Intelligent classroom interaction supervision method

Abstract

The invention relates to a method for supervising intelligent classroom interaction. The method comprises the following steps: dividing a classroom into micro-interaction units, capturing behavior characteristics in each unit, establishing behavior vectors, and recording specific participants and interaction modes; constructing an interaction unit time sequence, forming a behavior log graph, and performing time sequence modeling on multi-unit behavior vectors; identifying a state evolution curve of an individual, establishing an interaction state map, and identifying a state mutation point or a continuous low value interval as a supervision trigger point; identifying the logic integrating degree of the behaviors through a context analysis model; based on a behavior intention reasoning module, judging whether the behavior is in cooperation, coping, interference or simulation; introducing a system to judge abnormal states of a high-frequency low-value behavior, a content-free but high-frequency behavior and a theme-deviating behavior; continuously observing the change of the state of the student after intervention, and performing associated learning on intervention measures and effects to form a teacher personal intervention model library; the system recommends the most effective supervision measures to assist teachers to make decisions.

Description

Technical Field

[0001] The present invention relates to a method for supervising smart classroom interaction. Background Art

[0002] While current methods for monitoring classroom interactions have gained initial popularity and application with the development of educational informatization and artificial intelligence, they still suffer from numerous shortcomings and drawbacks in real-world teaching scenarios, failing to meet the high-quality demands for a deep understanding and precise monitoring of the teaching process. First, most current monitoring methods remain at a superficial statistical stage, focusing on the number of participants or responses. Their underlying logic is based on the existence of data rather than its meaningfulness. Consequently, the system can often only determine whether a student participated in a particular interaction, but cannot determine whether the interaction was relevant to the teaching content, whether it reflected cognitive depth, or whether there was a genuine thought process. In a classroom voting task, as long as a student clicks an option, the system will count it as valid participation. Regardless of whether the click was a blind choice or whether the student understood the question, the system cannot further distinguish between them. The resulting classroom portrait is highly distorted. Second, existing methods generally ignore the structural nature of the interaction process. That is, they focus more on the interaction results and rarely analyze the interaction paths, interaction methods, and multi-round communication relationships. This makes the system unable to identify low-quality, one-way interactions that appear active but lack feedback loops, and even more unable to determine whether an interaction has generated cognitive resonance between teachers and students or authentic knowledge collaboration between students. In addition, most current mainstream systems use the entire class or a certain link as a unit for behavior monitoring. This time-based division method ignores the natural heterogeneity of teaching activities and cannot divide the classroom into micro-interaction units that are truly teaching-meaningful. As a result, teaching activities are artificially coarsened and the behavior monitoring results lack pertinence and comparability.

[0003] In terms of behavior recognition algorithm models, most current smart classroom monitoring systems use rule-based matching or shallow machine learning models to achieve interaction recognition and behavior classification. These systems lack a deep understanding of classroom context and semantic modeling capabilities, making it difficult to effectively process natural language expressions and student behavioral intentions in complex teaching contexts. Recognition accuracy particularly declines in non-standard interactive scenarios such as subjective questions, open-ended discussions, and group collaboration. In a class where students are expected to express their opinions independently, their responses often exhibit a high degree of diversity due to individual language abilities, expression habits, and levels of subject understanding. Traditional systems are unable to make in-depth semantic judgments on these expressions and can only make rough assessments based on sentence length or keyword matching. These systems can easily misjudge perfunctory responses that are poor in content but well-formed as high-quality participation, leading to systematic biases in behavior assessment and classroom quality judgments. Furthermore, existing supervision methods have a serious lack of understanding of students' interactive motivations. Most systems only focus on the behavior itself and ignore the motivations behind it. They are unable to judge whether a behavior is due to active cognitive participation, herd imitation, mechanical operation or deliberate perfunctory. This is a serious gap in teaching regulation, because a behavior may seem to be participating, but in fact it may have no teaching value, and may even mislead teachers to make wrong judgments about the students' true status.

[0004] In addition, the current system also suffers from a broken supervisory feedback chain. Even if some systems can identify students' atypical behaviors or abnormal states, they often lack automatic feedback mechanisms and intervention strategy recommendations. Teachers still need to intervene based on experience. The system cannot achieve a complete closed loop from problem identification to problem solving, nor can it provide personalized, data-driven teaching decision support. Especially when faced with the evolution of student behavior trends, most current methods can only handle the judgment of single behaviors, and lack the ability to model behavioral continuity and state changes. They are unable to identify the trajectories of students whose behaviors are gradually degenerating or whose participation forms are sliding towards low quality, and are even unable to predict the risks of their future development that may seriously deviate from teaching objectives. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for supervising smart classroom interaction, thereby solving some of the drawbacks and shortcomings pointed out in the background technology.

[0006] The method for supervising smart classroom interaction is characterized by comprising the following steps:

[0007] S1. Classroom behavior modeling using interactive granularity units:

[0008] S1.1. Divide the class into micro-interaction units, which include a question and answer session, a discussion round, and an operational task;

[0009] S1.2. Capture the behavioral characteristics of each unit, including response time, number of responders, and complexity of interaction paths; establish a behavior vector for each unit, recording the specific participants and interaction methods; construct a time series of interaction units to form a behavior log map;

[0010] S2. Dynamic interactive state recognition and evolution trajectory tracking:

[0011] S2.1. Conduct temporal modeling of each student's multi-unit behavior vectors; identify individual state evolution curves, including high engagement → boredom and silence → low engagement → passive activation behavior patterns;

[0012] S2.2. Establish a student interaction status map, including enthusiastic, follower, indifferent, and confrontational types; identify status mutation points or persistent low value intervals as regulatory trigger points;

[0013] S3. Introducing a collaborative recognition mechanism for abnormal interactive behavior determination:

[0014] S3.1. Identify the logical fit of behavior through a contextual analysis model that integrates the teacher's context and the previous and subsequent student behaviors;

[0015] S3.2, based on the behavioral intention inference module, determines whether the behavior is motivated by cooperation, coping, interference, or imitation; and labels the behavior with interaction motivations, including active interaction, ambiguous response, perfunctory response, and malicious interference;

[0016] S3.3. Introduce the system to determine: abnormal states of high-frequency low-value behavior, low-content but high-frequency behavior, and off-topic behavior;

[0017] S4. Adopting dynamic supervision strategy for adaptive optimization:

[0018] S4.1. Every supervisory action taken by the teacher, including issuing reminders to a student or changing the interaction method, is recorded. Continuous observation is made to see whether the student's status improves or their engagement recovers after the intervention.

[0019] S4.2. Associate intervention measures with their effects to form a library of teacher-specific intervention models. In similar situations, the system recommends the most effective supervisory measures to assist teachers in decision-making.

[0020] Furthermore, the classroom behavior modeling method of the interactive granularity unit includes:

[0021] The introduction of micro-interaction unit modeling enables coarse-grained supervision logic based on entire lessons or chapters. A micro-interaction unit is defined as the smallest teaching interaction unit in the classroom with a clear start and end time, a concrete interaction theme, and specific participants. This includes a question and response, a round of student discussion, and a question or operation task. By jointly detecting voice changes, interactive system response events, and content switching behaviors in the classroom timeline, nodes with behavioral boundary characteristics in the teaching process are extracted to construct multiple continuous interaction segments of the classroom.

[0022] After the teacher asks a question, the starting time of the utterance T can be detected s , and when the student makes the first response or completes the relevant operation, the response time point T is recorded r ; Based on the activity inhibition effect in the comprehensive interaction, the normalized response time index is defined:

[0023]

[0024] in:

[0025] R t is the normalized response time index, which represents the average response agility of the student group to a certain interactive unit; T r Indicates the timestamp of the first valid student response; T s It is the timestamp of the teacher or system issuing the interactive instruction; |P a | represents the number of students who actively responded in the unit; α is the response adjustment parameter, which is used to control the impact of the number of responses on the normalization degree of response time; ln(1+|P a |) is a logarithmic smoothing function introduced to alleviate the excessive pull on the denominator caused by the growth of the group population.

[0026] Furthermore, the classroom behavior modeling method of the interactive granularity unit includes:

[0027] The deep behavioral feature extraction of the internal structure focuses on the complexity of the interaction path and the degree of multi-round transformation of knowledge transfer. In discussion or collaborative tasks, the interaction involves multiple rounds of speeches, alternating feedback between teachers and students, and horizontal questions between students. The path complexity index C is introduced to quantify the complexity. i , defined as follows:

[0028]

[0029] in:

[0030] C i represents the weighted value of the path complexity of all speech behaviors in the i-th interaction unit; n is the total number of speeches or operations in the interaction unit; h iIndicates the activity intensity of the i-th behavior, which is calculated based on the speech duration, language density, and speech energy factors; d i It is the time difference between the current behavior and the previous behavior, reflecting the degree of coherence between behaviors; δ is the influencing factor of the behavior interval. The higher the value, the stronger the penalty for speech interruption.

[0031] Furthermore, the classroom behavior modeling method of the interactive granularity unit includes:

[0032] The key parameters of each interaction unit are encoded into a single high-dimensional vector to construct a standard behavior representation:

[0033] V u = <R t ,N r ,C i ,μ(P),θ(M),γ(F s )>

[0034] in:

[0035] V u is the structured behavior vector of the u-th interaction unit; R t is the response time index; N r The number of responses is the number of students who participated in substantive interactions in the unit; C i is the complexity of the interaction path; μ(P) is the participant distribution function, which measures the balance of participation between teachers and students, and among students; θ(M) is the interaction mode weight function, in which different types of interaction modes, including voice questions, multiple-choice questions, and drag-and-drop tasks, are given different regulatory weights; γ(F s ) is the asynchronous participation factor, marking whether the interaction has late submission, repeated submission, or passive execution of non-real-time behavior.

[0036] Furthermore, the high-dimensional vector uniformly models all interactive units in each class, compares students horizontally and classrooms horizontally, and supports longitudinal trend analysis; all behavior vectors are combined in time series to construct the classroom interaction evolution trajectory S = {V1, V2, ..., V n}, and draw classroom behavior log maps based on the trajectories for multi-dimensional supervision visualization, abnormal behavior prediction, and teaching strategy feedback and evaluation.

[0037] Furthermore, the abnormal interactive behavior determination mechanism method of introducing collaborative identification:

[0038] Based on the contextual semantic analysis model, the logical fit of the behavior is identified. The teacher's current teaching context information, including teaching content, question topics, and task instructions, is captured in real time. The student interaction record sequence within the time window before and after the behavior is combined to construct a behavior time sequence chain and calculate the semantic consistency score between the student behavior and the teacher's current context through a time-sensitive weight function. The consistency function is defined as:

[0039]

[0040] in:

[0041] S c represents the semantic fit score between a student's behavior and the current teacher's teaching content, with a value range of 0 and 1. The closer to 1, the better the fit. β is the semantic deviation penalty coefficient, which is used to adjust the degree of decline in overall fit when there is semantic deviation in multiple dimensions. ω k is the time-sensitive weight of the k-th context dimension feature item; is a semantic distance function used to measure the behavior B generated by students at time t t and teacher contextual content C t The degree of content difference or disconnection on the kth semantic dimension.

[0042] Furthermore, the abnormal interactive behavior determination mechanism method of introducing collaborative identification:

[0043] Based on the behavioral intention reasoning module, we conduct motivation analysis on students' interactive behaviors and comprehensively consider the semantic intensity of speech content, the changing trend of behavioral patterns, and the multi-dimensional characteristics of expression frequency and methods to construct the behavioral intention mapping function as follows:

[0044]

[0045] in:

[0046] I d is the inferred behavior motivation label value, which is mapped to discrete labels by the nonlinear function ψ(·), including cooperative, coping, interfering, and imitating. In the integral expression, t0 to t1 represents the time window in which the behavior occurs; Γ(B t ) is the semantic validity function of the behavior content, which is used to determine whether the behavior contains substantive content related to the teaching objectives; Δ(B t ,B t-1 ) indicates whether the behavior continuously imitates or repeats the previous behavior in time; λ1 and λ2 are the behavior content contribution weight and behavior variability weight, respectively, which are used to regulate the system when judging motivation.

[0047] Furthermore, the abnormal interactive behavior determination mechanism method of introducing collaborative identification:

[0048] Each student's interactive behavior is labeled with a corresponding motivation label, including active interaction, vague response, perfunctory response, and malicious interference, to establish a classification system for behavior quality. Using labeled behaviors as input, the abnormal aggregation analysis function is used to identify whether students have continuous or structural deviant behaviors. The abnormal behavior intensity function is constructed as follows:

[0049]

[0050] in:

[0051] E s is the cumulative intensity score of abnormal behaviors identified in the class cycle; u represents the u-th interactive behavior, n is the total number of students' behaviors in the class; χ u It is the interactive activity factor of the behavior, which is used to reflect the intensity of the behavior in terms of time and frequency; is the semantic fit score of the behavior; To judge the value of the behavior intention; function It is a nonlinear penalty function defined for disruptive and perfunctory behaviors. η is the behavioral risk adjustment coefficient, which is used to adjust the tolerance and sensitivity of the system.

[0052] When E s When the cumulative value exceeds the preset safety threshold, the student is marked as a potential abnormal interactor and an early warning is triggered on the teacher side to assist in timely teaching intervention, including questioning guidance, private chat and roll call, and interactive adjustment operations.

[0053] The supervision method of smart classroom interaction of the present invention has the following beneficial effects: First, by introducing the micro-interaction unit modeling method, the method transforms the traditional coarse-grained supervision model with the entire class as the analysis unit into a fine-grained analysis logic based on specific teaching events, making the recording and analysis of teaching behaviors more targeted and structured, and significantly improving the expression accuracy and analyzability of classroom behavior data; constructing a multi-dimensional behavior feature extraction system, including response time, number of participants, path complexity, interaction mode, and participant distribution indicators, which can achieve multi-angle measurement of interaction quality, breaking through the previous limitations of only counting low-dimensional data such as the number of answers and click frequency, and effectively portraying the true depth of students' participation.

[0054] Secondly, by introducing the contextual semantic analysis and behavioral intention reasoning module, it is possible to accurately identify whether the interaction is in line with the topic and whether the behavior is real. It can distinguish between positive interaction, vague response, perfunctory response, and malicious interference behavior types, improve the classroom supervision's ability to identify the value of teaching, and prevent pseudo-active phenomena from misleading teaching judgments; establish an abnormal behavior aggregation recognition mechanism and a behavior intensity scoring function, sustainably monitor the evolution trend of individual behavior, and form a structured early warning model for high-frequency, low-quality, repetitive imitation, and off-topic behaviors, realizing a leap from single-point identification to behavior chain prediction, and enhancing the foresight and intervention efficiency of teaching management.

[0055] Finally, this method also has data-driven adaptive capabilities. By recording teacher intervention behaviors and student status changes, it builds a personalized regulation model library, promotes the teaching strategy from experience-based to data-feedback-based, and truly realizes the intelligent closed loop of supervision-feedback-optimization, providing a visual, structured, and intelligent solution for teaching evaluation, classroom regulation, and personalized teaching support in the smart classroom environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is a flow chart of the supervision method for smart classroom interaction of the present invention.

[0057] Figure 2 This is a flow chart of the classroom behavior modeling method for interactive granularity units of the present invention.

[0058] Figure 3 The present invention introduces a flow chart of the abnormal interactive behavior determination mechanism method for collaborative identification. DETAILED DESCRIPTION

[0059] The following is a detailed description of the specific embodiments of the present invention with reference to the accompanying drawings.

[0060] The entire class or the entire teaching content in the traditional sense is deconstructed in fine granularity and reconstructed into a set of micro-interaction units with time boundaries and behavioral boundaries. Each unit represents an independent interactive behavior with teaching value. Specifically, S1.1 is to divide a question and answer, a round of group discussion or a student operation task teaching segment into micro-interaction units with clear starting and ending points through the semantic feature recognition of classroom interactive events. The system defines the time range and topic focus of each unit by identifying the starting time of the teacher's question, the student's first response or the behavioral boundary of the task completion mark, thereby dividing a class into quantifiable fine-grained behavioral units; then, S1.2 further captures the key behavioral characteristics of each unit, including response time (that is, the time interval from the teacher initiating the interaction to the student's first response, used to evaluate the response efficiency), the number of responders (referring to the number of students who have made substantial interactive behaviors in the unit, used to measure the breadth of participation) and the complexity of the interaction path ( The system evaluates whether there are multiple rounds of speeches, whether there are horizontal interactions between students, and whether feedback closed-loop interaction depth elements are generated). Based on the above characteristics, the system establishes a standardized behavior vector for each micro-interaction unit. The behavior vector not only contains quantitative parameters such as time, frequency, and complexity, but also records the specific participant identity (student ID, teacher role), interaction method (voice speech, text answering, operation task), and further organizes the behavior vectors of all units into an interaction unit time series in chronological order, and finally generates a complete classroom behavior log map. This map not only reflects the structure and rhythm of the entire classroom interaction, but also serves as the core data basis for subsequent teaching analysis, abnormal behavior identification, and interaction strategy feedback, to achieve in-depth supervision and quantitative expression of classroom behavior.

[0061] The interactive behavior during the entire class is modeled and analyzed as a continuously changing dynamic process. The introduction of the time dimension realizes the identification of behavioral evolution and the intelligent judgment of intervention timing. Among them, S2.1 is specifically reflected in the time series modeling of the behavioral vector sequence formed by each student in multiple micro-interaction units in the classroom. The system will organize these behavioral vectors into a continuous data stream in chronological order. With the help of time sliding windows, time series clustering or behavioral sequence encoding methods, the state change trend of each student in the classroom process is captured, and then the state evolution curve of the individual student is drawn. The curve can present multiple change paths. The high participation state gradually transitions to burnout silence (manifested by a decrease in response frequency and reduced interaction quality), or is activated by a certain teaching event from a long-term low participation state and briefly upgraded to a passive interaction state. The above evolutionary trajectory provides teachers with a more explanatory behavioral portrait than static labels; then in S2.2, Based on the performance characteristics of students in time-series behavior data, the system further divides students into behavioral types and establishes a student interaction status map. Common types include enthusiastic (active speaking, continuous response), follower (responding under the teacher's guidance, but lacking initiative), indifferent (long-term low-frequency participation or empty behavior) and confrontational (showing resistance, interference, and detached interaction). Through map visualization, teachers can quickly grasp the behavioral distribution of different student groups in the classroom. At the same time, the system will use statistical mutation recognition algorithms or behavior density analysis to identify mutation points in state evolution (sudden silence, sudden activity) or continuous low value intervals (a student has not generated any interaction for a long time), and regard these changes as supervision trigger points to prompt teachers to intervene or guide, thereby realizing dynamic identification of student participation status, risk prediction and strategy support, and building a smart classroom behavior supervision mechanism driven by student state evolution.

[0062] Determine whether the behavior has teaching value, whether it conforms to the current classroom context, and whether the motivation behind it is positive. Identify real abnormal interactions through the fusion of multi-source behavior and semantic information. First, in S3.1, the system uses a contextual analysis model that integrates the teacher's context and the previous and subsequent student behaviors to identify whether the student's current behavior is logically consistent, that is, to determine whether the behavior responds to the task or question raised by the teacher, whether it is at the right time node, and whether it is semantically consistent with the teaching content. When making the identification, the system will combine the teacher's current teaching content, the classroom semantic focus, and the student's behavior in the previous few minutes. Record, build contextual semantic association map, calculate the semantic deviation of student behavior in this context, and thus draw a conclusion on whether the behavior is timely and appropriate to the topic; then in S3.2, the system introduces the behavior intention reasoning module to further explore the motivation behind the interactive behavior that has been identified as formally compliant, and systematically analyzes the richness of the behavior content, the depth of the language structure, the similarity between the behavior and others' behavior, repetitiveness, and the frequency of behavior initiation, and infers whether the behavior is cooperative (for promoting classroom communication), coping (perfunctory response to complete the task), disruptive (deliberately disrupting the interactive order), or imitative The system then labels the behavior as active interaction (active, valuable, and responsive), vague response (partially relevant but with deviation in expression), perfunctory response (low quality, high repetition, empty content), or malicious interference (obviously disrupting classroom order). This labeling system lays the foundation for subsequent behavior quality analysis and early warning mechanisms. Finally, in S3.3, the system further introduces aggregation logic to the overall behavior data for abnormality judgment, by identifying the following three typical abnormal behavior states: First, high-frequency low-value behavior, that is, although students participate many times, their behavioral motivations are mostly perfunctory. Or imitation, lack of substantial contribution; second, no content but high frequency behavior, that is, there are a large number of clicks, speeches or submissions, but the content of the behavior is empty or irrelevant to teaching; third, off-topic behavior, that is, semantically irrelevant to the current teaching content, and there is an active departure from the classroom context. The system identifies the above behavior pattern as an abnormal state, and forms a behavior risk score through a multi-dimensional analysis model. When the score exceeds the threshold, it will issue an early warning to the teacher, thereby realizing the intention recognition and value judgment beneath the surface of the behavior, and building an intelligent supervision mechanism that can identify non-teaching value participation, effectively improving the quality of classroom interaction and the accuracy of teacher intervention.

[0063] In combination with the above, the causal relationship between the intervention behaviors taken by teachers in the classroom and the changes in students' status is captured digitally and continuously learned, so that the supervision system not only has the ability to observe and identify, but also can make intelligent responses and strategy recommendations in future situations. First, in S4.1, the system will record all the supervision behaviors taken by teachers in the classroom. These behaviors include but are not limited to reminders, questions, task adjustments, and changes in interaction forms for a certain student. Each supervision behavior will be marked and stored in the behavior log together with the triggering reason, implementation time, target student, and classroom link information. The system will then continue to track the changes in student status after the supervision behavior occurs, especially focusing on the changing trends of key indicators in the student's behavior vector, whether the participation frequency has increased, whether the response time has shortened, whether the quality of speech has increased, and whether the interaction motivation label has changed from negative to positive. Through this process, the system completes the direct transformation from intervention behavior to student change. Causal chain modeling; on this basis, S4.2 archives all historical supervision behaviors and their corresponding results and establishes a behavior-result mapping model. The system is trained based on long-term accumulated data to form a personalized intervention strategy model library for teachers. The model can capture the effective or ineffective means used by each teacher when facing a specific type of student status (long-term indifference, sudden confrontation, and staged coping), and establish their personal stylized intervention portrait. When similar situations or behavior patterns appear in future classrooms, the system will call the model and actively recommend to teachers the most effective supervision measures in history. It is recommended to use voice prompts instead of text reminders, or guide students to complete collaborative tasks instead of independent answers, so as to achieve precision and personalization of supervision suggestions, and finally build an intelligent teaching decision-making assistance system with teacher behavior as the core and adaptive evolution, so that supervision can move from one-way judgment to closed-loop optimization, and continuously improve the efficiency of classroom interaction and the scientific nature of teacher decision-making.

[0064] Example 1:

[0065] In a 40-minute math class at a smart middle school, teacher Lin was using the smart classroom platform to teach a chapter on the application of derivatives. The class consisted of 35 students, 32 of whom had brought and used smart learning devices, and the system was fully integrated with the platform. At the 15th minute, Lin introduced a real-time interactive task: asking students to think about and answer the question: "What is the function f(x) = x?" 3 -3x is the extreme point in the interval [-2,2], and explains the meaning of the extreme value. The platform system detects that Mr. Lin’s voice command is issued at 15 minutes and 5 seconds, which is recorded as the starting timestamp of this interactive unit T s = 905 seconds (from the beginning of the class), the student terminal starts counting down. The time it takes for the system to receive the first student's effective response is 15 minutes and 20 seconds, that is, T r=920 seconds. This response comes from student Xiao Ming, who submitted the function derivative and accurately pointed out the extreme point.

[0066] The system detected that within 40 seconds, 13 students submitted answers that could be considered valid participants. Among these 13 students, 5 submitted correct answers. The remaining 8 students submitted answers that were not completely correct, but were well-structured and well-formatted. They were therefore identified as active responses by the platform algorithm. a |=13, which represents the actual number of actively participating students in this micro-interaction unit. To avoid unreasonable compression of response time due to an excessively large denominator when there are many participants, the system designs a logarithmic function for smoothing and introduces an adjustment parameter α as a control coefficient for the weight of the number of participants on the denominator. In this teaching scenario, the system sets a default α=0.35, and this value is set in the range of α∈[0.2,0.5]. When you want to enhance the impact of the number of participants on the response speed, you can increase it moderately, otherwise reduce it. According to the defined normalized response time formula:

[0067]

[0068] Substituting the data into:

[0069]

[0070] Since ln(14)≈2.639, we can calculate:

[0071]

[0072] Therefore, the normalized response time R of the interaction unit is t ≈7.8. This value can be used as a reference for evaluating the agility of interactive response. It is recorded in the behavior vector of the micro-interaction unit and compared horizontally with the response of subsequent interactive units to analyze fluctuations in student activity.

[0073] At the 28th minute of the course, Teacher Lin organized a group collaborative discussion titled "The Application of Newton's Method in Real Scenarios," intending to stimulate students' high-level thinking and cooperative communication awareness through a task-driven approach. This interactive session was systematically divided into the fourth micro-interaction unit, starting at T s =1680 seconds, end time T e = 1740 seconds, with a duration of 60 seconds. The system detected a total of 9 valid speech records during this period, from representative members of 4 groups and Teacher Lin himself. Since this discussion task is an open question + cross-group feedback, it has the typical characteristics of teacher-student alternation, multiple rounds of speeches, and horizontal interaction between students. The path complexity index C was calculated for this micro-interaction unit. i , thereby more truly quantifying the depth and structural quality of this discussion interaction.

[0074] By definition, the path complexity index is expressed as follows:

[0075]

[0076] The system records n = 9 consecutive speeches, and for each action the system calculates the corresponding activity intensity h i This value is standardized into a floating-point number between [0.5, 1.0] after comprehensive evaluation based on the student's speech energy (obtained by sound intensity sampling), speech duration (in seconds), and semantic density (estimated sentence effective component density through natural language processing models). The first speech was by student A, who explained the Newtonian modeling problem of the velocity change of an object in an inhomogeneous medium. The speech was clear, took 13 seconds, and the language structure was complete. The system calculated h1 = 0.92. The second speech was by student B, who quickly added a sentence without considering whether the resistance was approximately constant. Although compact, the semantics were simple, and the score was h2 = 0.61. And so on. The system also detects the time interval d between each behavior. i , which is the number of seconds from the end of the i-1th speech to the beginning of the i-th speech. The interval between the first and second speeches was 2.5 seconds, d2 = 2.5. Due to a delay in the group discussion, there was a 6.8-second silence between the fourth and fifth speeches, d5 = 6.8, significantly lengthening the behavior chain. To strengthen the penalty for silent periods or interruptions in speech, the system uses a control factor δ, which ranges from δ∈[0.05, 0.2]. In this class, δ was set to 0.12. This value is used to adjust the inhibitory effect of time intervals on complexity scores.

[0077] Taking the fifth speech as an example, the activity level of this behavior is h5 = 0.83, and the time interval is d5 = 6.8. Substituting some terms into the formula, we get:

[0078]

[0079] It can be seen that even if the quality of the behavior content is high, the contribution of this item to the overall complexity is greatly weakened due to the long time interval, which reflects the algorithm's focus on coherence. The system calculates all 9 speeches in this way, and the sum of each result is used to obtain the path complexity index C of the micro-interaction unit. i ≈4.96, which is marked in the behavior vector of the unit for subsequent comparative analysis.

[0080] 35 minutes into the class, Teacher Lin initiated the last interactive unit of the lesson: a drag-and-drop operation task based on derivative image recognition. Students were required to match the image curve with its corresponding first-order derivative function image. The platform recorded the start and end time of this unit as 2100 seconds to 2160 seconds, a total of 60 seconds. The system detected that 21 students submitted drag-and-drop results, of which 18 completed the operation, and 3 submitted but the system determined that the format was incorrect or the content was irrelevant and was classified as an invalid response. The number of valid responses was N. r =18, the system synchronously recognizes the complexity of the interaction path C in this unit i =2.42, the value is low, mainly because this task is a one-way submission type, which does not form a teacher-student dialogue or an interactive chain between students. At the same time, the response time index R t It is 8.1, slightly higher than the average level, indicating that the task requires a certain amount of cognitive processing and students generally take a long time to complete it.

[0081] Next, the system encodes the core behavior parameters of the interactive unit into a unified high-dimensional behavior vector V5 to support subsequent horizontal comparison and vertical evolution analysis. The vector structure is defined as follows:

[0082] V u = <R t ,N r ,C i ,μ(P),θ(M),γ(F s )>

[0083] In this scenario, it is known that R t =8.1, N r =18, C i = 2.42. The remaining parameters are calculated and explained below: The participant distribution function μ(P) is used to measure the balance of teacher-student participation. In this unit, students perform tasks, and the teacher only assigns tasks and does not participate in interactive behaviors. Therefore, the system sets this function to be the discreteness index of the participant structure vector, with a defined range of μ(P)∈[0,1]. Values closer to 0 indicate that the interaction is concentrated among a few people, while values closer to 1 indicate that the teacher-student distribution is evenly distributed. In this task, students participate widely, and the teacher does not interact directly. Therefore, μ(P)≈0.84 is calculated.

[0084] The interaction mode weight function θ(M) is used to represent the weight of the current task in supervision. Different interaction forms in the platform have basic weight settings, with a value range of θ(M)∈[0.5,1.5], where voice questions are set to 1.2, multiple-choice questions are 1.0, image dragging is 1.1, and long text answers are 1.4. In this unit, due to the image operation interaction with medium cognitive load, the system assigns θ(M)=1.1. Asynchronous participation factor γ(F s) is used to mark whether there is non-real-time behavior in the interaction. If there is late submission, repeated submission, or asynchronous answering, the indicator will be increased. The value range is γ(F s )∈[0,1], 0 is completely real-time, 1 is completely asynchronous. In this task, two students completed the submission after the interaction time due to delayed network reasons, and one student repeated the operation twice. The system determines that the asynchronous feature accounts for a high proportion in this task. Let γ(F s )=0.67.

[0085] The final behavior vector of this interactive unit is:

[0086] V5=<8.1,18,2.42,0.84,1.1,0.67>

[0087] The vector is then incorporated into the classroom behavior sequence trajectory S = {V1, V2, V3, V4, V5} by the system. Each vector records the structured data of an independent interaction unit. The system uses the sequence to draw a classroom behavior log map. The map uses time as the horizontal axis and dimensional projection characteristics (response time, path complexity, participation density) as the vertical axis, showing the overall interaction trend of the classroom. For example, in the first two units, students are highly active and teachers are highly involved. The last two units turn to task-based operations, with path complexity decreasing but the number of participants increasing. The system identifies the structural evolution of the transition from a guided classroom to an operational classroom, and issues prompts and suggestions for individual units with high asynchronous rates.

[0088] Example 2:

[0089] 38 minutes into the class, Teacher Lin assigned an open-ended reflection task: "Please summarize in one sentence the limitations of Newton's method for finding function extrema." This task aims to test students' understanding of the applicable conditions and potential risks of Newton's method, such as improper initial value selection and zero second-order derivatives, which can lead to algorithm failure. This task was delivered via voice and displayed simultaneously in the text prompt area on the student's terminal. The system identified the task's teaching context as a mathematical method reflection-type open-ended subjective question and constructed its semantic model content vector C. t , including keywords: Newton's method, extreme value, restriction, condition, failure situation, and marking the context timestamp t = 2280 seconds.

[0090] Within 90 seconds after the interactive task was issued, a total of 23 students submitted their answers via voice or text. Among them, student Zhang's answer was: Newton's method can quickly find the minimum value of a function, especially when the derivative is zero. The system determined that his language was fluent and the format was compliant, but there was a semantic deviation in the content. Because the context intended to require students to point out the limitations of Newton's method, Zhang's statement overly focused on its advantages. The system needed to determine whether this behavior was a vague response or superficial participation.

[0091] The abnormal interactive behavior collaborative identification mechanism first evaluates the logical fit between the behavior and the current teaching context through the contextual semantic analysis model, and constructs a temporal behavior chain {B t-2 ,B t-1 ,B t ,B t+1}, and found that Zhang's answers in the first two interactive units also showed clichés and lacked targeted content, further increasing the possibility of semantic disconnection in the current behavior. The system then scored the degree of fit based on the semantic consistency function:

[0092]

[0093] Substituting parameters into explanations is as follows: Assume that the number of semantic dimensions m = 4, which are topic keywords, intention direction, concept depth, and logical coherence. The system calculates the content difference degree of each dimension through BERT+LDA semantic modeling. They are 0.3, 0.6, 0.7, and 0.5 respectively, indicating that Zhang's answer is acceptable in terms of topic fit, but there are significant deviations in the dimensions of intention direction and conceptual depth. The system assigns time-sensitive weights ω to each dimension. k They are 0.2, 0.4, 0.3, and 0.1 respectively, which means that the current system pays more attention to the semantic fit of intention direction and concept depth to prevent students from getting away with simple keyword matching. At the same time, the semantic deviation penalty coefficient β is set to 1.8, which is taken from the recommended range β∈[1.0,2.5] and can be set according to the classroom tolerance.

[0094] Substitute the data into the formula:

[0095]

[0096] Finally, the system determines the semantic fit score S of the behavior c ≈0.50, which is lower than the preset effective interaction threshold (the system threshold is 0.65), and is marked as low-fit behavior. Combined with its behavioral motivation label (previously high repetitiveness and general content), it is classified as vague response + low content value. This label will be used as a reference factor in subsequent abnormal behavior analysis.

[0097] After determining the semantic fit, the system activates the behavioral intention reasoning module to identify the true motivation behind Zhang's behavior, so as to distinguish whether his answer is due to genuine participation, strategic coping, imitation of others, or deliberate interference, thereby improving the accuracy of identifying abnormal interactive behavior. In the interactive time window corresponding to this task, Zhang completed his answer between the 2280th and 2340th seconds, and the system extracted his behavioral content vector B in this time interval. t Combined with the previous behavior B t-1, evaluate its semantic change trend and the substance of the expression content, and then call the behavioral motivation function:

[0098]

[0099] In this scenario, the system defines the time window t0 = 2280 seconds, t1 = 2340 seconds, which is the complete process interval for Zhang to complete the reflection task. First, Γ(B t ) is evaluated, which represents the semantic validity score of Zhang's answer. The system scores based on three dimensions: vocabulary coverage, key concept matching, and logical structure clarity. The result is Γ(B t )=0.38, which is lower than the standard effective behavior threshold (0.60) set in the system, indicating that although the behavior is complete, the substance is thin. Then the system evaluates Δ(B t ,B t-1 ), that is, the semantic repetition between the speech in the previous round (the short answer Newton method in the third interactive unit is very useful), with the help of semantic vector similarity comparison and text structure matching results, the score is Δ(B t ,B t-1 )=0.84, indicating that Zhang’s current behavior is to repeat the previous statement in a different way.

[0100] The settings of the behavior weight coefficients λ1 and λ2 directly affect the system's judgment of content quality and repetitiveness. Based on the task-oriented characteristics of classroom interaction goals, this task is reflective and critical thinking. The system recommends increasing attention to content and setting the value range λ1∈[0.6,0.9] and λ2∈[0.1,0.4]. Here, λ1=0.75 and λ2=0.25 are set to ensure that the semantic quality weight is much higher than the proportion of repetition penalty.

[0101] After substituting the data into the behavioral motivation integral formula:

[0102]

[0103] The integration time interval is 60 seconds, and the result is ψ(0.495×60)=ψ(29.7). This value is fed into the nonlinear classification function ψ(·) as the intermediate expression vector. The system uses a sigmoid-like mapping function with a threshold activation mechanism to convert the continuous value into a discrete behavioral motivation label. The final output result is coping, which means that the system identifies this behavioral motivation as tending to complete tasks with low investment, which is consistent with Zhang's past behavioral trajectory of repetitive content and weak semantic expression.

[0104] The tag value I d = Coping style and previous semantic fit score S c= 0.498 is also fed into the behavioral risk model to generate an anomaly label: ambiguous response + coping motivation. The system identifies this as a typical example of low content value + high repetition risk. On the teacher's teaching dashboard, Zhang's profile picture will be marked with a yellow alert status, accompanied by system suggestions: It is recommended to assign a targeted follow-up task, guiding him to illustrate examples of failure scenarios of Newton's method in determining extreme values, or to transform his answers into voting material for discussion to stimulate cognitive conflict.

[0105] Previously, the system used the semantic fit function to determine that the answer score of the Newton method's limited subjective questions was only S. c =0.498. The motivational reasoning module calculates the student's motivation type as coping based on its semantic validity and similarity to previous behaviors. It then labels the interaction behavior with a combination of fuzzy response and coping motivation. The system inputs this label, along with the previous four interactions, into the abnormal behavior clustering identification mechanism and calls the abnormal behavior intensity function to calculate the degree of behavioral deviation of the student throughout the entire class cycle. To achieve accurate measurement, the system activates the following function:

[0106]

[0107] During the 40-minute teaching process of this class, Zhang participated in 6 interactive behaviors (i.e., n = 6), including 1 voice question answering, 1 multiple-choice question submission, 2 image drag-and-match tasks, 1 group discussion speech, and 1 subjective reflection question answering. The system annotated these 6 behaviors with the following data: In each behavior, the activity factor χ u It is calculated based on the participation position of the behavior (whether to answer the question quickly), the response speed (whether to submit within the time limit specified by the system), and the duration of the behavior. The system value range is χ u ∈[0.5,1.5], indicating that from low activity to high activity, Zhang’s specific interaction activity factors are 1.0, 0.8, 0.9, 1.2, 1.1, and 1.3, respectively, and the corresponding semantic fit scores are They are 0.72, 0.65, 0.61, 0.52, 0.47 and 0.498 respectively; and the system has identified Zhang's behavioral motivation types in the previous chapter as: imitation type, coping type, fuzzy response type, imitation type, coping type, and coping type.

[0108] To evaluate the severity of the intention behind the behavior, the system introduces a nonlinear penalty function The scores are given according to the classification penalty coefficient table, where active interaction is assigned to 0, vague response is assigned to 0.6, coping is assigned to 1.0, interference is assigned to 1.5, and imitation is assigned to 0.8. The score range is It is also used as a nonlinear weight input for abnormal risk; the system sets the behavioral risk adjustment coefficient η = 2.0 for the current class according to the complexity of the class, and the adjustable range is set to η∈[1.0,3.0]. This value controls the system's sensitivity to high-risk labels. The higher the value, the lower the tolerance.

[0109] Substitute the above data into the formula one by one and perform integral approximate summation (considered as discrete accumulation):

[0110] First interaction:

[0111] Second interaction:

[0112] Third interaction:

[0113] Fourth interaction:

[0114] 5th interaction:

[0115] 6th interaction:

[0116] Add up the items to get the total abnormal behavior intensity value:

[0117] E s ≈0.108+0.093+0.159+0.222+0.194+0.217=0.993

[0118] The system preset safety threshold is T safe =0.85, the threshold can be adjusted in the range of [0.7, 1.2], and is configured according to the class type and teaching fault tolerance requirements. In this example, due to E s =0.993>0.85, the system identifies Zhang as a potential abnormal interactor and triggers a mild warning on the teacher side, marking his profile picture in orange and providing intervention suggestions: if the student's behavior is labeled as coping or imitative for three consecutive times and the semantic fit is continuously lower than 0.55, it is recommended to conduct guided intervention by calling on the student to ask questions, limiting the answer format, and switching the interaction form.

[0119] At the same time, this information is also recorded by the system into Zhang's classroom behavior profile and used as a model learning sample in the next round of prediction and recommendation. For example, in the next lesson with a similar structure, if the system identifies Zhang's ambiguous responses and high-frequency interactions within the first 10 minutes of class, the risk factor will be increased in advance, helping teachers identify potential behavioral trends of formal participation but cognitive dissonance.

[0120] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. The supervision method of smart classroom interaction is characterized by The following steps are involved: S1. Classroom behavior modeling using interactive granularity units: S1.

1. Divide the class into micro-interaction units, which include a question and answer session, a discussion round, and an operational task; S1.

2. Capture the behavioral characteristics of each unit, including response time, number of responders, and complexity of interaction paths; establish a behavior vector for each unit, recording the specific participants and interaction methods; construct a time series of interaction units to form a behavior log map; S2. Dynamic interactive state recognition and evolution trajectory tracking: S2.

1. Conduct temporal modeling of each student's multi-unit behavior vectors; identify individual state evolution curves, including high engagement → boredom and silence → low engagement → passive activation behavior patterns; S2.

2. Establish a student interaction status map, including enthusiastic, follower, indifferent, and confrontational types; identify status mutation points or persistent low value intervals as regulatory trigger points; S3. Introducing a collaborative recognition mechanism for abnormal interactive behavior determination: S3.

1. Identify the logical fit of behavior through a contextual analysis model that integrates the teacher's context and the previous and subsequent student behaviors; S3.2, based on the behavioral intention inference module, determines whether the behavior is motivated by cooperation, coping, interference, or imitation; and labels the behavior with interaction motivations, including active interaction, ambiguous response, perfunctory response, and malicious interference; S3.

3. Introduce the system to determine: abnormal states of high-frequency low-value behavior, low-content but high-frequency behavior, and off-topic behavior; S4. Adopting dynamic supervision strategy for adaptive optimization: S4.

1. Every supervisory action taken by the teacher, including issuing reminders to a student or changing the interaction method, is recorded. Continuous observation is made to see whether the student's status improves or their engagement recovers after the intervention. S4.

2. Associate intervention measures with their effects to form a library of teacher-specific intervention models. In similar situations, the system recommends the most effective supervisory measures to assist teachers in decision-making.

2. The method for supervising smart classroom interaction according to claim 1 is characterized in that The classroom behavior modeling method of the interactive granularity unit includes: The modeling method of micro-interaction units is introduced to enable coarse-grained supervision logic based on the entire class or chapter; the micro-interaction unit is defined as the smallest teaching interaction unit in the classroom with a clear start and end time, a concrete interaction theme and a specific participating subject; through the joint detection of voice changes, interactive system response events, and content switching behaviors in the classroom timeline, nodes with behavioral boundary characteristics in teaching are extracted.

3. The method for supervising smart classroom interaction according to claim 2 is characterized in that The classroom behavior modeling method of the interactive granularity unit includes: The deep behavioral feature extraction of the internal structure focuses on the complexity of the interaction path and the degree of multi-round transformation of knowledge transfer. In discussion or collaborative tasks, the interaction involves multiple rounds of speeches, alternating feedback between teachers and students, and horizontal questions between students. The path complexity index C is introduced to quantify the complexity. i , defined as follows: in: C i represents the weighted value of the path complexity of all speech behaviors in the i-th interaction unit; n is the total number of speeches or operations in the interaction unit; h i Indicates the activity intensity of the i-th behavior, which is calculated based on the speech duration, language density, and speech energy factors; d i It is the time difference between the current behavior and the previous behavior, reflecting the degree of coherence between behaviors; δ is the influencing factor of the behavior interval. The higher the value, the stronger the penalty for speech interruption.

4. The method for supervising smart classroom interaction according to claim 3 is characterized in that The classroom behavior modeling method of the interactive granularity unit includes: The key parameters of each interaction unit are encoded into a single high-dimensional vector to construct a standard behavior representation: V u = <R t ,N r ,C i ,μ(P),θ(M),γ(F s )> in: V u is the structured behavior vector of the u-th interaction unit; R t is the response time index; N r The number of responses is the number of students who participated in substantive interactions in the unit; C i is the complexity of the interaction path; μ(P) is the participant distribution function, which measures the balance of participation between teachers and students, and among students; θ(M) is the interaction mode weight function, in which different types of interaction modes, including voice questions, multiple-choice questions, and drag-and-drop tasks, are given different regulatory weights; γ(F s ) is the asynchronous participation factor, marking whether the interaction has late submission, repeated submission, or passive execution of non-real-time behavior.

5. The method for supervising smart classroom interaction according to claim 4 is characterized in that The high-dimensional vector uniformly models all interactive units in each class, compares students horizontally and classes horizontally, and supports longitudinal trend analysis; all behavior vectors are combined in time series to construct the classroom interaction evolution trajectory S = {V1, V2, ..., V n }, and draw classroom behavior log maps based on the trajectories for multi-dimensional supervision visualization, abnormal behavior prediction, and teaching strategy feedback and evaluation.

6. The method for supervising smart classroom interaction according to claim 1 is characterized in that The abnormal interactive behavior determination mechanism method introducing collaborative identification: The context-based semantic analysis model identifies the logical fit of behavior and captures the teacher's current teaching context information in real time, including teaching content, question topics, and task instructions. Combined with the sequence of student interaction records in the time window before and after the behavior, the behavior time sequence chain is constructed and the semantic consistency score between the student behavior and the teacher's current context is calculated through a time-sensitive weight function.

7. The method for supervising smart classroom interaction according to claim 6 is characterized in that The abnormal interactive behavior determination mechanism method introducing collaborative identification: Based on the behavioral intention reasoning module, we conduct motivation analysis on students' interactive behaviors and comprehensively consider the semantic intensity of speech content, the changing trend of behavioral patterns, and the multi-dimensional characteristics of expression frequency and methods to construct the behavioral intention mapping function as follows: in: I d is the inferred behavior motivation label value, which is mapped to discrete labels by the nonlinear function ψ(·), including cooperative, coping, interfering, and imitating. In the integral expression, t0 to t1 represents the time window in which the behavior occurs; Γ(B t ) is the semantic validity function of the behavior content, which is used to determine whether the behavior contains substantive content related to the teaching objectives; Δ(B t ,B t-1 ) indicates whether the behavior continuously imitates or repeats the previous behavior in time; λ1 and λ2 are the behavior content contribution weight and behavior variability weight, respectively, which are used to regulate the system when judging motivation.

8. The method for supervising smart classroom interaction according to claim 7 is characterized in that The abnormal interactive behavior determination mechanism method introducing collaborative identification: Each student's interactive behavior is labeled with a corresponding motivation label, including: active interaction, vague response, perfunctory response, and malicious interference, in order to establish a classification system for behavior quality; the labeled behavior is used as input, and the abnormal aggregation analysis function is used to identify whether the student has continuous and structural deviant behavior.

Images