A self-learning optimization method based on student demand response strategy

By collecting multimodal data and semantic clustering, combined with the selection of teaching scenario parameters, the learning needs of students are quantified, and a self-learning iterative closed loop is constructed. This solves the problem of insufficient identification of students' personalized learning needs in intelligent teaching systems and achieves dynamic adaptation and accurate response.

CN123118317APending Publication Date: 2026-11-17CHONGQING DANDELION FUTURE TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611276302.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-21
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing intelligent teaching systems cannot accurately identify students' personalized learning needs and lack dynamic response capabilities, resulting in poor teaching adaptability, lagging response strategies, and an inability to achieve individualized instruction.

Method used

By collecting multimodal data and aligning it with the timeline, combined with semantic clustering and demand weight quantification, a response strategy sequence is generated. The strategy deviation is then quantified through student interaction feedback data, a self-learning iterative closed loop is constructed, and the strategy template library is updated.

Benefits of technology

It enables precise responses to students' learning needs, improves the adaptability of teaching and the accuracy of responses, dynamically optimizes strategy matching, and adapts to changes in personalized learning needs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118317A_ABST
    Figure CN123118317A_ABST
Patent Text Reader

Abstract

This invention discloses a self-learning optimization method based on student demand response strategies, relating to the field of educational artificial intelligence. The method includes: collecting multimodal data such as text, voice, exercise answers, and interface operations; aligning these data with timestamps to obtain a set of student demand vectors; clustering the demand vector set into demand clusters; calculating demand intensity weights; matching and filtering based on teaching scenarios to obtain a sequence of demand response strategies with time windows; executing the strategies and collecting interactive feedback to form a feedback sample dataset; comparing the samples with the expected baseline to obtain feedback deviation and generating a learning effect evaluation vector; associating the evaluation vector with the strategy sequence; updating the hit rate and deviation of the strategy template library; adjusting template priorities; and completing iterative optimization of the template library. The advantages of this invention are: quantifying students' personalized learning needs, overcoming the shortcomings of fixed and homogeneous responses in traditional intelligent teaching, achieving dynamic adaptive teaching tailored to individual needs, and significantly improving the accuracy and adaptability of student learning demand responses.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of educational artificial intelligence, and in particular to a self-learning optimization method based on student demand response strategies. Background Technology

[0002] With the widespread adoption of educational informatization and intelligent education technologies, personalized and autonomous learning has become a core trend in education development in the new era, and various intelligent learning systems are widely used in teaching scenarios. However, traditional intelligent learning models often adopt homogeneous teaching logic, focusing on fixed teaching processes and uniform resource delivery, generally neglecting students' personalized learning needs, knowledge gaps, learning pace, and autonomous learning requirements. This results in problems such as poor teaching adaptability, delayed dynamic response, and insufficient targeted learning optimization.

[0003] Current mainstream intelligent teaching response methods generally suffer from low levels of intelligence, poor adaptability, and an inability to iterate autonomously. Most existing methods rely solely on single text or exercise data to determine student needs, resulting in limited data dimensions that fail to comprehensively capture students' true learning status and implicit needs. Furthermore, they lack refined needs clustering and intensity quantification mechanisms, leading to insufficient accuracy in needs identification. Simultaneously, traditional strategy matching is often triggered by fixed rules, failing to dynamically adjust based on real-time teaching scenarios, easily resulting in a disconnect between response strategies and classroom progress or students' actual learning. In addition, existing methods lack a complete feedback loop and self-learning optimization mechanism, mechanically executing preset strategies without quantifying their effectiveness or correcting deficiencies based on student feedback. Template strategies remain static and unchanging, failing to adapt to the personalized and dynamic learning needs of different students, easily leading to delayed responses, mismatched matching, and poor adaptability. Overall, the flexibility, accuracy, and iterative evolution capabilities of teaching responses exhibit significant shortcomings. Summary of the Invention

[0004] To improve existing methods, a self-learning optimization method based on student demand response strategies is proposed. This method quantifies students' personalized learning needs and dynamically matches appropriate response strategies to the teaching scenario. It overcomes the shortcomings of fixed and homogeneous responses in traditional intelligent teaching, realizes dynamic adaptive teaching that is tailored to individual needs, and significantly improves the accuracy and adaptability of responding to students' learning needs.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A self-learning optimization method based on student demand response strategies includes: Multimodal input data is collected in real time through the teaching interface, including text questions, voice questions, exercise answer records and interface operation behavior sequences. The data is then aligned and mapped to a unified timeline by timestamps, and the intent tags at each moment are extracted and combined into a set of student demand vectors in chronological order. The student demand vector set is clustered into demand clusters according to intent labels. The demand intensity weight of each demand cluster is calculated. Candidate strategies are matched from the preset response strategy template library according to the weight and filtered according to the current teaching scenario state parameters to generate a demand response strategy sequence containing response action type and time window. Execute strategy items sequentially within each time window according to the strategy sequence. After each item is executed, collect subsequent operation changes, interaction response time and interface focus switching trajectory as real-time feedback. After all items are executed, summarize them to form a feedback sample dataset. The feedback sample dataset is compared item by item with the preset expected feedback baseline. The feedback deviation of each strategy item is calculated. Combined with the interactive content features, a learning effect evaluation vector is generated to characterize the effect of each strategy item on the learning state correction. A mapping is established between the learning effect evaluation vector and the demand response strategy sequence. The historical hit rate and historical feedback deviation of each template in the template library are updated according to the mapping. The template priority is adjusted according to the updated hit rate, and the adjusted template library is used for the next iteration.

[0006] Preferably, the step of aligning and mapping to a unified timeline by timestamp specifically includes: Extract the receiving timestamp of the text question data, the start and end timestamps of the voice segment in the voice question data, the start and submission timestamps of the answer record of the exercise, and the trigger timestamps of each operation event in the sequence of interface operation behavior, as the time points to be aligned; Using the earliest time among all the time points to be aligned as the reference zero point, each original timestamp is converted into a relative time offset relative to the reference zero point, forming a time alignment sequence for each modality of data; The time-aligned sequence is divided into continuous time slices according to the preset time window length. The frequency of occurrence of each intent tag and the number of interaction content features in each time slice are counted. The statistical results are summarized into the intent tag vector corresponding to the time slice. The intent label vectors of all time slices are concatenated in chronological order to form the student demand vector set.

[0007] Preferably, the step of clustering the student demand vector set into demand clusters according to intent labels and calculating the demand intensity weight of each demand cluster specifically includes: Semantic similarity is calculated for intent tags at each time point. Tags with semantic similarity greater than a preset threshold are grouped into the same demand cluster. Within the same demand cluster, the original intent tags and their associated interactive content features are retained. For each demand cluster, count the number of times the intent tag appears within the cluster and the total number of time slices spanned by the cluster. Divide the number of occurrences by the total number of time slices to obtain the time density value of the cluster. Extract the numerical components representing the frequency of student operations from the interactive content features within the cluster, arrange them in chronological order to form an operation frequency sequence, and calculate the average absolute value of the difference between adjacent values ​​in the sequence as the fluctuation coefficient of the cluster. The demand intensity weights of the demand clusters are obtained by weighting and summing the time density values ​​and volatility coefficients.

[0008] Preferably, the step of matching candidate strategies from a preset response strategy template library by weight sorting specifically includes: A pre-built response strategy template library is provided, with each strategy template containing a template identifier, a list of applicable requirement tags, a strategy action type field, and a strategy content data field. For each requirement cluster, extract the most frequently occurring idea map label in the cluster, match the idea map label with the list of applicable requirement labels for each template, and mark the template that matches successfully as the initial matching template; Obtain the status parameters of the current teaching scenario, including the current teaching chapter identifier and the class time progress value; perform compatibility verification on each initial matching template; and retain the templates that pass the verification as candidate templates. All candidate templates are sorted from high to low according to the demand intensity weight of the corresponding demand cluster, and a response time window parameter is assigned to each candidate template to form a candidate response strategy set.

[0009] Preferably, the collection of subsequent operation changes, interaction response duration, and interface focus switching trajectory as real-time feedback specifically includes: Record the strategy execution start timestamp when each strategy item is executed, and continuously collect the interface operation event stream during the execution process. The operation event stream includes mouse click, keyboard input and touch swipe events. After the strategy item is executed, record the strategy execution end timestamp, extract the target interface element identifier and operation action type of each operation event from the start to the end, and form the operation change record of the strategy item. The difference between the end timestamp and the start timestamp of the calculated strategy execution is used as the interaction response duration of the strategy item. The switching order of the focus elements of the interface is extracted from the operation event stream and recorded as the interface focus switching trajectory. The operation change records, interaction response durations, and interface focus switching trajectories are associated and stored as real-time feedback data entries for this strategy item. The entries of all strategy items together constitute the feedback sample dataset.

[0010] Preferably, the step of comparing the feedback sample dataset with the preset expected feedback baseline item by item specifically includes: Pre-set the expected feedback baseline for each strategy item type. The baseline includes the expected operation change mode, the upper limit of the expected interaction response time, and the range of the expected focus switching path length. Extract the operation change record of each strategy item from the feedback sample dataset, match the actual operation type and object level path of each operation event in the record with the expected type and expected path in the expected operation change pattern, count the number of mismatched events, and obtain the operation pattern deviation. If the interaction response duration of the strategy item is greater than the expected upper limit, the difference between the two is calculated as the duration deviation; if it is less than or equal to the upper limit, the duration deviation is set to zero. The actual number of switching steps of the interface focus switching trajectory of the strategy item is compared with the upper and lower limits of the expected path length range, and the focus switching deviation is calculated according to the degree of deviation. The feedback deviation of this strategy item is obtained by weighted summing of the operation mode deviation, duration deviation, and focus switching deviation.

[0011] Preferably, the generation of the learning performance evaluation vector representing the effect of each strategy term on the learning state correction specifically includes: The feedback deviations of all strategy items under the same idea map label are arranged in chronological order according to the response time windows of each strategy item, forming the deviation time series corresponding to that idea map label. Calculate the difference between two adjacent deviations in the time series of the deviation, count the number of positive differences and the number of negative differences, divide them by the total number of differences, and obtain the deviation deterioration ratio and deviation improvement ratio of the main plot label; Each strategy item under the idea map label is grouped according to the response action type, and the average value of the feedback deviation within each group is calculated as the correction contribution value of that action type to the idea map label. The deviation deterioration rate, deviation improvement rate, and correction contribution value of each main map label are arranged in the order of time slices to form a learning effect evaluation vector.

[0012] Preferably, the step of updating the historical hit rate and historical feedback deviation of each template in the mapping template library specifically includes: Extract the template identifier and corresponding feedback deviation of the strategy template that was actually hit in the current cycle from the mapping relationship, increment the historical hit count of each hit template by one, and add the current feedback deviation to the historical feedback deviation list of the template. For strategy templates that are not hit in this cycle, keep their historical hit count and historical deviation list unchanged, and increase the consecutive hit count of the template by one. For each template, divide the current historical hit count by the total number of matching attempts since the template was created to obtain the current historical hit rate, and use the average of all deviations in the historical feedback deviation list of the template as the average feedback deviation. The current historical hit rate and average feedback deviation of each template are added as new attribute fields and written into the template data of the corresponding template in the response strategy template library to complete the weighted update.

[0013] Preferably, the step of adjusting the template priority sorting according to the updated hit rate specifically includes: For each template in the response strategy template library, the historical hit rate is multiplied by the first weighting coefficient to obtain the hit rate score, the negative of the average feedback deviation of the template is multiplied by the second weighting coefficient to obtain the deviation score, and the two are added together to obtain the comprehensive priority score. All templates are sorted from highest to lowest based on their overall priority score to obtain the first sorted list. Then, they are sorted from highest to lowest based on their historical hit rate to obtain the second sorted list. For templates with the same overall priority score, their order is determined by their position in the second sorted list. The sorted templates are written sequentially into the priority index field of the template library, and a new matching priority number is assigned to each template, with higher priority numbers for templates placed earlier in the sequence. During the generation of demand response strategies for the next teaching cycle, templates are selected from the template library in descending order of matching priority number for matching, and the selected templates are used as part of the candidate response strategy set for the next cycle.

[0014] Compared with the prior art, the advantages of the present invention are: By collecting multimodal data across all dimensions and aligning it with the timeline, student learning needs are precisely analyzed. Combined with semantic clustering and demand weight quantification, the system accurately distinguishes the strength and fluctuation characteristics of student needs, avoiding the problems of incomplete or delayed demand identification. Simultaneously, it incorporates teaching scenario parameters to select appropriate strategies, ensuring that response strategies align with real-time teaching scenarios, balancing rationality and adaptability. Student interaction feedback data is used to quantify strategy execution deviations, and the effectiveness of multidimensional evaluation of strategies in correcting student learning states is assessed, overcoming the limitations of subjective evaluation. Most importantly, a complete self-learning iterative closed loop is constructed, continuously updating the hit rate and priority of the strategy template library, dynamically optimizing strategy matching logic, and continuously adapting to the personalized learning needs and changes in learning states of different students. This continuously improves the accuracy, adaptability, and effectiveness of intelligent teaching responses, achieving dynamic evolution and upgrading of teaching services. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of a self-learning optimization method based on a student demand response strategy proposed in this invention. Figure 2 This is a schematic diagram illustrating the computational demand intensity weighting of demand clusters proposed in this invention; Figure 3 This is a schematic diagram illustrating the real-time feedback acquisition proposed in this invention. Detailed Implementation

[0016] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0017] See Figure 1 As shown, a self-learning optimization method based on student demand response strategy includes: Multimodal input data is collected in real time through the teaching interface, including text questions, voice questions, exercise answer records and interface operation behavior sequences. The data is then aligned and mapped to a unified timeline by timestamps, and the intent tags at each moment are extracted and combined into a set of student demand vectors in chronological order. The student demand vector set is clustered into demand clusters according to intent labels. The demand intensity weight of each demand cluster is calculated. Candidate strategies are matched from the preset response strategy template library according to the weight and filtered according to the current teaching scenario state parameters to generate a demand response strategy sequence containing response action type and time window. Execute strategy items sequentially within each time window according to the strategy sequence. After each item is executed, collect subsequent operation changes, interaction response time and interface focus switching trajectory as real-time feedback. After all items are executed, summarize them to form a feedback sample dataset. The feedback sample dataset is compared item by item with the preset expected feedback baseline. The feedback deviation of each strategy item is calculated. Combined with the interactive content features, a learning effect evaluation vector is generated to characterize the effect of each strategy item on the learning state correction. A mapping is established between the learning effect evaluation vector and the demand response strategy sequence. The historical hit rate and historical feedback deviation of each template in the template library are updated according to the mapping. The template priority is adjusted according to the updated hit rate, and the adjusted template library is used for the next iteration.

[0018] Specifically, multimodal input data is acquired in real time through the teaching interface. This multimodal input data includes text question data, voice question data, exercise answer records, and interface operation behavior sequences. To achieve temporal consistency of the multimodal data, precise time alignment processing is performed on each modality, and key time points in all modalities are extracted, including the timestamp of receiving text question data, the start and end timestamps of voice segments, the start and submission timestamps of exercise answers, and the trigger timestamps of interface operation events. The set of all extracted time points is denoted as... The minimum value among them is taken as the reference zero point. =min(T), and then convert each original timestamp into a relative offset. The core conversion formula is: in, For any original timestamp, This is the offset time relative to the reference zero point. Thus, each event in each modal data obtains its position on a unified time axis, forming a time-aligned sequence.

[0019] The entire time axis is divided into continuous, non-overlapping time slices according to a preset time window length ΔT, and each slice is denoted as follows: The j-th slice covers the time interval. For each slice Count the occurrence frequency of each intent label in all aligned events within the slice, and denote the frequency of the m-th intent label as . Simultaneously, the interactive content features within the slice are summarized, such as the number of characters in the text input, the proportion of voice duration, and whether the answer is correct or not, to form a feature vector. By concatenating the frequency and feature values, the intent label vector corresponding to that slice is obtained. Its form is: Where M is the total number of preset intent tags, and L is the number of dimensions of the interactive content features. Finally, the intent tag vectors of all slices are concatenated in chronological order to form a complete student demand vector set V: The vector set V is arranged in chronological order, which not only preserves the temporal information of the demand, but also includes the intent distribution and interaction features within each window, providing a structured input for subsequent demand aggregation and strategy generation.

[0020] See Figure 2 As shown, semantic aggregation is performed on the intent tags at each time step in the demand vector set to eliminate redundant expressions of similar demands. Cosine similarity based on pre-trained word vectors is used as a semantic metric. For any two intent tags corresponding to text descriptions, the cosine of the angle between their vectors is calculated. When this value is greater than a preset threshold, the two are considered semantically similar and grouped into the same demand cluster. Within the same demand cluster, all original intent tags and their associated interaction content features are retained for subsequent detailed analysis.

[0021] For each demand cluster obtained after clustering The urgency level needs to be quantified, i.e., the demand intensity weight. First, calculate the time density value ρk of the cluster, defined as the number of occurrences of all intent tags within the cluster. Divide by the total number of time slices spanned by the cluster ,Right now: This value reflects the concentration of demand within a unit time window; a higher value indicates a denser demand. Secondly, the numerical components representing the frequency of student operations are extracted from the features of each interaction within this cluster and arranged chronologically as a sequence. The average of the absolute differences between two adjacent frequency values ​​in the sequence is calculated as the fluctuation coefficient of the cluster. : A higher volatility coefficient indicates more unstable student behavior and a more urgent need. Finally, the time density value and volatility coefficient are weighted and summed to obtain the demand intensity weight for this demand cluster. : Here, α and β are preset weighting coefficients used to balance the contributions of density and volatility.

[0022] After obtaining the intensity weights of each demand cluster, candidate strategies are further matched from a pre-defined response strategy template library. Each strategy template in the template library contains a template identifier, a list of applicable demand tags, a strategy action type field, and a strategy content data field. For each demand cluster, the most frequently occurring intent tag in the cluster is extracted as the intent graph tag, and it is matched against the list of applicable tags for each template in the template library. If a match is successful, the template is marked as the initial matching template. Subsequently, the status parameters of the current teaching scenario are obtained, including the current teaching chapter identifier and the class time progress value, and the initial matching template is checked for compatibility: for example, if the template content involves knowledge from subsequent chapters, it is not available in the current chapter; if the response action requires more time than the remaining class time, it is excluded. Templates that pass the check are retained as candidate templates. Finally, all candidate templates are matched according to the weights of their corresponding demand clusters. The candidate templates are sorted from highest to lowest, and specific response time window parameters are assigned to each candidate template based on the expected execution time and current timeline position of each strategy item. These parameters are then combined to form a set of candidate response strategies for use in subsequent execution stages.

[0023] See Figure 3 As shown, suppose P policy items are generated in the current cycle, and they are executed sequentially according to their order in the policy sequence. For the i-th policy item, the system timestamp is recorded at the start time of the execution action of that policy item. This triggers corresponding actions, such as displaying knowledge point explanation cards, adjusting the difficulty level of exercises, or playing micro-lesson video clips. During the execution of the strategy, the system continuously collects all operation event streams of students on the teaching interface through the underlying event listening interface, including mouse click events, keyboard input events, and touch swipe events.

[0024] Once all actions of this strategy are completed, record the execution end timestamp. The system extracts data from the stored event stream. to For all operation events within a given time period, extract the target UI element identifier and operation action type corresponding to each event to form an operation change record list for that strategy item. This list describes the specific changes in student interface interaction behavior during the execution of the policy.

[0025] At the same time, the system calculates the interaction response duration of the strategy item based on the two timestamps. The calculation formula is as follows: This duration reflects the time required for students to complete the corresponding operation from receiving the policy response, and is an important quantitative indicator for measuring the effectiveness of the response.

[0026] Furthermore, the system extracts the switching order of interface focus elements from the operation event flow. Specifically, whenever the mouse cursor or touch focus moves from one interface element (such as an input box, button, or area panel) to another, the system records the switching event and the element identifiers before and after it. By connecting all switching events in chronological order, the system obtains the interface focus switching trajectory corresponding to the strategy item. ,in This is the identifier for the j-th focused interface element. This represents the total number of focus transitions within the trajectory. The length of the trajectory. The complexity of student attention shifting between various components of the interface in response to stimuli is calculated as follows: That is, the number of times adjacent elements in the focus element sequence switch. When An excessively large value indicates that the student's attention is scattered, while an excessively small value may reflect insufficient response to the stimulus.

[0027] For each policy item i, record its operation changes. Interaction response time and focus switching trajectory The associated storage is a data entry that provides immediate feedback. After all P policy items have been executed, all items are combined into an ordered set, i.e., the feedback sample dataset D. This dataset maintains the execution order of each policy item, ensuring that subsequent evaluations can analyze the changing trends of feedback deviations over time. If a policy item is interrupted during execution due to an anomaly, the item is marked as invalid and recorded in the anomaly log, but this does not affect the subsequent processing of valid items.

[0028] Using the previously collected feedback sample dataset as input, the feedback deviation of each strategy item is calculated by comparing it item by item with a preset expected baseline. This is then aggregated to generate a learning effect evaluation vector, which characterizes the actual corrective effect of each strategy item on the student's learning state. A corresponding expected feedback baseline is pre-defined for each strategy item type. This baseline includes three parts: the expected operation change mode (i.e., the expected sequence of operation events generated by the student after the strategy item is executed and the corresponding interface object hierarchy path), and the expected upper limit of the interaction response time. This refers to the maximum acceptable time for a student to complete the expected operation, and the expected range of focus switching path length. This refers to the reasonable range of steps for switching the focus element on the interface.

[0029] For the i-th policy item in the feedback sample dataset, the system extracts the actual operation change record from that item. The actual operation type and object hierarchy path of each operation event in the record are matched one by one with the expected type and expected path in the expected operation change mode. The sum of the number of events with mismatched operation types and the number of events with mismatched object paths is used as the operation mode deviation. At the same time, if the interaction response time of this strategy item is... Greater than the expected upper limit The difference between the two is then calculated as the duration deviation. ;like ,but In addition, the actual number of steps in the focus switching trajectory of this strategy item. Within the expected range If a comparison is made, If the deviation exceeds this range, the focus switching deviation will be calculated based on the degree of deviation. The deviation is calculated as the ratio of the absolute value of the deviation to the range width. Finally, the three deviations are weighted and summed to obtain the feedback deviation of this strategy term. : in, , , The preset weighting coefficients are set to 0.4, 0.3, and 0.3 respectively, to balance the contribution of the three types of deviations to the overall evaluation.

[0030] After obtaining the feedback deviation of each strategy item, the system aggregates the deviations of all strategy items according to the same purpose map label, and arranges the aggregated deviations in chronological order of their response time windows to form the deviation time series corresponding to that purpose map label. For the j-th position in this sequence Calculate the difference between two adjacent deviations: statistics Number of >0 and quantity Divide each by the total number of differences N. 1. Obtain the deviation deterioration ratio of the concept map label. and deviation improvement ratio The former reflects the degree to which teaching effectiveness deteriorates with strategy implementation, while the latter reflects the degree of improvement. Subsequently, each strategy item under the subject map label is grouped according to its response action type, and the average feedback deviation within each group is calculated as the correction contribution value for that action type to the subject map label. Finally, the deviation deterioration ratio, deviation improvement ratio, and correction contribution value for each subject map label are arranged according to the time slices defined in the previous data collection steps to form a complete learning effectiveness evaluation vector. This vector, with its time-series components, characterizes the correction effect of different need types under different strategy actions, providing a basis for subsequent optimization and updating of strategy templates.

[0031] A mapping relationship is established between the previously generated learning performance evaluation vector and the actual demand response strategy sequence executed in the current cycle. From this mapping relationship, the template identifiers of all policy templates that were actually hit and executed in the current cycle, as well as the feedback deviation of the policy item corresponding to each hit template, are extracted. For each hit template, its historical hit count H is incremented by one, and the current feedback deviation is also incremented. Add the historical feedback deviation list to this template For strategy templates that are not hit within this period, their historical hit count and deviation list remain unchanged, while the system continues to count the consecutive misses of that template. Add one more time for subsequent determination of whether the template needs to be downgraded or temporarily frozen.

[0032] After completing the above statistical updates, the system recalculates the historical hit rate for each template. The calculation formula is as follows: Where H represents the historical hit count. For consecutive misses, This represents the total number of times the template has been skipped due to compatibility checks failing since its creation. (Denominator) This represents the total number of matching attempts made by the template since its creation. Simultaneously, the system calculates the average feedback deviation for this template. That is, the list of historical feedback deviations The arithmetic mean of all deviation values: in, This represents the total number of deviations in the list. If If empty, then Initialize to 0. The calculated... and As a new attribute field, it is written into the template data of the corresponding template in the template library to complete the weighted update.

[0033] Subsequently, the system adjusts the priority ranking of each template based on the updated hit rate. For each template, the historical hit rate is... Multiplied by the first weighting coefficient (Take 0.7) to obtain the hit rate score, and then calculate the negative of the average feedback bias of the template. Multiply by the second weighting factor (Take a value of 0.3) to obtain the deviation score, and add the two scores together to obtain the comprehensive priority score. : A higher score indicates that the template has both a high historical hit rate and a low feedback bias, and should be given priority. The system sorts all templates from highest to lowest based on their overall priority score to obtain a first sorted list; simultaneously, it sorts them from highest to lowest based on their historical hit rate to obtain a second sorted list. For templates with the same overall priority score, their order is determined by their position in the second sorted list, with higher hit rates taking precedence. The sorted templates are then written sequentially into the priority index field of the template library, and a new matching priority number is assigned to each template, with smaller numbers indicating higher priority.

[0034] When the demand response strategy generation is re-executed in the next teaching cycle, the system directly selects templates from the template library in ascending order of matching priority number for matching attempts. Templates that successfully pass the compatibility check and are selected are included in the candidate response strategy set for the next cycle. This process is continuously repeated with each teaching cycle, and the statistical parameters of each template are continuously accumulated and updated, and the priority ranking is dynamically adjusted accordingly, so that the strategy template library gradually approaches the optimal matching strategy in long-term operation.

[0035] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0036] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0037] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A self-learning optimization method based on student demand response strategies, characterized in that, include: Multimodal input data is collected in real time through the teaching interface, including text questions, voice questions, exercise answer records and interface operation behavior sequences. The data is then aligned and mapped to a unified timeline by timestamps, and the intent tags at each moment are extracted and combined into a set of student demand vectors in chronological order. The student demand vector set is clustered into demand clusters according to intent labels. The demand intensity weight of each demand cluster is calculated. Candidate strategies are matched from the preset response strategy template library according to the weight and filtered according to the current teaching scenario state parameters to generate a demand response strategy sequence containing response action type and time window. Execute strategy items sequentially within each time window according to the strategy sequence. After each item is executed, collect subsequent operation changes, interaction response time and interface focus switching trajectory as real-time feedback. After all items are executed, summarize them to form a feedback sample dataset. The feedback sample dataset is compared item by item with the preset expected feedback baseline. The feedback deviation of each strategy item is calculated. Combined with the interactive content features, a learning effect evaluation vector is generated to characterize the effect of each strategy item on the learning state correction. A mapping is established between the learning effect evaluation vector and the demand response strategy sequence. The historical hit rate and historical feedback deviation of each template in the template library are updated according to the mapping. The template priority is adjusted according to the updated hit rate, and the adjusted template library is used for the next iteration.

2. The self-learning optimization method based on student demand response strategy according to claim 1, characterized in that, The process of aligning and mapping to a unified timeline by timestamp specifically includes: Extract the receiving timestamp of the text question data, the start and end timestamps of the voice segment in the voice question data, the start and submission timestamps of the answer record of the exercise, and the trigger timestamps of each operation event in the sequence of interface operation behavior, as the time points to be aligned; Using the earliest time among all the time points to be aligned as the reference zero point, each original timestamp is converted into a relative time offset relative to the reference zero point, forming a time alignment sequence for each modality of data; The time-aligned sequence is divided into continuous time slices according to the preset time window length. The frequency of occurrence of each intent tag and the number of interaction content features in each time slice are counted. The statistical results are summarized into the intent tag vector corresponding to the time slice. The intent label vectors of all time slices are concatenated in chronological order to form the student demand vector set.

3. The self-learning optimization method based on student demand response strategy according to claim 1, characterized in that, The process of clustering student demand vector sets into demand clusters based on intent labels and calculating the demand intensity weight for each demand cluster specifically includes: Semantic similarity is calculated for intent tags at each time point. Tags with semantic similarity greater than a preset threshold are grouped into the same demand cluster. Within the same demand cluster, the original intent tags and their associated interactive content features are retained. For each demand cluster, count the number of times the intent tag appears within the cluster and the total number of time slices spanned by the cluster. Divide the number of occurrences by the total number of time slices to obtain the time density value of the cluster. Extract the numerical components representing the frequency of student operations from the interactive content features within the cluster, arrange them in chronological order to form an operation frequency sequence, and calculate the average absolute value of the difference between adjacent values ​​in the sequence as the fluctuation coefficient of the cluster. The demand intensity weights of the demand clusters are obtained by weighting and summing the time density values ​​and volatility coefficients.

4. The self-learning optimization method based on student demand response strategy according to claim 1, characterized in that, The step of matching candidate strategies from a preset response strategy template library by weight sorting specifically includes: A pre-built response strategy template library is provided, with each strategy template containing a template identifier, a list of applicable requirement tags, a strategy action type field, and a strategy content data field. For each requirement cluster, extract the most frequently occurring idea map label in the cluster, match the idea map label with the list of applicable requirement labels for each template, and mark the template that matches successfully as the initial matching template; Obtain the status parameters of the current teaching scenario, including the current teaching chapter identifier and the class time progress value; perform compatibility verification on each initial matching template; and retain the templates that pass the verification as candidate templates. All candidate templates are sorted from high to low according to the demand intensity weight of the corresponding demand cluster, and a response time window parameter is assigned to each candidate template to form a candidate response strategy set.

5. The self-learning optimization method based on student demand response strategy according to claim 1, characterized in that, The collection of subsequent operation changes, interaction response duration, and interface focus switching trajectory as real-time feedback specifically includes: Record the strategy execution start timestamp when each strategy item is executed, and continuously collect the interface operation event stream during the execution process. The operation event stream includes mouse click, keyboard input and touch swipe events. After the strategy item is executed, record the strategy execution end timestamp, extract the target interface element identifier and operation action type of each operation event from the start to the end, and form the operation change record of the strategy item. The difference between the end timestamp and the start timestamp of the calculated strategy execution is used as the interaction response duration of the strategy item. The switching order of the focus elements of the interface is extracted from the operation event stream and recorded as the interface focus switching trajectory. The operation change records, interaction response durations, and interface focus switching trajectories are associated and stored as real-time feedback data entries for this strategy item. The entries of all strategy items together constitute the feedback sample dataset.

6. The self-learning optimization method based on student demand response strategy according to claim 1, characterized in that, The step of comparing the feedback sample dataset with the preset expected feedback baseline item by item specifically includes: Pre-set the expected feedback baseline for each strategy item type. The baseline includes the expected operation change mode, the upper limit of the expected interaction response time, and the range of the expected focus switching path length. Extract the operation change record of each strategy item from the feedback sample dataset, match the actual operation type and object level path of each operation event in the record with the expected type and expected path in the expected operation change pattern, count the number of mismatched events, and obtain the operation pattern deviation. If the interaction response duration of the strategy item is greater than the expected upper limit, the difference between the two is calculated as the duration deviation; if it is less than or equal to the upper limit, the duration deviation is set to zero. The actual number of switching steps of the interface focus switching trajectory of the strategy item is compared with the upper and lower limits of the expected path length range, and the focus switching deviation is calculated according to the degree of deviation. The feedback deviation of this strategy item is obtained by weighted summing of the operation mode deviation, duration deviation, and focus switching deviation.

7. The self-learning optimization method based on student demand response strategy according to claim 1, characterized in that, The learning performance evaluation vector, which generates representations of the effect of each strategy term on the learning state correction, specifically includes: The feedback deviations of all strategy items under the same idea map label are arranged in chronological order according to the response time windows of each strategy item, forming the deviation time series corresponding to that idea map label. Calculate the difference between two adjacent deviations in the time series of the deviation, count the number of positive differences and the number of negative differences, divide them by the total number of differences, and obtain the deviation deterioration ratio and deviation improvement ratio of the main plot label; Each strategy item under the idea map label is grouped according to the response action type, and the average value of the feedback deviation within each group is calculated as the correction contribution value of that action type to the idea map label. The deviation deterioration rate, deviation improvement rate, and correction contribution value of each main map label are arranged in the order of time slices to form a learning effect evaluation vector.

8. The self-learning optimization method based on student demand response strategy according to claim 1, characterized in that, The specific details of updating the historical hit rate and historical feedback deviation of each template in the mapping template library include: Extract the template identifier and corresponding feedback deviation of the strategy template that was actually hit in the current cycle from the mapping relationship, increment the historical hit count of each hit template by one, and add the current feedback deviation to the historical feedback deviation list of the template. For strategy templates that are not hit in this cycle, keep their historical hit count and historical deviation list unchanged, and increase the consecutive hit count of the template by one. For each template, divide the current historical hit count by the total number of matching attempts since the template was created to obtain the current historical hit rate, and use the average of all deviations in the historical feedback deviation list of the template as the average feedback deviation. The current historical hit rate and average feedback deviation of each template are added as new attribute fields and written into the template data of the corresponding template in the response strategy template library to complete the weighted update.

9. The self-learning optimization method based on student demand response strategy according to claim 1, characterized in that, The specific steps of adjusting the template priority sorting according to the updated hit rate include: For each template in the response strategy template library, the historical hit rate is multiplied by the first weighting coefficient to obtain the hit rate score, the negative of the average feedback deviation of the template is multiplied by the second weighting coefficient to obtain the deviation score, and the two are added together to obtain the comprehensive priority score. All templates are sorted from highest to lowest based on their overall priority score to obtain the first sorted list. Then, they are sorted from highest to lowest based on their historical hit rate to obtain the second sorted list. For templates with the same overall priority score, their order is determined by their position in the second sorted list. The sorted templates are written sequentially into the priority index field of the template library, and a new matching priority number is assigned to each template, with higher priority numbers for templates placed earlier in the sequence. During the generation of demand response strategies for the next teaching cycle, templates are selected from the template library in descending order of matching priority number for matching, and the selected templates are used as part of the candidate response strategy set for the next cycle.