Cognitive level quantification method for large-model dynamic adaptive teaching strategy

Abstract

The invention relates to the technical field of intelligent teaching, and discloses a cognitive level quantification method for a large-model dynamic adaptive teaching strategy, and the method comprises the steps: collecting the multi-modal data of a learner in real time; preprocessing the multi-modal data to obtain preprocessed multi-modal data; inputting the preprocessed multi-modal data into a pre-constructed large model to obtain a target teaching strategy corresponding to the learner; the large model is used for calculating an index score corresponding to a preset dimension based on the preprocessed multi-modal data; determining a weight coefficient corresponding to the preset dimension based on the index score corresponding to the preset dimension; and determining a cognitive conflict level corresponding to the learner according to the index score corresponding to the preset dimension and the weight coefficient, and determining a target teaching strategy according to the cognitive conflict level corresponding to the learner. The teaching strategy can be immediately adjusted according to the cognitive conflict level judged in real time, the interaction behavior is responded in time, and the problems that behavior prediction lags behind and the strategy cannot be adjusted in real time are effectively solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent teaching, in particular to a cognitive level quantification method for dynamic adaptive teaching strategy of a large model. BACKGROUND

[0002] Among the existing intelligent teaching related technologies, two types of solutions are widely used. One is a knowledge question and answer system based on a large language model. The system adopts a static question and answer mode. After a learner inputs a specific question in text, the system first performs word segmentation, syntax analysis and semantic understanding on the question with the help of natural language processing technology, then calls a pre-trained large language model to retrieve relevant content in a knowledge base to generate a preliminary answer, and finally returns the answer and explanation such as formula derivation and example explanation in text form. However, the entire process can only realize one-way knowledge transmission and does not have dynamic interaction capability. The other is a personalized learning path recommendation technology. The technology relies on offline batch processing analysis. First, the behavior data such as accuracy and answering time of the learner are extracted from historical answer records and stored in a database. Then, analysis algorithms such as clustering and association rule mining are periodically run to identify learning patterns. Based on the analysis results, a learning order recommendation is generated. Finally, the recommendation is pushed to the learner interface in the form of a list or a graph. Similarly, the technology lacks real-time updating capability and cannot adjust the recommendation results according to the real-time state of the learner.

[0003] The above-mentioned methods lack dynamic adaptability. The personalized learning path recommendation technology relies on historical answer data for offline analysis, and the knowledge question and answer system does not involve real-time state capture of the learner. Therefore, it is difficult to capture the dynamic behaviors such as multi-round follow-up questions and strategy modification of the learner during the interaction process in a timely manner, and there is obvious lag in the judgment and behavior prediction of the cognitive state of the learner. It is difficult to adjust the teaching strategy according to the real-time changes of the learner. SUMMARY

[0004] Therefore, the embodiments of the present application provide a cognitive level quantification method for dynamic adaptive teaching strategy of a large model.

[0005] In a first aspect, the embodiments of the present application provide a cognitive level quantification method for dynamic adaptive teaching strategy of a large model, which comprises:

[0006] real-time collection of multi-modal data of the learner; pre-processing of the multi-modal data to obtain pre-processed multi-modal data; the multi-modal data includes knowledge structure level data, emotional state level data, human-computer communication level data and neural activity level data;

[0007] The preprocessed multimodal data is input into a pre-constructed large model to obtain the target teaching strategy corresponding to the learner. The large model is used to calculate the index scores corresponding to preset dimensions based on the preprocessed multimodal data; determine the weight coefficients corresponding to the preset dimensions based on the index scores corresponding to the preset dimensions; determine the cognitive conflict level of the learner based on the index scores and weight coefficients corresponding to the preset dimensions; and determine the target teaching strategy based on the cognitive conflict level of the learner. The preset dimensions include: knowledge conflict dimension, logical conflict dimension, and metacognitive conflict dimension.

[0008] In one possible implementation, the real-time acquisition of learners' multimodal data includes:

[0009] The collected sequence data of the learner's answer steps and the click stream data of the answer options, the knowledge graph nodes accessed by the learner and the connection relationships between the knowledge graph nodes, and the knowledge points that the learner has not mastered are used as the knowledge structure level data.

[0010] The collected data on the learner's confused facial expressions, the learner's mouse and / or touchscreen movement trajectory data, click frequency data, the learner's keyboard typing speed data, and the keyboard backspace key usage frequency data will be used as emotional state level data.

[0011] The collected voice data of the learner's problem-solving process, the text data of the learner's follow-up questions, the voice analysis results obtained by analyzing the voice data, and the text analysis results obtained by analyzing the text data are used as data at the human-computer communication level.

[0012] The neurophysiological characteristics data collected from the learners after they answered questions incorrectly within a preset time period are used as data at the neural activity level.

[0013] In one possible implementation, the preprocessing of the multimodal data to obtain preprocessed multimodal data includes:

[0014] The multimodal data is cleaned, standardized, and time-series aligned to obtain preprocessed multimodal data.

[0015] In one possible implementation, the step of calculating the index score corresponding to the preset dimension based on the preprocessed multimodal data includes:

[0016] Calculate the basic indicators corresponding to the knowledge structure level data, emotional state level data, human-computer communication level data, and neural activity level data under each preset dimension;

[0017] The index score corresponding to each preset dimension is calculated according to the basic index corresponding to the knowledge structure level data, the sentiment state level data, the human-computer communication level data and the neural activity level data corresponding to each preset dimension.

[0018] In a possible implementation, the determining of the weight coefficient corresponding to the preset dimension based on the index score corresponding to the preset dimension comprises:

[0019] The subjective weight corresponding to the preset dimension is determined based on the index score corresponding to the preset dimension and the Delphi method, the objective weight corresponding to the preset dimension is determined based on the entropy weight method, and the weight coefficient corresponding to the preset dimension is determined based on the subjective weight and the objective weight corresponding to the preset dimension.

[0020] In a possible implementation, the determining of the cognitive conflict level corresponding to the learner according to the index score and the weight coefficient corresponding to the preset dimension comprises:

[0021] A cognitive conflict index is determined according to the index score and the weight coefficient corresponding to the preset dimension.

[0022] A plurality of cognitive conflict index ranges and a cognitive conflict level corresponding to each cognitive conflict index range are obtained in advance;

[0023] A cognitive conflict index range to which the cognitive conflict index belongs is determined.

[0024] The cognitive conflict level corresponding to the cognitive conflict index range to which the cognitive conflict index belongs is determined as the cognitive conflict level corresponding to the learner.

[0025] In a possible implementation, the determining of the target teaching strategy according to the cognitive conflict level corresponding to the learner comprises:

[0026] A matching relationship between each cognitive conflict level and a corresponding teaching strategy is obtained in advance;

[0027] Based on the matching relationship, the teaching strategy corresponding to the cognitive conflict level corresponding to the learner is determined as the target teaching strategy.

[0028] In a second aspect, the embodiments of the present application provide a cognitive level quantification device for dynamic adaptive teaching strategy of a large model, the device comprising:

[0029] A first processing module is configured to collect multi-modal data of a learner in real time, and obtain preprocessed multi-modal data by preprocessing the multi-modal data; the multi-modal data comprises knowledge structure level data, sentiment state level data, human-computer communication level data and neural activity level data.

[0030] The second processing module is configured to input the preprocessed multi-modal data into a pre-constructed large model to obtain the target teaching strategy corresponding to the learner; the large model is configured to calculate an index score corresponding to a preset dimension based on the preprocessed multi-modal data; determine a weight coefficient corresponding to the preset dimension based on the index score corresponding to the preset dimension; determine a cognitive conflict level corresponding to the learner according to the index score corresponding to the preset dimension and the weight coefficient, and determine the target teaching strategy according to the cognitive conflict level corresponding to the learner; and the preset dimension includes a knowledge conflict dimension, a logic conflict dimension and a meta-cognition conflict dimension.

[0031] In a third aspect, an embodiment of the present application provides a computer device, including a memory and a processor, which are in communication connection with each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the large model dynamic adaptive teaching strategy cognitive level quantification method of the first aspect or any of the corresponding embodiments thereof.

[0032] In a fourth aspect, an embodiment of the present application provides a computer readable storage medium, which stores computer instructions, and the computer instructions are used to make a computer execute the large model dynamic adaptive teaching strategy cognitive level quantification method of the first aspect or any of the corresponding embodiments thereof.

[0033] The technical scheme provided by the present application has the following technical effects:

[0034] The multi-modal data collected by the embodiment of the present application covers four core layers of knowledge structure, emotional state, human-computer communication and neural activity. Among them, the knowledge structure layer data reflects the knowledge mastery degree of the learner, the emotional state layer data embodies the emotional fluctuation of the learner, the human-computer communication layer data captures the cognitive feedback in the interaction, and the neural activity layer data reveals the deep cognitive activity. Compared with the related art which only relies on text questions or historical answering data, this multi-dimensional data collection method can more completely depict the real-time cognitive state of the learner, and provide data support for subsequent accurate analysis.

[0035] Through preprocessing, the noise in the multi-modal data can be eliminated, the data format and dimension are unified, and the data quality and consistency are ensured. This link solves the problem of complex multi-modal data sources and heterogeneous formats, lays a reliable data foundation for subsequent index calculation and weight determination, and avoids the deviation of cognitive state judgment caused by data quality problems.

[0036] The preset dimensions include a knowledge conflict dimension, a logic conflict dimension, and a meta-cognition conflict dimension, which respectively correspond to core contradictions of learners in knowledge mastery, logic reasoning, and cognitive monitoring levels, and conform to a target hierarchical structure in the cognitive field. Compared with related technologies that only focus on a single judgment at the knowledge level, this layered quantification can more deeply and comprehensively disassemble cognitive contradictions of learners and accurately locate weak links in cognition.

[0037] The weight coefficient is determined based on the index score of the preset dimension, rather than using a fixed weight. For example, in a K12 mathematics basic teaching scenario, the knowledge conflict (such as concept misunderstanding) of the learner has a greater impact on the learning effect, and the system can automatically increase the weight of the knowledge conflict dimension through the index score feedback; in a higher physics logic reasoning teaching scenario, the logic conflict (such as a broken reasoning chain) is more critical, and the weight is inclined to the logic conflict dimension. Compared with the fixed analysis logic of related technologies, such as the historical correct rate being the core weight in the personalized learning path recommendation, this dynamic weight mechanism can more flexibly match the cognitive needs of different education scenarios and different learners, and improve the scene adaptability of cognitive conflict judgment.

[0038] The weight coefficient is determined based on the index score, which essentially reflects the contribution of each dimension to the current cognitive state through data feedback (for example, if the knowledge conflict index score of a learner is extremely high, it means that the contradiction at the knowledge level is the main cognitive problem at present, and the corresponding weight naturally increases). This data-driven weight determination method not only avoids the experience bias of purely subjective weights, but also avoids the data noise interference of purely objective weights, ensuring that the weight distribution can truly reflect the cognitive focus of the learner and provide a scientific basis for subsequent cognitive conflict level determination.

[0039] Real-time multi-modal data collection replaces offline storage of historical data, covering four levels of knowledge, emotion, communication, and neural, and real-time capturing of dynamic behaviors of learners, eliminating the limitations of relying on historical data from the data source.

[0040] Dynamic weight and multi-dimensional index calculation replace offline analysis of fixed algorithms, adjust the weight coefficient of each preset dimension based on real-time index score, and real-time quantify cognitive conflicts, avoiding the lag caused by periodic analysis, and ensuring that the judgment of the cognitive state is synchronized with the real-time behavior of the learner.

[0041] Teaching strategies can be immediately adjusted according to the real-time determined cognitive conflict level, and interactive behaviors can be responded to in a timely manner, effectively solving the problem of behavior prediction lag and inability to adjust strategies in real time. BRIEF DESCRIPTION OF DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the specific embodiments or related art of the present application, the drawings needed to be used in the description of the specific embodiments or related art will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0043] Figure 1 is a flowchart of the cognitive level quantification method of the large model dynamic adaptive teaching strategy of the embodiment of the present application;

[0044] Figure 2 is a flowchart of the overall idea of the embodiment of the present application;

[0045] Figure 3 is a structural block diagram of the cognitive level quantification device of the large model dynamic adaptive teaching strategy of the embodiment of the present application;

[0046] Figure 4 is a hardware structure schematic diagram of the computer device of the embodiment of the present application. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor belong to the scope of protection of the present application.

[0048] Among the existing intelligent teaching related technologies, two types of schemes are widely used. One is a knowledge question and answer system based on a large language model, and the other is a personalized learning path recommendation technology.

[0049] The knowledge question and answer system is based on a static question and answer mode of a large language model (such as ChatGPT), which provides answer explanations through natural language processing technology.

[0050] The flow of the knowledge question and answer system is as follows:

[0051] (1) User asks a question: the learner inputs a question through text.

[0052] (2) Natural language processing (NLP): the system performs word segmentation, syntax analysis and semantic understanding on the question.

[0053] (3) Large language model calling: based on a pre-trained model (such as ChatGPT), relevant content in the knowledge base is retrieved to generate a preliminary answer.

[0054] (4) Answer explanation output: Return answers and explanations in text form (such as formula derivation, example explanation), without dynamic interaction capability.

[0055] Personalized learning path recommendation: Use historical answer data (such as EdNet, ASSISTments dataset) to generate learning paths, rely on offline batch analysis.

[0056] The process of personalized learning path recommendation is as follows:

[0057] (1) Data collection and storage: Extract learner behavior data (such as accuracy, answer time) from historical answer records and store them in the database.

[0058] (2) Offline batch analysis: Periodically (such as daily / weekly) run analysis algorithms (such as clustering, association rule mining) to identify learning patterns.

[0059] (3) Learning path generation: Recommend learning order based on analysis results (such as "learn algebra first, then learn geometry").

[0060] (4) Recommendation result display: Push to the learner interface in the form of list or graph, without real-time update capability.

[0061] The disadvantages of the above two methods are:

[0062] Weak heuristic teaching ability: Unable to simulate the heuristic behavior of human teachers (such as contradiction questioning, analogy guidance), only provide one-way knowledge transfer, making it difficult for learners to trigger cognitive conflicts (such as logical contradictions, knowledge gaps), and lack of motivation for deep learning.

[0063] Lack of dynamic adaptability: Relying on historical data for offline analysis, lacking lightweight fusion of real-time multi-modal data (such as eye tracking, EEG signals), and not exploring deep cognitive levels, resulting in inability to capture learners' multi-round questioning, strategy modification and other interactive behaviors in time, and lagging behavior prediction.

[0064] The embodiment of the present application provides a cognitive level quantification method of large model dynamic adaptive teaching strategy, to solve the above problems, it should be noted that the steps shown in the flowchart of the drawing can be executed in a computer device such as a group of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that here.

[0065] Figure 1 is the flowchart of the cognitive level quantification method of large model dynamic adaptive teaching strategy of the embodiment of the present application, as shown in the figure, the flowchart includes the following: Figure 1 ​

[0066] S101: Real-time acquisition of learners' multimodal data. Preprocessing of the multimodal data to obtain preprocessed multimodal data.

[0067] In this embodiment, multimodal data includes: knowledge structure level data, emotional state level data, human-computer communication level data, and neural activity level data. As an example, learners include students. Learner multimodal data can be collected in real-time from the learning platform / system's backend log system.

[0068] S102: Input the preprocessed multimodal data into a pre-built large model to obtain the learner's corresponding target teaching strategy; the large model is used to calculate the index scores corresponding to the preset dimensions based on the preprocessed multimodal data; determine the weight coefficients corresponding to the preset dimensions based on the index scores corresponding to the preset dimensions; determine the learner's corresponding cognitive conflict level based on the index scores and weight coefficients corresponding to the preset dimensions; and determine the target teaching strategy based on the learner's corresponding cognitive conflict level.

[0069] In this embodiment, the preset dimensions include: knowledge conflict dimension, logical conflict dimension, and metacognitive conflict dimension. This embodiment does not limit the specific training method for the large model; those skilled in the art can choose according to actual needs.

[0070] In one possible implementation, the real-time acquisition of learners' multimodal data in S101 includes:

[0071] S1011: Collect learners' answer step sequence data and answer option click stream data, knowledge graph nodes accessed by learners and the connection relationships between knowledge graph nodes, and knowledge points that learners have not mastered as knowledge structure level data.

[0072] In this embodiment, the learning platform's backend log system can record the learner's step sequence (such as the application of formulas and the order of derivation in solving math problems) and option click flow (such as the option switching trajectory in multiple-choice questions) in real time. A knowledge graph interaction tracking tool is used to capture the knowledge graph nodes visited by the learner (such as the "Mechanical Motion" node) and the connections between nodes (such as the association between the "Mechanical Motion" node and the "Inertia" node). A real-time knowledge assessment module, such as a pop-up quiz on related knowledge points immediately after answering questions, identifies and records knowledge points that the learner has not mastered, such as those with a test accuracy rate below 60%.

[0073] S1012: Collect learners' confused facial expressions, learners' mouse and / or touchscreen movement trajectory data, click frequency data, learners' keyboard typing speed data, and keyboard backspace key usage frequency data as emotional state level data.

[0074] In this embodiment, a built-in camera can be used to capture a video stream of the learner's face at a frequency of 15 frames per second. Confused facial expression data, such as frame sequences of furrowed brows and downturned corners of the mouth, can be extracted using video analysis tools. An input device behavior recording plugin can be used to collect in real time mouse movement trajectories (e.g., irregular wandering paths), click frequency (e.g., number of clicks per minute and intervals), touchscreen operation force (when pressure-sensitive devices are supported), keyboard typing speed (e.g., number of characters input per minute), and backspace key usage frequency (e.g., number of backspaces per paragraph of text input).

[0075] S1013: Collect the voice data of the learner's problem-solving process, the text data of the learner's follow-up questions, and the voice analysis results obtained from analyzing the voice data and the text analysis results obtained from analyzing the text data as data for human-computer communication.

[0076] In this embodiment, the learning platform's voice recording function can be used to collect real-time audio data of learners verbally explaining their problem-solving approaches. The question-and-answer interaction module records text data of learners' follow-up questions (e.g., "Why is this formula used this way?"). The audio data is then processed using the MFCC feature extraction algorithm for noise reduction and feature analysis, yielding audio analysis results such as emotional intensity and speech rate. The text data is processed using the BERT semantic understanding model for word segmentation, stop word removal, and logical relationship recognition, yielding text analysis results (e.g., problem focus, logical contradictions).

[0077] S1014: Collect neurophysiological characteristic data of learners within a preset time after they answer questions incorrectly as data at the level of neural activity.

[0078] In this embodiment, a portable electroencephalography (EEG) device (such as EmotivEpoc+) can be used to acquire EEG signals from the anterior cingulate cortex within 50-150 milliseconds after the learner answers an incorrect question and receives an "error" notification from the system. Simultaneously, near-infrared light intensity signals at wavelengths of 760nm and 850nm from the learner's prefrontal cortex are acquired using an fNIRS device. These EEG signals and near-infrared light intensity signals are then used as neurophysiological characteristic data.

[0079] In this embodiment of the application, as an example, the learner's answer step sequence data and answer option click stream data can be extracted from the learning platform's interaction logs. The knowledge graph nodes accessed by the learner and the connections between these nodes are recorded. Core knowledge points that the learner has not yet mastered are extracted from the knowledge assessment test results. The collected data is then used as knowledge structure level data.

[0080] In the embodiments of the present application, as an example, the confusion expression data of the learner can be specifically recognized by combining a built-in camera with a pre-trained facial expression recognition (FER) model and a facial action coding system (FACS), and the mouse and / or touch screen movement trajectory data of the learner, the mouse and / or touch screen click frequency data of the learner, the keyboard typing speed data of the learner, and the keyboard backspace key usage frequency data of the learner can be extracted from the high-precision input device log.

[0081] In the embodiments of the present application, as an example, the voice data of the problem-solving thought of the learner can be specifically collected by a recording device, the text data of the self-explanation of the learner and the text data of the question of the learner can be collected by an interactive interface text input box, the feedback content data of the logical explanation of the learner and the instruction data of whether the learner requires step-by-step explanation can be extracted from the dialogue interaction log, the voice analysis result can be obtained by using NLP technology to perform core concept recognition and logical relationship analysis on the voice data, and the text analysis result can be obtained by using NLP technology to perform core concept recognition and logical relationship analysis on the text data. The above collected data can be used as human-computer communication level data.

[0082] In the embodiments of the present application, as an example, the error-related negativity (ERN) signal data of the learner within 50-150 ms after the answer error can be specifically collected by a portable EEG device, the prefrontal theta band electroencephalogram signal data of the learner can be collected, the 760 nm wavelength near-infrared light intensity data can be collected by an fNIRS device, and the 850 nm wavelength near-infrared light intensity data can be collected. The above collected data can be used as neural activity level data.

[0083] In one possible implementation, the pre-processing of the multi-modal data in S101 to obtain the pre-processed multi-modal data comprises:

[0084] The multi-modal data is subjected to data cleaning, standardization and time sequence alignment to obtain the pre-processed multi-modal data.

[0085] In the embodiments of the present application, data cleaning can eliminate invalid and abnormal data, correct data bias, improve data quality and reliability, and lay a foundation for subsequent standardization, time alignment and index calculation. As an example, data cleaning can include: removing abnormal data caused by device failure, such as EEG signal mutation and invalid click logs caused by accidental touch. Noise reduction processing is performed on voice data, and natural language preprocessing such as word segmentation and stop word removal is performed on text data. MTCNN algorithm is used for face detection, alignment and standardization of video frames, and faceless frames are removed.

[0086] In the embodiments of the present application, in view of the characteristics of different sources of multi-modal data, large dimension difference and scattered value range, unified mapping and numerical calibration are used to convert each type of data into standardized data that can be directly used for cross-modal fusion analysis. As an example, data standardization can include: for EEG signals at the level of neural activity (such as ERN amplitude in a 80-150ms time window, original value may be -50μV~0μV), near-infrared intensity data collected by fNIRS device (760nm and 850nm wavelength signals, original value may be 1000~5000 count units), using Z-score standardization method, calculating the difference between each data point and the mean value of the modal data, and then dividing by the standard deviation of the modal data, converting the data into standard normal distribution data with mean value of 0 and standard deviation of 1, eliminating the influence of dimension difference of different physiological signals on subsequent index calculation.

[0087] In the embodiments of the present application, time alignment can eliminate the asynchrony of different modal data in the time dimension, bind scattered multi-source data to a unified time axis, and ensure the time correlation of multi-modal data. As an example, time alignment can include: binding different modal data by timestamp, such as corresponding EEG signals, facial expression data and mouse click behavior at a certain time to the same time node, to ensure the consistency of multi-modal data in the time dimension.

[0088] In one possible implementation, the large model is used to calculate the index score corresponding to the preset dimension based on the preprocessed multi-modal data, and the index score corresponding to the preset dimension is calculated based on the preprocessed multi-modal data, specifically including:

[0089] Calculate the basic index corresponding to the knowledge structure layer data, the emotional state layer data, the human-computer communication layer data and the neural activity layer data under each preset dimension respectively.

[0090] According to the basic index corresponding to the knowledge structure layer data, the emotional state layer data, the human-computer communication layer data and the neural activity layer data under each preset dimension respectively, the index score corresponding to each preset dimension is calculated.

[0091] In the embodiments of the present application, the index score corresponding to the preset dimension includes: an index score corresponding to a knowledge conflict dimension, an index score corresponding to a logic conflict dimension, and an index score corresponding to a meta-cognition conflict dimension.

[0092] As an example, the calculation method of the index score corresponding to the knowledge conflict dimension is as follows:

[0093] (1) At the knowledge structure level, the concept connection error rate and the path contradiction are calculated, and the two are weighted and summed according to a preset proportion to obtain the knowledge structure dimension score. The knowledge structure dimension score is taken as the basic index corresponding to the knowledge structure level data.

[0094] Specifically, the connection set E_u of the learner's knowledge graph and the connection set E_s of the standard knowledge graph are obtained, and the formula is as follows: The concept connection error rate is calculated. At the same time, the standard reasoning path set S_{std} and the learner's reasoning path set S_{stu} are collected, the Jaccard similarity (J(S_{std}, S_{stu})) of the two is calculated, and the path contradiction is obtained by subtracting the similarity from 1. The two indexes are weighted and summed according to a preset proportion (such as 50% each) to obtain the knowledge structure dimension score.

[0095] Concept connection error rate:

[0096] Data sources: learner's answer step sequence, option click stream recorded by the learning platform, and standard knowledge graph (G_s=(V_s, E_s)) and learner's personal knowledge graph (G_u=(V_u, E_u)).

[0097] Calculation logic: First, the associated cognition of the learner to the knowledge concept is extracted from the answer step and the click stream to construct the connection set E_u of the personal knowledge graph. By comparing the connection set E_s of the standard knowledge graph, the number of error connections belonging to E_u but not belonging to E_s is counted. Finally, the concept connection error rate is obtained by dividing the number of error connections by the total number of connections in E_u, and the formula is as follows:

[0098] Path contradiction:

[0099] Data sources: The standard reasoning path set S_{std} generated by the system, such as "concept A→theorem B→formula C→answer D", the operation log (including the object of each step, timestamp) and the oral thought transcription text (extracted concepts and steps after NLP processing) of the learner's problem solving process.

[0100] Computing logic: The actual reasoning path of the learner is sorted into a set S_{stu}, and the Jaccard similarity between S_{std} and S_{stu} is calculated first (the number of intersection elements divided by the number of union elements, formula: J(S_{std}, S_{stu}) = |S_{std}∩S_{st u}| / |S_{std}∪S_{stu}|. Then subtract 1 from the similarity to get the path contradiction, that is, path contradiction = 1-J(S_{std}, S_{stu}.

[0101] Knowledge structure dimension score (basic index):

[0102] Computing logic: The above two indicators are weighted and summed according to the preset proportion (such as 40% for concept connection error rate and 60% for path contradiction). If the concept connection error rate is 0.3 and the path contradiction is 0.5, then the knowledge structure dimension score = 0.3 x 40% + 0.5 x 60% = 0.42.

[0103] (2) The emotional state level, calculate the confusion expression duration ratio, get the frustration probability through the random forest classification model and mark high frustration, and fuse the two according to the preset weight to get the emotional state dimension score. The emotional state dimension score is taken as the basic index corresponding to the data of the emotional state level.

[0104] Specifically, the total confusion expression time (T_{confusion}) and the total learning time (T_{total}) are extracted from the video stream, and the confusion expression duration ratio is calculated by the formula: C = (T_{confusion} / T_{total}) x 100%. The human-computer interaction log feature vector is input into the trained random forest classification model, and the frustration probability is output. If the probability is greater than or equal to 0.7, it is marked as high frustration (assigned value 1), otherwise it is marked as low frustration (assigned value 0). The confusion ratio (normalized to 0-1) and the frustration mark are fused according to the preset weight (such as 7:3) to get the emotional state dimension score (A).

[0105] Confusion expression duration ratio (C):

[0106] Data source: 15 frames / second real-time video stream collected by the built-in camera, total learning time T_{total}.

[0107] Computing logic: Locate the face region and align and normalize frame by frame with MTCNN algorithm, identify the facial action units related to confusion (such as frowning eyebrows, drooping mouth corners) based on FACS theory through CNN, and then use the SVM classifier (threshold (θ = 0.6)) to determine whether each frame is in a state of confusion. Calculate the total time segment T_{confusion} corresponding to all confusion state frames, and finally divide T_{confusion} by T_{total} and multiply by 100% to get the confusion expression duration ratio, the formula is: C = (T_{confusion} / T_{total}) × 100%.

[0108] Frustration probability (F) and high frustration label:

[0109] Data source: High-precision human-computer interaction log (mouse / touch screen movement trajectory, click frequency, high-speed random point mode, keyboard typing speed, backspace key usage frequency, window switching frequency).

[0110] Computing logic: Extract the feature vector ([f_1, f_2,..., f_n]) such as operation confusion degree, operation efficiency, and avoidance behavior from the log in a sliding time window (such as 30 seconds). Input the feature vector into the pre-trained random forest classification model, and the model outputs the frustration probability (F = 1 / (1+e^{-(w_0+w_1f_1+...+w_nf_n)})) through the Sigmoid function according to the learned feature weights (w_i). If (F>0.7), it is marked as high frustration (assigned value 1), otherwise it is marked as low frustration (assigned value 0).

[0111] Emotional state dimension score (basic indicator):

[0112] Computing logic: Fusion according to the preset weight (such as confusion expression duration ratio 70% and high frustration label 30%). First, normalize the confusion expression duration ratio to the [0, 1] interval (such as 30% to 0.3), and if the high frustration label is 1, then the emotional state dimension score = 0.3 × 70% + 1 × 30% = 0.51.

[0113] (3) Human-computer communication level, calculate self-explanation contradiction and focus concentration of follow-up questions, normalize the two and take equal weight to get the human-computer communication dimension score. Take the human-computer communication dimension score as the basic indicator corresponding to the human-computer communication level data.

[0114] Specifically, the BERT model is used to embed the self-explanation text E and the answer text A respectively, calculate the cosine similarity of the two vectors, and then get the self-explanation contradiction through SC = 1-cosine similarity. Perform text analysis on the question set in the question and answer log and map it to the knowledge graph node, and use the Herfindahl-Hirschman index HHI = Σ(si ) 2 Calculate the focus concentration of follow-up questions. Sum the two indicators after normalization with equal weight to obtain the human-machine communication dimension score (C).

[0115] Self-explanation contradiction (SC):

[0116] Data source: learner's self-explanation text (E) and final answer text (A) after solving the problem.

[0117] Calculation logic: Convert E and A into vectors respectively using the BERT sentence embedding model. Calculate the cosine similarity of the two vectors (the closer the value to 1, the higher the logical consistency). Subtract this similarity from 1 to get the self-explanation contradiction, formula: SC = 1 - CosineSimilarity(Embedding(E), Embedding(A)). If the cosine similarity is 0.4, then SC = 0.6.

[0118] Focus concentration of follow-up questions (FC):

[0119] Data source: learner's follow-up question set Q = {q1, q2,..., qn} in the question and answer system log, knowledge graph.

[0120] Calculation logic: Use NLP technology to perform text clustering on the question set, extract the core theme and map it to the concept nodes of the knowledge graph to form a node set C. Calculate the proportion s_i of the number of times each node is asked to the total number of questions. Calculate the concentration by HHI index, formula: HHI = Σ(s i ) 2 , the value of HHI is the focus concentration of follow-up questions FC, the higher the value, the more concentrated the questions, for example, s_1 = 0.6, s_2 = 0.3, s_3 = 0.1, then FC = 0.6 2 + 0.3 2 + 0.1 2 = 0.46.

[0121] Human-machine communication dimension score (basic indicator):

[0122] Calculation logic: Normalize SC and FC by Min-Max standardization (map to [0, 1]) respectively, and sum them with equal weight (50% each). If the normalized SC = 0.5 and FC = 0.4, then the human-machine communication dimension score = 0.5 x 50% + 0.4 x 50% = 0.45.

[0123] (4) At the level of neural activity, extract the intensity of ERN and calculate the prefrontal activation level, and then fuse them with pre-set weight after normalization to obtain the neural activity dimension score. The neural activity dimension score is taken as the basic indicator corresponding to the neural activity level data.

[0124] Specifically, the signal in the 80-150ms time window after the answer error is intercepted from the EEG signal, and the average amplitude or peak negative amplitude in the window is calculated as the ERN intensity. The signal collected by the fNIRS device is processed to obtain the mean value of the task period μ_{task} and the baseline period μ_{baseline}, and the prefrontal activation level is calculated by Δ[HbO]_p=μ_{task}-μ_{baseline}. The ERN intensity and the prefrontal activation level are normalized and fused according to the preset weight (such as 6:4) to obtain the neural activity dimension score (N).

[0125] Error-related negativity (ERN) intensity:

[0126] Data source: The electroencephalogram signal in the [t_error-200ms, t_error+500ms] time window after the learner answers the question incorrectly is collected by the portable EEG device.

[0127] Calculation logic: The electroencephalogram signals of multiple "mistake" trials are superimposed and averaged to eliminate noise. In the averaged ERP waveform, the average amplitude or peak negative amplitude in the 80-150ms time window is measured, and the absolute value is taken as the ERN intensity (such as the average amplitude is -15μV, and the absolute value 15μV is the ERN intensity).

[0128] Prefrontal cortex activation level Δ[HbO]_p:

[0129] Data source: The near-infrared light intensity signals of the prefrontal cortex at 760nm and 850nm wavelengths, and the mean values of the task period and the baseline period of the oxygenated hemoglobin concentration (μ_{task}, μ_{baseline}) collected by the fNIRS device.

[0130] Calculation logic: The original light intensity signal is band-pass filtered (to remove physiological noise) and PCA motion artifact corrected. The light intensity signal is converted to the oxygenated hemoglobin concentration change Δ[HbO] based on the modified Beer-Lambert law. The μ_{task} of the last 15 seconds of the task block (such as the problem solving period) and the μ_{baseline} of the baseline block (rest period) are calculated respectively. The difference between the two is the prefrontal cortex activation level, formula: Δ[HbO]_p=μ_{task}-μ_{baseline}, such as μ_{task}=0.8μmol / L (μ_{baseline}=0.3μmol / L, then Δ[HbO]_p=0.5μmol / L.

[0131] Neural activity dimension score (basic index):

[0132] Computing logic: After standardizing ERN intensity and Δ[HbO]_p with Z-score (mean 0, standard deviation 1) respectively, fuse them according to preset weight (e.g. ERN intensity 40%, Δ[HbO]_p 60%). If standardized ERN intensity is 0.6 and Δ[HbO]_p is 0.4, then the neural activity dimension score = 0.6 x 40% + 0.4 x 60% = 0.48.

[0133] According to the weight vector preset by the education scenario, multiply the four level scores by the corresponding weight vector, and input them into the machine learning fusion model in the art to output the index score corresponding to the knowledge conflict dimension in the [0, 1] interval.

[0134] Specifically, according to the weight vectors of knowledge structure W_S, affective state W_A, human-computer communication W_C, and neural activity W_N preset by the education scenario (such as K12 mathematics or higher physics), multiply S, A, C, and N by the corresponding weights, and input them into the machine learning fusion model (such as random forest or neural network) to output the index score corresponding to the knowledge conflict dimension in the [0, 1] interval.

[0135] As an example, the calculation method of the index score corresponding to the logical conflict dimension:

[0136] (1) At the knowledge structure level, count the number of broken reasoning chains and calculate the generation rate of contradictory reasoning paths, and normalize the two as basic indicators.

[0137] Specifically, analyze the learner's problem-solving step sequence, count the number of missing or skipped steps as the number of broken reasoning chains (PC0). Call the AI logic detection tool to scan all reasoning paths, identify the number of contradictory reasoning, and calculate the generation rate of contradictory reasoning paths (PC1) by "number of contradictory reasoning ÷ total number of reasoning". Normalize PC0 and PC1 to the 0-1 interval as the basic indicators at the knowledge structure level.

[0138] Number of broken reasoning chains (PC0):

[0139] Data sources: Text records or operation logs of learner's problem-solving steps, standard problem-solving steps preset by the system (such as "review the question → set up formula → substitute data → calculate → verify" for a math problem).

[0140] Computing logic: Compare the learner's steps with the standard steps, identify missing key steps (such as calculating directly without setting up formula) or skipped steps (such as jumping from reviewing the question to calculating directly), and count the total number of such steps, which is the number of broken reasoning chains (e.g. 5 steps in standard steps, 2 missing steps and 1 skipped step in learner's steps, then PC0 = 3).

[0141] Generation rate of contradictory reasoning paths (PC1):

[0142] Data source: Learner's all reasoning path records (such as thought text during problem solving process, step selection), AI logic detection tool (such as rule engine).

[0143] Calculation logic: Identify contradictory expressions in reasoning path with rule engine (such as asserting "object is at rest if not subject to force" and "object is still moving if not subject to force" at the same time), count the number of contradictory reasoning. Divide the number of contradictory reasoning by the total number of reasoning to get the contradictory reasoning path generation rate (such as total reasoning number 10 times, contradictory reasoning 3 times, then PC1 = 3 / 10 = 0.3).

[0144] Normalization processing:

[0145] Calculation logic: Use Min-Max standardization to map PC0 (assumed value range 0-10) and PC1 (value range 0-1) to [0, 1]. If PC0 = 3 (Xmin = 0, Xmax = 10), then normalized PC0* = (3-0) / (10-0) = 0.3. PC1 = 0.3 (Xmin = 0, Xmax = 1), normalized PC1* = 0.3.

[0146] (2) At the level of emotional state, calculate cognitive input degree and logic anxiety index, and normalize the two as basic indicators respectively.

[0147] Through facial recognition technology, the duration of frowning action during problem solving is extracted, and compared with the total problem solving time to get the cognitive input degree (PC2). The text feedback of learners is analyzed for emotion, and the voice interaction is detected for stress, and the average emotional intensity is obtained as the logic anxiety index (PC3). Normalize PC2 and PC3 to the range of 0-1 as the basic indicators of emotional state level.

[0148] Cognitive input degree (PC2):

[0149] Data source: Duration of frowning action extracted by facial recognition technology, total problem solving time.

[0150] Calculation logic: Through camera + AI expression analysis tool, the total duration of frowning action of learners during problem solving is counted. Divide the frowning duration by the total problem solving time to get the cognitive input degree, such as frowning duration 2 minutes, total problem solving time 10 minutes, then PC2 = 2 / 10 = 0.2.

[0151] Logic anxiety index (PC3):

[0152] Data source: Text sentiment analysis results (such as frequency of anxiety words in answer feedback text), voice stress detection results (such as voice fundamental frequency fluctuation amplitude).

[0153] Computational logic: Take the continuous mean of the emotional intensity of the text sentiment analysis output (such as the 0-1 interval, 0 for no anxiety, 1 for extreme anxiety) and the emotional intensity of the voice stress detection output as the logical anxiety index (such as the mean of the text emotional intensity 0.4, the mean of the voice emotional intensity 0.5, then PC3 = (0.4 + 0.5) / 2 = 0.45).

[0154] Normalization processing:

[0155] Computational logic: Map PC2 (value range 0-1) and PC3 (value range 0-1) to [0, 1] using Min-Max normalization, if the index itself is in the 0-1 interval, the original value can be directly retained, such as PC2 = 0.2, then PC2* = 0.2, PC3 = 0.45, then PC3* = 0.45.

[0156] (3) Human-computer communication level, statistical logic explanation feedback denial times, recognition logic re-ask rate and statistical requirement step-by-step explanation times, and the three are normalized as basic indicators.

[0157] Specifically, traverse the dialogue log, count the number of negative responses of the learner to the logical explanation as the logical explanation feedback denial times (PC4). Use NLP semantic similarity algorithm to identify repeated logical questions, calculate the logical re-ask rate (PC5). Count the number of times the learner explicitly requires step-by-step explanation (PC6). Normalize PC4, PC5, and PC6 to the 0-1 interval as the basic indicators of the human-computer communication level.

[0158] Logical explanation feedback denial times (PC4):

[0159] Data source: Human-computer dialogue log (such as the response text of the learner to the system logical explanation).

[0160] Computational logic: Traverse the log, count the number of times the learner explicitly denies or indicates that he does not understand (such as "I didn't understand" and "That's not right"), and directly use it as PC4 (such as 4 times of negative response, then PC4 = 4).

[0161] Logical re-ask rate (PC5):

[0162] Data source: Set of logical questions in human-computer dialogue log, Sentence-BERT semantic similarity model.

[0163] Computational logic: Calculate the semantic similarity between questions using Sentence-BERT (threshold 0.8 to determine repeated questions), and count the number of repeated logical questions. Divide the number of repeated questions by the total number of logical questions to get the logical re-ask rate (such as a total of 15 logical questions, 5 repeated, then PC5 = 5 / 15 ≈ 0.33).

[0164] PC6: Number of times of asking for step-by-step explanation

[0165] Data source: Human-computer dialogue log.

[0166] Calculation logic: The number of times that the statistical learner explicitly issues instructions such as "step-by-step explanation" and "step-by-step explanation" is directly used as PC6 (e.g. 3 times, PC6 = 3).

[0167] Calculation logic: Min-Max standardization is used, assuming that PC4 takes a value range of 0-10, PC5 takes a value range of 0-1, and PC6 takes a value range of 0-8. If PC4 = 4, then PC4* = (4-0) / (10-0) = 0.4. PC5 = 0.33, then PC5* = 0.33. PC6 = 3, then PC6* = (3-0) / (8-0) = 0.375.

[0168] (4) At the level of neural activity, obtain the frontal lobe activity and calculate the decision hesitation time, and normalize them as the basic indicators respectively.

[0169] Specifically, the EEG signal is analyzed in the frequency domain, and the power value of the frontal lobe theta band is extracted as the frontal lobe activity (PC7). The actual decision time of the key logical judgment node is recorded, and compared with the average reaction time of the node to obtain the decision hesitation time (PC8). PC7 and PC8 are normalized to the interval of 0-1, as the basic indicators at the level of neural activity.

[0170] Frontal lobe activity (PC7):

[0171] Data source: EEG signal of frontal lobe collected by EEG device, frequency domain analysis tool.

[0172] Calculation logic: Power spectrum analysis is performed on the EEG signal, and the power value of the theta band (4-8Hz) is extracted. By comparing the theta band power in the resting state, the power increase (e.g. resting power 20μV 2 / Hz, task period power 35μV 2 / Hz, then the power increase = 15μV 2 / Hz) is calculated, which is the frontal lobe activity PC7.

[0173] Decision hesitation time (PC8):

[0174] Data source: Behavior log (actual decision time of pre-set key logical judgment node), average reaction time of the node (based on historical data statistics).

[0175] Calculation logic: The actual decision time is subtracted from the average reaction time to obtain the decision hesitation time, e.g. actual decision time 8 seconds, average reaction time 5 seconds, then PC8 = 8-5 = 3 seconds.

[0176] Normalization:

[0177] Calculation logic: Min-Max normalization, assuming PC7 ranges from 0 to 50 μV 2 / Hz, and PC8 ranges from 0 to 10 seconds. If PC7 = 15 μV 2 / Hz, then PC7* = (15-0) / (50-0) = 0.3. PC8 = 3 seconds, then PC8* = (3-0) / (10-0) = 0.3.

[0178] Call the preset weight vector, multiply the normalized nine indicators with the corresponding weight vector, and sum them up to get the index score corresponding to the logic conflict dimension in the interval [0, 1].

[0179] Call the preset weight vector w_0 = 0.125, w_1 = 0.125, w_2 = 0.125, w_3 = 0.125, w_4 = 0.083, w_5 = 0.083, w_6 = 0.084, w_7 = 0.125, w_8 = 0.125, multiply the normalized nine indicators with the corresponding weight, and sum them up to get the index score corresponding to the logic conflict dimension in the interval [0, 1].

[0180] As an example, the calculation method of the index score corresponding to the meta-cognitive conflict dimension:

[0181] (1) At the knowledge structure level, calculate the target deviation degree and plan execution deviation rate, and normalize them as the basis indicators respectively.

[0182] Specifically, compare the actual learning path with the preset target path, count the number of actual covered knowledge points (A) and the number of preset target knowledge points (P), and calculate the target deviation degree by GDI = (A-P) / P. Record the planned time T_{plan}, the actual time T_{actual}, the number of correct answers C_{correct} and the total number of answers C_{total}, and calculate the plan execution deviation rate by PED = (T_{actual} / T_{plan}) × (C_{correct} / C_{total}). Normalize the two indicators to the interval [0, 1].

[0183] Target deviation degree (GDI):

[0184] Data source: number of knowledge points covered by actual learning path of learners (A, such as visiting 8 knowledge points), number of knowledge points in preset target path (P, such as ideal path of 3 knowledge points).

[0185] Calculation logic: divide (A-P) by P to get the target deviation degree, formula: GDI = (A-P) / P, for example, A = 8, P = 3, then GDI = (8-3) / 3 ≈ 1.67.

[0186] Plan Execution Deviation (PED):

[0187] Data source: planned time (T_planned, e.g. 20 minutes), actual time (T_actual, e.g. 40 minutes), correct answer number (C_correct, e.g. 5), total answer number (C_total, e.g. 10).

[0188] Calculation logic: first calculate the time deviation coefficient (T_actual / T_planned) and the correct rate coefficient (C_correct / C_total), then multiply them to get the plan execution deviation rate, the formula is: PED = (T_actual / T_planned) x (C_correct / C_total), for example, time coefficient = 2, correct rate coefficient = 0.5, then PED = 2 x 0.5 = 1.

[0189] Normalization processing:

[0190] Calculation logic: Min-Max standardization is adopted, assuming that the GDI value range is 0-3 and the PED value range is 0-2. If GDI = 1.67, then the normalized target deviation degree GDI* = (1.67-0) / (3-0)≈0.56. PED = 1, then the normalized plan execution deviation rate PED* = (1-0) / (2-0) = 0.5.

[0191] (2) At the emotional state level, calculate the emotion regulation failure index and the reflection log keyword richness, and normalize them as basic indicators respectively.

[0192] Specifically, MFCC feature extraction is used to extract voice tone features, RoBERTa-Large model is used to identify text emotion, preset emotion weight (e.g. anger = 0.6, anxiety = 0.4) and emotion frequency f_i are used, emotion regulation failure index is calculated by (ERF = sum w_i x f_i). TF-IDF algorithm is used to analyze reflection log, the number of deep words (N_depth), the total number of words (N_total) and the number of unique words (N_unique) are counted, and the keyword richness is calculated by RR = (N_depth / N_total) x log(N_unique). Normalize the two indicators to the interval of 0-1.

[0193] Emotion Regulation Failure Index (ERF):

[0194] Data source: emotion frequency (e.g. anger appears 3 times, anxiety appears 2 times) obtained by voice tone analysis (MFCC feature extraction), emotion type obtained by text emotion recognition (RoBERTa-Large model), and preset emotion weight (anger = 0.6, anxiety = 0.4).

[0195] ERF: ERF is calculated by summing the product of each emotion weight and its corresponding frequency. The formula is: ERF = Σ(w_i x f_i). For example, if the frequency of anger is 3 and the frequency of anxiety is 2, then ERF = 3 x 0.6 + 2 x 0.4 = 2.6.

[0196] Reflection log keyword richness (RR):

[0197] Data source: Learner's reflection log text, TF-IDF algorithm.

[0198] Calculation logic: Count the number of depth words such as "reason analysis" and "improvement method" in the log N_{depth words} and the total number of words N_{total words}, calculate the proportion of depth words. Count the number of unique words in the log N_{unique words} and calculate its logarithm. Multiply the two to get RR. The formula is: RR = (N_{depth words} / N_{total words}) x log(N_{unique words}). For example, if the proportion of depth words is 20% and the number of unique words is 15, then RR = 0.2 x log(15) ≈ 0.236.

[0199] Normalization processing:

[0200] Calculation logic: Min-Max standardization, assuming ERF value range 0-5, RR value range 0-1. If ERF = 2.6, then ERF* = (2.6-0) / (5-0) = 0.52. RR = 0.236, then RR* = 0.236.

[0201] (3) Human-computer communication level, calculate target correction frequency and check behavior abnormality index, normalize the two as basic indicators.

[0202] Specifically, count the number of target modifications N_{modifications} and the total conversation duration T_{conversation} within the conversation duration, and calculate the target modification frequency by (GRF = N_{modifications} / T_{conversation}). Analyze the click heat map data to get the number of repeated clicks N_{repeated clicks}, the total number of clicks N_{total clicks}, the dwell time T_{dwell} and the preset threshold T_{threshold}, and calculate the check behavior abnormality index by (AC = (N_{repeated clicks} / N_{total clicks}) x (T_{dwell} / T_{threshold})). Normalize the two indicators to the range of 0-1.

[0203] Target modification frequency (GRF):

[0204] Data source: Conversation duration (T_{conversation}, such as 10 minutes), number of times the learner modifies the learning target (N_{modifications}, such as 3 times).

[0205] Calculation logic: Divide N_{corrected} by T_{session} to get the target corrected frequency, formula is: GRF = N_{corrected} / T_{session}, if N_{corrected} = 3, T_{session} = 10 minutes, then GRF = 3 / 10 = 0.3 times / minute.

[0206] Check abnormal behavior index (AC):

[0207] Data sources: Click heat map (repeat click times N_{repeat click}, total click times N_{total click}), answer area dwell time (T_{dwell}, such as 60 seconds), normal dwell time threshold (T_{threshold}, such as 20 seconds).

[0208] Calculation logic: First calculate the repeat click rate (N_{repeat click} / N_{total click}) and the dwell time exceeding coefficient (T_{dwell} / T_{threshold}), and multiply the two to get AC, formula is: AC = (N_{repeat click} / N_{total click}) x (T_{dwell} / T_{threshold}), if repeat click rate is 50%, dwell time exceeds threshold by 3 times, then AC = 0.5 x 3 = 1.5.

[0209] Normalization processing:

[0210] Calculation logic: Min-Max standardization is adopted, assuming that GRF takes value range 0-0.5 times / minute, AC takes value range 0-3. If GRF = 0.3, then GRF* = (0.3-0) / (0.5-0) = 0.6. AC = 1.5, then AC* = (1.5-0) / (3-0) = 0.5.

[0211] (4) Neural activity level, calculate prefrontal activation index, normalize this index as basic index.

[0212] Specifically, collect prefrontal theta power spectral density PSD_θ during task period and baseline value PSD_{baseline} during resting period, and calculate prefrontal activation index by (ACCA = PSD_θ / PSD_{baseline}).

[0213] Prefrontal activation index (ACCA):

[0214] Data sources: prefrontal theta wave (4-8 Hz) power spectral density (PSD_θ, such as 12 μV 2 / Hz) collected by EEG device, theta wave power spectral density (PSD_{baseline}, such as 5 μV 2 / Hz) in resting state.

[0215] Calculation logic: divide (PSD_ theta) by (PSD_{baseline}) to get the pre- ACCA = PSD_ theta / PSD_{baseline} (e.g. PSD_ theta = 12, PSD_{baseline} = 5, then ACCA = 12 / 5 = 2.4).

[0216] Normalization processing:

[0217] Calculation logic: Min-Max normalization is used, assuming that ACCA ranges from 0 to 4. If ACCA = 2.4, then ACCA* = (2.4-0) / (4-0) = 0.6.

[0218] All normalized indicators are input into a hierarchical attention network model in the art, and the indicator score corresponding to the meta-cognitive conflict dimension in the interval [0, 1] is calculated by the model.

[0219] Specifically, the above four basic indicators are input into the HAN model after normalization processing, and the indicator score corresponding to the meta-cognitive conflict dimension is dynamically calculated.

[0220] In one possible implementation, the large model is used to determine the weight coefficient corresponding to the preset dimension based on the indicator score corresponding to the preset dimension, and the weight coefficient corresponding to the preset dimension is determined based on the indicator score corresponding to the preset dimension, specifically including:

[0221] The subjective weight corresponding to the preset dimension is determined based on the indicator score corresponding to the preset dimension and the Delphi method, the objective weight corresponding to the preset dimension is determined based on the entropy weight method, and the weight coefficient corresponding to the preset dimension is determined based on the subjective weight and the objective weight corresponding to the preset dimension.

[0222] In the embodiments of the present application, the subjective weight is determined by the Delphi method: an expert group composed of educational psychologists and cognitive science experts is formed, and three rounds of questionnaires are designed. The initial weights of the knowledge conflict dimension, the logic conflict dimension and the meta-cognitive conflict dimension are determined through the first round of open questionnaire, such as the initial weights of the knowledge conflict dimension, the logic conflict dimension and the meta-cognitive conflict dimension are a = 0.4, b = 0.3, g = 0.3. The opinions of the experts in the first round are summarized, induced and counted in the last two rounds, and the revised opinions are fed back to the experts anonymously to correct the opinions, and the feedback is repeated until a consensus is reached, and the expert consensus weight a_{subjective}, b_{subjective}, g_{subjective} is obtained.

[0223] The objective weight is determined by the entropy weight method: 500 groups of learner experimental data covering K12 and higher education scenarios are collected, and the information entropy of each preset dimension is calculated:

[0224] E j j represents the information entropy of the preset dimension j, p ijis the standardized value of the ith sample in the jth dimension, and n is the number of samples. The objective weights of each preset dimension are calculated according to the information entropy: α objective, β objective, γ objective. The lower the entropy value, the higher the weight.

[0225] As an example, a higher proportion of subjective weights obtained by the Delphi method is given, such as 0.6, and the objective weights obtained by the entropy weight method are also included, such as 0.4. The final weight coefficients α, β, and γ can be calculated according to the weight coefficient = subjective weight × 0.6 + objective weight × 0.4, ensuring that the weights take into account both the objective data differentiation and the compliance with educational theory, and satisfy α + β + γ = 1.

[0226] In one possible implementation, the large model is used to determine the cognitive conflict level of the learner according to the indicator scores and weight coefficients corresponding to the preset dimensions, and the cognitive conflict level of the learner is determined according to the indicator scores and weight coefficients corresponding to the preset dimensions, specifically including:

[0227] Determine the cognitive conflict index according to the indicator scores and weight coefficients corresponding to the preset dimensions.

[0228] Obtain a plurality of cognitive conflict index ranges and a cognitive conflict level corresponding to each cognitive conflict index range.

[0229] Determine the cognitive conflict index range to which the cognitive conflict index belongs.

[0230] Determine the cognitive conflict level corresponding to the cognitive conflict index range to which the cognitive conflict index belongs as the cognitive conflict level of the learner.

[0231] In the embodiments of the present application, the cognitive conflict index is a quantitative result of the cognitive level of the learner, and the calculation formula of the cognitive conflict index is: CCI = α·KC + β·PC + γ·MC. Wherein, CCI represents the cognitive conflict index, KC represents the indicator score corresponding to the knowledge conflict dimension, PC represents the indicator score corresponding to the logical conflict dimension, MC represents the indicator score corresponding to the meta-cognitive conflict dimension, α represents the weight coefficient corresponding to the knowledge conflict dimension, β represents the weight coefficient corresponding to the logical conflict dimension, and γ represents the weight coefficient corresponding to the meta-cognitive conflict dimension. For example, KC = 0.5, PC = 0.6, MC = 0.4, α = 0.385, β = 0.306, and γ = 0.309, CCI = 0.5 × 0.385 + 0.6 × 0.306 + 0.4 × 0.309 ≈ 0.498.

[0232] In the embodiments of the present application, a plurality of cognitive conflict index ranges and cognitive conflict levels corresponding to each cognitive conflict index range can be preset in the system configuration file, such as: 0≤CCI<0.35 corresponds to a low cognitive conflict level, 0.35≤CCI<0.7 corresponds to a moderate cognitive conflict level, and 0.7≤CCI≤1 corresponds to a high cognitive conflict level. The description of each level is clear, such as the low conflict level: the cognitive contradiction is weak and does not deviate from the comfort zone. Different teaching strategies corresponding to different cognitive conflict levels can also be preset.

[0233] The cognitive conflict index range to which the CCI belongs can be determined by comparing the calculated CCI value with the preset cognitive conflict index range through a conditional judgment statement (such as if-elif-else) (such as CCI=0.498, which satisfies 0.35≤0.498<0.7), and the cognitive conflict index range to which the CCI belongs is determined to be the range corresponding to the moderate cognitive conflict level.

[0234] According to the mapping relationship between the range and the level, the level corresponding to the range is determined as the cognitive conflict level of the learner, which can be stored in the learner cognitive state database.

[0235] In one possible implementation, the large model is used to determine the target teaching strategy according to the corresponding cognitive conflict level of the learner, and the target teaching strategy is determined according to the corresponding cognitive conflict level of the learner, which specifically includes:

[0236] Obtaining a matching relationship between each cognitive conflict level and the corresponding teaching strategy preset in advance.

[0237] Based on the matching relationship, the teaching strategy corresponding to the cognitive conflict level corresponding to the learner is determined as the target teaching strategy.

[0238] A matching table of cognitive conflict levels and corresponding teaching strategies is preset in the system strategy library, for example: low cognitive conflict level corresponds to implicit guidance strategy, implicit guidance strategy includes pushing similar concept comparison questions, proposing opposite hypotheses for learners to verify, highlighting nodes in knowledge graph that have not been noticed, etc. can automatically retrieve similar concepts or counterexamples related to the current learning content in the knowledge graph, and generate comparison questions. Propose hypotheses opposite to the current knowledge logic to guide learners to analyze errors. Highlight the associated nodes in the knowledge graph that the learner has not noticed to reveal potential knowledge gaps and encourage learners to move out of their comfort zone. The medium cognitive conflict level corresponds to the step-by-step reasoning chain strategy, which includes generating a sequence of Socratic guidance questions, interactive animation display of logical chains, and providing reasoning step templates, etc. Large Language Model (LLM) can be used to generate a sequence of Socratic guidance questions, each question corresponding to a reasoning step. Extract core logical relationships from knowledge graphs and fact graphs to display the reasoning process in an interactive animation format. When detecting that the learner skips reasoning, automatically pop up a reasoning breakpoint completion prompt to help organize the logic. The high cognitive conflict level corresponds to the AR visualization combined task degradation strategy, which includes calling ARKit / ARCore 3D models, splitting multi-step problems into single-step problems, and changing short answer questions into multiple-choice questions, etc. ARKit / ARCore SDK can be called to load preset 3D models to convert abstract knowledge (such as chemical bonds and principles of buoyancy) into interactive and tangible content. Replace the original task with a low-difficulty subtask that maintains the core concept, such as changing a multi-step function application problem into a single-step function evaluation problem or an open-ended essay outline into a multiple-choice outline. Break down complex tasks into small sub-steps and provide input boxes and guidance prompts for each step to reduce cognitive load. As an example, the matching relationship is stored in JSON format, which supports dynamic updates.

[0239] In the embodiments of the present application, the Transformer model can be used to fuse multi-modal data in real time to output the CCI value, and perceive and define the conflict degree.

[0240] Multi-modal data acquisition and preprocessing:

[0241] The system collects the learner's behavior data (such as mouse clicks, keyboard inputs, gaze points), physiological data (such as eye movement trajectories, EEG signals, heart rate), and other environmental data (such as screen brightness, noise level) in real time. Each modality of data is cleaned, standardized, and time-aligned to ensure data quality.

[0242] Multi-modal data encoding and fusion:

[0243] Data from different modalities are converted into unified vector representations through embedding layers, and vector normalization processing is performed, such as encoding behavior data (e.g., mouse clicks, keyboard inputs, gaze points, and question frequencies) into time series features and encoding physiological data (e.g., brain wave changes) into frequency domain features. Subsequently, the self-attention mechanism of the Transformer is used to capture the relationship between modal data and achieve efficient fusion of cross-modal information.

[0244] Transformer model design and training:

[0245] Encoder structure: A multi-head self-attention mechanism is designed to fully exploit the relevance of multi-modal data and extract cognitive conflict features of learners.

[0246] Model input: Multi-modal data are encoded into continuous sequence inputs, each input containing the behavior, physiology, and environment state of the learner at a certain time point.

[0247] Output layer: The CCI value is output through a fully connected layer, reflecting the current cognitive conflict level of the learner.

[0248] Real-time inference and dynamic monitoring:

[0249] Real-time calculation: The Transformer model processes data streams in real time and outputs dynamically updated CCI values.

[0250] Threshold triggering: Set the threshold of cognitive conflict, when the CCI value exceeds the preset range, trigger the subsequent teaching intervention strategy.

[0251] Traditional cognitive conflict detection is mostly limited to a single knowledge dimension (such as concept errors) or relies on subjective questionnaires (such as Likert scales), and cannot quantitatively detect conflicts in logical reasoning and metacognition in real time. Based on the SOLO and Bloom theories, a three-level conflict model of knowledge-logic-metacognition is constructed and quantified, and multi-dimensional detection is achieved through the following technologies: Knowledge conflict (KC) uses dynamic knowledge graph analysis combined with semantic similarity to trigger knowledge conflict alerts. Logical conflict (PC) is matched in real time through a rule engine to break the cause-effect chain, combined with step correction frequency to activate logical conflict markers. Metacognition conflict (MC) uses the synergy of planning-execution bias rate and reflection frequency to capture monitoring failure features in behavior sequences through the Transformer.

[0252] The accuracy of multi-modal data fusion is improved. Related technical defects: single modal detection (such as EEG or eye movement only) is susceptible to interference. Most systems use fixed weights, ignoring individual differences. This application uses a hierarchical attention network (HAN) to achieve multi-modal weighted fusion: eye movement fixation entropy and EEG form cross-validation, and behavior data and text perplexity are time-aligned. Dynamic weight adjustment mechanism: calculate the information entropy of each dimension by mixing weights, ensure the compliance of educational theory by Delphi method (expert consensus), and dynamically adjust by entropy weight method (data driven).

[0253] The technical solution of the application multi-modal cognitive conflict quantification model: integrate neural signals, emotional data and behavior logs, dynamically calculate the CCI index (formula: CCI = a · KC + b · PC + g · MC) through the hierarchical attention network, and the weight coefficients (a, b, g) are determined by the Delphi method and entropy weight method, and the grading standard is low, medium and high conflict (threshold value is dynamically adjusted according to experimental data).

[0254] Hierarchical dynamic strategy adaptation mechanism: dynamically match multi-granularity strategies based on CCI value, trigger implicit guidance (such as analogy prompt) when CCI is low, generate step-by-step reasoning chain when CCI is medium, trigger AR visualization support (such as 3D chemical bond reorganization) when CCI is high, and degrade task complexity.

[0255] The range of CCI value is normalized to the interval [0, 1]. Low cognitive conflict (Low CCI): 0 ≤ CCI < 0.35. The conflict level is not high enough to effectively promote deep thinking and concept change. Learners may be in the comfort zone, not enough challenges, or have a superficial understanding of the problem. Moderate cognitive conflict: 0.35 ≤ CCI < 0.7: ideal conflict level within the recent development zone. Learners encounter meaningful challenges, and their cognitive structure is impacted but not collapsed, with the strongest learning motivation and thinking activity, which is the best teaching intervention window. High cognitive conflict: 0.7 ≤ CCI ≤ 1. The conflict level is too high, exceeding the cognitive load limit of the learner. It may cause frustration, anxiety and confusion (cognitive paralysis), and learning efficiency is reduced, and support needs to be provided immediately to prevent disengagement.

[0256] Low CCI triggers implicit guidance, not to directly solve the conflict, but to gradually "create" or "reveal" the conflict, push the learner out of the comfort zone, and trigger their thinking, rather than directly giving the answer. The goal is to stimulate curiosity and suggest potential knowledge gaps. At this stage, the system automatically retrieves similar concepts or problems in the knowledge graph to give hints in the form of questions, propose a contrary hypothesis, and let students think about their mistakes or highlight a node in the knowledge graph that is related to the student's current thinking but has not been noticed by the student.

[0257] The purpose of generating a step-by-step reasoning chain when CCI is in the middle is to provide a procedural scaffold in the "zone of proximal development" of the student's thinking. Help students sort out chaotic logic, demonstrate the correct thinking process and steps, guide them to build their own solutions, rather than directly tell the results. Promote the development of metacognitive strategies (such as planning, monitoring). Use large language models to generate a series of Socratic guiding questions, each question guiding a reasoning step, the system extracting logical chains from the knowledge graph and the logic graph, and showing each step of reasoning in the form of interactive animation.

[0258] When high CCI triggers AR visualization support and degrades task complexity, it reduces cognitive load and prevents learners from giving up due to excessive frustration. Abstract concepts are made concrete through AR visualization, and complex tasks are broken down into manageable subtasks through degradation, helping students establish a basic understanding and successful experience, and rebuild confidence. In the question bank, preset the metadata of the simplified version or decomposition steps for each question, and the system matches and calls through the rule engine. Integrate ARKit / ARCore SDK, preset a 3D model library, when high conflict is detected, call the corresponding model and send it to the display device through the API, automatically replace a multi-step calculation comprehensive question with a single-step calculation question with the same core concept, change an open-ended short answer question into a multiple-choice or fill-in-the-blank question, reduce the burden of expression, automatically decompose the task into smaller steps, and provide input boxes and prompts for each step.

[0259] Figure 2 The overall idea flowchart of the embodiments of the present application is as follows: Figure 2As shown, the learner's neural signals (such as error-related negativity, prefrontal cortex activation level), eye movement trajectories and other physiological data are collected by the EEG device, the eye tracker captures the eye movement trajectory, and the behavior data (answering steps, option click flow) and emotional data (confusion expression, frustration behavior) are synchronously acquired, and the real-time transmission and convergence of various data are realized through API interaction; then, the converged multi-modal data is subjected to data cleaning (removal of device failure abnormal data, text noise reduction, video face frame removal, etc.), to ensure data quality; then, the multi-modal fusion link is entered, and through the self-attention mechanism of the Transformer architecture, different modal data is encoded into a unified vector and cross-modal correlation is mined, laying a foundation for subsequent conflict analysis; then, through the conflict classification module, according to Biggs' SOLO theory and Bloom's cognitive domain target hierarchical structure, the cognitive conflict is subdivided into three categories: knowledge conflict, logic conflict, and meta-cognition conflict, and the cognitive conflict index (CCI) is calculated in combination with the multi-dimensional indicators of each conflict type (such as KC's concept connection error, PC's reasoning chain breakage, MC's target deviation degree); finally, based on the cognitive conflict index CCI (normalized to the [0, 1] interval), the dynamic adaptation of teaching strategies is realized, that is, low CCI triggers implicit guidance, medium CCI generates step-by-step reasoning chain, high CCI triggers AR visualization support and task degradation, forming a complete technical link of data acquisition, processing, fusion, conflict quantification, and strategy adaptation, solving the problems of weak heuristics and insufficient dynamic adaptability of existing teaching technologies.

[0260] In the embodiment, a cognitive level quantization device for dynamically adaptive teaching strategies of large models is also provided, which is used to realize the above-mentioned embodiments and preferred embodiments, and will not be described again. As used below, the term "module" can be a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware, or a combination of software and hardware is also possible and contemplated.

[0261] Figure 3 is a structural block diagram of the cognitive level quantization device for dynamically adaptive teaching strategies of large models in the embodiment of the application.

[0262] The embodiment provides a cognitive level quantization device for dynamically adaptive teaching strategies of large models, as Figure 3 As shown, comprising:

[0263] The first processing module 11 is configured to collect multi-modal data of the learner in real time. The multi-modal data is preprocessed to obtain preprocessed multi-modal data. The multi-modal data includes knowledge structure level data, emotional state level data, human-computer communication level data, and neural activity level data.

[0264] The second processing module 12 is configured to input the preprocessed multi-modal data into a pre-constructed large model to obtain a target teaching strategy corresponding to the learner; the large model is configured to calculate an index score corresponding to a preset dimension based on the preprocessed multi-modal data; determine a weight coefficient corresponding to the preset dimension based on the index score corresponding to the preset dimension; determine a cognitive conflict level corresponding to the learner according to the index score corresponding to the preset dimension and the weight coefficient, and determine the target teaching strategy according to the cognitive conflict level corresponding to the learner; and the preset dimension includes a knowledge conflict dimension, a logic conflict dimension and a meta-cognition conflict dimension.

[0265] In one possible implementation, the first processing module 11 includes:

[0266] The first processing unit is configured to take the collected answer step sequence data and answer option click stream data of the learner, the knowledge graph nodes visited by the learner and the connection relationship between the knowledge graph nodes, and the knowledge points not mastered by the learner as the knowledge structure level data.

[0267] The second processing unit is configured to take the collected confusion expression data of the learner, the mouse and / or touch screen movement trajectory data, click frequency data of the learner, keyboard typing speed data of the learner, and keyboard backspace key usage frequency data as the emotional state level data.

[0268] The third processing unit is configured to take the collected voice data of the problem-solving thought of the learner, the text data of the question asked by the learner, the voice analysis result obtained by analyzing the voice data, and the text analysis result obtained by analyzing the text data as the human-computer communication level data.

[0269] The fourth processing unit is configured to take the collected neurophysiological feature data of the learner within a preset time after the answer error as the neural activity level data.

[0270] In one possible implementation, the first processing module 11 further includes:

[0271] The fifth processing unit is configured to perform data cleaning, standardization and time sequence alignment on the multi-modal data to obtain preprocessed multi-modal data.

[0272] In one possible implementation, the large model dynamic self-adaptive teaching strategy cognitive level quantization device further includes a first sub-model unit.

[0273] The first sub-model unit is configured to calculate a basic index corresponding to the knowledge structure level data, the emotional state level data, the human-computer communication level data and the neural activity level data under each preset dimension.

[0274] The index score corresponding to each preset dimension is calculated according to the basic index corresponding to the knowledge structure level data, the sentiment state level data, the human-computer communication level data and the neural activity level data under each preset dimension.

[0275] In a possible implementation, the cognitive level quantification apparatus for dynamically adapting the teaching strategy of the large model further includes a second sub-model unit.

[0276] The second sub-model unit is configured to determine a subjective weight corresponding to each preset dimension based on the index score corresponding to the preset dimension and the Delphi method, determine an objective weight corresponding to each preset dimension based on the entropy weight method, and determine a weight coefficient corresponding to each preset dimension based on the subjective weight and the objective weight corresponding to the preset dimension.

[0277] In a possible implementation, the cognitive level quantification apparatus for dynamically adapting the teaching strategy of the large model further includes a third sub-model unit.

[0278] The third sub-model unit is configured to determine a cognitive conflict index according to the index score corresponding to each preset dimension and the weight coefficient.

[0279] A plurality of cognitive conflict index ranges and a cognitive conflict level corresponding to each cognitive conflict index range are obtained in advance.

[0280] A cognitive conflict index range to which the cognitive conflict index belongs is determined.

[0281] A cognitive conflict level corresponding to the cognitive conflict index range to which the cognitive conflict index belongs is determined as the cognitive conflict level corresponding to the learner.

[0282] In a possible implementation, the cognitive level quantification apparatus for dynamically adapting the teaching strategy of the large model further includes a fourth sub-model unit.

[0283] The fourth sub-model unit is configured to obtain a matching relationship between each cognitive conflict level and a corresponding teaching strategy set in advance.

[0284] Based on the matching relationship, a teaching strategy corresponding to the cognitive conflict level corresponding to the learner is determined as a target teaching strategy.

[0285] The cognitive level quantification apparatus for dynamically adapting the teaching strategy of the large model in the embodiment is presented in the form of a functional unit. The unit herein refers to an ASIC circuit, a processor and a memory executing one or more software or fixed programs, and / or other devices that can provide the above functions.

[0286] Further function descriptions of the above modules and units are the same as those of the above corresponding embodiments, and will not be described here.

[0287] The embodiment of the present application further provides a computer device having the aboveFigure 3 The device shown is a cognitive level quantification device for a large-scale dynamic adaptive teaching strategy.

[0288] Please see Figure 4 , Figure 4 This is a schematic diagram of the hardware structure of a computer device according to an embodiment of this application, such as... Figure 4 As shown, the computer device includes one or more processors 10, memory 20, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as display devices coupled to the interfaces). In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 4 Take a processor 10 as an example.

[0289] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GDA), or any combination thereof.

[0290] The memory 20 stores instructions executable by at least one processor 10 to cause at least one processor 10 to perform the method shown in the above embodiments.

[0291] The memory 20 can include a program storage area and a data storage area. The program storage area can store an operating system, application programs required for at least one function. The data storage area can store data created by the use of the computer device according to the presentation of a small program landing page, and the like. In addition, the memory 20 can include a high-speed random access memory, and can also include a non-transitory memory such as at least one disk storage device, a flash memory device, or other non-transitory solid state memory device. In some alternative embodiments, the memory 20 can optionally include memory that is remotely located with respect to the processor 10, and these remote memories can be connected to the computer device through a network. Examples of the network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof. The memory 20 can include a volatile memory such as a random access memory. The memory can also include a non-volatile memory such as a flash memory, a hard disk, or a solid state disk. The memory 20 can also include a combination of the above-mentioned types of memory.

[0292] The computer device also includes a communication interface 30 for the computer device to communicate with other devices or communication networks.

[0293] The embodiments of the present application also provide a computer readable storage medium. The above-mentioned method according to the embodiments of the present application can be implemented in hardware, firmware, or recorded in a storage medium, or implemented through computer code originally stored in a remote storage medium or non-transitory machine readable storage medium and downloaded to a local storage medium, so that the method described herein can be processed by such software on a storage medium using a general purpose computer, a special purpose processor, or programmable or special purpose hardware. The storage medium can be a disk, an optical disk, a read-only memory, a random access memory, a flash memory, a hard disk, or a solid state disk, etc. Further, the storage medium can also include a combination of the above-mentioned types of memory. It can be understood that the computer, the processor, the microprocessor controller, or the programmable hardware includes a storage component that can store or receive software or computer code, when the software or computer code is accessed and executed by the computer, the processor, or the hardware, the method shown in the above embodiments is implemented.

[0294] Part of the present application can be applied as a computer program product, for example, computer program instructions, when executed by a computer, through the operation of the computer, can invoke or provide methods and / or technical solutions according to the present application. Those skilled in the art should understand that the form of computer program instructions in computer readable medium includes but is not limited to source files, executable files, installation package files and the like, and accordingly, the way of computer program instructions executed by computer includes but is not limited to: the computer directly executes the instructions, or the computer compiles the instructions and then executes the corresponding compiled program, or the computer reads and executes the instructions, or the computer reads and installs the instructions and then executes the corresponding installed program. Here, the computer readable medium can be any available computer readable storage medium or communication medium accessible to the computer.

[0295] Although the embodiments of the present application are described in conjunction with the drawings, various modifications and changes can be made by those skilled in the art without departing from the spirit and scope of the present application, and such modifications and changes fall within the scope defined by the appended claims.

Claims

1. A cognitive level quantification method for a large model dynamic adaptive teaching strategy, characterized in that, The method comprises: Real-time acquisition of multi-modal data of a learner; preprocessing of the multi-modal data to obtain preprocessed multi-modal data; the multi-modal data comprises knowledge structure level data, emotional state level data, human-machine communication level data and neural activity level data; Inputting the preprocessed multi-modal data into a pre-constructed large model to obtain a target teaching strategy corresponding to the learner; the large model is used to calculate an index score corresponding to a preset dimension based on the preprocessed multi-modal data; determine the weight coefficient corresponding to the preset dimension based on the index score corresponding to the preset dimension; determine the cognitive conflict level corresponding to the learner according to the index score and weight coefficient corresponding to the preset dimension, and determine the target teaching strategy according to the cognitive conflict level corresponding to the learner; the preset dimension includes a knowledge conflict dimension, a logic conflict dimension and a meta-cognition conflict dimension.

2. The method of claim 1, wherein, The real-time acquisition of multi-modal data of a learner comprises: The answer step sequence data and answer option click stream data of the learner, the knowledge graph nodes accessed by the learner and the connection relationship between the knowledge graph nodes, and the knowledge points not mastered by the learner are taken as the knowledge structure level data; The confusion expression data of the learner, the mouse and / or touch screen movement trajectory data, click frequency data of the learner, keyboard typing speed data, and keyboard backspace key usage frequency data are taken as the emotional state level data; The speech data of the learner's problem solving thought, the text data of the learner's follow-up question, the speech analysis result obtained by analyzing the speech data, and the text analysis result obtained by analyzing the text data are taken as the human-machine communication level data; The neurophysiological characteristic data of the learner within a preset time after an answer error is taken as the neural activity level data.

3. The method of claim 1, wherein, The preprocessing of the multi-modal data to obtain preprocessed multi-modal data comprises: Data cleaning, standardization and time sequence alignment are performed on the multi-modal data to obtain preprocessed multi-modal data.

4. The method of claim 1, wherein, The calculation of the index score corresponding to the preset dimension based on the preprocessed multi-modal data comprises: Calculate the basic index corresponding to the knowledge structure level data, the emotional state level data, the human-machine communication level data and the neural activity level data under each preset dimension respectively; According to the basic index corresponding to the knowledge structure level data, the emotional state level data, the human-machine communication level data and the neural activity level data under each preset dimension respectively, the index score corresponding to each preset dimension is calculated.

5. The method of claim 1, wherein, The determination of the weight coefficient corresponding to the preset dimension based on the index score corresponding to the preset dimension comprises: Determine the subjective weight corresponding to the preset dimension based on the index score corresponding to the preset dimension and the Delphi method, determine the objective weight corresponding to the preset dimension based on the entropy weight method, and determine the weight coefficient corresponding to the preset dimension based on the subjective weight and the objective weight corresponding to the preset dimension.

6. The method of claim 1, wherein, The cognitive conflict level corresponding to the learner is determined according to the index score corresponding to the preset dimension and the weight coefficient, and the method comprises the steps of: determining a cognitive conflict index according to the index score corresponding to the preset dimension and the weight coefficient; obtaining a plurality of cognitive conflict index ranges and a cognitive conflict level corresponding to each cognitive conflict index range; determining the cognitive conflict index range to which the cognitive conflict index belongs; determining the cognitive conflict level corresponding to the cognitive conflict index range to which the cognitive conflict index belongs as the cognitive conflict level corresponding to the learner.

7. The method of claim 1, wherein, The target teaching strategy is determined according to the cognitive conflict level corresponding to the learner, and the method comprises the steps of: obtaining a matching relationship between each cognitive conflict level and a corresponding teaching strategy set in advance; based on the matching relationship, determining the teaching strategy corresponding to the cognitive conflict level corresponding to the learner as the target teaching strategy. 8.A device for quantifying a cognitive level of a dynamic self-adaptive teaching strategy of a large model, characterized in that, The device comprises: a first processing module configured to collect multi-modal data of a learner in real time, and to obtain pre-processed multi-modal data by preprocessing the multi-modal data; the multi-modal data comprises knowledge structure level data, emotional state level data, human-computer communication level data and neural activity level data; a second processing module configured to input the pre-processed multi-modal data into a pre-constructed large model to obtain a target teaching strategy corresponding to the learner; the large model is configured to calculate an index score corresponding to a preset dimension based on the pre-processed multi-modal data, to determine a weight coefficient corresponding to the preset dimension based on the index score corresponding to the preset dimension, to determine a cognitive conflict level corresponding to the learner according to the index score corresponding to the preset dimension and the weight coefficient, and to determine a target teaching strategy according to the cognitive conflict level corresponding to the learner; the preset dimension comprises a knowledge conflict dimension, a logic conflict dimension and a meta-cognition conflict dimension.

9. A computer device, comprising: comprise: a memory and a processor, which are communicatively connected, the memory stores computer instructions, and the processor executes the computer instructions to perform the method in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer instructions, and the computer instructions are used to make a computer execute the method in any one of claims 1 to 7.

Images