Intelligent tutoring system, learning path optimization method, device, and medium

Abstract

The application provides an intelligent teaching assistant system, a learning path optimization method, equipment and a medium. Multi-modal sensors are used to collect multi-channel learning behavior data of students when they interact with an intelligent teaching assistant terminal for personalized learning. The sequence dependency relationship between learning content units is obtained based on a preset teaching knowledge graph. For each teaching stage, learning behavior features are extracted from the multi-channel learning behavior data, a structured feature vector is constructed based on the inherent attributes of the learning content units, and a joint representation vector is generated through multi-modal fusion. The joint representation vectors are aggregated to generate a teaching stage representation vector. The teaching effectiveness is evaluated based on the teaching stage representation vector to obtain a learning state deviation. The learning state deviation is used to filter teaching stages with substandard teaching effectiveness, and the learning path is optimized and adjusted based on the sequence dependency relationship. The application can realize closed-loop dynamic optimization of the intelligent learning path.

Description

Technical Field

[0001] This application relates to the field of intelligent education technology, and more specifically, to an intelligent teaching assistant system, a learning path optimization method, a device, and a medium. Background Technology

[0002] In the field of intelligent education technology, personalized learning path optimization is a core element in achieving individualized instruction and improving teaching efficiency and quality. Traditional learning path planning often relies on students' static profile information, such as entrance test scores, learning style questionnaires, or single-dimensional learning history data, such as answer accuracy and time spent. These methods generate an initial learning sequence for students through predefined rules or simple decision tree models. However, this static path planning lacks the ability to perceive and provide feedback on students' real-time learning status during the teaching process. It cannot adaptively adjust according to students' dynamically changing knowledge mastery, engagement, and cognitive load, resulting in insufficient accuracy and timeliness of the learning path.

[0003] With the development of educational technology, some existing solutions attempt to incorporate more student behavioral data to optimize learning paths. For example, analyzing students' online learning logs, such as video viewing completion rates and assignment submission times, can infer their learning interest and difficulties. However, these methods have significant limitations: First, they rely on a single data dimension, mostly superficial behavioral indicators such as clickstreams, making it difficult to capture the deeper cognitive states and emotional investment students experience during learning, such as confusion, focus, or fatigue. Second, existing methods typically treat learning behaviors and knowledge content in isolation, failing to deeply integrate behavioral data with subject-specific logic. This can easily lead to knowledge gaps or skipped learning, ultimately harming the overall teaching effectiveness. Therefore, how to achieve intelligent learning path optimization and thus improve the overall teaching effectiveness of personalized tutoring has become a challenge for the industry. Summary of the Invention

[0004] This application provides an intelligent learning path optimization method, an intelligent teaching assistant system, equipment, and medium, which can realize closed-loop dynamic optimization of intelligent learning paths, thereby improving the overall teaching effect of personalized teaching and tutoring.

[0005] Firstly, this application provides an intelligent learning path optimization method, which includes the following steps:

[0006] Multi-modal sensors are used to collect multi-channel learning behavior data of students when they interact with intelligent teaching assistant terminals for personalized learning. Based on a preset teaching knowledge graph, the sequence dependency relationship between learning content units corresponding to multiple teaching stages in the personalized learning path is obtained, and the sequence dependency relationship reflects the logic of subject knowledge.

[0007] For each teaching stage, the learning behavior features of students on different interaction channels are extracted from the multi-channel learning behavior data. A structured feature vector of each learning content unit is constructed based on the inherent attributes of the learning content unit. At the same time, a student behavior feature vector of each learning content unit is constructed based on the learning behavior features.

[0008] The structured feature vector and the student behavior feature vector are fused in a multimodal manner to generate a joint representation vector for each learning content unit. For each teaching stage, the joint representation vectors of all learning content units within that teaching stage are aggregated to generate a teaching stage representation vector.

[0009] The teaching effectiveness of the teaching guidance is evaluated based on the teaching stage representation vector to obtain the teaching effectiveness evaluation value of each teaching stage. The teaching effectiveness evaluation value is then compared with a predetermined threshold to obtain the learning state deviation degree of each teaching stage.

[0010] The deviation of each learning state is used to filter out the teaching stages in the teaching and tutoring process where the teaching effect is not up to standard. Based on the sequence dependency relationship and the teaching stages where the teaching effect is not up to standard, the learning path is optimized and adjusted.

[0011] Preferably, the multi-channel learning behavior data collected by multimodal sensors during personalized learning interactions between students and intelligent teaching assistant terminals specifically includes:

[0012] Use visual sensors to capture students' facial expression data and eye movement data;

[0013] Use audio sensors to record students' voice input data and intonation variation data;

[0014] Use interactive sensors to record students' touch input and keyboard input data;

[0015] The facial expression data, eye movement data, voice input data, tone change data, touch input data, and keyboard input data are integrated into the multi-channel learning behavior data.

[0016] Preferably, based on a pre-defined teaching knowledge graph, obtaining the sequence dependencies between learning content units corresponding to multiple teaching stages in a personalized learning path specifically includes:

[0017] Extract the prerequisite and prerequisite knowledge points for each learning content unit from the teaching knowledge graph;

[0018] Establish the sequential association between each learning content unit and other learning content units based on the aforementioned prerequisite and subsequent knowledge points;

[0019] The sequential association is taken as the sequence dependency.

[0020] Preferably, for each teaching stage, extracting student learning behavior characteristics across different interaction channels from the multi-channel learning behavior data specifically includes:

[0021] Separate visual channel data, audio channel data, and interactive channel data from the multi-channel learning behavior data;

[0022] Feature extraction is performed on the visual channel data to obtain the student's attention concentration index;

[0023] Feature extraction is performed on the audio channel data to obtain the students' comprehension and expression accuracy index;

[0024] Feature extraction is performed on the interactive channel data to obtain the student's response speed index;

[0025] Collect students' learning outcome data for each learning content unit, and convert the learning outcome data into knowledge mastery indicators;

[0026] The attention concentration index, comprehension and expression accuracy index, response speed index, and knowledge mastery index are combined to form the learning behavior characteristics.

[0027] Preferably, constructing a structured feature vector for each learning content unit based on its inherent attributes specifically includes:

[0028] Select the core inherent attributes of the learning content units;

[0029] The selected core inherent attributes are converted into numerical values.

[0030] The numerical attributes of each learning content unit are arranged in a fixed order to construct a structured feature vector.

[0031] Preferably, the multimodal fusion of the structured feature vector and the student behavior feature vector to generate a joint representation vector for each learning content unit specifically includes:

[0032] The structured feature vector and the student behavior feature vector are input into a preset feature fusion unit;

[0033] In the feature fusion unit, a vector concatenation operation is performed on the structured feature vector and the student behavior feature vector;

[0034] The joint representation vector is generated from the result of the vector concatenation operation.

[0035] Preferably, for each teaching stage, the joint representation vector of all learning content units within that teaching stage is aggregated to generate the teaching stage representation vector, specifically including:

[0036] Collect the joint representation vectors of all learning content units within this teaching phase;

[0037] The importance weight of each learning content unit is determined based on the sequence dependency relationship;

[0038] Perform a weighted average pooling operation based on the importance weights on the joint representation vector;

[0039] The teaching stage representation vector is generated from the result of the weighted average pooling operation.

[0040] Secondly, this application provides an intelligent teaching assistant system, including an intelligent learning path optimization unit, wherein the intelligent learning path optimization unit specifically includes:

[0041] The data acquisition module is used to collect multi-channel learning behavior data of students when they interact with the intelligent teaching assistant terminal through multi-modal sensors. Based on the preset teaching knowledge graph, it obtains the sequence dependency relationship between the learning content units corresponding to multiple teaching stages in the personalized learning path, wherein the sequence dependency relationship reflects the subject knowledge logic.

[0042] The processing module is used to extract the learning behavior features of students on different interaction channels from the multi-channel learning behavior data for each teaching stage, construct a structured feature vector for each learning content unit based on the inherent attributes of the learning content unit, and construct a student behavior feature vector for each learning content unit based on the learning behavior features.

[0043] The processing module is further configured to perform multimodal fusion of the structured feature vector and the student behavior feature vector to generate a joint representation vector for each learning content unit. For each teaching stage, the module aggregates the joint representation vectors of all learning content units within that teaching stage to generate a teaching stage representation vector.

[0044] The processing module is also used to evaluate the teaching effect of the teaching guidance based on the teaching stage representation vector, obtain the teaching effect evaluation value of each teaching stage, and compare the teaching effect evaluation value with a predetermined threshold to obtain the learning state deviation degree of each teaching stage.

[0045] The execution module is used to filter out the teaching stages that do not meet the teaching effectiveness standards during the teaching and tutoring process using the deviation degree of each learning state, and optimize and adjust the learning path based on the sequence dependency relationship and the teaching stages that do not meet the teaching effectiveness standards.

[0046] Thirdly, this application provides a computer device, the computer device including a memory and a processor, the memory storing code, and the processor being configured to acquire the code and execute the above-described intelligent learning path optimization method.

[0047] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described intelligent learning path optimization method.

[0048] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects:

[0049] This application achieves precise diagnosis and dynamic optimization of learning paths by multimodal fusion of the structured features of learning content units and real-time student behavioral features, followed by phased aggregation and evaluation. Combined with the logical sequence dependencies of subject knowledge reflected in the teaching knowledge graph, this enables precise diagnosis and dynamic optimization. First, multimodal sensors collect multi-channel learning behavior data, and the sequential dependencies between learning content units are extracted based on the teaching knowledge graph. This establishes a two-way mapping mechanism between learning content units and students' actual interactive behaviors, while fully respecting the inherent logic of subject knowledge. This provides a reliable data foundation with knowledge structure constraints for subsequent feature fusion and effect evaluation. Second, by constructing structured feature vectors for learning content units and student behavioral feature vectors separately and performing multimodal fusion to generate joint representation vectors, the joint representation vectors of all learning content units within the same teaching stage are aggregated to form the teaching stage. The representation vector enables multi-granular semantic expression from the micro-content unit level to the macro-teaching stage level, effectively capturing the fine-grained behavioral differences of students across different interaction channels and the deep connections between content attributes, thus improving the completeness and discriminative power of the representation. Then, based on the representation vector of the teaching stage, the teaching effect is quantitatively evaluated and the learning state deviation is calculated. This objectively identifies students' mastery deviations and weaknesses in specific teaching stages, avoiding the assessment lag and one-sidedness caused by traditional methods relying solely on final grades or single indicators. Finally, the learning state deviation is used to accurately screen teaching stages with substandard teaching effects, and targeted optimization and adjustments are made in conjunction with the sequence dependencies of the knowledge graph. This allows the learning path to maintain the logical coherence of subject knowledge while dynamically compensating for and reinforcing students' individual weaknesses, significantly improving the adaptability and teaching efficiency of personalized learning. In summary, this application's solution can achieve closed-loop dynamic optimization of intelligent learning paths based on multimodal behavioral perception and knowledge logic constraints, thereby improving the overall teaching effect of personalized tutoring. Attached Figure Description

[0050] Figure 1 This is an exemplary flowchart of an intelligent learning path optimization method according to some embodiments of this application;

[0051] Figure 2 This is a flowchart illustrating the extraction of learning behavior features according to some embodiments of this application;

[0052] Figure 3 This is a schematic diagram of the process for generating teaching stage representation vectors according to some embodiments of this application;

[0053] Figure 4 This is a schematic diagram of the structure of an intelligent learning path optimization unit according to some embodiments of this application;

[0054] Figure 5 This is a schematic diagram of the structure of a computer device implementing an intelligent learning path optimization method according to some embodiments of this application. Detailed Implementation

[0055] To better understand the technical solution of this application, the technical solution of this application will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0056] refer to Figure 1 The figure is an exemplary flowchart of an intelligent learning path optimization method according to some embodiments of this application. The intelligent learning path optimization method mainly includes the following steps:

[0057] In step S101, multi-channel learning behavior data of students during personalized learning interactions with the intelligent teaching assistant terminal are collected by a multi-modal sensor. Based on a preset teaching knowledge graph, the sequence dependency relationships between learning content units corresponding to multiple teaching stages in the personalized learning path are obtained, wherein the sequence dependency relationships reflect the subject knowledge logic.

[0058] In some embodiments, the collection of multi-channel learning behavior data of students during personalized learning interactions with intelligent teaching assistant terminals using multi-modal sensors can be achieved in the following ways: for example, using a visual sensor to capture students' facial expression data and eye movement trajectory data; using an audio sensor to record students' voice input data and tone change data; using an interaction sensor to record students' touch input data and keyboard input data; integrating the facial expression data, eye movement trajectory data, voice input data, tone change data, touch input data, and keyboard input data into the multi-channel learning behavior data, while simultaneously recording auxiliary data such as temperature and environmental noise.

[0059] It should be noted that the teaching knowledge graph in this application is a structured graph constructed based on the subject knowledge logic. This teaching knowledge graph can be implemented using knowledge graph construction tools or databases. For example, knowledge graph construction tools can be Neo4j, Protege, or KgBuilder, which can directly map the subject knowledge logic. The database can be a relational database (such as MySQL), using tables to store knowledge point attributes and relationships. Knowledge point attributes include, for example, name, type, and difficulty. Relationships include, for example, knowledge point A is prerequisite knowledge for knowledge point B. The subject logic is reflected through field associations. This is only an example and is not intended to limit the scope of this application.

[0060] In some embodiments, obtaining the sequence dependency relationships between learning content units corresponding to multiple teaching stages in a personalized learning path based on a preset teaching knowledge graph can be achieved in the following ways:

[0061] First, extract the prerequisite knowledge points and subsequent knowledge points for each learning content unit from the teaching knowledge graph. The prerequisite knowledge points are the pre-existing knowledge that must be mastered to learn the current content unit, and the subsequent knowledge points are the follow-up knowledge that can be connected after the current content unit is learned.

[0062] Secondly, based on the aforementioned prerequisite and prerequisite knowledge points, establish the sequential relationship between each learning content unit and other learning content units. Establishing the sequential relationship involves transforming the extracted prerequisite relationships into a clear learning order. For example, if unit A is a prerequisite knowledge point of unit B, then the learning order of A is determined to be before B. This will not be elaborated further here.

[0063] Finally, the sequential association is taken as the sequence dependency relationship. That is, based on the logic of subject knowledge, the sequential learning association of learning content units is clarified from the teaching knowledge graph, and a sequence dependency relationship is finally formed.

[0064] In step S102, for each teaching stage, the learning behavior features of students on different interaction channels are extracted from the multi-channel learning behavior data. A structured feature vector of each learning content unit is constructed based on the inherent attributes of the learning content unit. At the same time, a student behavior feature vector of each learning content unit is constructed based on the learning behavior features.

[0065] In some embodiments, reference Figure 2 The figure is a flowchart illustrating the extraction of learning behavior features according to some embodiments of this application. The extraction of student learning behavior features on different interaction channels from the multi-channel learning behavior data can be achieved in the following manner:

[0066] In step S1021, visual channel data, audio channel data, and interactive channel data are separated from the multi-channel learning behavior data, wherein the visual channel data includes facial expression data and eye movement trajectory data, etc.

[0067] In step S1022, feature extraction is performed on the visual channel data to obtain the student's attention concentration index. This can be achieved in the following way:

[0068] First, the visual channel data is preprocessed;

[0069] First, preprocess the acquired facial expression data (e.g., image sequences) and eye-tracking data (e.g., screen coordinate sequences):

[0070] For facial expression data, it is necessary to eliminate the influence of uneven ambient light, locate the eye and facial muscle areas (such as the corners of the mouth and eyebrows) using a face detection model (such as MTCNN), and remove blurry frames caused by large head movements.

[0071] For eye-tracking data: invalid coordinates during blinking need to be removed. This can be achieved through coordinate calibration, which involves binding eye-tracking coordinates to the learning content display area of ​​the smart teaching assistant terminal to filter out abnormal trajectory points that exceed the learning interface range.

[0072] Secondly, attention-related features are extracted from different dimensions;

[0073] From the two types of preprocessed data, extract the core features directly related to attention, namely:

[0074] Facial expression characteristics: Statistically measure the duration of focused facial expressions, such as stable eye contact, relaxed facial muscles, and no frequent frowning / yawning / head-down movements, per unit of time; if distracting expressions are detected, such as frequent face touching or wandering eyes, the duration of a single instance exceeds 10 seconds, it is marked as an attention interruption event, and the number of interruptions per unit of time is counted.

[0075] Eye movement trajectory characteristics: Calculate the effective fixation rate, which is the duration of eye fixation in the learning content area / total interaction time; fixation stability, which is the standard deviation of the continuous fixation duration in the same learning content area. The smaller the standard deviation, the more stable the fixation; and ineffective saccade rate, which is the number of rapid saccades in non-learning areas per minute, such as frequently looking out of the screen.

[0076] Feature quantization and weight allocation:

[0077] The above features are converted into quantitative scores of 0-100, and weights are assigned according to the degree of influence of each feature on attention: for example, the weight of effective gaze percentage can be 40%; if the percentage is ≥80%, 100 points are awarded, and the score decreases by 20 points for every 10% decrease.

[0078] The weight of the time spent focusing on facial expressions can be 30%: if the proportion is ≥75%, you get 100 points, and the score decreases by 25 points for every 10% decrease;

[0079] The number of attention interruption events can be weighted at 15%: 0 interruptions earn 100 points, and each additional interruption reduces the score by 30 points;

[0080] The weight of invalid video scanning rate can be 15%: ≤2 times / minute gets 100 points, and decreases by 25 points for every additional 2 times.

[0081] Attention Concentration Index Calculation and Correction

[0082] The initial score can be calculated using a weighted summation formula: Initial score = (effective gaze percentage score × 40%) + (focused expression duration percentage score × 30%) + (attention interruption event score × 15%) + (ineffective video scanning rate score × 15%). This can be further adjusted based on the learning scenario: for example, if the current learning content is a high-difficulty unit, the initial score can be increased by 5%-10% (as brief eye adjustments are normal under high-difficulty content); if it is a low-difficulty unit (difficulty level ≤ 2), the initial score can be decreased by 3%-5% (distractions in low-difficulty content better reflect attention problems). The final adjusted value is the student's attention concentration index. The above is merely an example and does not constitute a specific limitation on this application.

[0083] In step S1023, feature extraction is performed on the audio channel data to obtain the student's comprehension and expression accuracy index; feature extraction is performed on the interactive channel data to obtain the student's response speed index.

[0084] In specific implementation, feature extraction of the audio channel data to obtain the student's comprehension and expression accuracy index can be performed in the following way:

[0085] First, the audio data is preprocessed. The collected student voice input data and intonation change data are preprocessed to remove environmental noise and invalid voice segments, and the voice signal is converted into an analyzable digital signal.

[0086] Then, extract the core features, that is, extract the content matching degree from the voice input data: after converting the student's voice into text, compare it with the standard answer or key information of the knowledge point (such as core concepts and formula expressions), and count the proportion of matching keywords and logical expressions.

[0087] Extracting the rationality of expression from intonation variation data involves analyzing effective speech rate (effective speech rate = number of matching keywords / expression duration; ≥0.8 keywords / second and content matching degree ≥80% are considered fluent) and tone stability. Analyzing the effective speech rate, for example, if it is too slow and the content matching degree is <60%, it may indicate poor comprehension; if it is too fast and the content matching degree is <60%, it may indicate rote memorization. Analyzing tone stability involves, for example, extracting the tone variance of key information segments through audio analysis (after normalization, ≤0.1 indicates reasonable emphasis, >0.3 indicates irregular fluctuations), rather than subjectively judging whether there is reasonable emphasis.

[0088] Finally, feature quantification and index calculation are performed: content matching degree and expression rationality are quantified into scores of 0-100, and then weighted according to their influence on comprehension and expression, for example, content matching degree accounts for 60% and expression rationality accounts for 40%, and the final result is the student's comprehension and expression accuracy index.

[0089] In addition, in specific implementation, feature extraction of the interaction channel data to obtain the student's response speed index can be performed in the following way:

[0090] First, the interactive data is preprocessed: the touch input and keyboard input data collected by the interactive sensors are filtered out, and invalid data, such as accidental touches, repeated inputs, and meaningless rapid clicks, are removed. The valid input records corresponding to the interactive commands of the smart teaching assistant terminal are retained, and the command issuance time and student input completion time of each record are associated.

[0091] Secondly, extract core response features: For each valid interaction record, calculate the response time interval, which is the difference between the time when the intelligent teaching assistant terminal issues the interaction command and the time when the student completes the input and submits it; calculate the average response time, the standard deviation of the response time (which can reflect the stability of the response speed), and the number of timeout responses (i.e., the number of times the input is not completed beyond the preset reasonable response time).

[0092] Finally, the indicators are quantified and integrated: the average response time, response speed stability, and number of timeout responses are compared with preset reasonable thresholds and quantified, and then calculated by weighting. For example, the average response time accounts for 50% of the weight, the response speed stability accounts for 30% of the weight, and the number of timeout responses accounts for 20% of the weight. The final weighted calculation result is the response speed indicator.

[0093] In step S1024, the learning outcome data of students in each learning content unit is collected, and the learning outcome data is converted into a knowledge mastery index; the attention concentration index, comprehension and expression accuracy index, response speed index and knowledge mastery index are combined into the learning behavior characteristics.

[0094] In practice, student learning outcome data for each learning content unit is collected, and this learning outcome data is converted into a knowledge mastery index. This can be achieved in the following ways:

[0095] First, collect learning outcome data: for each learning content unit, collect multiple types of student outcome data, including homework for that unit, such as answers to exercises, in-class quizzes, answers to objective / subjective questions corresponding to knowledge points, knowledge point application tasks, such as simple practical operations, case analysis reports, and real-time answer data from interactions with the intelligent teaching assistant terminal, to ensure coverage of both theoretical answers and simple applications.

[0096] Then, the collected learning outcome data is transformed into a knowledge mastery index. For example, the data is preprocessed to remove invalid data, such as cheating on assignments / tests. The scoring standards for different outcome types are standardized, such as converting percentage and five-point systems to 0-100 points. Then, weights are allocated according to outcome type, such as 40% for in-class quizzes, 30% for homework, 20% for instant quizzes, and 10% for application tasks. The comprehensive score for the unit is calculated. Combined with the analysis of incorrect answers to knowledge points in the unit, if the comprehensive score is ≥80 points and the error rate of core knowledge points is ≤5%, it can be marked as high mastery (90-100 points); if the score is 60-79 points and the error rate of core knowledge points is 5%-15%, it is medium mastery (60-89 points); if the score is <60 points or the error rate of core knowledge points is >15%, it is low mastery (0-59 points). The final comprehensive score is the knowledge mastery index for the unit. The above is only a simple example and is not intended to limit this application.

[0097] Furthermore, in specific implementation, the combination of the attention concentration index, comprehension and expression accuracy index, response speed index, and knowledge mastery index into the learning behavior characteristics can be achieved in the following ways:

[0098] First, standardize the quantitative criteria for indicators: normalize the four indicators of attention concentration, comprehension and expression accuracy, response speed and knowledge mastery to a unified score range of 0-100, eliminate the differences in the original quantitative dimensions of different indicators, and ensure that they can be directly used in combined calculations;

[0099] Then, weights are assigned according to the importance of the indicators: based on the degree of influence of learning behavior on teaching effectiveness, corresponding weights are assigned to the four indicators. For example, knowledge mastery directly reflects the learning outcome, so the weight is set at 40%; attention concentration and comprehension and expression accuracy are core process indicators, each set at 20%; and response speed is an auxiliary process indicator, set at 10%. It should be noted that the weight allocation should conform to the logic of result-oriented + process-supported. This is only an example and is not intended to limit the specific application.

[0100] Finally, the learning behavior features are generated by weighted integration: the normalized scores of the four indicators are multiplied by their corresponding weights and then summed to obtain the comprehensive learning behavior feature scores of students in a single teaching stage; or the data dimensions of each indicator after weighting are retained to form a multi-dimensional learning behavior feature vector. The final result can be directly used for subsequent fusion operations with structured feature vectors.

[0101] In some embodiments, the structured feature vector of each learning content unit can be constructed based on the inherent properties of the learning content unit in the following ways:

[0102] First, screen the core inherent attributes of the learning content unit: First, determine the core attributes inherent to the learning content unit that do not change with the student's learning behavior, the knowledge type, the density of knowledge point associations, and the content delivery format. Core attributes include, for example, knowledge difficulty, such as basic, intermediate, and extended levels; knowledge type, such as concept, formula derivation, and case application; the density of knowledge point associations, such as the number of core knowledge points that overlap between this learning content unit and other units; and the content delivery format, such as text, video, and practical exercises.

[0103] Secondly, the selected core inherent attributes are numerically converted: non-quantifiable attributes are converted into calculable numerical forms. For example, knowledge difficulty is assigned as Basic Level = 1, Advanced Level = 2, and Extended Level = 3; knowledge type is encoded using one-hot encoding, such as Concepts = [1,0,0], Formulas = [0,1,0], and Applications = [0,0,1]; association density is directly rounded down by the number of cross-knowledge points, such as 3 cross-knowledge points = 3; and content delivery format is assigned as Text = 1, Video = 2, and Practical = 3.

[0104] Finally, the numerical attributes of each learning content unit are arranged in a fixed order to construct a structured feature vector. This structured feature vector is a one-dimensional vector. The fixed order is, for example, difficulty, type encoding, association density, and content carrying format. The following example uses a basic concept unit, which includes two intersecting knowledge points in text form. The six positions of the structured feature vector correspond to four inherent attributes (where "knowledge type" requires multi-dimensional encoding), and the order is as follows:

[0105] The knowledge difficulty level is ranked 1st, the knowledge type (one-hot encoding, assuming 3 categories: concept, formula, and application) is ranked 2nd-4th, the number of overlapping knowledge points is ranked 5th, and the content format is ranked 6th. Then, values ​​are assigned to each attribute according to rules. For a basic-level concept-based unit with 2 overlapping knowledge points and text format:

[0106] Knowledge difficulty levels: By convention, basic level = 1, intermediate level = 2, and advanced level = 3, therefore the first element of the vector should be 1;

[0107] Knowledge type (one-hot encoding): The concept class is defined as [1,0,0], the formula class as [0,1,0], and the application class as [0,0,1]. Therefore, the second bit of the vector is filled with 1, the third bit is filled with 0, and the fourth bit is filled with 0.

[0108] Number of overlapping knowledge points: Take the actual number 2 directly, and fill in 2 in the 5th position of the vector;

[0109] Content delivery format: text = 1, video = 2, practical operation = 3. Therefore, the 6th bit of the vector is filled with 1. Finally, its structured feature vector can be represented as [1,1,0,0,2,1], which intuitively reflects the inherent attributes of the learning content unit. This is only an example and is not intended to limit the specific application.

[0110] In step S103, the structured feature vector and the student behavior feature vector are fused in a multimodal manner to generate a joint representation vector for each learning content unit. For each teaching stage, the joint representation vectors of all learning content units within that teaching stage are aggregated to generate a teaching stage representation vector.

[0111] In some embodiments, the structured feature vector and the student behavior feature vector are fused in a multimodal manner to generate a joint representation vector for each learning content unit. Specifically, this can be achieved in the following way:

[0112] The structured feature vector and the student behavior feature vector are input into a preset feature fusion unit. Specifically, before input, the structured feature vector and the student behavior feature vector are standardized to the 0-1 range respectively.

[0113] In the feature fusion unit, a vector concatenation operation is performed on the structured feature vector and the student behavior feature vector. Before concatenation, the structured feature vector can be assigned a weight of 0.3 and the student behavior feature vector a weight of 0.7. After weighting, the vectors are concatenated. At the same time, the correlation coefficient between the two types of features is calculated. When the correlation coefficient is ≥0.6, the weights are adjusted to 0.4:0.6. In specific implementation, the two vectors can be directly connected in a fixed order. For example, the structured vector is [1,1,0,0,2,1] of length 6 and the student behavior feature vector is [85,70,90,80] of length 4. After concatenation, a combined vector of length 6+4=10 is formed [1,1,0,0,2,1,85,70,90,80]. This is only an example and is not intended to limit the specific scope of this application.

[0114] The joint representation vector is generated from the result of the vector concatenation operation. That is, after the concatenated combined vector is standardized, for example, the numerical range of each dimension is unified, the joint representation vector is obtained. The joint representation vector contains both the inherent attributes of the learning content and the behavioral characteristics of the student. It can be directly used as the input for subsequent learning path optimization, personalized recommendation and other models, which will not be elaborated here.

[0115] It should be noted that in the above embodiments, the structured feature vector (except for the one-hot encoding of knowledge types) and the student behavior feature vector can be subjected to Min-Max normalization processing respectively, so as to uniformly map the values ​​of the two types of vectors to the 0-1 interval, thereby eliminating the difference in units. This will not be elaborated here.

[0116] refer to Figure 3 The figure is a flowchart illustrating the generation of teaching stage representation vectors according to some embodiments of this application. For each teaching stage, the joint representation vectors of all learning content units within that teaching stage are aggregated. The generation of teaching stage representation vectors can be achieved through the following steps:

[0117] In step S1031, the joint representation vectors of all learning content units within this teaching phase are collected;

[0118] In step S1032, the importance weight of each learning content unit is determined based on the sequence dependency relationship. That is, the stronger the fundamental nature, scope of influence, and irreplaceability of the learning content unit in the sequence dependency relationship, the higher the weight. This can be achieved in the following ways:

[0119] First, analyze the position and influence of the learning content unit in the sequence dependency relationship. Specifically, count the number of subsequent units for each learning content unit (i.e. how many learning content units directly or indirectly depend on it). The more subsequent units there are, the more foundation the learning content unit is for subsequent content, and the higher the initial value of its importance weight. For example, if learning content unit A has 5 subsequent units and learning content unit B has 2, then the initial weight of A is higher than that of B.

[0120] Secondly, the strength of dependency is distinguished. Specifically, if learning content unit X is a strong prerequisite or a necessary prerequisite unit for learning content unit Y (i.e., Y cannot be learned without mastering X), then X's weight is added to the dependency relationship of Y; if it is only a weak prerequisite or an auxiliary prerequisite (i.e., to help understand Y), then the added weight is lower, thereby strengthening the weight of the core basic unit.

[0121] Then, in conjunction with the depth adjustment of the sequence dependency relationship, specifically, the root node unit at the very front of the dependency chain (without a preceding unit, which is the starting point of the entire knowledge system) is given the highest basic weight because it affects all subsequent content; as the dependency chain deepens (the later the learning content unit), the weight decreases proportionally, for example, by 10% for each level deeper, to avoid the subsequent learning content unit having too high a weight.

[0122] Finally, the normalization process is used to obtain the importance weight of each learning content unit. Specifically, the weight values ​​calculated above are uniformly mapped to the range of 0-1, and the importance weight of each learning content unit is obtained. The importance weight of each learning content unit reflects its coreness in the sequence dependency relationship and is also convenient for subsequent calculations.

[0123] In step S1033, a weighted average pooling operation based on the importance weights is performed on the joint representation vector; the teaching stage representation vector is generated from the result of the weighted average pooling operation.

[0124] Since a teaching phase contains multiple learning content units, each learning content unit corresponds to a joint representation vector. This joint representation vector integrates the inherent attributes of the unit with student behavioral characteristics. Furthermore, the importance weight of each learning content unit has been determined through sequence dependency relationships, and the importance weight reflects the core status of the unit within the learning phase.

[0125] In practice, firstly, weighted average pooling is performed. The joint representation vectors of all units within the teaching stage are weighted and averaged according to their respective importance weights. That is, the joint representation vector of each unit is first multiplied by its own importance weight, then all the product results are added together, and finally divided by the sum of the weights to ensure that the result is reasonably scaled. For example, if there is a learning content unit A with a weight of 0.6 and a joint representation vector of V1, and a learning content unit B with a weight of 0.4 and a joint representation vector of V2, then the weighted average result is (0.6×V1 + 0.4×V2) / (0.6+0.4).

[0126] Then, a teaching stage representation vector is generated, which is the result of the weighted average above. The teaching stage representation vector integrates the key features of all units in the learning stage and highlights the influence of important units through weights. It can be directly used to describe the comprehensive characteristics of the entire teaching stage (such as the distribution of knowledge difficulty, the overall performance of students, etc., which will not be elaborated here).

[0127] In step S104, the teaching effectiveness of the teaching guidance is evaluated based on the teaching stage representation vector to obtain the teaching effectiveness evaluation value of each teaching stage. The teaching effectiveness evaluation value is then compared with a predetermined threshold to obtain the learning state deviation of each teaching stage.

[0128] The core logic of obtaining teaching effectiveness evaluation values ​​based on teaching stage representation vectors is as follows: extract key information reflecting teaching effectiveness from the stage representation vectors that integrate content attributes, student behavior, and unit importance, and quantify it into a score. This can be achieved in the following way:

[0129] First, we analyze the core dimensions of the teaching stage representation vector. Since the teaching stage representation vector contains multiple dimensions, the core dimensions directly related to teaching effectiveness include: the average overall knowledge mastery within the stage, the student behavior achievement rate for key units, and the learning completion rate of core knowledge points. Specifically, the average overall knowledge mastery within the stage is calculated by weighted averaging of the knowledge mastery indicators and their importance weights for all learning content units within that teaching stage. The student behavior achievement rate for key units includes, for example, the average achievement of attention concentration and comprehension accuracy. The learning completion rate of core knowledge points is determined by quantifying students' understanding of the core knowledge points within the teaching stage. The degree of mastery of knowledge points, combined with the importance weight of the unit in which the knowledge point is located, is used to obtain the comprehensive result. Specifically, the core knowledge points and their respective learning content units are identified. Then, the completion rate of each knowledge point is quantified. That is, for each core knowledge point, the completion rate is calculated based on the student's performance in the corresponding learning content unit, such as the accuracy rate of related exercises and the quality of application task completion. Finally, the overall completion rate is calculated by weighting the importance weight of each unit. That is, the completion rate of each core knowledge point is multiplied by the importance weight of its respective learning content unit, and the result is obtained by summing them. The final result is the learning completion rate of the core knowledge points in this teaching stage.

[0130] Secondly, the core dimensions mentioned above are quantified and weighted. For example, the values ​​of the core dimensions are mapped to a score of 0-100. For example, the average score for overall knowledge mastery is 80 points and the score for key unit behavior achievement rate is 75 points. Then, the weighted sum is calculated according to the weight of each core dimension's impact on teaching effectiveness. For example, mastery accounts for 50%, behavior achievement rate accounts for 30%, and core knowledge point completion accounts for 20%.

[0131] Finally, a teaching effectiveness evaluation value is generated, which is the result of weighted summation as the teaching effectiveness evaluation value for this teaching stage. The teaching effectiveness evaluation value ranges from 0 to 100, with higher values ​​indicating better teaching effectiveness for this stage.

[0132] In some embodiments, comparing the teaching effectiveness evaluation value with a predetermined threshold to obtain the deviation of the learning state at each teaching stage can be done in the following ways:

[0133] First, set a predetermined threshold for each teaching stage, for example, a predetermined threshold of 70 points, which represents the minimum performance standard that should be achieved in that stage;

[0134] Then, the original deviation is obtained by subtracting the predetermined threshold from the teaching effectiveness evaluation value of this stage;

[0135] Divide the original deviation by the predetermined threshold and multiply by 100% to obtain the standardized learning state deviation. For example, when the teaching effect evaluation score is 63 points and the preset threshold is 70 points, the learning state deviation = (63-70) / 70×100%=-10%). A negative learning state deviation indicates that the expected results have not been achieved. The larger the absolute value of the negative value, the more serious the deviation. A positive value indicates that the effect is better than expected.

[0136] In step S105, the teaching stages where the teaching effect is not up to standard are screened out using the deviation degree of each learning state. Based on the sequence dependency relationship and the teaching stages where the teaching effect is not up to standard, the learning path is optimized and adjusted.

[0137] In practice, the deviation of each learning state is used to filter out the teaching stages in the teaching and tutoring process where the teaching effect is not up to standard. That is, the standard for judging non-compliance is set: slight deviation (-5%~-2%) and severe deviation (<-5%) are both judged as non-compliance stages, that is, the actual effect is more than 2% lower than expected. The learning state deviation of all teaching stages is traversed, and the stages that meet the standard are marked as non-compliance stages, which are the key targets for subsequent path optimization.

[0138] It should be noted that optimizing the learning path based on sequence dependencies (i.e., the sequential dependency chain of learning content units) and substandard teaching stages is essentially about locating the root cause and scope of the problem from the dependencies, and then adjusting the path accordingly to fix the defects.

[0139] In some embodiments, optimizing and adjusting the learning path based on the sequence dependencies and teaching stages where the teaching effect is unsatisfactory can be achieved in the following ways:

[0140] First, identify the dependencies of the unmet goals stage. Specifically, analyze the prerequisites and subsequent impacts of the unmet goals stage through sequential dependencies (e.g., stage A → stage B → stage C). For example, if stage B is unmet, first determine whether it is due to the failure to master the prerequisite stage A (B depends on A); at the same time, clarify that the failure of B may affect the learning of the subsequent stage C (C depends on B).

[0141] Then, adjust the learning path according to the scenario. Specifically, if the unmet stage is a core preparatory stage (e.g., if A is unmet, leading to a chain reaction of unmet needs in B and C), prioritize adjusting the path to consolidate A (only consolidate weak knowledge points, 15 minutes) → relearn B (60% of core knowledge points + consolidating 40% of weak knowledge points, 30 minutes) → then learn C. First, fix the basic deficiencies to avoid further obstacles in subsequent learning. If the unmet stage is an intermediate stage (e.g., B is unmet but A is met): insert a B reinforcement unit (targeted exercises, 15 minutes) into the learning path, such as targeted exercises and simplified explanations. After supplementing the learning, proceed to C to avoid skipping problems.

[0142] If the stage that fails to meet the standard is an isolated stage, that is, without any prerequisites, such as if failing to meet the standard in stage D does not affect other stages, then the teaching method for that stage should be replaced or the difficulty reduced to suit the characteristics of the students before continuing.

[0143] The above adjustments allow the learning path to align with the logic of knowledge dependence, prioritizing the resolution of deficiencies at key points and ensuring that subsequent learning is built on a solid foundation. Further details will not be elaborated here.

[0144] On the other hand, in some embodiments, this application provides an intelligent teaching assistant system, which includes an intelligent learning path optimization unit, referencing... Figure 4 The figure is a schematic diagram of the structure of an intelligent learning path optimization unit 400 according to some embodiments of this application. The intelligent learning path optimization unit 400 includes: a data acquisition module 401, a processing module 402, and an execution module 403, which are described below:

[0145] The acquisition module 401 in this application is mainly used to acquire multi-channel learning behavior data of students when interacting with the intelligent teaching assistant terminal through multi-modal sensors. Based on the preset teaching knowledge graph, it obtains the sequence dependency relationship between learning content units corresponding to multiple teaching stages in the personalized learning path, wherein the sequence dependency relationship reflects the subject knowledge logic.

[0146] Processing module 402 in this application is used to extract the learning behavior features of students on different interaction channels from the multi-channel learning behavior data for each teaching stage, construct a structured feature vector of each learning content unit based on the inherent attributes of the learning content unit, and construct a student behavior feature vector of each learning content unit based on the learning behavior features.

[0147] The processing module 402 is further configured to perform multimodal fusion of the structured feature vector and the student behavior feature vector to generate a joint representation vector for each learning content unit, and for each teaching stage, to aggregate the joint representation vectors of all learning content units within that teaching stage to generate a teaching stage representation vector.

[0148] The processing module 402 is further configured to evaluate the teaching effectiveness of the teaching guidance based on the teaching stage representation vector, obtain the teaching effectiveness evaluation value of each teaching stage, and compare the teaching effectiveness evaluation value with a predetermined threshold to obtain the learning state deviation degree of each teaching stage.

[0149] The execution module 403 in this application is mainly used to filter out the teaching stages in the teaching and tutoring process where the teaching effect is not up to standard by using the deviation degree of each learning state, and to optimize and adjust the learning path based on the sequence dependency relationship and the teaching stages where the teaching effect is not up to standard.

[0150] In addition, this application also provides a computer device, the computer device including a memory and a processor, the memory storing code, and the processor being configured to acquire the code and execute the above-described intelligent learning path optimization method.

[0151] In some embodiments, reference Figure 5 The figure is a schematic diagram of the structure of a computer device implementing an intelligent learning path optimization method according to some embodiments of this application. The intelligent learning path optimization method in the above embodiments can... Figure 5 The computer device shown is used to implement this, and the computer device 500 includes at least one processor 501, a communication bus 502, a memory 503, and at least one communication interface 504.

[0152] Processor 501 can be a general-purpose central processing unit (CPU) or an application-specific integrated circuit (ASIC).

[0153] The communication bus 502 can be used to transmit information between the aforementioned components.

[0154] Memory 503 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CDROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disks or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. Memory 503 may exist independently and be connected to processor 501 via communication bus 502. Memory 503 may also be integrated with processor 501.

[0155] The memory 503 stores program code for executing the scheme of this application, and its execution is controlled by the processor 501. The processor 501 executes the program code stored in the memory 503. The program code may include one or more software modules. In the above embodiments, the intelligent learning path optimization method can be implemented by the processor 501 and one or more software modules in the program code in the memory 503.

[0156] Communication interface 504 uses any transceiver-like device to communicate with other devices or communication networks, such as Ethernet, radio access network (RAN), wireless local area networks (WLAN), etc.

[0157] In a specific implementation, as one example, a computer device may include multiple processors, each of which may be a single-core (single CPU) processor or a multi-core (multi CPU) processor. Here, a processor may refer to one or more devices, circuits, and / or processing cores used to process data (e.g., computer program instructions).

[0158] The aforementioned computer device can be a general-purpose computer device or a special-purpose computer device. In specific implementations, the computer device can be a desktop computer, a portable computer, a network server, a handheld digital assistant (PDA), a mobile phone, a tablet computer, a wireless terminal device, a communication device, or an embedded device. This application does not limit the type of computer device.

[0159] In addition, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described intelligent learning path optimization method.

[0160] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0161] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for optimizing intelligent learning paths, characterized in that, Includes the following steps: Multi-modal sensors are used to collect multi-channel learning behavior data of students when they interact with intelligent teaching assistant terminals for personalized learning. Based on a preset teaching knowledge graph, the sequence dependency relationship between learning content units corresponding to multiple teaching stages in the personalized learning path is obtained, and the sequence dependency relationship reflects the logic of subject knowledge. For each teaching stage, the learning behavior features of students on different interaction channels are extracted from the multi-channel learning behavior data. A structured feature vector of each learning content unit is constructed based on the inherent attributes of the learning content unit. At the same time, a student behavior feature vector of each learning content unit is constructed based on the learning behavior features. The structured feature vector and the student behavior feature vector are fused in a multimodal manner to generate a joint representation vector for each learning content unit. For each teaching stage, the joint representation vectors of all learning content units within that teaching stage are aggregated to generate a teaching stage representation vector. The teaching effectiveness of the teaching guidance is evaluated based on the teaching stage representation vector to obtain the teaching effectiveness evaluation value of each teaching stage. The teaching effectiveness evaluation value is then compared with a predetermined threshold to obtain the learning state deviation degree of each teaching stage. The deviation of each learning state is used to filter out the teaching stages in the teaching and tutoring process where the teaching effect is not up to standard. Based on the sequence dependency relationship and the teaching stages where the teaching effect is not up to standard, the learning path is optimized and adjusted. The process of obtaining the sequence dependencies between learning content units corresponding to multiple teaching stages in a personalized learning path, based on a pre-set teaching knowledge graph, includes the following steps: Extract the prerequisite and prerequisite knowledge points for each learning content unit from the teaching knowledge graph; Based on the aforementioned prerequisite and subsequent knowledge points, establish the sequential association between each learning content unit and other learning content units, wherein other learning content units are those that have prerequisite knowledge points associated with the current learning content unit and those that have subsequent knowledge points associated with the current learning content unit. The sequential association is taken as the sequence dependency; The process of constructing a structured feature vector for each learning content unit based on its inherent attributes includes the following steps: Select the core inherent attributes of the learning content units; The selected core inherent attributes are converted into numerical values. The numerical attributes of each learning content unit are arranged in a fixed order to construct a structured feature vector; Among them, the student behavior feature vector is a multi-dimensional learning behavior feature vector formed by normalizing and weighting each learning behavior feature, and retaining the data dimensions after weighting each indicator. The process of multimodal fusion of the structured feature vector and the student behavior feature vector to generate a joint representation vector for each learning content unit includes the following steps: The structured feature vector and the student behavior feature vector are input into a preset feature fusion unit; In the feature fusion unit, a vector concatenation operation is performed on the structured feature vector and the student behavior feature vector; The joint representation vector is generated from the result of the vector concatenation operation.

2. The intelligent learning path optimization method according to claim 1, characterized in that, Multi-modal sensors are used to collect multi-channel learning behavior data of students during personalized learning interactions with intelligent teaching assistant terminals, including the following steps: Use visual sensors to capture students' facial expression data and eye movement data; Use audio sensors to record students' voice input data and intonation variation data; Use interactive sensors to record students' touch input and keyboard input data; The facial expression data, eye movement data, voice input data, tone change data, touch input data, and keyboard input data are integrated into the multi-channel learning behavior data.

3. The intelligent learning path optimization method according to claim 1, characterized in that, For each teaching stage, the learning behavior characteristics of students on different interaction channels are extracted from the multi-channel learning behavior data, including the following steps: Separate visual channel data, audio channel data, and interactive channel data from the multi-channel learning behavior data; Feature extraction is performed on the visual channel data to obtain the student's attention concentration index; Feature extraction is performed on the audio channel data to obtain the students' comprehension and expression accuracy index; Feature extraction is performed on the interactive channel data to obtain the student's response speed index; Collect students' learning outcome data for each learning content unit, and convert the learning outcome data into knowledge mastery indicators; The attention concentration index, comprehension and expression accuracy index, response speed index, and knowledge mastery index are combined to form the learning behavior characteristics.

4. The intelligent learning path optimization method according to claim 1, characterized in that, For each teaching stage, the joint representation vector of all learning content units within that teaching stage is aggregated to generate the teaching stage representation vector, including the following steps: Collect the joint representation vectors of all learning content units within this teaching phase; The importance weight of each learning content unit is determined based on the sequence dependency relationship; Perform a weighted average pooling operation based on the importance weights on the joint representation vector; The teaching stage representation vector is generated from the result of the weighted average pooling operation.

5. An intelligent teaching assistant system, comprising an intelligent learning path optimization unit, which performs intelligent learning path optimization using the method described in any one of claims 1 to 4, characterized in that, The intelligent learning path optimization unit specifically includes: The data acquisition module is used to collect multi-channel learning behavior data of students when they interact with the intelligent teaching assistant terminal through multi-modal sensors. Based on the preset teaching knowledge graph, it obtains the sequence dependency relationship between the learning content units corresponding to multiple teaching stages in the personalized learning path, wherein the sequence dependency relationship reflects the subject knowledge logic. The processing module is used to extract the learning behavior features of students on different interaction channels from the multi-channel learning behavior data for each teaching stage, construct a structured feature vector for each learning content unit based on the inherent attributes of the learning content unit, and construct a student behavior feature vector for each learning content unit based on the learning behavior features. The processing module is further configured to perform multimodal fusion of the structured feature vector and the student behavior feature vector to generate a joint representation vector for each learning content unit. For each teaching stage, the module aggregates the joint representation vectors of all learning content units within that teaching stage to generate a teaching stage representation vector. The processing module is also used to evaluate the teaching effect of the teaching guidance based on the teaching stage representation vector, obtain the teaching effect evaluation value of each teaching stage, and compare the teaching effect evaluation value with a predetermined threshold to obtain the learning state deviation degree of each teaching stage. The execution module is used to filter out the teaching stages that do not meet the teaching effectiveness standards during the teaching and tutoring process using the deviation degree of each learning state, and optimize and adjust the learning path based on the sequence dependency relationship and the teaching stages that do not meet the teaching effectiveness standards.

6. A computer device comprising a memory and a processor, the memory storing code, characterized in that, The processor is configured to acquire the code and execute the intelligent learning path optimization method according to any one of claims 1 to 4.

7. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the intelligent learning path optimization method according to any one of claims 1 to 4.

Images