comparing the cognitive load demand values ​​before and after modal lightweighting adjustment in Title 4, provided in an embodiment of the present invention.

[0068] Figure 7 This is a schematic diagram comparing the fatigue interference coefficients and attribution confidence scores at each level, provided in an embodiment of the present invention.

[0069] Figure 8 This is a schematic diagram of the scatter distribution of response time and cognitive load values ​​for each question provided in the embodiments of the present invention;

[0070] Figure 9 This is a schematic diagram showing an overview of the targeted practice question generation parameter configuration provided in this embodiment of the invention. Detailed Implementation

[0071] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0072] Example

[0073] According to an embodiment of the present invention, a method for automatic generation and adaptive difficulty adjustment of multimodal classroom exercises based on cognitive load management is described in [reference needed]. Figure 1 This includes the following steps:

[0074] Step 1: Obtain the multi-level dependency chain topology description of the target knowledge point.

[0075] Obtain all in-degree paths of the target knowledge point in the multidisciplinary knowledge network, filter cross-disciplinary dependency chains with a depth of more than two levels, extract the knowledge point identifier, the subject identifier, and the current mastery value of each level node in each cross-disciplinary dependency chain, and generate a multi-level dependency chain topology description of the target knowledge point.

[0076] Step 1 may specifically include:

[0077] Step S101: Perform a reverse traversal from the multidisciplinary knowledge network with the target knowledge point as the endpoint to obtain the set of all in-degree paths leading to the target knowledge point.

[0078] Step S102: For each path in the in-degree path set, count the level depth and the number of subject identifiers involved in the path, retain paths with a depth of more than two levels and no less than two subject identifiers, and obtain the cross-disciplinary dependency chain set.

[0079] Step S103: For each interdisciplinary dependency chain in the interdisciplinary dependency chain set, extract the knowledge point identifier, subject identifier, and current mastery value of each level node in the direction from the bottom to the target knowledge point, organize the node information of each level in topological order, and generate a multi-level dependency chain topology description.

[0080] It should be noted that the in-degree path mentioned above refers to all directed paths from other nodes along directed edges to the target knowledge point. Filtering interdisciplinary dependency chains with a depth exceeding two levels means retaining dependency paths that contain at least three different levels of nodes and span two or more subject identifiers. The current mastery level is a value calculated by accumulating the student's historical answer data for the corresponding knowledge point, ranging from 0 to 1.

[0081] Taking "Physics: Application of Equations of Uniformly Accelerated Linear Motion" as an example, this paper illustrates the specific structure of a multi-level dependency chain topology. The system performs a reverse traversal of the multidisciplinary knowledge network and identifies the following three-level interdisciplinary dependency chain:

[0082] Level 1 (Bottom Level): Chinese Language - Condition Extraction, Knowledge Point Identifier KN-CHN-001, Subject Identifier CHN, Current Mastery Value ;

[0083] Level 2 (Intermediate Level): Mathematics - Solving Linear Equations in One Variable, Knowledge Point Identifier: KN-MATH-012, Subject Identifier: MATH, Current Mastery Level: ;

[0084] Level 3 (Target Level): Physics - Application of Equations of Uniformly Accelerated Linear Motion, Knowledge Point Identifier: KN-PHY-007, Subject Identifier: PHY, Current Mastery Level: .

[0085] The aforementioned dependency chain means that when students solve application problems involving uniformly accelerated linear motion in physics, they must first extract valid conditions from the text of the problem (relying on their language skills for extracting conditions), then substitute the extracted conditions into the equations for algebraic solution (relying on their mathematical equation-solving skills), and finally complete the calculation and deduction of physical quantities. These three levels belong to the three subjects of language, mathematics, and physics, forming a typical three-level interdisciplinary dependency chain. The multi-level dependency chain topology is organized in the topological order of level 1, level 2, and level 3, including the knowledge point identifier, subject identifier, and mastery value for each node.

[0086] Step 2: Generate diagnostic practice pair sequences arranged in dependency chain topological order.

[0087] For each multi-level dependency chain to be diagnosed, diagnostic practice pairs are generated layer by layer from bottom to top. Each diagnostic practice pair consists of a control question and a probe question: the control question only examines the knowledge point at the current level and is presented in the typical modality of the subject at that level; the probe question examines the connection and application between the knowledge point at the current level and the node at the previous level and presents the connection process in a cross-modal form. Diagnostic practice pairs at all levels are arranged in topological order of the multi-level dependency chain, and the diagnostic practice pair sequence is output.

[0088] It should be noted that the aforementioned control questions are presented in a typical modality for that subject level, meaning that the corresponding primary presentation modality is selected based on the subject's attributes. For example, control questions for language arts knowledge points are presented in a pure text modality, requiring students to extract valid conditions from paragraphs; control questions for mathematics knowledge points are presented in a text and formula modality, requiring students to complete equation transformations or numerical calculations; control questions for physics knowledge points are presented in a text and diagram modality, requiring students to read physical quantities from diagrams and substitute them into formulas. The core function of control questions is to isolate the assessment: the questions only involve the knowledge point at the current level itself and do not include application requirements for knowledge points at adjacent levels, thus ensuring that the answers to control questions only reflect the mastery of the knowledge point at the current level and are not affected by other levels. The probe questions present the connection process in a cross-modal form. This means that the typical modalities of two adjacent subject levels are combined in the same question to present the application scenario of knowledge connection. Its core function is connection diagnosis: on the premise that the control question has verified the mastery of the knowledge point at the current level, the probe questions further examine whether the output of the knowledge point at the current level can be smoothly transferred to the application scenario of the knowledge point at the next higher level, so that the combination of the correct and incorrect marks of the two questions can distinguish between the two types of "knowledge ontology defects" and "cross-level connection defects".

[0089] It should be noted that the generation of the diagnostic practice pairs mentioned above is based on a pre-configured question template library. This library stores control and probe question templates for each knowledge point. Each template includes the question's skeleton structure, a range of fillable parameters, and a corresponding modal configuration scheme. When generating diagnostic practice pairs, the library matches the template corresponding to the current level of knowledge point, fills in the specific parameters, and outputs the complete question.

[0090] It should be noted that for complex target knowledge points containing multiple interdisciplinary dependency chains, the diagnostic practice sequence arrangement strategy is as follows: within the same dependency chain, the sequences are arranged in a bottom-up order; if multiple independent dependency chains exist, the shallower chains are completed first, followed by the deeper chains, so that the cognitive load is lower in the early stages of the sequence and gradually increases in the later stages, matching the students' warm-up-consolidation-challenge cognitive rhythm. When multiple dependency chains share a bottom-level node, the diagnostic practice pair for the shared node is generated only once and appears once in the sequence. Its attribution conclusion applies to all dependency chains sharing that node, thus avoiding unnecessary cognitive consumption caused by repeated diagnoses.

[0091] The construction method and data structure definition of the question template library:

[0092] The question template library adopts a two-stage architecture of offline construction and online query. In the offline construction stage, subject matter experts design several sets of control question templates and probe question templates for each knowledge point, categorized by granularity. Each template consists of the following fields:

[0093] (1) Knowledge point identifier field (kn_id): consistent with the node identifier in the multidisciplinary knowledge network, used for retrieval and matching, the type is string, such as "KN-MATH-012".

[0094] (2) Template type field (tpl_type): The value is “control” or “probe”, which is used to distinguish the template function type.

[0095] (3) Question skeleton structure field (skeleton): The question skeleton is stored in a natural language string with placeholders. The placeholders are marked with double curly braces, for example, "Given that the initial velocity of an object is {{v0}} m / s and the acceleration is {{a}} m / s², find the velocity after {{t}} seconds."

[0096] (4) Populateable parameter range field (param_ranges): Stores the parameter value range corresponding to each placeholder in key-value pairs, including the numerical range (min, max), step size (step) and numerical type (integer or decimal), for example {"v0":{"min": 1, "max": 20, "step": 1, "type": "int"}}.

[0097] (5) Modal configuration scheme field (modal_config): Stores the sequence of modal types used by the template in the form of a list, such as ["text", "formula"], and specifies the position anchor and presentation order of each modal element in the question.

[0098] (6) Cognitive Load Precalibration Field (L_preset): Stores the cognitive load baseline value obtained by the pre-experiment measurement when the parameters are taken at the default median value. It is used to estimate the cognitive load requirement value of the questions before the actual parameters are filled.

[0099] (7) Difficulty_range: Records the range of difficulty levels for which the template is applicable. For example, [2, 4] means that the template is applicable to question generation requirements of difficulty level 2 to 4.

[0100] (8) Subject connection identifier field (bridge_kn_ids): This field is only included in the test question template. It stores the identifiers of adjacent level knowledge points connected by the test question in the form of a list, which is used to accurately locate cross-level connection relationships when multiple dependency chains exist.

[0101] Template matching algorithm:

[0102] When generating diagnostic practice pairs online, the template matching algorithm performs the following steps:

[0103] Step M1: Using the knowledge point identifier kn_id of the current level and the template type tpl_type as the primary keys, perform an exact query in the question template library to obtain a set of candidate templates. If the candidate set is empty, then fall back to the general template library of the higher-level subject category to perform fuzzy matching.

[0104] Step M2: For the candidate set For each template, calculate the distance between its difficulty label field and the target difficulty level corresponding to the current level mastery value. Target Difficulty Level Template difficulty range median Candidate templates are sorted in ascending order of distance, and the template with the smallest distance is selected as the matching result. If multiple templates have the same distance, the first one is selected in ascending order of template number.

[0105] Step M3: Fill the param_ranges field of the selected template with parameters according to the following rules: Based on the relative position of the target difficulty level and the difficulty range, select parameter values ​​proportionally within the min to max range of the parameter range. The specific calculation is as follows:

[0106]

[0107] Step M4: Combine the skeleton string after parameter filling with the modal configuration scheme to output the complete question object, including the question stem, the list of modal elements, and an estimated cognitive load requirement. The estimated cognitive load requirement is based on L_preset, and the baseline value is linearly scaled according to the actual parameter complexity (number of digits, number of reasoning steps). The correction coefficient is obtained by fitting pre-experimental data.

[0108] It should be noted that the aforementioned question template library was pre-built by the subject-specific teaching and research teams according to the data structure described above during the initial deployment of the system, covering the main knowledge points of each subject (Chinese, mathematics, physics, chemistry, English, etc.) across all educational levels (primary, middle, and high school). It can be continuously expanded through a manual review mechanism: after teachers upload new questions to the system backend, the system automatically parses the question structure and generates a corresponding template draft, which is then added to the library after teacher confirmation.

[0109] Step 3: Collect diagnostic response data and calculate effective residual cognitive occupancy.

[0110] The diagnostic practice test sequence is presented to students sequentially from the bottom to the top, and the correct / incorrect marks, response times, and operation intervals for each diagnostic question are recorded. The cognitive load value generated by each completed question is exponentially weighted and accumulated according to time decay to calculate the effective residual cognitive load at the current moment.

[0111] Step 3 may specifically include:

[0112] Step S301: According to the topological order of the diagnostic practice pair sequence, push diagnostic questions to the students' classroom practice terminals in sequence. Record the students' correct or incorrect marks for each question, the response time from question presentation to answer submission, and the operation interval between submission times of two adjacent questions.

[0113] Step S302: After completing each question, calculate the effective residual cognitive occupancy for all questions completed before the current time. :

[0114]

[0115] in, This represents the number of questions completed at the current moment. The question number and Take from 1 to , For the first The cognitive load of the question. For the first From the time the question was completed to the current time The time interval (unit: minutes). This is a parameter for the working memory half-life (unit: minutes). and If the two terms have the same dimensions, their ratio is a dimensionless number, and the exponent term is dimensionless. The score is a dimensionless score. and Dimensions are consistent. When When it increases, the first The contribution of the question to the current residual occupancy in half-life The periodic decay is reduced to half of its original value.

[0116] It should be noted that the above cognitive load values The result is obtained by weighted summation of the number of modal elements in the question, the modal interaction complexity, and the number of operation steps required to answer the question. The calculation method is as follows:

[0117]

[0118] in, For the first The number of modal elements contained in the question. For the first The modal interaction complexity score of the question. For the first The number of operation steps required to answer the question. , , For pre-calibrated constant weights, satisfying To enable direct weighted summation of the three components, the following steps are performed before calculation: , , Mean normalization based on the range is performed separately, mapping each component to a uniform range of 0 to 1. This includes the modal interaction complexity score. The number of operation steps is pre-defined based on the number of interaction jumps between different modal elements in the problem. The number of independent actions that students need to complete in the question-answering process is pre-calculated.

[0119] Objective calibration methods for cognitive load components:

[0120] Cognitive load value The three components were objectively calibrated using the following preliminary experimental procedure:

[0121] Calibration Process 1: Modal Interaction Complexity Scoring The calibration (external cognitive load);

[0122] Step P1: The expert group counts the number of information jump links between modal elements for each modal combination scheme in the question template library. That is, the number of times students need to switch between different modalities during the answering process (for example, switching from textual question stem to formula and then to diagram is counted as 2 jumps), and this is recorded as the jump count. .

[0123] Step P2: Organize 50 to 100 students matched with the target age group to participate in the pre-experiment. Use the Primary-Secondary Task Paradigm to measure the secondary task response time when answering each question. The longer the secondary task response time, the higher the external cognitive load.

[0124] Step P3: Based on the number of jumps Using as the independent variable and sub-task reaction time as the dependent variable, a linear regression equation is fitted. After normalizing the regression predicted values ​​to the range of 0 to 1, they are used as... The pre-calibrated value is stored in the L_preset field of the template library.

[0125] Calibration Process Two: Number of Operation Steps The calibration (internal cognitive load);

[0126] Step P4: The expert group breaks down the standard answer path for each question template step by step, and counts the number of independent cognitive operations required from reading the question to submitting the answer (including information extraction, formula retrieval, numerical substitution, transformation and derivation, result verification, etc.), which are recorded as the number of operation steps. .

[0127] Step P5: Under pure text unimodal conditions (without external load interference), organize students of the same size as in Step P2 to measure the average response time of each question corresponding to each number of operation steps. Verify the monotonous positive correlation between the number of operation steps and the measured cognitive load, and map the number of operation steps to the range of 0 to 1 using range normalization. The calibration value.

[0128] Calibration process three: Weighting coefficients Calibration;

[0129] Step P6: Organize 120 to 200 students from all levels of primary school to high school to complete 30 to 50 questions with different modal configurations and number of operation steps in a formal practice scenario, and collect accuracy, response time and subjective fatigue score (Borg fatigue scale, 9-point system) simultaneously.

[0130] Step P7: Based on each question , , Using subjective fatigue rating as the dependent variable and employing it as the independent variable, a multiple linear regression was performed. The resulting regression coefficients, after normalization, are... , , The initial calibration value. The current calibration result reference range is... , , This implementation method takes , , The regression coefficients differ among students of different age groups. The system labels and stores the corresponding weights for each of the four educational stages: lower elementary, upper elementary, junior high, and senior high.

[0131] Working memory half-life parameters The initial values ​​are preset based on the students' age groups: 5-7 minutes for lower elementary grades (grades 1-3), 7-9 minutes for upper elementary grades (grades 4-6), 9-12 minutes for junior high school, and 12-15 minutes for senior high school. These preset ranges are based on experimental research in cognitive psychology regarding the working memory capacity of children and adolescents, and were determined by fitting preliminary experimental data from 120-200 students of different age groups. The specific method of the preliminary experiment was as follows: The actual number of valid answers and the time sequence for each student were recorded in the practice sequence. The moment when the accuracy rate began to decline was taken as the working memory saturation marker, and an exponential decay model was fitted using the least squares method. In The parameters are summarized by age group, and the mean ± standard deviation range is taken as the preset interval for each school stage.

[0132] Step P8: Safety Ratio Coefficient The experimental method for determining this is as follows: A scatter plot was drawn with the horizontal axis representing the accuracy rate of subsequent questions and the vertical axis representing the accuracy rate. The accuracy rate began to show a statistically significant decline (more than 10% decrease relative to the baseline). )of The threshold, the range of the mean ± standard deviation of the threshold, is the corresponding The value range is 0.6 to 0.85, and this embodiment takes... As a conservative estimate, the goal is to maximize effective practice capacity while retaining a 25% cognitive buffer margin.

[0133] Working memory half-life parameters The initial values ​​are preset based on the students' age groups: 5 to 7 minutes for lower elementary grades (grades 1-3), 7 to 9 minutes for upper elementary grades (grades 4-6), 9 to 12 minutes for junior high school, and 12 to 15 minutes for senior high school. The median of these preset intervals is used as the average of the preset values. The initial value is determined and continuously updated using the least squares method based on the actual decay trajectory in subsequent practice sequences to approximate the student's individual working memory characteristics.

[0134] Step 4: Calculate the current available cognitive capacity and mark the high-risk fatigue levels.

[0135] The current available cognitive capacity is obtained by subtracting the effective residual cognitive occupancy from the student's initial cognitive capacity threshold. The cognitive load requirements of the control and probing questions in the next level of diagnostic practice are compared with the current available cognitive capacity. When the cognitive load requirement of any question exceeds the safe range, the level is marked as a high-risk fatigue level.

[0136] Step 4 may specifically include:

[0137] Step S401: Calculate the current available cognitive capacity. :

[0138]

[0139] in, As the initial cognitive capacity threshold for students, This refers to the effective residual cognitive occupancy. , and All are with Dimensionless score values ​​with consistent dimensions can be directly subtracted.

[0140] Step S402: Obtain the cognitive load requirement value of the control questions in the next level of diagnostic practice. Cognitive load requirement of detection questions Judge separately and Does it exceed ,in This is a preset safety ratio coefficient. When... or Upon establishment, this level was marked as a high-risk fatigue level.

[0141] It should be noted that the above-mentioned preset safety ratio coefficient The value ranges from 0.6 to 0.85, and is used to reserve a certain cognitive margin on the basis of available capacity, so as to avoid completely consuming the students' available cognitive resources.

[0142] It should be noted that the above-mentioned initial cognitive capacity threshold for students This is determined based on the students' age group and historical practice performance data. In one embodiment, The result was calculated based on the product of the student's average maximum number of valid answers and the average cognitive load per question in the first three practice sessions. ,in This represents the average maximum number of valid answers across the first three practice sessions. This represents the average cognitive load per question across the first three practice sessions. and Using the same normalization scheme, The score is a dimensionless score, and Consistent units of measurement. For new users who do not yet have historical practice data, Preset default values ​​are used for different age groups: 3.0 for lower elementary grades, 4.0 for upper elementary grades, 5.0 for junior high school, and 6.0 for senior high school. These default values ​​are based on the statistical average of large-scale user data. Current available cognitive capacity. The actual size is affected by three factors: the first is The first is the individual baseline, namely the upper limit of the student's own working memory capacity; the second is... The cumulative level, i.e., the sum of residual cognitive load of completed questions in the current practice sequence, is directly related to the number of completed questions, the cognitive load value of each question, and the time interval between the completion of each question and the current moment; the third is the safety ratio coefficient. ,pass Retain a cognitive buffer margin on top of the available capacity to prevent students' cognitive resources from being exhausted to the limit.

[0143] In this embodiment, to reduce the cognitive load on students during the diagnostic process, modal lightweighting is performed on diagnostic exercises marked as high-risk for fatigue before presentation. Modal lightweighting includes two processes: First, high-cognitive-occupancy modal elements in control and probe questions are replaced with equivalent low-occupancy modal elements, such as replacing animated demonstrations with static step-by-step diagrams and complex interactive operations with selection / fill-in operations. Second, information presented in parallel across modalities is converted into a sequentially presented form, allowing students to process only a single modal of information at any given time. After these two processes, the cognitive load requirement of the diagnostic exercise pairs is reduced to within a safe range of available capacity, while maintaining the knowledge points tested and the diagnostic effectiveness of the questions.

[0144] Step 5: Perform fatigue compensation correction and perform hierarchical attribution determination.

[0145] Fatigue compensation correction is performed on the responses at each level of the diagnostic exercise sequence. A fatigue interference coefficient is calculated based on the effective residual cognitive occupancy at each level, and this coefficient is used to correct response time differences and attribution confidence. Based on the corrected data, a layer-by-layer attribution determination is performed to identify the bottleneck level and bottleneck type.

[0146] Step 5 may specifically include:

[0147] Step S501, perform diagnostic exercises on each level in the sequence. Obtain the effective residual cognitive occupancy at the time of response at this level. Calculate the fatigue interference coefficient :

[0148]

[0149] in, As the initial cognitive capacity threshold for students, hierarchical The effective residual cognitive occupancy at the moment of response, i.e., in step S302 The value to be calculated before the question at this level is presented. and Both have the same dimensions and are dimensionless score values; the ratio between the two is... It is a dimensionless coefficient, with a value ranging from 0 to 1. The larger the value, the higher the degree of fatigue interference when answering at that level.

[0150] Step S502, for each level Obtain the response time of the control question. (Unit: seconds) and response time of the probe question (Unit: seconds), calculate the difference in response time after correction. :

[0151]

[0152] in, The baseline response time parameter (in seconds) represents the expected additional time baseline for the detection problem relative to the control problem under fatigue-free conditions. This is a dimensionless fatigue interference coefficient. The unit is seconds, and If the dimensions are the same, they can be subtracted directly. The expected time dilation caused by fatigue disturbance is subtracted from the original difference to obtain the corrected response time difference. (Unit: seconds) Reflects the actual increase in time spent by students in the interdisciplinary knowledge connection stage after eliminating the effects of fatigue.

[0153] Step S503: For levels within the high-risk fatigue level, multiply the attribution confidence of their incorrect responses by a fatigue discount factor. Reduce the weights and obtain the attribution confidence after weight reduction. Fatigue discount factor. The calculation method is as follows:

[0154]

[0155] in, This is the discount adjustment constant, with a value ranging from 0.3 to 0.7; This is the lower limit of the discount factor, ranging from 0.2 to 0.4. The larger The smaller the value, the greater the weighting of the confidence level for misattribution at that level indicates more severe fatigue. For levels not within the high-risk fatigue level, the attribution confidence level remains unchanged. After weighting, the attribution confidence level is equal to the initial attribution confidence level.

[0156] It should be noted that the initial attribution confidence level mentioned above is pre-configured based on the combination of correct and incorrect labels for the control and probe questions: when the control question is correct but the probe question is incorrect, the initial attribution confidence level is set to a higher preset value; when the control question itself is incorrect, the initial attribution confidence level is set to a lower preset value; when both questions are correct, the initial attribution confidence level is set to the highest preset value. Each preset value is a constant in the range of 0 to 1, and is stored in the system configuration after pre-calibration.

[0157] Step S504, based on the corrected response time difference And after weighting down the attribution confidence, perform a bottom-up attribution determination for each level:

[0158] If a certain level's control question is correct but the probe question is incorrect, and the attribution confidence of that level after weight reduction is not lower than the reliable judgment threshold, then the interdisciplinary connection between that level and the upper level is determined to be a bottleneck, and the bottleneck type is marked as interdisciplinary connection defect type.

[0159] If a question at a certain level is incorrect, and the attribution confidence level after weight reduction is not lower than the reliable judgment threshold, then the knowledge point at that level is determined to be a bottleneck, the bottleneck type is marked as knowledge ontology defect type, and the attribution judgment to higher levels is terminated. The reason for termination is that when the underlying knowledge point ontology is not mastered, all upper-level connection judgments relying on that knowledge point are unreliable.

[0160] If both the control and detection questions at a certain level are correct, and the difference in response time after correction is... If the time consumption does not exceed the preset abnormal time threshold, it is determined that there is no bottleneck at this level, and the attribution determination is performed on the next higher level.

[0161] It should be noted that the aforementioned reliability threshold is a preset constant, ranging from 0.5 to 0.8. When the attribution confidence of a certain level after weighting is lower than the reliability threshold, it indicates that it is affected by fatigue interference, and the performance of that level is insufficient to support a reliable attribution conclusion.

[0162] In this embodiment, to avoid overlooking real knowledge bottlenecks due to overcompensation, levels with attribution confidence below the reliability threshold after weight reduction are marked as levels to be verified. This is based on the working memory half-life parameter. Effective residual cognitive occupancy when responding to questions at the level to be verified Calculate the required shortest recovery interval :

[0163]

[0164] in, This refers to the working memory half-life parameter (unit: minutes). This represents the effective residual cognitive occupancy when responding at the level to be verified. As the initial cognitive capacity threshold for students, To preset the safety ratio coefficient, To verify the cognitive load requirement of the diagnostic practice pair. , All are dimensionless score values, and the ratio between the two is a dimensionless number. The logarithmic term is dimensionless. and The units are consistent, and the unit is minutes. The above formula indicates that... The minimum waiting time required to decay to a level where the validation diagnostic practice pairs can be safely performed. When It has not exceeded At this time, no recovery interval needs to be inserted, and verification can be performed immediately.

[0165] In diagnostic exercises, the insertion length at subsequent positions in the sequence should not be less than [a certain value]. A cognitive recovery interval is established. No new practice questions are presented to students during this interval. After recovery, a simplified verification diagnostic exercise is added to the level to be verified for secondary confirmation. The simplified verification diagnostic exercise retains only the core assessment elements of control and probe questions, and the modality configuration uses a low cognitive occupancy scheme. The attribution results of the secondary confirmation overwrite the initial attribution conclusions.

[0166] Step 6: Generate targeted multimodal practice questions and implement adaptive difficulty adjustment.

[0167] Based on the bottleneck level and bottleneck type determined in the final attribution results of each level, targeted multimodal practice questions are generated and the difficulty level is adjusted.

[0168] Step 6 may specifically include:

[0169] Step S601: For the level where the bottleneck type is knowledge ontology defect type, match the single-point special practice template corresponding to the knowledge point of the level from the question template library, generate single-point special practice questions with the typical modality of the subject of the level, and reduce the difficulty level to the level corresponding to the current mastery value of the level.

[0170] Step S602: For levels where the bottleneck type is interdisciplinary connection deficiency, bridging practice templates involving knowledge points of adjacent two levels are matched from the question template library to generate bridging multimodal practice questions. These bridging multimodal practice questions connect knowledge points of two levels through progressive contextual transitions: first, independent application scenarios of the lower-level knowledge points are presented; then, application elements of the upper-level knowledge points are gradually introduced within the same scenario; and finally, the corresponding mapping relationship between the two disciplinary concepts is explicitly presented in the transitional scenarios.

[0171] It should be noted that the above correspondence between difficulty levels and mastery values ​​is a pre-configured mapping table. The lower the mastery value, the lower the corresponding difficulty level, ensuring that the difficulty of the exercises matches the students' current actual ability level.

[0172] It should be noted that the aforementioned progressive contextual transition refers to presenting the connection between two levels of knowledge in a step-by-step manner within the same exercise. The first stage only involves calculations or operations of the basic knowledge points, while the second stage introduces the application requirements of the higher-level knowledge points based on the results of the first stage, and the conceptual mapping between the two stages is indicated by textual annotations or graphic arrows.

[0173] In this embodiment, to control the presentation pace of exercises to match the student's current cognitive state, the generated targeted multimodal exercises are assembled and output according to presentation pace control parameters. These parameters include the display duration of each modal element, the modality switching interval, and the interaction waiting time. The display duration is determined based on the information density of the modal elements and the current available cognitive capacity. Certainly, the higher the information density or The lower the value, the longer the display duration. The modal switching interval is determined based on the cognitive transition requirements between adjacent modal elements.

[0174] In this embodiment, to improve the accuracy of fatigue prediction in subsequent practice sequences, the time-series data of the actual effective residual cognitive occupancy at each level in the sequence during this diagnostic practice are recorded, and the actual decay trajectory is compared with the current working memory half-life parameter. The predicted trajectory is compared, and the working memory half-life parameter is updated using the least squares method. The value of is determined. Simultaneously, the student's initial cognitive capacity threshold is updated based on the maximum number of valid answers received during this diagnosis. Updated working memory half-life parameters and students' initial cognitive capacity threshold Used for calculating the effective residual cognitive occupancy in the student's subsequent practice sequences and determining the risk of fatigue.

[0175] Comparative analysis with existing technologies:

[0176] Existing methods for automatically generating multimodal classroom exercises have systemic shortcomings in three dimensions: difficulty adaptation, fatigue modeling, and multimodal design. This invention proposes corresponding solutions to address these shortcomings, as detailed below.

[0177] I. Difficulty Matching Problem

[0178] Existing methods for generating practice questions typically set difficulty based solely on students' individual mastery of target knowledge points or their historical accuracy rate, neglecting the transmission effect of deficiencies in lower-level knowledge points on higher-level performance within a multi-level interdisciplinary dependency chain. When a student makes a mistake on a practice question, existing methods generally only perform a single attribution—directly attributing the error to insufficient mastery of the target knowledge point—without distinguishing whether the error stems from a deficiency in the target knowledge point itself (knowledge point deficiency type) or a lack of interdisciplinary connection between the underlying prerequisite knowledge points and the target knowledge point (interdisciplinary connection deficiency type). This confusion leads to a systematic deviation in the knowledge point positioning of subsequently generated practice questions: for connection deficiency bottlenecks, existing methods tend to reduce the difficulty of the target knowledge point itself, rather than specifically strengthening cross-level connection capabilities, ensuring that practice questions consistently fail to accurately target students' true skill gaps.

[0179] This invention constructs a multi-level dependency chain topology description through steps 1 and 2, and replaces single-attribution with layer-by-layer diagnostic practice pairs (control questions + probe questions). Control questions isolate and examine single knowledge points, while probe questions examine interdisciplinary applications at adjacent levels. The combination of correct and incorrect labels for both can accurately distinguish the level and type of bottleneck. Step 6 generates single-point specialized practice questions or bridging multimodal practice questions based on the bottleneck type, and maps the difficulty level to the actual mastery value of that level. This achieves accurate positioning and difficulty matching of students' true ability bottleneck nodes, overcoming the difficulty misalignment problem caused by single-attribution in existing methods.

[0180] II. The Impact of Learning Fatigue on Learning Outcomes

[0181] Current methods independently assess the cognitive load of each practice question, assuming that students maintain the same cognitive capacity throughout the practice sequence. This ignores the residual workload of completed questions in working memory and its non-linear decay over time. While this assumption is acceptable for short practice sessions (e.g., 1-3 questions), the cumulative effect of working memory residuals is significant in typical 15-25 minute continuous practice scenarios in primary and secondary school classrooms. Even if the cognitive load of individual questions remains within the theoretical threshold in the later stages of practice, the actual available cognitive capacity has been significantly reduced due to accumulated fatigue, leading to decreased attention and a sharp drop in accuracy. More seriously, current methods misjudge this fatigue-induced performance decline as a knowledge deficit, causing a systematic error in the direction of difficulty adjustment.

[0182] This invention introduces an effective residual cognitive occupancy through step 3. The exponential decay cumulative model, using the working memory half-life parameter The cognitive load contribution of completed questions is weighted by time decay to track the actual remaining state of students' working memory resources in real time. Step 4 is based on... Calculate the current available cognitive capacity Before presenting new questions, the system anticipates high-risk fatigue levels and implements modal lightweighting adjustments to proactively control cognitive consumption during the diagnostic process and prevent diagnostic operations from exacerbating fatigue accumulation. Step 5 introduces a fatigue interference coefficient. Fatigue compensation correction is applied to the response time difference, and a fatigue discount factor is used. The attribution confidence of high-risk fatigue levels is downweighted to prevent fatigue-induced performance decline from being misjudged as a knowledge bottleneck; for levels requiring verification with insufficient confidence after downweighting, a hypothesis based on... A second confirmation is added after calculating the shortest recovery interval, using a dual mechanism to ensure that the attribution conclusions simultaneously meet the requirements of fatigue exclusion and true bottleneck detection. These measures systematically overcome the negative impact of learning fatigue on diagnostic accuracy and practice effectiveness at three levels: cognitive state monitoring, fatigue feedforward control, and attribution posterior correction.

[0183] III. Advantages of Multimodal Practice Questions

[0184] In existing methods for generating practice questions, the introduction of multimodal elements often lacks coordinated design with the knowledge point dependency structure and students' cognitive state: on the one hand, the combination of each modality is fixed or random, failing to reflect the modal adaptability of different subject knowledge points; on the other hand, while multimodal parallel presentation enhances the richness of information expression, it also significantly increases the additional cognitive burden caused by modality switching (i.e., the modal interaction complexity component in cognitive load). This actually exacerbates learning difficulties when students' cognitive resources have already been depleted.

[0185] This invention demonstrates three systematic advantages in multimodal design: First, modality and subject-specific adaptability—control questions select typical modalities based on subject attributes (text and formulas for mathematics, text and diagrams for physics), while probe questions combine typical modalities from adjacent subject levels to present cross-disciplinary connection scenarios. This makes the modality configuration itself an explicit expression of the subject's knowledge structure, reducing students' cognitive consumption in unnecessary modality transitions. Second, dynamic modal adjustment of cognitive state perception—when a certain level is marked as high-risk for fatigue, the system replaces modal elements with high cognitive occupancy coefficients (such as animated demonstrations and complex interactions) with low-occupancy equivalent forms (such as static step-by-step diagrams and selection / filling operations), and converts cross-modal parallel presentation into a temporally serialized form. Without sacrificing diagnostic effectiveness, it proactively reduces the modal cognitive load requirement of the questions to a safe capacity range. Third, the progressive contextual transition design of bridging multimodal exercises—bridging exercises designed to address bottlenecks in interdisciplinary connections—explicitly present the corresponding mapping relationship between two disciplinary concepts in the same question through a phased and progressive approach. Multimodal annotations (text annotations and graphic arrows) visualize the process of connecting abstract interdisciplinary concepts, specifically strengthening students' interdisciplinary transfer ability at key junctions of dependency chains.

[0186] This invention continuously calculates the effective residual cognitive occupancy during the presentation of diagnostic practice sequences. Based on an exponentially weighted accumulation method, the cognitive load value of completed questions is decayed and accumulated according to the working memory half-life parameter, thereby obtaining the actual occupancy level of the student's working memory resources at the current moment. Furthermore, by calculating the currently available cognitive capacity and comparing it with the cognitive load requirement value of subsequent questions, high-risk levels of fatigue are identified before presentation, avoiding the direct presentation of diagnostic practice pairs that exceed the student's current capacity.

[0187] In the attribution phase, a fatigue interference coefficient is used to correct for response time differences, eliminating the interference of fatigue-induced time dilation on the assessment of interdisciplinary connectivity. A fatigue discount factor is used to reduce the attribution confidence of high-risk fatigue levels, preventing levels severely affected by fatigue from being directly attributed to knowledge deficiencies. These two corrections ensure that the results of the hierarchical attribution determination reflect students' knowledge mastery levels, rather than performance decline due to cognitive resource depletion. For levels requiring verification with insufficient attribution confidence after weighting, verification is performed by inserting a cognitive recovery interval and adding secondary confirmation, avoiding overcorrection that could lead to the omission of genuine bottlenecks.

[0188] Finally, based on the bottleneck levels and types determined after fatigue compensation correction, single-point specialized practice questions or bridging multimodal practice questions are generated respectively. This ensures that the difficulty level and knowledge point positioning of subsequent practice questions accurately correspond to the students' actual ability bottleneck nodes, thereby improving the relevance of practice within the limited classroom practice time.

[0189] Example 2

[0190] This embodiment uses the complete process of a junior high school student (hereinafter referred to as Student A) in a physics classroom exercise as an example to illustrate how the method of the present invention runs end-to-end and demonstrates its effect.

[0191] Background and initial parameter settings:

[0192] Student A is a second-year junior high school student. The target knowledge point for this exercise is "Physics: Application of Equations of Uniformly Accelerated Linear Motion" (knowledge point identifier KN-PHY-007). Based on Student A's historical exercise data, the system initializes the following parameters: working memory half-life parameter. (minutes); Initial cognitive capacity threshold (Based on calculations from the first three practice sessions) , Preset safety ratio coefficient Cognitive load weighting , , Discount Adjustment Constant Discount factor lower limit Reference response duration parameter (seconds); Reliability judgment threshold 0.6; Abnormal time consumption threshold 45 (seconds).

[0193] Step 1 execution process: Obtain multi-level dependency chain topology description

[0194] The system performs a reverse traversal with KN-PHY-007 as the endpoint and identifies the following three-level interdisciplinary dependency chain:

[0195] Level 1 (Bottom Level): Chinese Language - Condition Extraction (KN-CHN-001, Subject CHN, Mastery Value) → Level 2 (Intermediate Level): Mathematics - Solving Linear Equations in One Variable (KN-MATH-012, Subject MATH, Mastery Value) → Level 3 (Target Level): Physics - Application of Equations of Uniformly Accelerated Linear Motion (KN-PHY-007, Subject PHY, Mastery Value) )

[0196] The chain has a depth of 3 levels and spans three subject identifiers, meeting the screening criteria, and was generated as a multi-level dependency chain topology description.

[0197] Step 2 execution process: Generate diagnostic practice pair sequences

[0198] The system generates a diagnostic practice pair for each of Level 1 and Level 2 from the bottom up, forming a sequence of diagnostic practice pairs containing 4 questions:

[0199] The first pair (Level 1, Language Arts - Condition Extraction): The control question is a pure text modality, presenting a physics question stem paragraph containing multiple known quantities and implicit conditions. Students are required to extract the valid conditions and fill in a list, which only tests the ability to extract conditions. The exploration question is a cross-modal combination of text and formula. Based on the extracted conditions, students are required to fill in the conditions into the skeleton of a given algebraic expression, which tests the ability to convert and connect the extracted conditions into mathematical symbols.

[0200] The second pair (Level 2, Mathematics - Solving Linear Equations in One Variable): The control questions are text and formula-based, directly presenting a linear equation in one variable and requiring students to solve for the unknowns, only testing their equation-solving ability; the exploration questions are a cross-modal combination of text, formulas, and diagrams, with the diagrams including a schematic of uniformly accelerated motion, requiring students to read the initial velocity and acceleration values ​​from the diagrams and then write equations to solve for the motion time, testing their interdisciplinary ability to connect equation solving with reading physical motion diagrams.

[0201] Step 3 execution process: Collect response data and calculate effective residual cognitive occupancy.

[0202] Student A completed four diagnostic questions in sequence. The answer data and cognitive load values ​​are shown below (number of modal elements for each question). Modal interaction complexity and number of operation steps (All ranges have been normalized to the range of 0 to 1)

[0203] Question 1 (First pair of control questions): Correct. Second, , , , .

[0204] Question 2 (First pair of probe questions): Correct. Second, , , , .

[0205] Question 3 (Second pair of control questions): Correct. Second, , , , .

[0206] Question 4 (Second pair of probe questions): Incorrect. Second, , , , .

[0207] At the time after completing question 4 ( Taking this as an example, to calculate the effective residual cognitive occupancy, let's assume that the time interval between completing each question is approximately equal to the corresponding response time plus 0.5 minutes of presentation and switching time. minute, minute, minute, :

[0208]

[0209] Step 4: Calculate available cognitive capacity and mark fatigue risk.

[0210] Taking the moment before question 4 is presented (questions 1 to 3 have been completed) as an example for calculation:

[0211]

[0212]

[0213] Cognitive load requirement for question 4 Safety capacity threshold .because The second pair of probe questions did not trigger the fatigue high-risk flag and were presented normally in the original modal configuration (text, formula and diagram combination).

[0214] Step 5 Execution Process: Fatigue Compensation Correction and Layered Attribution Determination

[0215] For Level 1 (first pair): Question 1, the control question, has no completed questions before it is presented. Fatigue interference coefficient The response time for the control question is 42 seconds, and the response time for the probe question is 68 seconds. The difference in response time after correction is... The timeout period is within 45 seconds, which is below the abnormal time threshold. The control question is correct, the detection question is correct, and... The threshold was not exceeded, so there is no bottleneck in level 1.

[0216] For Level 2 (Second Pair): Questions 1 and 2 were completed before Question 3 was presented. Fatigue interference coefficient The response time for the control question is 38 seconds, and the response time for the probe question is 95 seconds. The difference in response time after correction is... The time consumption exceeded the abnormal time threshold of 45 seconds, indicating a significant increase in time consumption for interdisciplinary collaboration. The second level was not marked as high-risk for fatigue. (lower), fatigue discount factor The control question was correct, and the probe question was incorrect. The initial attribution confidence score was set to a high preset value of 0.80. The weighted attribution confidence score was then reduced. It is higher than the reliable judgment threshold of 0.6.

[0217] Attribution conclusion: The bottleneck is the interdisciplinary connection between Level 2 (Mathematics - Solving Linear Equations in One Variable) and Level 3 (Physics - Application of Equations of Uniformly Accelerated Linear Motion). The bottleneck type is the interdisciplinary connection defect type. After weighting, the attribution confidence is 0.763, which can be directly used for generating practice questions in step 6.

[0218] Step 6: Execution process: Generate targeted multimodal practice questions

[0219] Level 2 Mastery Value The mapping table corresponds to difficulty level 4 (transfer application level). The system matches bridging practice templates from the question template library and generates bridging multimodal practice questions with progressive contextual transitions, as follows:

[0220] Phase 1 (Pure Mathematics, Text and Formula Modal): Giving the kinematic equations Known , , Find the duration of the exercise. This stage uses only text and formula modalities to confirm that the student's equation transformation and numerical calculation abilities are intact, and the solution results are then presented. Retain as known quantities for the second stage.

[0221] Phase Two (Interdisciplinary Integration, Text, Formulas, and Step-by-Step Diagram Modal): Presenting a labeled... , , and A schematic diagram of uniformly accelerated linear motion, with red arrows and text annotations indicating "the first stage required". This refers to the motion time of the objects in this diagram, forming a visual anchor point for interdisciplinary concept mapping; students are then required to use... Calculate motion displacement This completes the interdisciplinary transfer from solving equations to calculating kinematic displacements.

[0222] Presentation rhythm control parameters: The first stage formula display lasts for 8 seconds, the second stage graphic display lasts for 12 seconds, the interval between the two stages is 3 seconds, and the interaction waiting time is based on the current available cognitive capacity. Determined dynamically.

[0223] Comparison of running effects:

[0224] If the existing method is used, the error in question 4 (the second pair of probe questions) will be directly attributed to insufficient mastery of the knowledge point of "Application of Physics Uniformly Accelerated Linear Motion Equations". Subsequent practice questions will reduce the overall difficulty of physics motion equation questions (such as reducing the number of parameters and shortening the reasoning chain). However, the real bottleneck for students lies in the interdisciplinary transfer of the mathematical equation solution to the physical motion context. Lowering the difficulty of physics questions does not have a targeted strengthening effect on this connection obstacle, making it difficult to overcome the bottleneck in subsequent practice. This implementation method accurately locates the bottleneck as an interdisciplinary connection deficiency by diagnosing layer by layer (level 2 control questions are correct, probe questions are incorrect, and confidence is 0.763 after fatigue elimination). The generated bridging multimodal practice questions explicitly show the correspondence between the mathematical equation solution and the physical motion parameters with step-by-step diagrams and diagram arrows. At difficulty level 4 (transfer application level), it specifically strengthens the interdisciplinary transfer ability, accurately hitting the students' real ability weaknesses and achieving higher practice relevance and learning efficiency within the limited classroom practice time.

[0225] Example 3: Two-level dependency chain between Chinese and Mathematics, a bottleneck scenario due to deficiencies in knowledge ontology.

[0226] This embodiment takes the process of a sixth-grade elementary school student (hereinafter referred to as student B) in a math classroom exercise as an example, with the target knowledge point being "Mathematics - Application of Fraction Division", to illustrate the operating logic of the method of the present invention under the conditions of two-level dependency chain and knowledge ontology defect bottleneck.

[0227] Initial parameter settings:

[0228] Student B is an upper elementary school student: (minute); ( , ); ; , , ; ; ; (seconds); Reliability judgment threshold 0.6; Abnormal time consumption threshold 35 (seconds).

[0229] Step 1: Dependency chain topology description;

[0230] The system identified a two-level interdisciplinary dependency chain: Level 1 (bottom level): Language Arts - Application Problem Condition Analysis (KN-CHN-003, CHN, → Level 2 (Target Level): Mathematical Fraction Division Application (KN-MATH-021, MATH) ).

[0231] Step 2: Generate diagnostic practice pair sequences;

[0232] One diagnostic practice pair (2 questions) was generated: Control questions (Level 1): Plain text modality, requiring students to extract valid conditions from the text of application problems containing percentage relationships and fill in an information extraction table. ; Probe Question (Level 1): Cross-modal combination of text and formula; after recognizing the fractional conditions, fill the conditions into the skeleton of the fractional division equation. .

[0233] Step 3: Collect data;

[0234] Control question (Question 1): Incorrect. Second, ; Probe question (Question 2): Incorrect. Second, Before Question 2 is presented , Safety limit It does not trigger the high-risk fatigue marker.

[0235] Step 5: Attribution determination;

[0236] Level 1 control questions have no preceding questions. , The control question itself is flawed; the initial attribution confidence level was set to a low preset value of 0.65, and the attribution confidence level was reduced after weighting. (Reliable judgment threshold) The knowledge point ontology of level 1 is determined to be the bottleneck, the bottleneck type is knowledge ontology defect type, and the attribution judgment to level 2 is terminated.

[0237] Step 6: Generate targeted practice questions;

[0238] Level 1 Mastery Value This corresponds to difficulty level 3 (comprehensive application level). Single-point specialized practice questions for language arts application problems are generated in plain text mode, requiring students to gradually extract the total quantity, percentage quantity, and corresponding quantity from paragraphs containing score information. The difficulty level is reduced to level 3.

[0239] Comparison with Example 2: This example demonstrates a two-level (not three-level) dependency chain, a knowledge ontology-deficient bottleneck, and the primary school level ( Shorter The three dimensions of difference (smaller) verify that the method of the present invention can accurately locate the bottleneck level at a shallower dependency chain depth, and can promptly terminate the attribution process to a higher level and avoid the accumulation of errors when the control problem itself is wrong.

[0240] Example 4: The three-tiered dependency chain of chemistry, mathematics, and Chinese language, highlighting the bottleneck in interdisciplinary connection at the high school level.

[0241] This example uses a first-year high school student (hereinafter referred to as Student C) as an example, with the target knowledge point being "Chemistry - Quantitative Calculation of Chemical Equations". It focuses on explaining the characteristics of parameters in the high school stage and the identification of bottlenecks in interdisciplinary connection, as well as the generation of bridging practice questions.

[0242] Initial parameter settings:

[0243] Student C is a high school student: (minute); ( , ); ; , , ; ; ; (seconds); Reliability judgment threshold 0.6; Abnormal time consumption threshold 50 (seconds).

[0244] Step 1: Dependency chain topology description;

[0245] A three-level interdisciplinary dependency chain was identified: Level 1 (bottom layer): Language and Chemistry contextual condition extraction (KN-CHN-005, CHN, → Level 2 (Intermediate Level): Mathematical Proportional Relationship Calculation (KN-MATH-031, MATH) → Level 3 (Target Level): Chemistry · Chemical Equation Econometrics Calculation (KN-CHM-008, CHM) ).

[0246] Step 2: Generate diagnostic practice pair sequences;

[0247] Two diagnostic practice pairs (4 questions) are generated, with the cognitive load requirement for each question as follows: Level 1 control questions. Level 1 Detection Question Level 2 control question Level 2 Detection Question .

[0248] Steps 3-4: Data collection and fatigue risk assessment;

[0249] Answer status: Question 1 is correct ( Second, Question 2 is correct. Second, Question 3 is correct. Second, Question 4 is incorrect. Second, Before Question 4 is presented: , , No high risk of fatigue was triggered. This example reflects the high school stage. Larger Due to its relatively long duration, even with a high load in question 4, the safety limit was not exceeded.

[0250] Step 5: Attribution determination;

[0251] Level 1: Both the control and detection questions are correct. Second Seconds, no bottleneck detected. Level 2: , , Second Seconds; Level 2 fatigue high risk not triggered. If the control question is correct but the probe question is incorrect, the attribution confidence level will be reduced after weighting. The conclusion of the attribution is that there is a bottleneck in the interdisciplinary connection between Level 2 (mathematics and proportional relationship calculation) and Level 3 (chemistry and chemical equation quantitative calculation).

[0252] Step 6: Generate bridging exercises;

[0253] Level 2 Mastery Value The difficulty level is set at 4. The first stage (pure mathematics) requires students to solve proportional equations based on the known masses and molar masses of the reactants. The second stage (interdisciplinary integration) uses a diagrammatic arrow to indicate that "the mass ratio obtained in the first stage is the numerical representation of the stoichiometric ratio," visualizing the correspondence between mathematical proportional relationships and stoichiometric ratios.

[0254] Comparison with Example 2: This example covers the combination of chemistry, mathematics and Chinese language subjects, which verifies the generalization ability of the present invention under non-physics science subject target knowledge points; at the same time, it demonstrates the characteristics of a more relaxed fatigue threshold and more sufficient overall capacity management under the parameter combination of high school level, reflecting the cross-level adaptability of the parameter system of the present invention.

[0255] Example 5: Scenario of multiple parallel dependency chains sharing the underlying node

[0256] This embodiment takes a junior high school student (hereinafter referred to as student D) as an example, and the target knowledge point is "Physics and the comprehensive application of Ohm's law in circuits". It focuses on explaining the deduplication diagnosis mechanism when multiple parallel dependency chains share the same underlying node.

[0257] Initial parameter settings:

[0258] Student D: (minute); ( , ); ; , , ; ; ; (seconds); Reliability judgment threshold 0.6; Abnormal time consumption threshold 45 (seconds).

[0259] Step 1: Dependency chain topology description;

[0260] The system identified two dependency chains that share the same underlying node:

[0261] Dependency chain A (chain depth 3): Condition extraction for Chinese language and circuit problems (KN-CHN-006) → Mathematics: Analysis of Linear Function Graphs (KN-MATH-018) → Comprehensive Application of Ohm's Law in Physics and Circuits (KN-PHY-019) ).

[0262] Dependency chain B (chain depth 2): Language Arts / Circuit Problem Condition Extraction (KN-CHN-006, shared node) → Comprehensive Application of Ohm's Law in Physics and Circuits (KN-PHY-019) ).

[0263] Step 2: Diagnose practice pairs of sequences (deduplication strategy);

[0264] Based on the permutation strategy, the shallower dependency chain B is completed first, followed by dependency chain A; the diagnostic practice pair for the shared node KN-CHN-006 is generated only once, and the attribution conclusions apply to both chains simultaneously. The final sequence contains 4 questions (6 would be needed without deduplication, reducing the number of questions by 33%):

[0265] Question 1: Shared Node Control Problem (Extracting only Chinese language conditions); Question 2: Dependency Chain B Probing Problem (Connecting Chinese language to direct application of Ohm's Law in physics); Question 3: Dependency Chain A Level 2 Control Problem (Pure mathematics, analysis of linear function graphs); Question 4: Dependency Chain A Level 2 Probing Problem (Connecting mathematical graph analysis to reading Ohm's Law graph in physics).

[0266] Steps 3-5: Attribution determination results;

[0267] Question 1 is correct, Question 2 is correct, Question 3 is correct, Question 4 is incorrect. (seconds). Level 1 attribution of dependency chain B (shared node): Both control and probe questions are correct, and there is no bottleneck in the decision. This conclusion also applies to level 1 of dependency chain A. Level 2 attribution of dependency chain A: Control question is correct, probe question is incorrect. After calculation, the time consumption exceeded the abnormal threshold and the confidence level met the standard. It was determined that the bottleneck between Level 2 (Mathematics - Linear Function Image Analysis) and Level 3 (Physics - Comprehensive Application of Ohm's Law in Circuits) was a bottleneck with interdisciplinary connection defects.

[0268] Step 6: Generate bridging exercises;

[0269] Level 2 Mastery Value This corresponds to difficulty level 4. Generate bridging multimodal practice problems: The first stage uses purely mathematical diagrams (…). - (Graph of a straight line equation) Students are required to read the slope and intercept; in the second stage, in the context of Ohm's law graphs, a red arrow is used to indicate that "the slope is the resistance value". Students are required to use the results of image analysis to calculate the circuit resistance and visualize the interdisciplinary transfer path between mathematical analysis and physical quantity extraction through step-by-step diagrams. The difficulty level is set at 4.

[0270] A comparison with the aforementioned embodiments illustrates that this embodiment covers scenarios with multiple parallel dependency chains and shared underlying nodes, verifying the effectiveness of the shared node deduplication strategy in reducing unnecessary cognitive consumption while ensuring the integrity of end-to-end diagnosis. It also demonstrates the actual operational effect of the strategy of arranging multiple dependency chains by depth, from shallow to deep. This embodiment, along with Embodiment 2 (three-level chain / interdisciplinary connection defect type / physics-mathematics-Chinese / middle school), Embodiment 3 (two-level chain / knowledge ontology defect type / mathematics-Chinese / primary school), and Embodiment 4 (three-level chain / interdisciplinary connection defect type / chemistry-mathematics-Chinese / high school), collectively covers three topological structures (two-level / three-level / multi-chain shared nodes), two bottleneck types, four subjects (physics / mathematics / chemistry / Chinese), and three educational stages (primary / middle / high school), verifying the applicability and consistency of this invention in various practical teaching scenarios.

[0271] Example 6

[0272] This embodiment proposes a multimodal classroom exercise automatic generation and difficulty adaptive adjustment system based on cognitive load management, including:

[0273] The dependency chain topology extraction module is used to obtain the cross-disciplinary multi-level dependency chain of the target knowledge point in the multi-disciplinary knowledge network, extract the knowledge point identifier, the subject identifier, and the current mastery value of each level node, and generate a multi-level dependency chain topology description.

[0274] The diagnostic practice pair generation module is used to generate diagnostic practice pairs consisting of control and probe questions in a bottom-up manner according to the multi-level dependency chain topology, and outputs a sequence of diagnostic practice pairs.

[0275] The module for collecting answer data and calculating cognitive load is used to present diagnostic questions to students in sequence and collect the correct and incorrect marks, response time and operation interval time. It calculates the effective residual cognitive load at the current moment based on an exponential weighted accumulation method.

[0276] The cognitive capacity assessment and risk labeling module is used to calculate the current available cognitive capacity, compare the cognitive load requirement of subsequent diagnostic practice pairs with the current available cognitive capacity, and label the high-risk level of fatigue.

[0277] The fatigue compensation attribution module is used to correct the response time difference and attribution confidence of each level based on the fatigue interference coefficient, perform layer-by-layer attribution judgment, and determine the bottleneck level and bottleneck type.

[0278] The exercise generation and difficulty adjustment module is used to generate targeted multimodal exercises based on bottleneck level and bottleneck type, adjust the difficulty level to match the student's current mastery level, and assemble and output according to the presentation rhythm control parameters.

[0279] By way of example, the system proposed in this invention may also include various features and combinations thereof in the method embodiments, which will not be elaborated here.

[0280] See Figures 2-8 The following application examples are proposed:

[0281] A middle school eighth-grade student (Student A) was required to complete a physics application problem involving "buoyancy calculation" during a class exercise. The backend processing server identified a multi-disciplinary, multi-level dependency structure for this target knowledge point from a multi-disciplinary knowledge network: buoyancy calculation relies on the ability to solve mathematical equations, which in turn relies on the ability to extract contextual information from language. Student A's historical practice data was pre-stored in the system, and the initial cognitive capacity threshold was set. The working memory half-life parameter was estimated based on historical data from the first three practice sessions. The initial values ​​are preset according to the secondary school stage. The classroom practice terminal provides student A with a multimodal question presentation interface, and the background processing server is responsible for all data processing tasks such as dependency chain analysis, cognitive load calculation, and exercise question generation.

[0282] The backend processing server performs a reverse traversal of the multidisciplinary knowledge network with "buoyancy calculation" as the endpoint, obtaining all in-degree paths. After screening, dependency chains with a depth exceeding two levels and spanning two or more subject identifiers are retained, ultimately extracting a three-layer cross-disciplinary dependency chain. This dependency chain, from the bottom layer to the target knowledge point, is as follows: Language condition extraction (level 1) → Mathematical equation solving (level 2) → Physics buoyancy calculation (level 3, i.e., the target knowledge point). The knowledge point identifier, the subject identifier, and the current mastery value of student A for each level node are shown in the table below.

[0283] Table 1. Multilevel Dependency Chain Topology Description

[0284]

[0285] The system, following a topological order from Level 1 to Level 2, matches templates layer by layer from the question template library and fills in parameters to generate two diagnostic practice pairs, totaling four diagnostic questions, forming a diagnostic practice pair sequence. Level 1 control questions are presented in a text modal, testing the extraction of linguistic conditions; Level 1 probe questions are presented in a text-formula cross-modal, testing the process of extracting conditions from the linguistic question stem and connecting them to the formulation of mathematical equations. Level 2 control questions are presented in a text-formula modal, testing the solution of mathematical equations; Level 2 probe questions are presented in a text-formula-diagram combination modal, testing the process of applying the equation solution to a physical buoyancy scenario. The modal configuration and cognitive load requirements for each diagnostic question are shown in the table below.

[0286] Table 2 Diagnostic Exercise Sequence Configuration

[0287]

[0288] The classroom practice terminal pushes four diagnostic questions to Student A sequentially according to the topology. The backend processing server records the correctness mark, response time, and operation interval for each question. After each question is completed, the student... (The sentence is incomplete and requires more context to translate accurately.) Calculate the effective residual cognitive occupancy at the current moment. ,in For the first The cognitive load of the question. For the first The time interval between completing the question and the current moment. Set the value to 12 minutes (preset value for middle school level). From the formula The calculated components were weighted and summed after mean normalization. The collected data for each question and the effective residual cognitive occupancy after completing question 3 are shown in the table below.

[0289] Table 3. Diagnostic response data and effective residual cognitive occupancy.

[0290]

[0291] After completing question 3, the backend processing server uses the formula Calculate the currently available cognitive capacity, where Assuming it is the initial cognitive capacity threshold for student A, To achieve the effective residual cognitive occupancy after completing Question 3, we need to obtain... The system then obtains the cognitive load requirement value for Level 2 probe question (question 4). And take the preset safety ratio coefficient. Calculate the safety limit .because Level 2 was marked as a high-risk fatigue level. The system then performed modal lightweighting adjustment on Question 4: the dynamic demonstration in the diagram was replaced with a static step-by-step diagram, and the cross-modal parallel information was converted into a time-series serialized presentation. After adjustment, the cognitive load requirement value of Question 4 was reduced to 0.41, which meets the safety range requirements.

[0292] Table 4. Results of Cognitive Capacity Assessment and Fatigue Risk Marking

[0293]

[0294] The backend processing server performs fatigue compensation corrections for Level 1 and Level 2 respectively. The effective residual cognitive load at the time of response for Level 1 is... (No preceding questions yet), fatigue interference coefficient Level 2 answering time (after completing question 2 and before question 3 is presented) effective residual cognitive occupancy Fatigue interference coefficient Take the baseline response duration parameter. Calculate the difference in corrected response time between the two levels in seconds. .

[0295] Level 1: The preset abnormal time consumption threshold is 28 seconds. The threshold was not exceeded. Level 2: The actual response time for Question 4 after lightweighting adjustments was recorded as 74 seconds. The time elapsed was within seconds, which did not exceed the abnormal time consumption threshold.

[0296] Level 2 is marked as a high-risk fatigue level, and a discount adjustment constant is applied. Discount factor lower limit Fatigue discount factor Question 3 (control question) is incorrect. The initial attribution confidence score was preset to 0.55. The attribution confidence score after weighting is... The value is below the reliability threshold of 0.60, so Level 2 is marked as a level to be verified. The required shortest recovery interval is calculated using the formula. (Simplified version of the cognitive load requirement for the verification and diagnostic exercise) ,Right now The level of performance has not exceeded the requirements for safe execution verification, so no recovery interval is needed. A simplified version of the verification diagnostic exercise pair can be directly added. Verification results show that the control question for the knowledge point (equation solving) corresponding to question 3 still has errors after secondary confirmation. The secondary confirmation attribution confidence level is 0.72, which is higher than the reliable judgment threshold. The final judgment level 2 is a knowledge ontology defect-type bottleneck. The attribution judgment results for each level are summarized below.

[0297] Table 5. Results of Hierarchical Attribution Determination

[0298]

[0299] Level 2 (equation solving) has been identified as a knowledge ontology-deficient bottleneck, with a current mastery level of 0.65. The system matches a single-point practice template for equation solving from the question template library, generates single-point practice questions using a typical mathematical modality (text + formula), and lowers the difficulty level to a low-to-medium difficulty level (Level 2) based on the mapping relationship corresponding to the mastery level of 0.65. Since the attribution determination for Level 2 has terminated the progression to Level 3 (buoyancy calculation), bridging practice questions are not generated for Level 3 at this time. Simultaneously, the system adjusts the difficulty level based on the currently available cognitive capacity. Determine the presentation rhythm control parameters: appropriately extend the display duration, set the modal switching interval to 3 seconds, and set the interaction waiting time to 5 seconds.

[0300] Table 6 Results of Targeted Practice Questions Generation

[0301]

[0302] After completing this diagnosis, the system records time-series data of the actual effective residual cognitive occupancy at each level, and compares the actual decay trajectory with data based on... The predicted trajectory is compared with the predicted trajectory over a period of minutes, and then updated using the least squares method. Simultaneously updated based on the maximum number of valid answers. The updated version and It will be used for cognitive load calculation and fatigue risk assessment in student A's subsequent practice sequence.

[0303] The entire data flow starts from the three-layer dependency chain topology description extracted in step 1. Step 2 generates a sequence of diagnostic practice pairs with modal configurations. Step 3 collects correct / incorrect labels and response times, and calculates the effective residual cognitive occupancy in real time. Step 4 triggers fatigue high-risk labeling based on available cognitive capacity and drives modal lightweight adjustment. Step 5 applies the fatigue interference coefficient to the attribution confidence weighting and finally identifies the knowledge ontology defect bottleneck through secondary confirmation. Step 6 generates single-point specialized practice questions with difficulty matching based on the bottleneck type and mastery value, and configures the presentation rhythm parameters. The output data of each step directly serves as the input of the next step, ensuring data continuity and logical consistency throughout the entire process from dependency chain identification to targeted practice question output.

[0304] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for automatic generation and adaptive difficulty adjustment of multimodal classroom exercises based on cognitive load management, characterized in that: include: Perform a reverse traversal of the multidisciplinary knowledge network with the target knowledge point as the endpoint, filter in-degree paths with a depth of more than two levels and spanning two or more subject identifiers, and obtain a set of cross-disciplinary dependency chains. Extract the knowledge point identifier, subject identifier, and current mastery value of each node at each level in each interdisciplinary dependency chain, organize the node information at each level in topological order, and generate a multi-level dependency chain topology description of the target knowledge point. Diagnostic practice pairs are generated layer by layer from bottom to top according to the multi-level dependency chain topology. Each diagnostic practice pair consists of a control question and a probe question. The control question only examines the knowledge point at the current level itself and is presented in the typical modality of the subject at that level. The probe question examines the connection and application of the knowledge point at the current level with the node at the previous level and is presented in a cross-modal form that combines the typical modalities of two adjacent subject levels. The diagnostic practice pairs at each level are arranged in the multi-level dependency chain topology order, and the diagnostic practice pair sequence is output. The diagnostic practice test sequence is presented to students sequentially for them to answer. The correct / incorrect markings, response times, and operation intervals for each question are collected. After each question is completed, the cognitive load values ​​of the completed questions are exponentially weighted and accumulated using the working memory half-life parameter to calculate the effective residual cognitive load at the current moment. : ； in, This represents the number of questions completed. For the first The cognitive load of the question. For the first The time interval between the completion of the problem and the current time. For working memory half-life parameters; The initial cognitive capacity threshold of students Subtract the effective residual cognitive occupancy Obtain the current available cognitive capacity Obtain the cognitive load requirement value for each question in the next level of diagnostic practice, and then compare the cognitive load requirement value with the current available cognitive capacity and the preset safety ratio coefficient. product The comparison is performed, and when the cognitive load requirement of any question exceeds the product, the level is marked as a high-risk fatigue level. Modal lightweighting adjustment is then performed on diagnostic practice pairs marked as high-risk fatigue levels before presentation. This modal lightweighting adjustment includes: replacing modal elements with high cognitive occupancy coefficients with equivalent modal elements with low occupancy coefficients, and converting cross-modal parallel presentation information into a temporally serialized form, thereby reducing the cognitive load requirement of the diagnostic practice pairs to a lower level. The knowledge points tested and the diagnostic effectiveness remain unchanged; For each level The effective residual cognitive occupancy before the presentation of the control question at this level. With the initial cognitive carrying threshold The ratio is used to calculate the fatigue interference coefficient. ; using the fatigue interference coefficient according to Fatigue compensation correction is performed on the response time differences at each level to obtain the corrected response time differences. ,in and These are the response times for the detection and control questions, respectively. The baseline response time parameter is used; for levels marked as high-risk fatigue levels, the response time is determined based on the fatigue interference coefficient. The fatigue discount factor is calculated, and the initial attribution confidence of incorrect answers is weighted downwards to obtain the weighted attribution confidence. This is the discount adjustment constant. The lower limit of the discount factor is used; based on the corrected response time difference and the weighted attribution confidence, a bottom-up attribution determination is performed to determine the bottleneck level and bottleneck type. Based on the bottleneck level and bottleneck type, a corresponding template is matched from the question template library to generate targeted multimodal practice questions, and the difficulty level is adjusted to the level corresponding to the current mastery value of the bottleneck level.

2. The method according to claim 1, characterized in that, The multidisciplinary knowledge network is stored in a directed graph structure, where nodes represent knowledge points of each discipline and directed edges represent prerequisite dependencies between knowledge points. The construction method of the multidisciplinary knowledge network is as follows: subject experts mark the granularity of knowledge points and cross-disciplinary prerequisite dependencies according to the curriculum standards to form an initial directed graph structure. After the system is running, it performs statistical analysis on historical answer data based on the Bayesian knowledge tracing model. When there is a significant positive correlation between a student's answer performance on a certain target knowledge point and the mastery of a certain prerequisite knowledge point, the weight of the corresponding directed edge is automatically supplemented or strengthened. Each node stores a knowledge point identifier, its subject identifier, a list of recommended modality types, and the current student's mastery value. The mastery value is updated in real time by the Bayesian knowledge tracing model based on the student's historical answer sequence.

3. The method according to claim 1, characterized in that, The multimodal exercises cover the following modalities: text modality, which presents the question stem and conditions in natural language; formula modality, which presents quantitative relationships using mathematical symbols and equations, suitable for mathematical and science knowledge points; graphic modality, which presents spatial relationships or physical scenarios using static graphics; step-by-step graphic modality, which presents the process deduction using multiple static images arranged chronologically; and interactive operation modality, which requires students to complete the answer through drag-and-drop, filling, or clicking operations. The cognitive occupancy coefficient of each modality type is stored as a preset constant in the system configuration, and the cognitive occupancy coefficient of the interactive operation modality is higher than that of the text modality, formula modality, graphic modality, and step-by-step graphic modality.

4. The method according to claim 1, characterized in that, The effective residual cognitive occupancy The calculation method is as follows: ； in, This represents the number of questions completed at the current moment. For the first The cognitive load of the question. For the first The time interval between the completion of the problem and the current time. For working memory half-life parameters; The cognitive load value The calculation method is as follows: ； in, For the first The number of modal elements contained in the question. For the first The modal interaction complexity score of the question. For the first The number of operation steps required to answer the question. , , For pre-calibrated constant weights and satisfying , , , Mean normalization based on the range was performed before calculation.

5. The method according to claim 4, characterized in that, The working memory half-life parameter The initial values ​​are preset according to the students' age group: 5 to 7 minutes for lower elementary grades, 7 to 9 minutes for upper elementary grades, 9 to 12 minutes for junior high school, and 12 to 15 minutes for senior high school. After each diagnostic exercise on the sequence is completed, the actual decay trajectory is compared with the current one. The predicted trajectory is compared, and the working memory half-life parameter is updated using the least squares method. The value of is determined; simultaneously, the initial cognitive capacity threshold is updated based on the maximum number of valid answers given by the students in this diagnosis. .

6. The method according to claim 1, characterized in that, After marking the high-risk fatigue level, the diagnostic practice pairs marked as high-risk fatigue levels undergo modal lightweighting adjustment before presentation. This modal lightweighting adjustment includes: replacing high-cognitive-occupancy modal elements in control and probe questions with equivalent low-occupancy modal elements, including replacing animated demonstrations with step-by-step diagrams and complex interactive operations with selection and completion operations; converting cross-modal parallel information into a sequentially presented form, allowing students to process only a single modal of information at any given time; after this modal lightweighting adjustment, the cognitive load requirement of the diagnostic practice pairs is reduced to within a safe range of available capacity, while the knowledge assessment points and diagnostic effectiveness remain unchanged.

7. The method according to claim 1, characterized in that, The method of correcting the response time difference and attribution confidence of each level using the fatigue interference coefficient includes: For each level Calculate the fatigue interference coefficient ,in hierarchical The effective residual cognitive occupancy before the control question is presented. As the initial cognitive threshold for students, The value ranges from 0 to 1; Calculate the difference in response time after correction ,in The response time of the probe question, To control the response time of the questions, The baseline for the expected additional time consumption of the detection problem relative to the control problem under fatigue-free conditions; For levels marked as high-risk fatigue levels, multiply the initial attribution confidence of their incorrect responses by a fatigue discount factor. The attribution confidence score after weight reduction is obtained, among which , This is a discount adjustment constant with a value ranging from 0.3 to 0.

7. The discount factor is the lower limit and ranges from 0.2 to 0.4; for levels not marked as high-risk fatigue levels, the attribution confidence remains unchanged. The initial attribution confidence is pre-configured based on the combination of correct and incorrect labels for control and probe questions: a higher preset value is used when the control question is correct but the probe question is incorrect, a lower preset value is used when the control question itself is incorrect, and the highest preset value is used when both questions are correct.

8. The method according to claim 7, characterized in that, The bottom-up, layer-by-layer attribution determination includes: If a certain level's control question is correct but the probe question is incorrect, and the attribution confidence after weight reduction is not lower than the reliable judgment threshold, then the interdisciplinary connection between this level and the upper level is determined to be a bottleneck, and the bottleneck type is marked as interdisciplinary connection defect type. If a certain level of control question is incorrect, and the attribution confidence after weight reduction of that level is not lower than the reliable judgment threshold, then the knowledge point ontology of that level is determined to be a bottleneck, the bottleneck type is marked as knowledge ontology defect type, and the attribution judgment to higher levels is terminated. If both the control and detection questions at a certain level are correct, and the difference in response time after correction does not exceed the preset abnormal time consumption threshold, then it is determined that there is no bottleneck at that level, and the attribution judgment is continued for the next level. The reliability threshold is a preset constant ranging from 0.5 to 0.

8. Levels with a reduced attribution confidence level below the reliability threshold are marked as levels to be verified, and the required shortest recovery interval is calculated using the following formula. : ； in, For working memory half-life parameters, This represents the effective residual cognitive occupancy when responding at the level to be verified. As the initial cognitive threshold for students, To preset the safety ratio coefficient, To verify the cognitive load requirement of the diagnostic practice pair, a cognitive recovery interval no shorter than the minimum recovery interval is inserted at a subsequent position in the diagnostic practice pair sequence. After recovery, a simplified version of the diagnostic practice pair is added to the level to be verified for secondary confirmation. The attribution results of the secondary confirmation cover the initial attribution conclusion.

9. The method according to claim 1, characterized in that, The process of generating targeted multimodal practice questions and adjusting their difficulty levels based on bottleneck levels and types includes: For levels where the bottleneck type is knowledge ontology deficiency, single-point specialized practice templates corresponding to the knowledge points of that level are matched from the question template library. Single-point specialized practice questions are generated using the typical modalities of the subject at that level, and the difficulty level is adjusted to the level corresponding to the current mastery value of that level. For levels where the bottleneck type is interdisciplinary connection deficiency, bridging practice templates involving knowledge points of adjacent two levels are matched from the question template library to generate bridging multimodal practice questions. The bridging multimodal practice questions connect the knowledge points of the two levels through progressive contextual transitions: the first stage only involves independent application scenarios of the lower-level knowledge points, and the second stage introduces application elements of the upper-level knowledge points based on the results of the first stage, and explicitly presents the corresponding mapping relationship between the two subject concepts in the transitional context in a multimodal manner with text annotations and graphic arrows. The difficulty levels are divided into five levels, and the correspondence between each difficulty level and the mastery value is as follows: Mastery values ​​in the range of 0 to 0.2 correspond to difficulty level 1, where questions require the identification and restatement of basic definitions of knowledge points; mastery values ​​in the range of 0.2 to 0.4 correspond to difficulty level 2, where questions require the direct application of formulas or rules in a single-step context; mastery values ​​in the range of 0.4 to 0.6 correspond to difficulty level 3, where questions involve multi-step derivation and parameters must be extracted from the context; mastery values ​​in the range of 0.6 to 0.8 correspond to difficulty level 4, where questions require the flexible application of knowledge points in new contexts and cross-step reasoning; and mastery values ​​in the range of 0.8 to 1.0 correspond to difficulty level 5, where questions involve the comprehensive application of multiple knowledge points and open-ended reasoning requirements. The parameters for difficulty adjustment include four dimensions: numerical complexity, number of reasoning steps, contextual abstraction, and modal combination complexity. When adjusting to a lower difficulty level, all four dimensions decrease simultaneously. When adjusting to a higher difficulty level, the four dimensions gradually increase sequentially from numerical complexity, number of reasoning steps, contextual abstraction, to modal combination complexity. The diagnostic practice pairs are arranged in the following order: within the same dependency chain, they are arranged in bottom-up order; when there are multiple independent dependency chains, the shallower chains are completed first, followed by the deeper chains; when multiple dependency chains share a bottom-level node, the diagnostic practice pair for the shared node is generated only once, and its attribution conclusion applies to all dependency chains that share that node. The student's initial cognitive capacity threshold Based on the student's average maximum number of valid answers in the first three practice sessions Compared with the average cognitive load per item The product is obtained by calculation, that is For new users who do not yet have historical practice data, The default values ​​are set according to age group: 3.0 for lower elementary grades, 4.0 for upper elementary grades, 5.0 for junior high school, and 6.0 for senior high school. After generating the targeted multimodal practice questions, the practice questions are assembled and output according to the presentation rhythm control parameters, which include the display duration of each modal element, the modality switching interval, and the interaction waiting time; the display duration is based on the information density of the modal elements and the current available cognitive capacity. Certainly, the higher the information density or The lower the value, the longer the display duration; the modal switching interval is determined based on the cognitive conversion requirements between adjacent modal elements.

10. A multimodal classroom exercise automatic generation and difficulty adaptive adjustment system based on cognitive load management, used to execute the method described in any one of claims 1 to 10, characterized in that, include: The dependency chain topology extraction module is used to filter cross-disciplinary dependency chains with a depth of more than two levels and spanning two or more subject identifiers from a multi-disciplinary knowledge network. It extracts the knowledge point identifier, the subject identifier, and the current mastery value of each level node and generates a multi-level dependency chain topology description of the target knowledge point. The diagnostic practice pair generation module is used to generate diagnostic practice pairs consisting of control questions and probe questions layer by layer from bottom to top according to the multi-level dependency chain topology description, and output a diagnostic practice pair sequence; The module for collecting answer data and calculating cognitive load is used to present the diagnostic practice pair sequence to students in sequence and collect the correct and incorrect labels, response time and operation interval time. Based on the working memory half-life parameter, the cognitive load value of the completed questions is exponentially weighted and accumulated to calculate the effective residual cognitive occupancy at the current moment. The cognitive capacity assessment and risk labeling module is used to obtain the current available cognitive capacity by subtracting the effective residual cognitive occupancy from the student's initial cognitive capacity threshold, compare the cognitive load requirement value of the subsequent diagnostic practice pair with the product of the current available cognitive capacity and the preset safety ratio coefficient, label the fatigue high-risk level, and perform modal lightweight adjustment on the diagnostic practice pair labeled as fatigue high-risk level. The fatigue compensation attribution module is used to calculate the fatigue interference coefficient based on the ratio of the effective residual cognitive occupancy of each level to the initial cognitive carrying threshold. The fatigue interference coefficient is used to correct the response time difference and reduce the attribution confidence of the high-risk fatigue level. Layer-by-layer attribution judgment is performed to determine the bottleneck level and bottleneck type. The exercise generation and difficulty adjustment module is used to match the corresponding template from the question template library according to the bottleneck level and bottleneck type to generate targeted multimodal exercise questions, adjust the difficulty level to the level corresponding to the current mastery value of the bottleneck level, and assemble and output according to the presentation rhythm control parameters.

Images