A teaching method and system for ideological and political education assisted by AI intelligent agents in higher vocational colleges

Abstract

This invention relates to the field of artificial intelligence-assisted teaching technology, specifically to a method and system for AI-assisted teaching in ideological and political education in higher vocational colleges. The method includes collecting passive behavior data and active participation records of students regarding knowledge nodes; identifying a set of silent students who have passive behavior records but no active participation; calculating cognitive attention scores based on dwell time, repeated browsing, and anonymous confusion marking; aggregating and generating cognitive hotspot scores and marking hotspot nodes; evaluating the quality of teacher questions and triggering follow-up questions when the quality is insufficient; calculating follow-up question relevance scores using hotspot nodes as the target domain and combining semantic similarity with cognitive hotspot scores; calculating the activation rate based on the proportion of silent students actively participating after follow-up questions, and iteratively updating the weight coefficients in the cognitive attention calculation. This invention enables precise allocation of follow-up question resources and adaptive evolution of model parameters, significantly improving the overall interactive coverage level of ideological and political education classrooms.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence-assisted teaching technology, specifically to a teaching method and system for ideological and political education using AI intelligent agents in higher vocational colleges. Background Technology

[0002] AI-assisted instruction has evolved from early rule-driven computer-aided instruction systems and expert-system-based intelligent tutoring systems to the current technological evolution of conversational educational agents with large language models as their core engine. At present, technological combinations such as retrieval-enhanced generation, multi-agent collaboration, and knowledge graphs have enabled educational agents to demonstrate significant capabilities in areas such as knowledge question answering, personalized content delivery, and learning path planning. Against this backdrop, AI agents are widely used in higher education classrooms to enhance the depth and breadth of teacher-student interaction, particularly in the field of ideological and political education, where various agent systems have emerged focused on dialogue guidance, classroom follow-up questioning, and learning analysis.

[0003] Existing AI-assisted ideological and political education teaching systems typically use students' active participation data as the core data source for perceiving students' cognitive state. Based on this, they generate suggestions for teachers to ask follow-up questions or teaching intervention strategies. Their advantage lies in their ability to conduct real-time cognitive analysis of students participating in discussions and generate targeted Socratic questioning chains, helping teachers guide superficial viewpoints to deeper value reflections, thereby improving the quality of classroom discussions to some extent. However, the aforementioned system has a structural deficiency: in actual ideological and political education classes, silent students who do not actively speak account for 40% to 60% of the total number of students. This group does not generate any active interaction data that can be collected by the system during the entire teaching process. As a result, the learning model constructed by the AI ​​agent can only cover the few students who actively speak, resulting in significant data bias. The follow-up questioning chain and interaction strategies generated based on this biased data are actually only aimed at students who have participated in the discussion. The implicit cognitive confusion and unexpressed value judgment doubts that are common among silent students are always outside the perception range of the AI ​​agent and cannot be transformed into targeted basis for classroom follow-up questioning and value guidance. Ultimately, this leads to a systemic deficiency in the AI ​​agent's teaching assistance effectiveness at the overall classroom coverage level.

[0004] Therefore, there is an urgent need for a method to solve the problems existing in the current technology. Summary of the Invention

[0005] To address the aforementioned shortcomings of existing technologies, this invention provides a teaching method and system for ideological and political education assisted by AI intelligent agents in higher vocational colleges, which can effectively solve the problems mentioned in existing technologies.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] This invention provides a teaching method assisted by AI intelligent agents for ideological and political education in higher vocational colleges, comprising the following steps:

[0008] S100. In ideological and political education courses, collect passive behavior data of each student on each knowledge node; the passive behavior data includes data on the duration of each student's stay on each knowledge node, data on the number of times they repeatedly browsed the knowledge node, and data on anonymous confusion submissions; and simultaneously collect records of each student's active participation.

[0009] S200. Within a preset time window, detect the number of active participation records for each student. When the number of active participation records for a student is zero and there is at least one non-zero record in the passive behavior data, the student is marked as a silent student. All students who meet the above conditions are aggregated to form a set of silent students.

[0010] S300. Based on the data of dwell time, number of repeated browsings and anonymous confusion submission markers of each student in the set of silent students at each knowledge node, calculate the cognitive attention score of each student at each knowledge node using a preset weighted calculation rule.

[0011] S400. Statistically aggregate the cognitive attention scores of each student in the set of silent students on the same knowledge node to obtain the cognitive hotspot score of each knowledge node; combine the cognitive hotspot scores of each knowledge node to generate a cognitive hotspot distribution vector; mark the knowledge nodes whose cognitive hotspot scores exceed the preset hotspot activation threshold as hotspot nodes.

[0012] S500: Real-time collection of teacher question text; calculation of question quality score based on the structural features of the teacher question text and historical activation statistics using a preset weighted calculation rule; triggering follow-up question generation process when the question quality score is lower than a preset question quality threshold.

[0013] S600. In response to the triggering of the follow-up question generation process, a set of candidate follow-up questions is generated with the hotspot node as the target domain; based on the semantic similarity between each candidate follow-up question and the hotspot node, the follow-up question relevance score of each candidate follow-up question is calculated using the cognitive hotspot score of the hotspot node as a weighting factor; the candidate follow-up question set is sorted in descending order according to the follow-up question relevance score, and candidate follow-up questions whose ranking does not exceed the preset push limit are pushed to the teacher terminal.

[0014] S700. After pushing the candidate follow-up question, detect whether each student in the set of silent students has actively participated in the hot spot node; calculate the activation rate of each hot spot node based on the proportion of students in the set of silent students whose cognitive attention score exceeds the preset attention threshold who have actively participated in the behavior; iteratively update the weight coefficients in the preset weighted calculation rule based on the activation rate, and apply the updated weight coefficients to the cognitive attention score calculation in the next preset time window.

[0015] Furthermore, the calculation of each student's cognitive attention score at each knowledge node using a preset weighted calculation rule specifically includes:

[0016] Calculate the cognitive attention score S(i,k) of each student i in the set of silent students at each knowledge node k using the following formula:

[0017] ;

[0018] Where t(i,k) represents the duration of student i's stay at knowledge node k. This represents the maximum dwell time of all students across all knowledge nodes during the current course. Here, r(i,k) is the normalized value of the dwell time data, used to measure the student's sustained engagement with the knowledge node; r(i,k) is the number of times student i repeatedly browses the knowledge node k. This represents the maximum number of repeated views within the current course. c(i,k) is the normalized value of the repeated browsing count data, used to measure the strength of students' willingness to actively review knowledge nodes; c(i,k) is the anonymous confusion submission label data of student i for knowledge node k, which takes a preset positive value or zero, used to mark the explicit confusion signal of students for knowledge nodes. , , The preset weight coefficients in the preset weighted calculation rules, and ;

[0019] The passive behavior data of students outside the set of silent students are not included in the calculation of the cognitive attention score S(i,k).

[0020] Furthermore, the statistical aggregation of the cognitive attention scores of each student in the set of silent students on the same knowledge node specifically includes:

[0021] The arithmetic mean of the cognitive attention scores S(i,k) of each student in the set of silent students at knowledge node k is calculated to obtain the cognitive hotspot score H(k) of knowledge node k, as shown in the following formula:

[0022] ;

[0023] in, S(i,k) represents the number of students in the set of silent students, and S(i,k) represents the cognitive attention score of student i in the set of silent students at knowledge node k.

[0024] The value of the cognitive hotspot score H(k) reflects the average implicit attention of the silent student set to knowledge node k. The higher the value of H(k), the stronger the concentration of cognitive confusion of knowledge nodes in the silent student set. When the cognitive hotspot score H(k) exceeds the preset hotspot activation threshold, the knowledge node is marked as a hotspot node, and the hotspot node constitutes the target domain when the candidate follow-up question set is generated.

[0025] Furthermore, the calculation of the question quality score based on the structural features of the teacher's question text and historical activation statistics specifically includes:

[0026] Calculate the question quality score Q using the following formula:

[0027] ;

[0028] Wherein, O is the openness score, obtained by classifying the question structure of the teacher's question text using a pre-trained classifier, with the lowest preset value for closed-ended questions, the middle preset value for semi-open-ended questions, and the highest preset value for fully open-ended questions; L is the cognitive level score, obtained by normalizing the teacher's question text after mapping it to the corresponding level of a preset cognitive classification system using a pre-trained classifier; P is the historical activation prediction score, obtained by retrieving historical course records, acquiring historical questions whose semantic similarity to the teacher's question text exceeds a preset similarity threshold, and normalizing the average activation rate generated by the historical questions on the set of historical silent students. , , The preset weight coefficients for each item and ;

[0029] When the question quality score Q is lower than the preset question quality threshold, it is determined that the teacher's question text has insufficient activation potential for the set of silent students, and the follow-up question generation process is triggered.

[0030] Furthermore, the step of calculating the relevance score of each candidate follow-up question based on the semantic similarity between each candidate follow-up question and the hotspot node, using the cognitive hotspot score as a weighting factor, specifically includes:

[0031] For the candidate follow-up question q' in the candidate follow-up question set and the hotspot node k, the follow-up question relevance score is calculated according to the following formula. :

[0032] ;

[0033] in, The semantic embedding vector for the candidate follow-up question q'; The semantic embedding vector of the hotspot node k; for and The inner product, and These are the modulus lengths of the two, for and Cosine similarity; The cognitive hotspot score of the hotspot node k is used as a weighting factor to give higher recommendation priority to candidate follow-up questions that semantically hit the cognitive confusion cluster of the silent student set.

[0034] When a candidate follow-up question q' is associated with multiple hotspot nodes, the follow-up question relevance score of the candidate follow-up question q' for each of the hotspot nodes is calculated separately. A weighted sum is performed, and the result of the weighted sum is used as the final follow-up relevance score of the candidate follow-up question q', which is then used in the descending sorting.

[0035] Furthermore, the iterative update of the weight coefficients based on the activation rate specifically includes:

[0036] When the activation rate is higher than the preset activation rate threshold, the difference between the activation rate and the preset activation rate threshold is used as a positive supervision signal, and the weight coefficient corresponding to the passive behavior data type with a higher activation contribution is increased through the gradient update rule.

[0037] When the activation rate is lower than or equal to a preset activation rate threshold, the absolute value of the difference between the activation rate and the preset activation rate threshold is used as a negative supervision signal, and the weight coefficient of the passive behavior data type corresponding to insufficient activation effect is reduced through the gradient update rule.

[0038] The updated weight coefficients satisfy the constraint that the sum of all terms equals a preset constant, and the updated weight coefficients take effect in the calculation of the cognitive attention score in the next preset time window.

[0039] Furthermore, the anonymous confusion submission tag data is collected in the following way: an anonymous confusion submission entry is set up on the learning interface corresponding to each knowledge node; when a student submits confusion text through the anonymous confusion submission entry, the confusion text is only recorded internally by the system and is not displayed in plain text to other students or the teacher's terminal; semantic clustering is performed on the confusion texts submitted by different students in the set of silent students under the same knowledge node, and the confusion texts with semantic similarity exceeding a preset clustering threshold are merged into the same confusion entry; the anonymous confusion submission tag data of the corresponding knowledge node is updated according to the number of merged confusion entries, and the updated anonymous confusion submission tag data participates in the calculation of the cognitive attention score in the next round.

[0040] Furthermore, when a student in the set of silent students actively participates after the candidate follow-up question is pushed, the student is marked as having actively participated and the number of active participation records is accumulated; at the beginning of the next preset time window, the student is removed from the set of silent students based on the updated number of active participation records; the cognitive attention score and cognitive hotspot score of each knowledge node are recalculated based on the removed set of silent students, and the cognitive hotspot distribution vector and the marking results of hotspot nodes are updated synchronously.

[0041] Furthermore, when a teacher selects and uses a candidate follow-up question from the pushed candidate follow-up questions on the teacher terminal, the candidate follow-up question is marked as a selected candidate follow-up question; the mapping relationship between the selected candidate follow-up question and the corresponding hot spot node, and the activation rate of the hot spot node are combined as a supplementary supervision signal to adjust the semantic embedding vector used in the subsequent follow-up question relevance score calculation process, so that candidate follow-up questions associated with the hot spot node and with a high historical selection rate will receive priority improvement in the subsequent push ranking.

[0042] A vocational college ideological and political education AI-assisted teaching system includes:

[0043] The data acquisition module is used to collect passive behavior data of each student at each knowledge node in ideological and political education courses; the passive behavior data includes data on the time each student spends at each knowledge node, data on the number of times they repeatedly browse, and data on anonymous confusion submissions; and simultaneously collects records of each student's active participation.

[0044] The silent student identification module is used to detect the number of active participation records of each student within a preset time window; when the number of active participation records of a student is zero and there is at least one non-zero record in the passive behavior data, the student is marked as a silent student, and all students who meet the above conditions are aggregated to form a silent student set.

[0045] The cognitive analysis module is used to calculate the cognitive attention score of each student at each knowledge node based on the data of dwell time, number of repeated browsings, and anonymous confusion submission tags of each student in the set of silent students, using a preset weighted calculation rule.

[0046] The cognitive hotspot construction module is used to statistically aggregate the cognitive attention scores of each student in the silent student set on the same knowledge node to obtain the cognitive hotspot score of each knowledge node; combine the cognitive hotspot scores of each knowledge node to generate a cognitive hotspot distribution vector; and mark the knowledge nodes whose cognitive hotspot scores exceed a preset hotspot activation threshold as hotspot nodes.

[0047] The question quality assessment module is used to collect teachers' question texts in real time; based on the structural features of the teachers' question texts and historical activation statistics, it calculates the question quality score using a preset weighted calculation rule; when the question quality score is lower than a preset question quality threshold, it triggers the follow-up question generation process.

[0048] The follow-up question generation module is used to generate a set of candidate follow-up questions in response to the triggering of the follow-up question generation process, with the hot spot node as the target domain; calculate the follow-up question relevance score of each candidate follow-up question based on the semantic similarity between each candidate follow-up question and the hot spot node, using the cognitive hot spot score of the hot spot node as a weighting factor; sort the candidate follow-up question set in descending order according to the follow-up question relevance score, and push the candidate follow-up questions whose ranking does not exceed the preset push limit to the teacher terminal;

[0049] The activation feedback and model update module is used to detect whether each student in the set of silent students actively participates at the hot spot nodes after the candidate follow-up question is pushed; calculate the activation rate of each hot spot node based on the proportion of students in the set of silent students whose cognitive attention score exceeds a preset attention threshold who actively participate; iteratively update the weight coefficients in the preset weighted calculation rules based on the activation rate, and apply the updated weight coefficients to the cognitive attention score calculation in the next preset time window.

[0050] The technical solution provided by this invention has the following advantages compared with the known prior art:

[0051] This invention systematically identifies and constructs a set of silent students by collecting passive behavioral data such as dwell time, repeated browsing, and anonymous submission of questions, combined with active participation records. Based on this, it calculates cognitive attention scores and cognitive hotspot scores, enabling the quantification of the real cognitive confusion of the silent student group that has long been ignored by existing AI systems. This effectively solves the problem of bias in learning analysis caused by the bias of active participation data in traditional methods, and significantly improves the coverage of AI-assisted teaching across the entire class.

[0052] This invention introduces a closed-loop feedback mechanism based on the activation rate of hotspot nodes. It dynamically adjusts the weighting coefficient of passive behavior data in the calculation of cognitive attention based on the actual active participation of silent students after follow-up questioning. This mechanism enables the system to automatically optimize model parameters for the behavioral characteristics of different courses and student groups, eliminating the need for repeated manual parameter tuning and significantly improving the method's generalization and adaptability in different vocational college ideological and political education classrooms. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0054] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0055] Figure 2 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0057] The present invention will be further described below with reference to embodiments.

[0058] Example 1:

[0059] Reference Figure 1 This embodiment provides a teaching method assisted by AI intelligent agents for ideological and political education in higher vocational colleges, which specifically includes the following steps:

[0060] S100. Data collection: In ideological and political education courses, collect passive behavior data of each student towards each knowledge node, and simultaneously collect records of each student's active participation.

[0061] A knowledge node is the smallest addressable learning unit of course content. In actual deployment, a knowledge node can correspond to a chapter or paragraph in the courseware, a time segment in the video, or an interactive question group. The specific division is configured by the teacher in the system management terminal before class, and the system records the interaction data between each student and each node using the node number as an index.

[0062] Passive behavior data includes three categories: ① Dwell time data t(i,k), which is the cumulative dwell time (in seconds) of student i on the page or media clip corresponding to knowledge node k; ② Repeat browsing count data r(i,k), which is the number of times student i revisits knowledge node k within a course; ③ Anonymous confusion submission tag data c(i,k), which is the semaphore by which student i actively tags confusion for knowledge node k through the anonymous confusion submission portal (see the following description for specific collection methods). The above three types of data are reported in real time by the learning client and aggregated and stored by the server. The data collection itself uses conventional front-end tracking and server-side logging techniques.

[0063] Active participation records include, but are not limited to, explicit participation events that can be identified by the system, such as students speaking, answering questions, voting, and posting comments in the classroom interaction system. Each active participation behavior corresponds to a record written into the student's participation log.

[0064] S200. Silent student identification: Within a preset time window, detect the number of active participation records for each student; when a student's active participation record count is zero and there is at least one non-zero record in their passive behavior data, mark the student as a silent student; and summarize all students who meet the above conditions to form a silent student set.

[0065] The preset time window can be set to the duration of a single lesson (e.g., 45 minutes), or teachers can flexibly configure it as a rolling window of several minutes on the system management terminal to adapt to different classroom rhythms.

[0066] The key to this step lies in defining silence: simply not speaking is insufficient to classify a student as silent; simultaneously, at least one non-zero record must exist in the passive behavior data, meaning the student actually interacted with a knowledge node (e.g., spent some time on a node, repeatedly viewed a node, or submitted anonymous questions), but did not actively participate. This dual condition ensures the accurate identification of students truly in a state of implicit confusion, while excluding absent students who did not participate at all or those who did not report data due to network failures, thus preventing them from interfering with the accuracy of subsequent hotspot analysis.

[0067] S300. Cognitive attention score calculation: Based on the data of each student's dwell time, number of repeated browsings, and anonymous confusion submission tags at each knowledge node in the silent student set, the cognitive attention score of each student at each knowledge node is calculated using a preset weighted calculation rule.

[0068] Calculate the cognitive attention score S(i,k) of each student i in the set of silent students at each knowledge node k using the following formula:

[0069] ;

[0070] Where t(i,k) is the time student i spends on knowledge node k; Tmax is the maximum time all students spend on all knowledge nodes in the current course. The normalized value of dwell time, ranging from [0,1], is used to measure the duration of students' attention to knowledge nodes. The longer the dwell time, the more attention resources students invest in that node, which is an indirect signal of potential cognitive confusion.

[0071] r(i,k) represents the number of times student i repeatedly browses knowledge node k; Rmax: the maximum number of times the knowledge node is repeatedly browsed within the current course. The normalized value for the number of repeated views, ranging from [0,1], is used to measure the strength of students' willingness to actively review knowledge nodes. Repeated views often mean that students have difficulty understanding the content of the node and need to confirm it multiple times to advance their learning.

[0072] c(i,k) represents the anonymous confusion submission tag data of student i for knowledge node k, taking a preset positive value (e.g., 1.0) or zero. When a student actively tags confusion for this node through the anonymous confusion submission entry, c(i,k) takes a preset positive value; otherwise, it takes zero. This item is the most direct explicit confusion signal among the three types of data, so it can be appropriately assigned a higher weight in the weight setting.

[0073] , , These are the preset weighting coefficients in the preset weighted calculation rules, and The initial values ​​can be set to equal weights (i.e., ... Alternatively, settings can be differentiated based on prior experience, and the activation rate feedback mechanism of the S700 steps can be automatically iterated and optimized.

[0074] It should be noted that passive behavior data from students outside the silent student group are not included in the calculation of cognitive attention scores. This limitation ensures that subsequent hotspot analysis focuses on the group that truly experiences implicit confusion, and avoids the behavioral data of students who have actively participated from shifting the distribution of hotspots.

[0075] S400, Cognitive Hotspot Construction: Statistically aggregate the cognitive attention scores of each student in the set of silent students on the same knowledge node to obtain the cognitive hotspot score of each knowledge node; combine the cognitive hotspot scores of each knowledge node to generate a cognitive hotspot distribution vector; mark the knowledge nodes whose cognitive hotspot scores exceed the preset hotspot activation threshold as hotspot nodes.

[0076] The arithmetic mean of the cognitive attention scores S(i,k) of each student in the set of silent students at knowledge node k is used to obtain the cognitive hotspot score H(k) of knowledge node k:

[0077] ;

[0078] in, Let S be the number of students in the silent student set, and S(i,k) be the cognitive attention score of student i in the silent student set at knowledge node k.

[0079] H(k) represents the average implicit attention of the set of silent students to knowledge node k. The higher H(k) is, the stronger the concentration of cognitive confusion about that knowledge node in the set of silent students, that is, more students have a high degree of implicit confusion about that node at the same time.

[0080] Arrange the H(k) values ​​of all knowledge nodes in the course in order of node number and combine them into a cognitive hotspot distribution vector. This vector intuitively reflects the spatial distribution of cognitive confusion among silent students across the entire course, and can be used for post-class teaching review and course content optimization.

[0081] When the H(k) of a certain knowledge node exceeds the preset hotspot activation threshold When this happens, the node is marked as a hotspot node. Hotspot activation threshold. The system administrator can configure this based on the course size and historical data; for example, the 75th percentile of the historical H(k) distribution can be used as a reference value. Hotspot nodes constitute the target domain for the subsequent follow-up question generation process, ensuring that follow-up question resources are concentrated on the knowledge gaps where confusion is most pronounced.

[0082] S500, Question Quality Assessment

[0083] The system collects teachers' question texts in real time; it calculates the question quality score based on the structural characteristics of the question texts and historical activation statistics using a preset weighted calculation rule; and it triggers the follow-up question generation process when the question quality score is lower than the preset question quality threshold.

[0084] Teachers' questions are collected in real time through their terminals, which can be PC-based teaching software or mobile applications, providing an input interface for the question text. The system immediately initiates a question quality assessment after the teacher submits the question, and the entire process is transparent to the teacher and does not interrupt the teaching flow.

[0085] Calculate the question quality score Q using the following formula:

[0086] ;

[0087] Here, O represents the openness score, obtained by classifying the question structure of the teacher's questions using a pre-trained classifier. Closed-ended questions (such as yes / no or true / false questions) are assigned a preset minimum value (e.g., 0); semi-open-ended questions (such as "why" questions with limited answer ranges) are assigned a preset middle value (e.g., 0.5); and fully open-ended questions (such as open-ended discussion questions like "What are some possible reasons for...?") are assigned a preset maximum value (e.g., 1.0). The O value measures the potential of the questions to stimulate deeper thinking among students at the structural level.

[0088] L represents the cognitive level score, obtained by normalizing the teacher's questions after mapping them to the corresponding levels of a pre-trained cognitive classification system using a pre-trained classifier. This pre-trained classification system can employ Bloom's Taxonomy of Objectives, progressively divided into six levels: memory, comprehension, application, analysis, evaluation, and creation. Lower-level questions (memory, comprehension) are mapped to lower score ranges, while higher-level questions (analysis, evaluation, creation) are mapped to higher score ranges. After normalization, L ranges from [0,1]. The L value measures the degree to which the questions stimulate students' higher-order thinking abilities.

[0089] P represents the historical activation prediction score. It is obtained by retrieving historical course records and acquiring historical questions whose semantic similarity to the teacher's question text exceeds a preset similarity threshold. The score is obtained by normalizing the average activation rate generated by the historical questions on the set of historical silent students. The P value is used to measure the historical driving effect of similar questions on silent students and is the key to connecting historical teaching data with current teaching decisions. If there is insufficient historical data in the early stage of system deployment, P can take a default intermediate value (such as 0.5). After historical data accumulates to a reliable sample size, it can be gradually replaced by the measured average value.

[0090] , , The preset weight coefficients for each item and It can be configured according to the course type and teacher style. Its meaning is similar to the weight coefficient in S(i,k), and it can also be adjusted through iterative optimization.

[0091] When Q is lower than the preset question quality threshold Time (e.g.) When Q is set to 0.6 (i.e., Q < 0.6 when the full score is 1.0), it is determined that the current teacher's question has insufficient activation potential for the set of silent students. The system then triggers the follow-up question generation process to actively intervene and provide assistance.

[0092] S600, Follow-up Question Generation and Push

[0093] In response to the triggering of the follow-up question generation process, a set of candidate follow-up questions is generated with hotspot nodes as the target domain. Based on the semantic similarity between each candidate follow-up question and the hotspot node, the follow-up question relevance score of each candidate follow-up question is calculated using the cognitive hotspot score of the hotspot node as a weighting factor. The follow-up questions are sorted in descending order of follow-up question relevance score, and the candidate follow-up questions whose ranking does not exceed the preset push limit are pushed to the teacher's terminal.

[0094] The generation of candidate follow-up questions is based on hotspot nodes as the target domain: the system inputs the node text and cognitive hotspot distribution vector of the hotspot nodes into the large language model, and the instruction model generates several candidate follow-up questions around the core content of each hotspot node. The design goal of the candidate follow-up questions is to guide silent students to actively express their core confusion about the hotspot nodes. The upper limit of the number of candidate follow-up questions generated is configured by the system, generally generating 20 to 50 candidate follow-up questions for subsequent screening to ensure diversity after ranking.

[0095] The core innovation of this step lies in the calculation of the relevance score for follow-up questions. For candidate follow-up questions q' and hotspot nodes k in the candidate follow-up question set, the follow-up question relevance score is calculated using the following formula. :

[0096] ;

[0097] in, The semantic embedding vector of the candidate follow-up question q' is obtained by encoding the candidate follow-up question text through a pre-trained semantic embedding model.

[0098] The semantic embedding vector of hotspot node k is obtained by encoding the text content of the hotspot node using the same encoding model, ensuring that the two vectors are in the same semantic space.

[0099] for and The inner product (dot product); and Let be the magnitudes of the two vectors, respectively. for and The cosine similarity, with a value range of [-1, 1], is used to measure the semantic fit between the candidate follow-up question q' and the hot spot node k. The higher the cosine similarity, the more accurately the candidate follow-up question can point to the knowledge content involved in the hot spot node.

[0100] H(k) is the cognitive hotspot score of hotspot node k. As a weighting factor, it makes the candidate follow-up questions of hotspot nodes with a higher degree of cognitive confusion in semantics receive higher recommendation priority. This concentrates the limited follow-up questioning opportunities on the knowledge-weak areas that need to be activated the most, rather than distributing them evenly among all hotspot nodes.

[0101] When a candidate follow-up question q' is associated with multiple hot nodes, R(q',k) is calculated for each hot node, and then a weighted sum is performed. The weighted sum is used as the final follow-up question relevance score for the candidate follow-up question q', and it is then sorted in descending order.

[0102] ;

[0103] in, To determine the final relevance score, the candidate follow-up questions are sorted in descending order based on this score. The top 3 candidate follow-up questions, ranked no higher than a preset limit, are then pushed to the teacher's terminal in a list format for selection. This design preserves the teacher's control over the teaching process; the system serves only as an auxiliary decision-making tool and does not forcibly replace the teacher's teaching behavior.

[0104] S700, Activation Feedback and Weight Iterative Update: After pushing candidate follow-up questions, detect whether each student in the silent student set actively participates at hotspot nodes; calculate the activation rate of each hotspot node based on the proportion of students in the silent student set whose cognitive attention score exceeds the preset attention threshold who actively participate; iteratively update the weight coefficients in the preset weighted calculation rules based on the activation rate, and apply the updated weight coefficients to the cognitive attention score calculation in the next preset time window.

[0105] The specific calculation logic for the activation rate is as follows: within a certain observation period after pushing candidate follow-up questions (configurable to 5 to 10 minutes), the cognitive attention score S(i,k) of the silent student set exceeds the preset attention threshold. The proportion of students (i.e. students judged by the system to have significant latent confusion about hot nodes) who actually engage in active participation (such as answering follow-up questions, initiating bullet comments, participating in voting, etc.) is the activation rate A(k) of hot node k.

[0106] The rules for iteratively updating the weight coefficients are as follows:

[0107] When the activation rate A(k) is higher than the preset activation rate threshold At that time, A(k) and The difference (i.e.) This serves as a positive monitoring signal, increasing the weight coefficients corresponding to passive behavior data types with high activation contribution through gradient update rules. For example, if anonymous confused submission data c(i,k) has the highest correlation with the activation result within the successful activation time window, then the weight coefficients are increased accordingly. ;

[0108] When the activation rate A(k) is less than or equal to At that time, with As a negative supervision signal, the weight coefficient of the corresponding passive behavior data type is reduced by the gradient update rule when the activation effect is insufficient;

[0109] The updated weight coefficients must always satisfy The constraints can be normalized to ensure they always hold. The updated weighting coefficients take effect in the calculation of cognitive attention scores within the next preset time window, enabling continuous adaptive optimization of the system's predictive effectiveness for different classroom passive behavior data.

[0110] In one specific embodiment, an anonymous entry point for submitting questions is set up on the learning interface corresponding to each knowledge node. Students can submit text of their questions about the current knowledge node through this entry point without revealing their identity, thereby eliminating students' psychological concerns about being evaluated by classmates or teachers and thus preventing them from expressing their questions. This is especially important for ideological and political education classes in higher vocational colleges where the classroom participation atmosphere is relatively conservative.

[0111] When students submit their questions through the anonymous question submission portal, the text is recorded only within the system and is not displayed in plain text to other students' or teachers' terminals, thus ensuring the students' anonymity.

[0112] Semantic clustering is performed on the confused texts submitted by different students in the set of silent students under the same knowledge node. Confused texts with semantic similarity exceeding a preset clustering threshold are merged into the same confused entry. The anonymous confused submission tag data c(i,k) of the corresponding knowledge node is updated based on the number of merged confused entries. The updated c(i,k) participates in the calculation of the cognitive attention score in the next round. By using semantic clustering instead of simple counting, redundancy of repeated expressions can be effectively removed, making c(i,k) more accurately reflect the true confusion density, rather than simply being affected by multiple repeated submissions.

[0113] In one specific embodiment, when a student in the silent student set actively participates after a candidate follow-up question is pushed, the student is marked as having actively participated, and the number of their active participation records is accumulated. At the start of the next preset time window, the activated students are removed from the silent student set based on the updated number of active participation records. This ensures that the silent student set always accurately reflects the group of students who are still in a state of implicit confusion within the current time window, rather than continuing to include the historical confusion data of activated students.

[0114] Based on the removed set of silent students, the cognitive attention score and cognitive hotspot score of each knowledge node are recalculated, and the cognitive hotspot distribution vector and the labeling results of hotspot nodes are updated simultaneously. This enables dynamic tracking and real-time response to the distribution of cognitive hotspots in the classroom, allowing the system to maintain accurate perception of the current silent student group throughout the entire class period.

[0115] In one specific embodiment, when a teacher selects and uses a candidate follow-up question from the pushed candidate follow-up questions on the teacher's terminal, the candidate follow-up question is marked as a selected candidate follow-up question. The mapping relationship between the selected candidate follow-up question and its corresponding hot spot node, as well as the activation rate of the hot spot node, are used as supplementary supervision signals to adjust the semantic embedding vector used in the subsequent follow-up question relevance score calculation process, so that candidate follow-up questions associated with hot spot nodes and with high historical selection rates will receive higher priority in subsequent push rankings.

[0116] This mechanism introduces explicit feedback on teacher preferences, allowing the follow-up recommendation system to incorporate supervisory signals from teachers' professional judgment, in addition to activation rate-driven automatic optimization, thus further improving recommendation quality. Teachers' selection behavior is simultaneously recorded in the system's historical database, providing richer historical samples for calculating the P item (historical activation prediction score) in the question quality score Q of subsequent courses.

[0117] Example 2:

[0118] Reference Figure 2 This embodiment provides an AI-assisted teaching system for ideological and political education in higher vocational colleges, including:

[0119] The data collection module is used to collect passive behavior data of students at each knowledge node in ideological and political education courses. Passive behavior data includes data on the duration of each student's stay at each knowledge node, the number of times they repeatedly browsed the knowledge node, and the data on anonymous submissions of questions. The module also collects records of each student's active participation.

[0120] The silent student identification module is used to detect the number of active participation records of each student within a preset time window; when the number of active participation records of a student is zero and there is at least one non-zero record in their passive behavior data, the student is marked as a silent student, and all students who meet the above conditions are aggregated to form a silent student set.

[0121] The cognitive analysis module is used to calculate the cognitive attention score of each student at each knowledge node based on the data of the time each student spends at each knowledge node, the number of times they repeatedly browse the knowledge node, and the anonymous confusion submission tag data of each student in the silent student set, using a preset weighted calculation rule.

[0122] The cognitive hotspot construction module is used to statistically aggregate the cognitive attention scores of each student in the silent student set on the same knowledge node to obtain the cognitive hotspot score of each knowledge node; combine the cognitive hotspot scores of each knowledge node to generate a cognitive hotspot distribution vector; and mark the knowledge nodes whose cognitive hotspot scores exceed the preset hotspot activation threshold as hotspot nodes.

[0123] The question quality assessment module is used to collect teachers' question texts in real time; based on the structural characteristics of the teachers' question texts and historical activation statistics, it calculates the question quality score using a preset weighted calculation rule; when the question quality score is lower than the preset question quality threshold, it triggers the follow-up question generation process.

[0124] The follow-up question generation module is used to generate a set of candidate follow-up questions in response to the triggering of the follow-up question generation process, with hot nodes as the target domain; based on the semantic similarity between each candidate follow-up question and the hot nodes, the follow-up question relevance score of each candidate follow-up question is calculated using the cognitive hotspot score of the hot nodes as a weighting factor; the candidate follow-up question set is sorted in descending order according to the follow-up question relevance score, and the candidate follow-up questions whose ranking does not exceed the preset push limit are pushed to the teacher's terminal.

[0125] The activation feedback and model update module is used to detect whether each student in the silent student set actively participates at hotspot nodes after the candidate follow-up questions are pushed; the activation rate of each hotspot node is calculated based on the proportion of students in the silent student set whose cognitive attention scores exceed the preset attention threshold who actively participate; the weight coefficients in the preset weighted calculation rules are iteratively updated based on the activation rate, and the updated weight coefficients are applied to the cognitive attention score calculation in the next preset time window.

[0126] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the protection scope of the technical solutions of the embodiments of the present invention.

Claims

1. A teaching method assisted by AI intelligent agents for ideological and political education in higher vocational colleges, characterized in that, Includes the following steps: S100. In ideological and political education courses, collect passive behavior data of each student on each knowledge node; the passive behavior data includes data on the duration of each student's stay on each knowledge node, data on the number of times they repeatedly browsed the knowledge node, and data on anonymous confusion submission markers. Simultaneously collect records of each student's active participation; S200. Within a preset time window, detect the number of active participation records for each student. When the number of active participation records for a student is zero and there is at least one non-zero record in the passive behavior data, the student is marked as a silent student. All students who meet the above conditions are aggregated to form a set of silent students. S300. Based on the data of dwell time, number of repeated browsings and anonymous confusion submission markers of each student in the set of silent students at each knowledge node, calculate the cognitive attention score of each student at each knowledge node using a preset weighted calculation rule. S400. Statistically aggregate the cognitive attention scores of each student in the set of silent students on the same knowledge node to obtain the cognitive hotspot score of each knowledge node; combine the cognitive hotspot scores of each knowledge node to generate a cognitive hotspot distribution vector; mark the knowledge nodes whose cognitive hotspot scores exceed the preset hotspot activation threshold as hotspot nodes. S500: Real-time collection of teacher question texts; calculation of question quality score based on the structural features of the teacher question texts and historical activation statistics using preset weighted calculation rules; When the question quality score is lower than the preset question quality threshold, the follow-up question generation process is triggered; S600. In response to the triggering of the follow-up question generation process, a set of candidate follow-up questions is generated with the hotspot node as the target domain; based on the semantic similarity between each candidate follow-up question and the hotspot node, the follow-up question relevance score of each candidate follow-up question is calculated using the cognitive hotspot score of the hotspot node as a weighting factor. The candidate follow-up questions are sorted in descending order according to the follow-up question relevance score, and the candidate follow-up questions whose ranking does not exceed the preset push limit are pushed to the teacher's terminal. S700. After pushing the candidate follow-up question, detect whether each student in the set of silent students has taken the initiative to participate at the hot spot node; Based on the proportion of students in the set of silent students whose cognitive attention scores exceed a preset attention threshold who generate the active participation behavior, the activation rate of each hot spot node is calculated; the weight coefficients in the preset weighted calculation rules are iteratively updated based on the activation rate, and the updated weight coefficients are applied to the cognitive attention score calculation in the next preset time window.

2. The method for AI-assisted teaching in ideological and political education in higher vocational colleges according to claim 1, characterized in that, The calculation of each student's cognitive attention score at each knowledge node using a preset weighted calculation rule specifically includes: Calculate the cognitive attention score S(i,k) of each student i in the set of silent students at each knowledge node k using the following formula: ； Where t(i,k) represents the duration of student i's stay at knowledge node k. This represents the maximum dwell time data for all students across all knowledge nodes during the current course. The normalized value of the dwell time data is used to measure the student's sustained engagement with the knowledge node; r(i,k) is the number of times student i repeatedly browses the knowledge node k. This represents the maximum number of repeated views within the current course. c(i,k) is the normalized value of the repeated browsing count data, used to measure the strength of students' willingness to actively review knowledge nodes; c(i,k) is the anonymous confusion submission label data of student i for knowledge node k, which takes a preset positive value or zero, used to mark the explicit confusion signal of students for knowledge nodes. , , The preset weight coefficients in the preset weighted calculation rules, and ; The passive behavior data of students outside the set of silent students are not included in the calculation of the cognitive attention score S(i,k).

3. The method for AI-assisted teaching in ideological and political education in higher vocational colleges according to claim 1, characterized in that, The statistical aggregation of the cognitive attention scores of each student in the set of silent students on the same knowledge node specifically includes: The arithmetic mean of the cognitive attention scores S(i,k) of each student in the set of silent students at knowledge node k is calculated to obtain the cognitive hotspot score H(k) of knowledge node k, as shown in the following formula: ； in, S(i,k) represents the number of students in the set of silent students, and S(i,k) represents the cognitive attention score of student i in the set of silent students at knowledge node k. The value of the cognitive hotspot score H(k) reflects the average implicit attention of the silent student set to knowledge node k. The higher the value of H(k), the stronger the concentration of cognitive confusion of knowledge nodes in the silent student set. When the cognitive hotspot score H(k) exceeds the preset hotspot activation threshold, the knowledge node is marked as a hotspot node, and the hotspot node constitutes the target domain when the candidate follow-up question set is generated.

4. The method for AI-assisted teaching in ideological and political education in higher vocational colleges according to claim 1, characterized in that, The calculation of the question quality score based on the structural features of the teacher's question text and historical activation statistics specifically includes: Calculate the question quality score Q using the following formula: ； Wherein, O is the openness score, obtained by classifying the question structure of the teacher's question text using a pre-trained classifier, with the lowest preset value for closed-ended questions, the middle preset value for semi-open-ended questions, and the highest preset value for fully open-ended questions; L is the cognitive level score, obtained by normalizing the teacher's question text after mapping it to the corresponding level of a preset cognitive classification system using a pre-trained classifier; P is the historical activation prediction score, obtained by retrieving historical course records, acquiring historical questions whose semantic similarity to the teacher's question text exceeds a preset similarity threshold, and normalizing the average activation rate generated by the historical questions on the set of historical silent students. , , The preset weight coefficients for each item and ; When the question quality score Q is lower than the preset question quality threshold, it is determined that the teacher's question text has insufficient activation potential for the set of silent students, and the follow-up question generation process is triggered.

5. A teaching method assisted by AI intelligent agents for ideological and political education in higher vocational colleges according to claim 1, characterized in that, The step of calculating the relevance score of each candidate follow-up question based on the semantic similarity between each candidate follow-up question and the hotspot node, using the cognitive hotspot score as a weighting factor, specifically includes: For the candidate follow-up question q' in the candidate follow-up question set and the hotspot node k, the follow-up question relevance score is calculated according to the following formula. : ； in, The semantic embedding vector for the candidate follow-up question q'; The semantic embedding vector of the hotspot node k; for and The inner product, and These are the modulus lengths of the two, for and Cosine similarity; The cognitive hotspot score of the hotspot node k is used as a weighting factor to give higher recommendation priority to candidate follow-up questions that semantically hit the cognitive confusion cluster of the silent student set. When a candidate follow-up question q' is associated with multiple hotspot nodes, the follow-up question relevance score of the candidate follow-up question q' for each of the hotspot nodes is calculated separately. A weighted sum is performed, and the result of the weighted sum is used as the final follow-up relevance score of the candidate follow-up question q', which is then used in the descending sorting.

6. A teaching method assisted by AI intelligent agents for ideological and political education in higher vocational colleges according to claim 1, characterized in that, The iterative update of the weight coefficients based on the activation rate specifically includes: When the activation rate is higher than the preset activation rate threshold, the difference between the activation rate and the preset activation rate threshold is used as a positive supervision signal, and the weight coefficient corresponding to the passive behavior data type with a higher activation contribution is increased through the gradient update rule. When the activation rate is lower than or equal to a preset activation rate threshold, the absolute value of the difference between the activation rate and the preset activation rate threshold is used as a negative supervision signal, and the weight coefficient of the passive behavior data type corresponding to insufficient activation effect is reduced through the gradient update rule. The updated weight coefficients satisfy the constraint that the sum of all terms equals a preset constant, and the updated weight coefficients take effect in the calculation of the cognitive attention score in the next preset time window.

7. A teaching method assisted by AI intelligent agents for ideological and political education in higher vocational colleges according to claim 1, characterized in that, The anonymous confusion submission tag data is collected in the following way: An anonymous confusion submission entry is set up in the learning interface corresponding to each knowledge node; when a student submits confusion text through the anonymous confusion submission entry, the confusion text is only recorded internally by the system and is not displayed in plain text to other students or the teacher's terminal; semantic clustering is performed on the confusion texts submitted by different students in the set of silent students under the same knowledge node, and the confusion texts with semantic similarity exceeding a preset clustering threshold are merged into the same confusion entry; the anonymous confusion submission tag data of the corresponding knowledge node is updated according to the number of merged confusion entries, and the updated anonymous confusion submission tag data participates in the calculation of the cognitive attention score in the next round.

8. A teaching method assisted by AI intelligent agents for ideological and political education in higher vocational colleges according to claim 1, characterized in that, When a student in the set of silent students actively participates after the candidate follow-up question is pushed, the student is marked as having actively participated and the number of active participation records is accumulated. At the beginning of the next preset time window, the student is removed from the set of silent students based on the updated number of active participation records. Based on the removed set of silent students, the cognitive attention score and cognitive hotspot score of each knowledge node are recalculated, and the cognitive hotspot distribution vector and the marking results of hotspot nodes are updated synchronously.

9. A teaching method assisted by AI intelligent agents for ideological and political education in higher vocational colleges according to claim 1, characterized in that, When a teacher selects and uses a candidate follow-up question from the candidate follow-up questions pushed to the teacher terminal, the candidate follow-up question is marked as a selected candidate follow-up question. The mapping relationship between the selected candidate follow-up question and the corresponding hot spot node, as well as the activation rate of the hot spot node, are used as supplementary supervision signals to adjust the semantic embedding vector used in the subsequent follow-up question relevance score calculation process, so that candidate follow-up questions that are associated with the hot spot node and have a high historical selection rate will receive priority improvement in the subsequent push ranking.

10. A vocational college ideological and political education AI-assisted teaching system, characterized in that, include: The data acquisition module is used to collect passive behavior data of students towards each knowledge node in ideological and political education courses. The passive behavior data includes data on the time each student spends at each knowledge node, the number of times they repeatedly browse, and the data on anonymous question submissions; records of each student's active participation are collected simultaneously. The silent student identification module is used to detect the number of active participation records of each student within a preset time window; when the number of active participation records of a student is zero and there is at least one non-zero record in the passive behavior data, the student is marked as a silent student, and all students who meet the above conditions are aggregated to form a silent student set. The cognitive analysis module is used to calculate the cognitive attention score of each student at each knowledge node based on the data of dwell time, number of repeated browsings, and anonymous confusion submission tags of each student in the set of silent students, using a preset weighted calculation rule. The cognitive hotspot construction module is used to statistically aggregate the cognitive attention scores of each student in the silent student set on the same knowledge node to obtain the cognitive hotspot score of each knowledge node; combine the cognitive hotspot scores of each knowledge node to generate a cognitive hotspot distribution vector; and mark the knowledge nodes whose cognitive hotspot scores exceed a preset hotspot activation threshold as hotspot nodes. The question quality assessment module is used to collect teachers' question texts in real time; based on the structural characteristics of the teachers' question texts and historical activation statistics, a question quality score is calculated using a preset weighted calculation rule. When the question quality score is lower than the preset question quality threshold, the follow-up question generation process is triggered; The follow-up question generation module is used to generate a set of candidate follow-up questions in response to the triggering of the follow-up question generation process, with the hot spot node as the target domain; and to calculate the follow-up question relevance score of each candidate follow-up question based on the semantic similarity between each candidate follow-up question and the hot spot node, with the cognitive hot spot score of the hot spot node as a weighting factor. The candidate follow-up questions are sorted in descending order according to the follow-up question relevance score, and the candidate follow-up questions whose ranking does not exceed the preset push limit are pushed to the teacher's terminal. The activation feedback and model update module is used to detect whether each student in the set of silent students has taken initiative to participate at the hot spot node after the candidate follow-up question is pushed. The activation rate of each hotspot node is calculated based on the proportion of students in the set of silent students whose cognitive attention scores exceed a preset attention threshold who generate the active participation behavior. The weight coefficients in the preset weighted calculation rules are iteratively updated based on the activation rate, and the updated weight coefficients are applied to the cognitive attention score calculation in the next preset time window.

Images