AI-based learning question personalized recommendation method

Abstract

The invention relates to the technical field of artificial intelligence, and discloses an AI-based learning question personalized recommendation method, which comprises the following steps of: 1, acquiring eye movement track data in an English learning process through a visual sensor, acquiring voice stream data through audio acquisition equipment, and capturing touch behavior sequence data through an interactive recording device; step 2, calculating a reading glancing path curvature based on the eye movement trajectory data, detecting a spoken language non-natural pause interval based on the voice stream data, and counting a writing modification frequency based on the touch behavior sequence data; and step 3, curvature of a reading glancing path is carried out. The technical scheme of multi-modal real-time perception and dynamic state modeling is adopted, eye movement tracks, voice streams and touch behavior data are captured in real time through the visual sensor, the audio acquisition device and the interactive recording device, dynamic state vectors are generated in combination with a reinforcement learning strategy model, and the technical effect of accurately monitoring instantaneous cognitive performance changes is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence, and particularly to a personalized learning question recommendation method based on AI. Background Art

[0002] In the field of English subject education, traditional personalized learning systems usually construct a knowledge mastery model based on learners' historical performance data and push questions through collaborative filtering or content recommendation algorithms. Existing technologies mostly focus on knowledge point correlation analysis and difficulty progression control, and can achieve basic-level adaptive recommendations, but there are significant limitations in practical applications.

[0003] For example, the Chinese invention application with the publication number CN116089705B discloses a personalized learning question recommendation method and system based on AI. The method includes: an information collection module obtains the information of students and transports the information to an information evaluation module; the information evaluation module extracts the school years of each student and the answers in the historical answer sheets of different subjects of each student in the corresponding school years.

[0004] For example, the Chinese invention application with the publication number CN112232610B discloses a personalized question recommendation method and system using a machine learning model. The method includes: collecting relevant questions of students, establishing a question bank, using a machine learning model to perform multi-objective estimation on each question in the question bank, scoring the questions in the question bank according to the estimated objectives and the contribution value of each question to the total score of the students, and selecting the first number of questions according to the scores to establish a first question set.

[0005] Deficiencies of the above patents: Existing AI recommendation systems mainly rely on historical learning data to construct a static ability model and lack the ability to monitor the real-time cognitive state of learners. When students' cognitive efficiency decreases due to instantaneous factors such as distraction, mood swings, or mental fatigue, the system still recommends questions according to the preset ability level, resulting in incorrect ability assessment and question matching. And this kind of perception defect will trigger a vicious cycle: the misjudged learning portrait leads to the disconnection between the recommended content and the real cognitive state, further aggravating the learners' sense of frustration and cognitive load, and ultimately causing a continuous decline in learning efficiency. Especially in the long-term learning scenario, this problem will lead to a significant reduction in the credibility and practicality of the recommendation system.

[0006] When the current system detects an abnormal learning state, it generally adopts a single-dimensional mechanical adjustment strategy and fails to establish a dynamic response system with multi-parameter coordination. The specific manifestations are as follows: Through coarse-grained intervention means such as simply reducing the difficulty or pausing learning, it is impossible to perform multi-dimensional strategy combinations according to individual learning goals, the integrity of the knowledge system, and the real-time psychological state. This rigid adjustment mode leads to strategy conflicts between short-term intervention and long-term teaching plans, resulting in fragmentation of knowledge coverage and uncontrollability of the learning progress. Especially in time-sensitive scenarios such as exam preparation sprints, this defect will directly weaken the core value of the personalized recommendation system and lead to systematic deviations in the learning path.

[0007] To this end, the present invention proposes an AI-based personalized recommendation method for learning questions to solve the above-mentioned problems. Summary of the Invention

[0008] In view of the deficiencies of the prior art, the present invention provides an AI-based personalized recommendation method for learning questions to solve the problems raised in the above background technology.

[0009] To achieve the above objectives, the present invention is realized through the following technical solutions: An AI-based personalized recommendation method for learning questions, including: Step 1, collect eye movement trajectory data during the English learning process through a visual sensor, obtain speech stream data through an audio acquisition device, and capture touch behavior sequence data through an interaction recording device; Step 2, calculate the curvature of the reading saccade path based on the eye movement trajectory data, detect the unnatural pause intervals in the spoken language based on the speech stream data, and count the writing modification frequency based on the touch behavior sequence data; Step 3, perform feature fusion on the curvature of the reading saccade path, the unnatural pause intervals in the spoken language, and the writing modification frequency with historical learning data to generate a dynamic state vector including grammar processing speed, cross-cultural understanding degree, and anxiety fluctuation coefficient; Step 4, input the dynamic state vector into the reinforcement learning strategy model, and combine the association strength of grammar knowledge points in the knowledge graph and the teaching progress constraint conditions to output a recommendation strategy including the question difficulty manifold, the proportion of question types distribution, and the feedback interval parameter; Step 5, screen target questions according to the question difficulty manifold and the proportion of question types distribution in the recommendation strategy, and attach a grammar structure hint box and cultural background notes to the target questions; Step 6, update the dynamic state vector based on the real-time interaction data of the target questions, and feedback the update result to the reinforcement learning strategy model for strategy optimization.

[0010] Preferably, in the above Step 1, the data collection and preprocessing further include: Sub-step 1.1, capture the eye movement coordinate sequence through a visual sensor at a sampling rate of 200Hz , calculate the saccade path curvature after smoothing the trajectory using Kalman filter : , Among them, is the first derivative, is the second derivative, is the angular acceleration in the direction, is the angular velocity in the direction, is the sum of the squares of the velocities; Sub-step 1.2, perform phoneme boundary detection on the speech stream obtained by the audio acquisition device. When the duration of the silent segment satisfies: , it is determined as an unnatural pause, Among them, is the average pause duration, is the standard deviation, is the fundamental frequency, is the fundamental frequency change rate threshold, is the differential time; Sub-step 1.3, extract the erasure action event from the touch behavior sequence. When the erasure displacement satisfies: and , it is counted as an effective modification, Among them, is the width of the writing area, is the displacement coefficient, is the touch pressure, is the touch pressure change rate threshold, is the differential time.

[0011] Preferably, in the said step 2, the feature extraction and calculation further include: Sub-step 2.1, perform saccade segment segmentation on the eye movement trajectory data obtained in step 1. When the distance between adjacent fixation points satisfies: , it is determined as an effective saccade segment, Among them, is the standard deviation of the fixation point distance, is the curvature change threshold; Sub-step 2.2, perform phoneme-level segmentation on the speech stream data obtained in step 1, and calculate the distribution of the silent segment duration : , and mark it as an unnatural pause, Among them, is the average fundamental frequency of the sentence, is the duration of the i-th silent segment, is the total number of silent segments in the speech stream, is the mutation coefficient, is the instantaneous change in the fundamental frequency of speech; Sub-step 2.3: Classify events for the touch behavior sequence obtained in step 1 and count the effective modification frequency: When and , wherein, is the effective modification frequency, is the number of erasure times that meet the conditions, is the total duration of the writing task, is the displacement ratio coefficient, is the minimum touch pressure threshold, is the horizontal displacement of the touch erasure action, is the peak touch pressure, is the width of the writing area.

[0012] Preferably, in the said step 3, feature fusion and state generation further include: Sub-step 3.1: Dynamically normalize the reading saccade path curvature C obtained in step 2: , wherein, is the normalized saccade path curvature, is the mean of the historical curvatures, is the standard deviation; Sub-step 3.2: Align the temporal sequence of the spoken non-natural pause interval obtained in step 2 with the touch modification frequency : , When the alignment error perform feature splicing, wherein, is the temporal alignment offset of the multimodal data, is the duration of the i-th detected non-natural pause, is the touch modification frequency after the time offset , is the time window threshold; Sub-step 3.3: Weight the fused features using the attention mechanism: , wherein, is the attention weight of the k-th feature, is the query vector, is the transposed key vector, is the dimension of the feature vector, When the attention score activates the cross-cultural understanding degree calculation.

[0013] Preferably, in step 4, the strategy generation and optimization further includes: Sub-step 4.1, query the knowledge graph based on the syntax processing speed in the dynamic state vector : When , retain the associated edge, where, is the quantization value of the knowledge point association strength, is the cosine similarity between knowledge point vectors, is the association strength threshold; Sub-step 4.2, construct a multi-objective optimization function to generate an initial strategy: , where, , is the anxiety fluctuation coefficient, is the maximum anxiety value, is the dynamic weight coefficient, is the knowledge point coverage gain term, is the strategy utility function, is the cognitive comfort compensation term; Sub-step 4.3, perform the teaching progress constraint check: When , increase the question type distribution weight, where, is the progress weight of knowledge point i, is the question difficulty, is the minimum coverage coefficient, is the standard learning duration benchmark value, is the total number of knowledge points to be covered on the current day.

[0014] Preferably, in step 5, the question screening and enhancement further includes: Sub-step 5.1, construct a candidate set according to the question difficulty manifold in the recommendation strategy : , where, is the set of candidate questions, is the question 's difficulty value, is the i-th question object in the question bank, is the difficulty tolerance, is the target difficulty value; Sub-step 5.2, allocate the number of questions according to the distribution ratio P of question types: When , truncate in descending order of priority, wherein, is the total number of preset questions, is the number of questions to be selected for the j-th question type, is the distribution ratio of the j-th question type, is the maximum number of questions recommended at one time threshold, is the total number of question type categories; Sub-step 5.3, add dynamic annotations to the screened questions: Otherwise when , wherein, is the grammar processing speed threshold, is the complete grammar hint, is the core point hint, is the dynamic hint type.

[0015] Preferably, in the said step 6, the status update and strategy optimization further include: Sub-step 6.1, calculate the response time deviation based on the real-time interaction data of the target questions: When , trigger status correction, wherein, is the response time deviation, is the response deviation threshold, is the expected response time based on historical data; Sub-step 6.2, update the grammar processing speed in the dynamic status vector: , wherein, is the updated grammar processing speed, is the grammar processing speed before update, is the historical weight coefficient, is the number of correct answers, is the total response time; Sub-step 6.3, calculate the strategy reward value and update the reinforcement learning model: When , trigger strategy reset, wherein, is the comprehensive strategy reward value, is the progress reward item, is the comfort reward item, is the progress reward weight, is the comfort reward weight, is the minimum reward threshold.

[0016] Preferably, in step 2, the reading saccade path curvature is obtained by calculating the angular change rate of the line connecting adjacent fixation points; In step 2, the detection of the oral unnatural pause interval adopts a combined criterion of phoneme boundary alignment and silent segment duration.

[0017] Preferably, in step 4, the knowledge graph node weight is dynamically adjusted according to the grammar knowledge point association strength and the error pattern propagation path; In step 4, the reinforcement learning policy model adopts a dual-objective reward function, including a knowledge point coverage gain term and a cognitive comfort compensation term.

[0018] Preferably, in step 5, the display duration of the grammar structure hint box is inversely proportional to the current cognitive load characteristics; In step 5, the detail level of the cultural background annotation is negatively correlated with the cross-cultural understanding degree.

[0019] The present invention provides an AI-based personalized learning question recommendation method. It has the following beneficial effects: 1. The present invention adopts a multi-modal real-time perception and dynamic state modeling technical solution. By using visual sensors, audio acquisition devices and interaction recording devices to capture eye movement trajectories, speech streams and touch behavior data in real time, and combining with a reinforcement learning policy model to generate a dynamic state vector, it achieves the technical effect of accurately monitoring the instantaneous cognitive efficiency change. Compared with the evaluation methods that rely on static historical data in the prior art, it solves the problem of misjudgment of abilities caused by instantaneous factors such as attention dispersion and mood fluctuations, enables the question recommendation to match the real cognitive state in real time, and avoids the decline in learning efficiency caused by the disconnection between the recommended content and the cognitive load.

[0020] 2. The present invention constructs an elastic recommendation system with multi-dimensional constraint coordination. By integrating the knowledge graph association strength, teaching progress constraints and real-time psychological state parameters, it designs a strategy generation mechanism for multi-objective optimization, achieving the technical effect of dynamically balancing short-term intervention and long-term teaching planning. Compared with the single-dimensional mechanical adjustment strategies in the prior art, it solves the problems of strategy conflicts and knowledge fragmentation caused by rigid intervention, realizes the adaptive coordination of difficulty, question type and feedback rhythm while ensuring the integrity of knowledge point coverage, and significantly improves the reliability of path planning in time-sensitive scenarios such as exam preparation. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 is the flowchart of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0022] To enable those skilled in the art to understand the solution of the present invention, the following will clearly and completely describe the technical solution in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments in the present invention, other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of the present invention.

[0023] The following will describe the present invention in detail with reference to the accompanying drawings: Embodiment

[0024] Please refer to the attached Figure 1 , the embodiment of the present invention provides a personalized recommendation method for learning questions based on AI, including: Step 1, collect eye movement trajectory data during the English learning process through a visual sensor, obtain speech stream data through an audio acquisition device, and capture touch behavior sequence data through an interaction recording device; Sub-step 1.1, capture the eye movement coordinate sequence through a visual sensor at a sampling rate of 200Hz , calculate the saccade path curvature after smoothing the trajectory using Kalman filtering : , wherein, is the first derivative, is the second derivative, is the angular acceleration in the direction, is the angular velocity in the direction, and is the sum of the squares of the velocities; Sub-step 1.2, perform phoneme boundary detection on the speech stream obtained by the audio acquisition device. When the duration of the silent segment satisfies: , it is determined as an unnatural pause, wherein, is the average pause duration, is the standard deviation, is the fundamental frequency, is the fundamental frequency change rate threshold, is the differential time; Sub-step 1.3, extract the erasure action event from the touch behavior sequence. When the erasure displacement satisfies: and , it is counted as an effective modification, wherein, is the width of the writing area, is the displacement coefficient, is the touch pressure, is the threshold of the touch pressure change rate, is the differential time.

[0025] Step 2: Calculate the curvature of the reading saccade path based on the eye movement trajectory data, detect the unnatural pause intervals of the spoken language based on the speech stream data, and count the writing modification frequency based on the touch behavior sequence data; In Step 2, the curvature of the reading saccade path is obtained by calculating the angular change rate of the line connecting adjacent fixation points; In Step 2, the detection of the unnatural pause intervals of the spoken language adopts the combined criterion of phoneme boundary alignment and silent segment duration; Sub-step 2.1: Perform saccade segment segmentation on the eye movement trajectory data obtained in Step 1. When the distance between adjacent fixation points satisfies: it is determined as a valid saccade segment, where, is the standard deviation of the fixation point distance, is the curvature change threshold; Sub-step 2.2: Perform phoneme-level segmentation on the speech stream data obtained in Step 1, and calculate the silent segment duration distribution: is marked as an unnatural pause, where, is the average fundamental frequency of the sentence, is the duration of the i-th silent segment, is the total number of silent segments in the speech stream, is the mutation coefficient, is the instantaneous change amount of the speech fundamental frequency; Sub-step 2.3: Classify the events of the touch behavior sequence obtained in Step 1, and count the effective modification frequency: When and , where, is the effective modification frequency, is the number of erasure times that meet the conditions, is the total duration of the writing task, is the displacement proportionality coefficient, is the minimum touch pressure threshold, is the horizontal displacement amount of the touch erasure action, is the touch pressure peak value, is the width of the writing area.

[0026] Step 3: Perform feature fusion on the reading saccade path curvature, oral unnatural pause intervals, and writing modification frequencies with historical learning data to generate a dynamic state vector containing grammar processing speed, cross-cultural understanding degree, and anxiety fluctuation coefficient; Sub-step 3.1: Dynamically normalize the reading saccade path curvature C obtained in Step 2: , where, is the normalized saccade path curvature, is the historical curvature mean, is the standard deviation; Sub-step 3.2: Align the oral unnatural pause intervals obtained in Step 2 with the touch modification frequency in time series: , When the alignment error perform feature splicing, where, is the time alignment offset of the multi-modal data, is the detected duration of the i-th unnatural pause, is the time offset of the touch modification frequency, is the time window threshold; Sub-step 3.3: Use the attention mechanism to weight the fused features: , where, is the attention weight of the k-th feature, is the query vector, is the transposed key vector, is the dimension of the feature vector, When the attention score activate the cross-cultural understanding degree calculation; Step 4: Input the dynamic state vector into the reinforcement learning policy model, and combine the association strength of grammar knowledge points in the knowledge graph and the teaching progress constraint conditions to output a recommended policy containing the question difficulty manifold, question type distribution ratio, and feedback interval parameters; In Step 4, the weights of the knowledge graph nodes are dynamically adjusted according to the association strength of grammar knowledge points and the error pattern propagation path; In Step 4, the reinforcement learning policy model adopts a dual-objective reward function, including a knowledge point coverage gain term and a cognitive comfort compensation term; Sub-step 4.1: Query the knowledge graph based on the grammar processing speed in the dynamic state vector: When , retain the associated edges, wherein, is the quantization value of the knowledge point association strength, is the cosine similarity between knowledge point vectors, is the association strength threshold; Sub-step 4.2, construct a multi-objective optimization function to generate an initial strategy: , wherein, , is the anxiety fluctuation coefficient, is the maximum anxiety value, is the dynamic weight coefficient, is the knowledge point coverage gain term, is the strategy utility function, is the cognitive comfort compensation term; Sub-step 4.3, perform teaching progress constraint verification: When , increase the question type distribution weight, wherein, is the progress weight of knowledge point i, is the question difficulty, is the minimum coverage coefficient, is the standard learning duration benchmark value, is the total number of knowledge points to be covered on the current day; Step 5, screen target questions according to the question difficulty manifold and question type distribution ratio in the recommended strategy, and attach a grammar structure hint box and cultural background notes to the target questions; In Step 5, the display duration of the grammar structure hint box is inversely proportional to the current cognitive load characteristics; In Step 5, the detail level of the cultural background notes is negatively correlated with the cross-cultural understanding degree; Sub-step 5.1, construct a candidate set according to the question difficulty manifold in the recommended strategy : , wherein, is the candidate question set, is the question 's difficulty value, is the i-th question object in the question bank, is the difficulty tolerance, is the target difficulty value; Sub-step 5.2, allocate the number of questions according to the question type distribution ratio P: When , truncate in descending order of priority, wherein, is the total number of preset questions, is the number of questions to be selected for the j-th question type, is the distribution ratio of the j-th question type, is the maximum number of questions recommended at one time threshold, is the total number of question type categories; Sub-step 5.3, add dynamic annotations to the screened questions: Otherwise, when , wherein, is the grammar processing speed threshold, is the complete grammar hint, is the core point hint, is the dynamic hint type; Step 6, update the dynamic state vector based on the real-time interaction data of the target question, and feedback the update result to the reinforcement learning policy model for policy optimization; Sub-step 6.1, calculate the response time deviation based on the real-time interaction data of the target question: When , trigger state correction, wherein, is the response time deviation, is the response deviation threshold, is the expected response time based on historical data; Sub-step 6.2, update the grammar processing speed in the dynamic state vector: , wherein, is the updated grammar processing speed, is the grammar processing speed before update, is the historical weight coefficient, is the number of correct answers, is the total response time; Sub-step 6.3, calculate the policy reward value and update the reinforcement learning model: When , trigger policy reset, wherein, is the comprehensive policy reward value, is the progress reward item, is the comfort reward item, is the progress reward weight, is the comfort reward weight, is the minimum reward threshold.

[0027] Step 1 constructs a three-dimensional cognitive state perception network by integrating visual, auditory, and touch three-modal sensing technologies. The Kalman filtering process of eye movement trajectories effectively eliminates environmental noise interference and ensures the accurate extraction of saccade path features; the phoneme-level segmentation of speech streams combined with fundamental frequency mutation detection can capture implicit anxiety features that cannot be recognized by traditional methods; the event classification algorithm for touch behavior accurately distinguishes valid modifications from misoperations through double-threshold determination of displacement and pressure. The collaborative acquisition mechanism of multi-source heterogeneous data breaks through the bottleneck of limited data dimensions of a single sensor and provides raw data with high signal-to-noise ratio for subsequent cognitive state modeling.

[0028] Step 2 establishes a feature calculation system based on cognitive mechanics. The dynamic threshold segmentation algorithm for saccade path curvature can automatically identify abnormal saccade patterns caused by distracted attention; the multi-dimensional joint criterion for speech silent segments realizes the quantitative assessment of mental workload; the spatio-temporal correlation statistical model of touch modification frequency can effectively characterize the intensity of cognitive conflict during the writing process. The feature extraction method integrating spatio-temporal characteristics significantly improves the sensitivity and specificity of instantaneous cognitive state detection.

[0029] Step 3 designs a multi-modal alignment mechanism with memory enhancement ability. The dynamic normalization process effectively eliminates feature offsets caused by individual differences through comparative analysis of historical reference values; the time series alignment algorithm uses least squares optimization to solve the fusion problem of mismatched sampling rates of multi-source data; the attention-weighted feature selection mechanism can automatically focus on key cognitive dimensions according to the current learning stage. The adaptive feature fusion framework enables the system to construct a dynamic vector reflecting the real cognitive state and provides a decision basis for personalized recommendation.

[0030] Step 4 constructs a deep coupling mechanism between teaching rules and machine learning. The dynamic pruning algorithm of the knowledge graph controls through the correlation strength threshold to ensure that the recommended content conforms to the logical structure of the knowledge system; the dual-objective optimization function innovatively introduces an anxiety coefficient adjustment factor to achieve the dynamic balance between teaching progress and psychological endurance; the progress constraint verification module adopts a coverage feedback mechanism to effectively prevent the omission of knowledge points. The strategy generation mode integrating domain knowledge makes the recommendation system have both educational scientificity and technological intelligence.

[0031] Step 5 realizes the intelligent matching of cognitive features and teaching resources. The dynamic difficulty manifold construction algorithm adjusts through the tolerance threshold to ensure the challenge of the questions and avoid cognitive overload; the priority truncation strategy for question type distribution ensures the integrity of knowledge coverage while controlling cognitive load; the hierarchical prompt system provides appropriate learning scaffolds according to the real-time ability assessment results. The elastic resource presentation method significantly improves the naturalness of human-computer interaction and teaching effectiveness.

[0032] Step 6: Establish a reinforcement learning mechanism with self-evolution ability. The real-time monitoring module for response time deviation can quickly identify the risk of policy failure; the sliding average update algorithm for grammar processing speed takes into account the continuity and suddenness of state changes; the multi-dimensional reward function realizes the intelligent exploration and utilization of the policy space through the threshold trigger mechanism. The closed-loop optimization architecture enables the recommendation system to have the ability of continuous evolution, ensuring the optimization of long-term teaching effects.

[0033] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An AI-based personalized recommendation method for learning questions, characterized in that, Including: Step 1: Collect eye movement trajectory data during the English learning process through a visual sensor, obtain speech stream data through an audio acquisition device, and capture touch behavior sequence data through an interaction recording device; Step 2: Calculate the curvature of the reading saccade path based on the eye movement trajectory data, detect unnatural pauses in speech based on the speech stream data, and count the writing modification frequency based on the touch behavior sequence data; Step 3: Perform feature fusion on the curvature of the reading saccade path, unnatural pauses in speech, and writing modification frequency with historical learning data to generate a dynamic state vector including grammar processing speed, cross-cultural understanding degree, and anxiety fluctuation coefficient; Step 4: Input the dynamic state vector into a reinforcement learning policy model, and combine the association strength of grammar knowledge points in the knowledge graph and the teaching progress constraint conditions to output a recommendation strategy including the question difficulty manifold, question type distribution ratio, and feedback interval parameters; Step 5: Screen target questions according to the question difficulty manifold and question type distribution ratio in the recommendation strategy, and attach grammar structure hint boxes and cultural background annotations to the target questions; Step 6: Update the dynamic state vector based on the real-time interaction data of the target questions, and feedback the update result to the reinforcement learning policy model for policy optimization.

2. The personalized recommendation method for learning questions based on AI according to claim 1, wherein In Step 1, the data collection and preprocessing further include: Sub-step 1.1, capture the eye movement coordinate sequence at a sampling rate of 200 Hz through a visual sensor , calculate the saccade path curvature after smoothing the trajectory using Kalman filtering : ， Among them, is the first derivative, is the second derivative, is the angular acceleration in the direction, is the angular velocity in the direction, is the sum of the squares of the velocities; Sub-step 1.2: Perform phoneme boundary detection on the speech stream obtained by the audio acquisition device. When the duration of the silent segment satisfies: , it is determined as a non-natural pause, Among them, is the average pause duration, is the standard deviation, is the fundamental frequency, is the fundamental frequency change rate threshold, is the differential time; Sub-step 1.3: Extract the erasure action event from the touch behavior sequence. When the erasure displacement satisfies: and which is counted as an effective modification Among them, is the width of the writing area, is the displacement coefficient, is the touch pressure, is the threshold value of the touch pressure change rate, is the differential time.

3. An AI-based personalized learning question recommendation method according to claim 1, characterized in that, In Step 2, the feature extraction and calculation further include: Sub-step 2.1, perform saccade segment segmentation on the eye movement trajectory data obtained in step 1, when the distance between adjacent fixation points satisfies: , it is determined as a valid saccade segment, Among them, is the standard deviation of the fixation point spacing, is the curvature change threshold; Sub-step 2.2: Perform phoneme-level segmentation on the speech stream data obtained in Step 1 and calculate the duration of the silent segments Distribution: , marked as non-natural pauses, Among them, is the average value of the fundamental frequency of the statement, is the duration of the i-th silent segment, is the total number of silent segments in the speech stream, is the mutation coefficient, is the instantaneous change in the fundamental frequency of the speech; Sub-step 2.3: Classify events for the touch behavior sequence obtained in Step 1 and count the effective modification frequency: When And , Among them, is the effective modification frequency, is the number of erasure times that meet the conditions, is the total duration of the writing task, is the displacement proportionality coefficient, is the minimum touch pressure threshold, is the horizontal displacement of the touch erasure action, is the peak touch pressure, is the width of the writing area.

4. An AI-based personalized learning question recommendation method according to claim 1, characterized in that In Step 3, the feature fusion and state generation further include: Sub-step 3.1: Dynamically normalize the curvature C of the reading saccade path obtained in Step 2; ， wherein, is the normalized saccade path curvature, is the historical curvature mean value, is the standard deviation; Sub-step 3.2: Align the non-natural pause intervals of the spoken language obtained in step 2 with the touch modification frequency in time series: ， When the alignment error feature stitching is performed, where is the time alignment offset of the multi-modal data, is the duration of the i-th detected unnatural pause, is the time offset and the touch modification frequency after that, is the time window threshold; Sub-step 3.3: Weight the fused features using an attention mechanism; ， Among them, is the attention weight of the k-th feature, is the query vector, is the transposed key vector, is the feature vector dimension, When the attention score activate the calculation of cross-cultural understanding degree.

5. A personalized recommendation method for learning questions based on AI according to claim 1, characterized in that In Step 4, the policy generation and optimization further include: Sub-step 4.1, based on the syntax processing speed in the dynamic state vector Query the knowledge graph: When , retain the associated edges Among them, is the quantization value of the knowledge point association strength, is the cosine similarity between knowledge point vectors, is the association strength threshold; Sub-step 4.2: Construct a multi-objective optimization function to generate an initial policy; ， Among them, , is the anxiety fluctuation coefficient, is the maximum anxiety value, is the dynamic weight coefficient, is the knowledge point coverage gain term, is the strategy utility function, is the cognitive comfort compensation term; Sub-step 4.3: Perform teaching progress constraint verification; When , increase the weight of the question type distribution Among them, is the progress weight of knowledge point i, is the question difficulty, is the minimum coverage coefficient, is the standard learning duration benchmark value, is the total number of knowledge points to be covered on the current day.

6. A personalized recommendation method for learning questions based on AI according to claim 1, characterized in that In Step 5, the question screening and enhancement further include: Sub-step 5.1: Based on the question difficulty manifold in the recommendation strategy Construct a candidate set: ， Among them, is the set of candidate questions, is the question 's difficulty value, is the i-th question object in the question bank, is the difficulty tolerance, is the target difficulty value; Sub-step 5.2: Allocate the number of questions according to the question type distribution ratio P; When , truncate in descending order of priority wherein, is the total number of preset questions, is the number of questions to be selected for the j-th question type, is the distribution ratio of the j-th question type, is the maximum number of questions recommended at one time threshold, is the total number of question type categories; Sub-step 5.3: Attach dynamic annotations to the screened questions; Otherwise, when , Among them, is the syntax processing speed threshold, is the complete syntax hint, is the core point hint, is the dynamic hint type.

7. A personalized recommendation method for learning questions based on AI according to claim 1, characterized in that, In Step 6, the state update and policy optimization further include: Sub-step 6.1: Calculate the response time deviation based on the real-time interaction data of the target questions; When , trigger status correction, Among them, is the response time deviation, is the response deviation threshold, is the expected response time based on historical data; Sub-step 6.2: Update the grammar processing speed in the dynamic state vector; ， Among them, is the grammar processing speed after update, is the grammar processing speed before update, is the historical weight coefficient, is the number of correct answers, is the total response time; Sub-step 6.3: Calculate the policy reward value and update the reinforcement learning model; When , the trigger strategy is reset, Among them, is the comprehensive strategy reward value, is the progress reward item, is the comfort reward item, is the progress reward weight, is the comfort reward weight, is the minimum reward threshold.

8. A personalized recommendation method for learning questions based on AI according to claim 1, characterized in that, In Step 2, the curvature of the reading saccade path is obtained by calculating the angular change rate of the line connecting adjacent fixation points; In Step 2, the detection of unnatural pauses in speech adopts a joint criterion of phoneme boundary alignment and silent segment duration.

9. The personalized recommendation method for learning questions based on AI according to claim 1, characterized in that, In Step 4, the node weights of the knowledge graph are dynamically adjusted according to the association strength of grammar knowledge points and the error pattern propagation path; In Step 4, the reinforcement learning policy model adopts a dual-objective reward function, including a knowledge point coverage gain term and a cognitive comfort compensation term.

10. A personalized recommendation method for learning questions based on AI according to claim 1, characterized in that, In Step 5, the display duration of the grammar structure hint box is inversely proportional to the current cognitive load characteristics; In Step 5, the detail level of the cultural background annotation is negatively correlated with the cross-cultural understanding degree.

Images