A large model driven teaching evaluation multi-agent method

Abstract

The application discloses a large model driven teaching evaluation multi-agent method, comprising: classroom multi-modal data acquisition and preprocessing; teaching behavior feature and teacher-student emotion feature extraction, constructing a classroom feature vector dataset; constructing a priori knowledge base to realize knowledge retrieval enhancement; teaching evaluation multi-agent construction and collaborative evaluation, obtaining each agent evaluation result; multi-dimensional teaching evaluation report generation and feedback; fusion of emotional feedback for continuous learning and adaptive optimization. The application solves the technical problems of the existing teaching evaluation multi-agent, such as lack of emotion perception ability, insufficient evaluation robustness, weak self-adaptive ability and the like, realizes comprehensive perception, accurate analysis and personalized feedback of teaching behavior, supports dynamic evolution of intelligent teaching evaluation, and significantly improves the scientificity and practicality of teaching evaluation.

Description

Technical Field

[0001] This invention relates to the field of large language model intelligent agents and teaching evaluation, specifically to a large model-driven multi-agent method for teaching evaluation. Background Technology

[0002] With the deep application of artificial intelligence technology in education, teaching evaluation is transforming from traditional manual assessment to intelligent assessment. Multi-agent systems, due to their advantages of distributed processing and efficient collaboration, are gradually becoming the core technological support for intelligent teaching evaluation. The core requirement of teaching evaluation is to achieve comprehensive perception and accurate analysis of teaching behavior, teaching content, teacher-student interaction, and other dimensions, providing personalized feedback for teaching improvement.

[0003] Existing multi-agent technologies for teaching evaluation primarily focus on assessing teaching behaviors and content, lacking the perception and analysis of teacher and student emotions. However, emotional states directly impact teaching effectiveness and learning quality; ignoring emotional factors can lead to biased and inaccurate evaluation results. Furthermore, existing technologies suffer from three core problems: First, the multi-agent system architecture lacks a targeted hierarchical design, with ambiguous division of responsibilities among agents, resulting in low collaboration efficiency and difficulty adapting to diverse teaching scenarios. Second, the agent evaluation process lacks sufficient prior knowledge support, making it prone to cognitive biases that create evaluation illusions, leading to insufficient credibility and interpretability. Third, the system lacks an effective continuous learning mechanism, making it difficult to dynamically adjust evaluation strategies based on different teacher styles, student characteristics, and teaching scenarios, resulting in weak adaptive capabilities.

[0004] In existing technologies, some solutions attempt to construct intelligent agents for teaching evaluation, but most are single-agent architectures, lacking multi-dimensional collaborative evaluation capabilities. Some multi-agent solutions fail to incorporate sentiment analysis and lack robust knowledge enhancement and continuous learning mechanisms, making it difficult to meet the precise evaluation needs of complex teaching scenarios. Therefore, how to construct a multi-agent system that improves the comprehensiveness, accuracy, and adaptability of teaching evaluation is a pressing technical problem that needs to be solved. Summary of the Invention

[0005] To overcome the aforementioned shortcomings and deficiencies of existing technologies, the present invention aims to provide a large-model-driven multi-agent method for teaching evaluation. This invention optimizes the multi-agent architecture design and constructs knowledge enhancement and continuous learning mechanisms to address issues such as one-sided evaluation, inefficient collaboration, insufficient credibility, and weak adaptive capabilities in existing technologies, thereby achieving comprehensive perception, accurate analysis, and personalized feedback of teaching behaviors.

[0006] The objective of this invention is achieved through the following technical solution:

[0007] A large-model-driven multi-agent method for instructional evaluation, comprising:

[0008] Multimodal data acquisition and preprocessing in the classroom;

[0009] Extract teaching behavior characteristics and teacher-student emotional characteristics to construct a classroom feature vector dataset;

[0010] Enhanced knowledge retrieval is achieved by constructing a prior knowledge base;

[0011] The teaching evaluation multi-agent construction and collaborative assessment were carried out to obtain the evaluation results of each agent.

[0012] Generation and feedback of multi-dimensional teaching evaluation reports;

[0013] Integrating continuous learning and adaptive optimization with emotional feedback.

[0014] Furthermore, the classroom multimodal data acquisition and preprocessing includes:

[0015] Data collection: Video data is collected through classroom recording equipment, audio data is collected through audio collection equipment, and text data and interactive data are collected through the teaching interaction platform;

[0016] Preprocessing: Frame extraction, noise reduction, and format standardization are performed on the acquired video data; noise reduction, frame segmentation, and feature alignment are performed on the acquired audio data; word segmentation, stop word removal, and semantic standardization are performed on the acquired text data.

[0017] further,

[0018] The teaching behavior characteristics include teacher lecturing behavior characteristics, student participation behavior characteristics, and teacher-student interaction behavior characteristics.

[0019] The emotional characteristics of teachers and students include facial expression emotional characteristics, voice emotional characteristics, and semantic emotional characteristics.

[0020] We construct a classroom feature vector dataset by weighted integration of teaching behavior characteristics and teacher and student emotional characteristics.

[0021] Furthermore, the construction of the prior knowledge base specifically includes:

[0022] Knowledge base construction: A structured knowledge base is constructed according to five knowledge types. The data sources for the five knowledge types include the corresponding subject curriculum standards, the corresponding lesson type teaching specifications, teacher and student emotion evaluation standards, typical teaching scenario emotion cases, and teaching evaluation cases. A knowledge graph is constructed using a MySQL database for storage.

[0023] Feature and Knowledge Encoding: The Sentence-BERT model is used to encode the classroom feature vectors in the classroom feature vector dataset and each knowledge vector in the knowledge base, generating a 768-dimensional vector;

[0024] Knowledge retrieval and matching involves calculating the cosine similarity between the classroom feature vector and the 768-dimensional vector, and selecting the top five knowledge items with the highest similarity as augmented knowledge.

[0025] Furthermore, the construction and collaborative assessment of multi-agent teaching evaluation specifically includes:

[0026] The teaching evaluation system is constructed using a multi-agent hierarchical structure, comprising three layers: a strategy and coordination layer, a teaching analysis and collaboration layer, and a data perception layer. The teaching analysis and collaboration layer includes a lesson type analysis agent, a teaching behavior analysis agent, and a teaching content evaluation agent. The strategy and coordination layer includes a comprehensive decision-making agent.

[0027] Clearly define the responsibilities of each intelligent agent, including its inputs, outputs, and core duties;

[0028] Establish a collaborative mechanism process of "data sharing, task allocation, parallel evaluation, result fusion, and conflict coordination";

[0029] The integrated decision-making agent assigns evaluation tasks based on the lesson type analysis structure, and each agent conducts evaluations in parallel and synchronizes intermediate results; a weighted voting method is used to merge the intermediate results to obtain the evaluation results of each agent.

[0030] Furthermore, the multi-dimensional teaching evaluation report includes four core contents: (1) Teaching behavior evaluation (40%), including teaching behavior, student participation, teacher-student interaction scores and problem analysis; (2) Teaching content evaluation (30%), including content suitability, coverage of key points and difficulties, courseware design scores and optimization suggestions; (3) Emotional state evaluation (20%), including the trajectory of teacher and student emotional changes and the analysis of the impact of emotions on teaching effectiveness; and (4) Comprehensive suggestions (10%), including targeted teaching improvement measures.

[0031] Furthermore, the continuous learning and adaptive optimization that integrates emotion feedback specifically includes:

[0032] Feedback data collection: Collect feedback from teachers and students on the evaluation report, the trajectory of changes in classroom emotions, and data related to teaching effectiveness;

[0033] Learning sample set construction: Concatenate classroom feature vector dataset, agent evaluation results, and feedback data to construct a continuous learning sample set;

[0034] Model parameter update: The model parameters of each agent are updated by combining federated learning and meta-learning.

[0035] Evaluation criteria iteration: The evaluation criteria and weight allocation of each agent are dynamically adjusted based on continuous learning results and feedback data;

[0036] Knowledge update: Incorporate new assessment experiences and feedback knowledge into the prior knowledge base to achieve knowledge iteration.

[0037] Furthermore, a method combining federated learning and meta-learning is adopted to update the model parameters of each agent, specifically as follows:

[0038] Federated learning initialization: Each agent acts as a local node and initializes the model parameters θlocal, while the comprehensive decision agent acts as the central node and initializes the global parameters θglobal.

[0039] Local training: Each node trains the model based on the local sample set S and updates the local parameters using the stochastic gradient descent algorithm;

[0040] Parameter aggregation: The central node collects parameters from each local node and aggregates global parameters using a weighted average algorithm;

[0041] Meta-learning optimization: The MAML algorithm is used to realize knowledge transfer and optimize the model's adaptive ability.

[0042] Furthermore, the feedback from teachers and students on the evaluation report was collected through a post-class questionnaire to assess teachers' acceptance of the evaluation report and students' ratings of their emotional experience in the classroom.

[0043] The teaching effectiveness data includes data such as the accuracy rate of subsequent assignments and unit test scores.

[0044] Furthermore, when the evaluation results of each agent deviate by more than 10%, a conflict coordination mechanism is triggered, which calls upon the relevant knowledge from the knowledge enhancement module to conduct a secondary evaluation and correct the deviation.

[0045] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0046] 1. Innovatively integrating the dimension of emotion analysis, constructing a multimodal emotion feature extraction mechanism, solving the problem of one-sided assessment caused by neglecting emotional factors in existing teaching evaluation, and achieving a comprehensive perception of the teaching ecosystem;

[0047] 2. Design a hierarchical multi-agent architecture and collaboration mechanism, clarify the responsibilities and boundaries of each agent, improve the collaboration efficiency of multiple agents, adapt to teaching and evaluation scenarios of multiple lesson types, and solve the problem of inefficient collaboration of existing multi-agent agents;

[0048] 3. Construct a knowledge enhancement mechanism for emotion fusion, incorporate emotion-related knowledge into a prior knowledge base, and enhance the credibility and interpretability of the assessment through joint retrieval to alleviate the problem of assessment illusion;

[0049] 4. Establish a continuous learning loop that integrates emotional feedback, combine federated learning and meta-learning to achieve dynamic system adaptation and self-optimization, solve the problem of weak adaptive capabilities of existing technologies, and support the long-term stable operation of the system in real teaching environments. Attached Figure Description

[0050] Figure 1 This is a flowchart illustrating an embodiment of the present invention.

[0051] Figure 2 This is a flowchart illustrating the multi-agent continuous learning and adaptation process according to an embodiment of the present invention. Detailed Implementation

[0052] The present invention is further described below through specific embodiments, but the scope of protection of the present invention is not limited thereto.

[0053] Example

[0054] like Figure 1 and Figure 2 As shown, a large-model-driven multi-agent method for instructional evaluation includes the following steps:

[0055] S1 classroom multimodal data acquisition and preprocessing to obtain multimodal data.

[0056] This step is used to complete the collection and standardization of multimodal classroom data, providing high-quality data support for subsequent feature extraction. It includes the following sub-steps:

[0057] The S101 data acquisition system collects classroom video data via a high-definition recording camera (1920×1080 resolution, 30fps frame rate), teacher voice data via a lavalier microphone (44.1kHz sampling rate, 16bit bit depth), student voice data via a desktop microphone array, and courseware display and teacher-student interactive text (including answer records and question content) data via the smart teaching platform. The acquisition time is 45 minutes per class period.

[0058] S102: Video data preprocessing, using a Gaussian filtering algorithm to remove video noise, the formula is:

[0059]

[0060] Where σ is 1.5, the video is frame extracted (1 frame is extracted every 2 frames) to obtain valid image frames and complete the standardization process of the video data;

[0061] S103: Audio data preprocessing, using Wiener filtering algorithm for noise reduction, the formula is:

[0062]

[0063] Among them, P(k) is the signal power spectrum, and P(k) is the noise power spectrum. After noise reduction, the audio is framed (frame length 25ms, frame shift 10ms) to achieve preliminary alignment of audio features.

[0064] S104: Preprocess the text data. Use a word segmentation tool to segment the text, remove stop words such as "de, le, ma", and retain the core vocabulary related to teaching to complete the semantic standardization processing of the text data.

[0065] S2: Extract teaching behavior features and teacher-student emotion features, and construct a classroom feature vector dataset.

[0066] This step extracts teaching behavior features and teacher-student emotion features based on the multi-modal data preprocessed in S1.

[0067] The teaching behavior features described in S201 include teacher teaching behavior features, student participation behavior features, and teacher-student interaction behavior features;

[0068] The teacher teaching behavior feature vector Bt = [t1, t2, t3], where t1 is the proportion of teaching duration, t2 is the average speaking speed (words per minute), and t3 is the number of times of highlighting key and difficult points;

[0069] The student participation behavior feature vector Bs = [s1, s2, s3], where s1 is the average number of student speeches, s2 is the answer correct rate, and s3 is the participation degree in group discussions;

[0070] The teacher-student interaction behavior feature vector Bi = [i1, i2, i3], where i1 is the number of teacher-student interactions, i2 is the proportion of interaction duration, and i3 is the average duration of teacher feedback;

[0071] Fuse the three types of behavior features to obtain the teaching behavior feature vector B. The formula is:

[0072]

[0073] Among them, α, β, and γ are weight coefficients, which are dynamically adjusted according to the course type requirements and satisfy α + β + γ = 1.

[0074] S202: Extract the facial expression emotion features of teachers and students. The facial expression emotion features are extracted using an improved CNN model (input image size 48×48, grayscale image), and the probability values of 7 basic emotions (happy, angry, sad, surprised, fearful, disgusted, neutral) are extracted to construct the teacher expression emotion feature vector E tf =[f happy ,f angry ,...,f neutral , and the student expression emotion feature vector E sf , and the model is trained using the FER-2013 dataset with an accuracy rate of 92.3%;

[0075] S203 extracts the emotional features of teachers' and students' speech, and extracts 12-Vimel frequency cepstral coefficients (MFCCs) using the following formula:

[0076]

[0077] Where x(n) is the audio frame signal, FFT is the Fast Fourier Transform, DCT is the Discrete Cosine Transform, the MFCC features are input into the SVM classifier, and the output is 7 types of emotion probability values, constructing the teacher and student speech emotion feature vector E. tv =[v happy ,v angry ,...,v neutral ]、E sv .

[0078] S204 semantic emotion feature extraction uses the BERT model (pre-trained model is bert-base-chinese). Input is teacher-student interaction text, outputting 7 emotion probability values, constructing a semantic emotion feature vector E. text =[t happy ,t angry ,...,t neutral ].

[0079] S205 integrates the above-mentioned multi-dimensional emotional features and uses a weighted fusion algorithm to generate the final emotional feature vectors E for teachers and students. t E s The formula is:

[0080]

[0081]

[0082] in, The weighting coefficients are set to 0.5, 0.3, and 0.2 respectively, satisfying the following conditions: ;

[0083] S206 will use the teaching behavior feature vector B and the teacher emotion feature vector E. t And student emotion feature vector E s By concatenating the vectors, we obtain the classroom feature vector X=[B,E]. t E s Construct a classroom feature dataset D={X1,X2,...,X...} n}, where n is the number of samples.

[0084] S3 constructs a prior knowledge base to enhance knowledge retrieval and improve assessment credibility, specifically including the following steps:

[0085] The S301 prior knowledge base is constructed by building a structured knowledge base according to five knowledge types. The data sources for these five knowledge types include: corresponding subject curriculum standards, corresponding lesson type teaching specifications, teacher and student emotion evaluation standards (referencing the PANAS emotion scale), typical teaching scenario emotion cases, and teaching evaluation cases. A MySQL database is used for storage, and a knowledge graph is constructed, where nodes represent knowledge types and edges represent relationships.

[0086] S302 Feature and Knowledge Encoding: The Sentence-BERT model is used to encode the classroom feature vector X and each piece of knowledge in the knowledge base, generating a 768-dimensional vector. The formula is as follows:

[0087]

[0088] Among them, Text refers to classroom-specific text or knowledge text.

[0089] S303 Knowledge retrieval matching calculates the cosine similarity between the classroom feature vector VecX and the knowledge vector VecK, using the following formula:

[0090]

[0091] S304: Knowledge enhancement application, selecting the top-5 knowledge items with the highest similarity as enhanced knowledge, and sorting them according to similarity weight, using the following formula:

[0092]

[0093] Where K is the i-th knowledge in the knowledge base; the filtered enhanced knowledge is pushed to the corresponding agent to correct the output of the evaluation model and optimize the evaluation results.

[0094] S4 Teaching Evaluation Multi-Agent Construction and Collaborative Assessment: Obtaining Evaluation Results for Each Agent; This step is used to construct a hierarchical multi-agent system to achieve multi-dimensional collaborative assessment, and specifically includes the following sub-steps:

[0095] The S401 intelligent agent is constructed in a layered manner, adopting a three-layer architecture design. The modules and intelligent agents deployed in each layer are as follows, including the three layers:

[0096] (1) Data perception layer: Deploy data acquisition module and preprocessing module, corresponding to each sub-step of step S1, responsible for data acquisition and preprocessing;

[0097] (2) Teaching analysis and collaboration layer: Deploy 3 types of intelligent agents, all based on the Qwen3-7B model, with the following parameters: learning rate 1e-5, batch size=8, and 10 training rounds; including lesson type analysis agent, teaching behavior analysis agent, and teaching content evaluation agent;

[0098] (3) Strategy and Coordination Layer: Deploys a comprehensive decision-making agent, built based on the GPT-4 model, and is responsible for coordinating the work of each agent.

[0099] S402 defines the responsibilities of each agent, clarifying their inputs, outputs, and core duties:

[0100] (1) Lesson type analysis agent: Input = classroom feature vector X, Output = lesson type classification result + corresponding evaluation criteria, Responsibility = identify lesson type and match evaluation system;

[0101] (2) Teaching behavior analysis agent: Input = teaching behavior feature vector B + lesson type evaluation criteria, output = teaching behavior effectiveness score + problem analysis, responsibility = evaluate the rationality and effectiveness of teaching behavior;

[0102] (3) Teaching content evaluation agent: Input = teaching content text + classroom feature vector X + lesson type evaluation criteria, output = content suitability score + optimization suggestions, responsibility = evaluate the suitability of teaching content with lesson type and student characteristics;

[0103] (4) Integrated decision-making agent: Input = output results of each agent + knowledge enhancement support, output = multi-dimensional evaluation report, responsibility = integration of evaluation results, coordination of conflicts, generation of evaluation report;

[0104] The S403 collaborative mechanism is established, creating a collaborative process of "data sharing, task allocation, parallel evaluation, result fusion, and conflict coordination." The comprehensive decision-making agent serves as the core of the collaboration, responsible for task scheduling and data synchronization for each agent.

[0105] The S404 integrated decision-making agent assigns evaluation tasks based on the lesson type analysis structure. Each agent conducts evaluations in parallel and synchronizes intermediate results. A weighted voting method is used to merge the intermediate results to obtain the evaluation results of each agent.

[0106] The formula is:

[0107]

[0108] Where R is the integrated evaluation score after fusion, w j R represents the weights of each agent. j The evaluation score is calculated for each agent. When the deviation of the evaluation results of each agent exceeds 10%, the conflict coordination mechanism is triggered, and the relevant knowledge of the knowledge enhancement module is called to conduct a second evaluation to correct the deviation.

[0109] S5 Multi-dimensional Teaching Evaluation Report Generation and Feedback.

[0110] This step is used to generate a multi-dimensional teaching evaluation report and complete the feedback push, specifically including the following sub-steps:

[0111] S501: Results fusion, the comprehensive decision-making agent integrates the evaluation results of each agent, knowledge enhancement support and sentiment analysis conclusions to ensure the comprehensiveness and accuracy of the results;

[0112] S502: Report generation, generating a multi-dimensional teaching evaluation report covering 4 core modules: (1) Teaching behavior assessment (40%): teaching behavior, student participation, teacher-student interaction scoring and problem analysis; (2) Teaching content assessment (30%): content suitability, coverage of key points and difficulties, courseware design scoring and optimization suggestions; (3) Emotional state assessment (20%): teacher and student emotional change trajectory, analysis of the impact of emotions on teaching effectiveness; (4) Comprehensive suggestions (10%): targeted teaching improvement measures;

[0113] S503: Feedback push, pushing evaluation reports and personalized teaching improvement suggestions to teachers and relevant teaching management departments to provide decision support for teaching optimization.

[0114] S6 integrates continuous learning and adaptive optimization with emotion feedback to build a continuous learning loop, enabling the system to dynamically adapt and self-optimize. This includes the following steps:

[0115] S601: Feedback Data Collection. Two types of feedback data are collected: (1) Teacher and Student Feedback: After class, teachers' approval of the evaluation report (0-10 points) and students' rating of their classroom emotional experience (0-10 points) are collected through questionnaires; (2) Related Data: Teaching effectiveness data such as the accuracy rate of subsequent assignments and unit test scores are collected to construct a feedback dataset F={F t ,F s ,F e}, where F t For teacher approval, F s Rate students' emotional experience, F e This is a correlation value for teaching effectiveness;

[0116] S602 Learning Sample Construction: The classroom feature dataset D, agent evaluation results R, and feedback dataset F are concatenated to construct the continuous learning sample set S, as shown in the formula:

[0117] The sample set is normalized using the min-max normalization algorithm, with the following formula:

[0118]

[0119] Where x is the original value of the sample, These are the minimum and maximum values ​​of the sample, respectively.

[0120] S603: Model parameter update, using a combination of federated learning and meta-learning to update the model parameters of each agent:

[0121] (1) Federated learning initialization: Each agent acts as a local node and initializes the model parameters θ. local The integrated decision-making agent acts as the central node, initializing the global parameter θ. global ;

[0122] (2) Local training: Each node trains the model based on its local sample set S, and updates the local parameters using the stochastic gradient descent (SGD) algorithm, as shown in the formula:

[0123]

[0124] Where t is the number of training epochs, η is the learning rate (0.001), and L is the loss function (using MSE loss).

[0125] (3) Parameter aggregation: The central node collects the parameters of each local node and aggregates the global parameters using a weighted average algorithm. The formula is as follows:

[0126]

[0127] Where m is the number of local nodes, n is the number of samples in the i-th node, and N is the total number of samples;

[0128] (4) Meta-learning optimization: The MAML algorithm is used to realize knowledge transfer and optimize the model's adaptive ability. The formula is:

[0129]

[0130] Where T represents the teaching scenario task, T represents the task training set, and α represents the meta-learning rate (taken as 0.01).

[0131] S604: Evaluation criteria iteration, dynamically adjusting the weights of the evaluation criteria based on continuous learning results and feedback data, using the following formula:

[0132]

[0133] Where ω is the original weight, β is the adjustment coefficient (taken as 0.1), and Corr(R,F) is the correlation coefficient between the evaluation result and the feedback data;

[0134] S605: Knowledge Update. This involves incorporating the experience, feedback, and optimized assessment criteria from this evaluation into the prior knowledge base to complete the knowledge iteration and provide more accurate knowledge support for subsequent evaluations.

[0135] This invention also provides a large-model-driven multi-agent system for teaching evaluation, wherein the system applies the above-mentioned large-model-driven multi-agent method for teaching evaluation, and includes:

[0136] Data acquisition module: used to execute step S101, collect classroom video, audio, text and interactive data to provide data support for subsequent feature extraction and evaluation;

[0137] Feature extraction module: used to execute steps S102-S104 and S201-S205, preprocess the collected multimodal data, extract teaching behavior features and teacher and student emotional features, and construct a classroom feature dataset;

[0138] Multi-agent collaborative evaluation module: used to execute steps S301-S304, including lesson type analysis agent, teaching behavior analysis agent, teaching content evaluation agent, and comprehensive decision-making agent, to achieve multi-dimensional collaborative evaluation;

[0139] Knowledge Enhancement Module: Used to execute steps S401-S404, construct a prior knowledge base for teaching evaluation, realize knowledge retrieval and fusion, and provide knowledge support for intelligent agent evaluation;

[0140] Continuous learning module: used to execute steps S501-S505, build a continuous learning closed loop, realize the updating of intelligent agent model parameters and the iteration of evaluation criteria, and improve the system's adaptive capability;

[0141] Evaluation report generation module: used to execute steps S601-S603, integrate the output results of each module, generate a multi-dimensional teaching evaluation report and provide personalized improvement suggestions.

[0142] Example 2

[0143] This method is applied to a real-world scenario, specifically in the teaching of a new lesson on "Solving Quadratic Equations in One Variable" in junior high school mathematics.

[0144] 1. Basic information about the implementation scenario

[0145] This embodiment is applied to a new lesson on "Solving Quadratic Equations in One Variable" in junior high school mathematics. The lesson lasts 45 minutes, is taught by a core junior high school mathematics teacher, and is taught to 40 second-year junior high school students. The classroom is equipped with a high-definition recording camera, a lavalier microphone, a desktop microphone array, and a smart teaching platform to meet the data collection requirements of step S1.

[0146] 2. Specific implementation process of each step

[0147] Step S1: Classroom Multimodal Data Acquisition and Preprocessing

[0148] Step S101: Data Acquisition. Acquire 45 minutes of classroom video data using a high-definition recording camera (1920×1080 resolution, 30fps frame rate), acquire teacher's voice data using a lavalier microphone (44.1kHz sampling rate, 16bit bit depth), acquire student's voice data using a desktop microphone array, and acquire courseware display text and teacher-student interaction text (including 12 classroom answer records and 8 teacher-student questions) through the smart teaching platform.

[0149] Step S102: Video data preprocessing, using Gaussian filtering algorithm (σ=1.5) to remove video noise, extracting 1 frame image every 2 frames, resulting in a total of 675 valid images;

[0150] Step S103: Audio data preprocessing, using Wiener filtering algorithm for noise reduction, dividing the audio into frames (frame length 25ms, frame shift 10ms), resulting in 18,000 audio frames;

[0151] Step S104: Text data preprocessing. The jieba word segmentation tool is used to segment the text, and after removing stop words, the core teaching vocabulary (such as "quadratic equation", "completing the square method", "discriminant") is retained to obtain standardized text data.

[0152] Step S2: Extraction of teaching behavior characteristics and teacher-student emotional characteristics

[0153] Step S201: Extract teaching behavior features, extract three types of behavioral features and construct vectors:

[0154] Teacher's teaching behavior feature vector B t =[0.65, 220, 4] (t1 = lecture time percentage 65%, t2 = average speaking speed 220 words / minute, t3 = number of times key points and difficulties are highlighted 4 times).

[0155] Student participation behavior feature vector B s =[1.2, 0.85, 0.9] (s1=average number of times students speak (1.2), s2=85% accuracy rate in answering questions, s3=90% participation rate in group discussions);

[0156] Teacher-student interaction behavior feature vector B i =[7, 0.2, 15] (i1 = 7 times of teacher-student interaction, i2 = 20% of the interaction time, i3 = 15 seconds of average teacher feedback time);

[0157] Weighting in new teaching scenarios The result of the fusion is B = 0.4 × B t +0.3×B s +0.3×B i ;

[0158] Step S202: Facial expression emotion feature extraction. The teacher and students' facial expressions and emotions are extracted using a CNN model, resulting in the teacher's facial expression emotion vector E. tf =[0.82, 0.03, 0.02, 0.05, 0.01, 0.02, 0.05] (82% probability of happiness, lower probabilities of other emotions), student's average emotional expression vector E sf =[0.75, 0.04, 0.03, 0.08, 0.02, 0.03, 0.05];

[0159] Step S203: Extract speech emotion features. Extract speech emotion using the MFCC+SVM model to obtain the teacher's speech emotion vector E. tv =[0.78, 0.04, 0.02, 0.06, 0.01, 0.03, 0.06], Student's average speech emotion vector E sv =[0.70, 0.05, 0.04, 0.10, 0.02, 0.04, 0.05];

[0160] Step S204: Semantic sentiment feature extraction. The semantic sentiment of the text is extracted using the BERT model to obtain the sentiment vector E of the teacher-student interaction text. text =[0.80, 0.03, 0.02, 0.07, 0.01, 0.02, 0.05]; In Example 1,

[0161] Step S205: Multi-dimensional feature fusion, with weights λ1=0.5, λ2=0.3, and λ3=0.2, to obtain the teacher and student emotion feature vectors E respectively. t E s ,

[0162] Step S206: Join B and E t E s Obtain the classroom feature vector X, and construct the classroom feature dataset D (n=675).

[0163] Step S3: Construction and application of knowledge enhancement mechanisms for emotion integration.

[0164] Step S301: Knowledge base construction. Based on the Junior High School Mathematics Curriculum Standards (2022 Edition) and the new lesson teaching specifications, a structured knowledge base containing 500+ emotion cases and 1000+ assessment cases will be constructed.

[0165] Step S302: Feature and knowledge encoding. The classroom feature vector X and knowledge base knowledge are encoded into a 768-dimensional vector using the Sentence-BERT model.

[0166] Step S303: Knowledge retrieval and matching, calculate cosine similarity, and select the top-5 matching knowledge (such as "the effect of teachers' happy emotions on students' participation in new lessons" and "evaluation criteria for key points and difficulties in new lessons of quadratic equations").

[0167] Step S304: Knowledge enhancement application, push the matching knowledge to each agent, and the teaching content evaluation agent will subsequently combine the knowledge optimization conclusions to supplement the analysis that "students' surprise is concentrated in the explanation of matching methods, suggesting that the knowledge point is relatively difficult".

[0168] Step S4: Construction and Collaborative Evaluation of a Multi-Agent System for Teaching Assessment

[0169] Step S401: Layered construction of intelligent agents, deployment of a three-layer architecture and four types of intelligent agents, model parameters set according to the instruction manual (Qwen3-7B fine-tuning, learning rate 1e-5, etc.).

[0170] Step S402: Define the responsibilities of each agent, and clarify the inputs, outputs and responsibilities of each agent according to the instruction manual;

[0171] Step S403: Establish a collaborative mechanism and initiate the process of "data sharing - task allocation - parallel evaluation - result fusion - conflict coordination";

[0172] Step S404: Multi-agent collaborative evaluation. The lesson type analysis agent outputs the classification results of "new lesson" and the corresponding evaluation criteria. After parallel evaluation by each agent, the teaching behavior analysis agent outputs a score of 8.6 points, and the teaching content evaluation agent outputs a score of 9.2 points. The comprehensive decision-making agent performs weighted fusion (with a weight of 0.33 for each agent) to obtain a comprehensive evaluation score of R=8.9 points. There is no evaluation conflict and no need to trigger a coordination mechanism.

[0173] Step S5: Generation and Feedback of Multi-Dimensional Teaching Evaluation Report

[0174] Step S501: Results fusion, the comprehensive decision-making agent integrates the evaluation results of each agent, knowledge enhancement support and sentiment analysis conclusions;

[0175] Step S502: Report generation. A multi-dimensional evaluation report is generated, which clarifies that the teaching behavior score is 8.6, the teaching content score is 9.2, the emotional state assessment is "positive", and the overall score is 8.9. Suggestions are made such as "optimizing the pace of explaining the knowledge points of the teaching method and paying attention to the understanding difficulties corresponding to students' surprise emotions".

[0176] Step S503: Feedback push, push the report to the instructors and the school's academic affairs office for teaching improvement and teaching management.

[0177] Step S6: Integrating continuous learning and adaptive optimization with emotion feedback

[0178] Step S601: Feedback data collection. After class, a questionnaire was collected to assess teacher approval (9.2 points), student emotional experience (8.8 points), and the accuracy rate of subsequent assignments (89%). A feedback dataset F=[9.2, 8.8, 0.89] was constructed.

[0179] Step S602: Construct learning samples by concatenating D, R (overall score 8.9) and F, and normalizing them to obtain the learning sample set S;

[0180] Step S603: Model parameter update. Federated learning (5 local nodes, total sample size 675) and meta-learning (α=0.01) are used to update the parameters of each agent. After the update, the model evaluation accuracy is improved by 2.1%.

[0181] Step S604: Iterate the evaluation criteria. Based on the correlation coefficient Corr(R,F)=0.87, adjust the weight ω=ω+0.1×0.87 to optimize the evaluation criteria for new lessons.

[0182] Step S605: Knowledge update, incorporate the experience and feedback from this assessment into the knowledge base, and add a case study on "the correlation between emotions and the difficulty of knowledge points in the new teaching of quadratic equations".

[0183] 3. Implementation Results

[0184] The experimental results of this embodiment show that, in the scenario of new junior high school mathematics lessons, the system achieves a teaching evaluation accuracy rate of 93.7%, an emotion analysis accuracy rate of 92.3%, and a teacher and student satisfaction rate of 89.6% for the evaluation report. Compared with existing teaching evaluation systems without emotion analysis, the system significantly improves the comprehensiveness and accuracy of the evaluation, and can effectively support teaching optimization and improvement.

[0185] It is understood that the application scenarios of this invention are not limited to junior high school mathematics new lessons, but can also be adapted to multiple subjects such as Chinese, English, and physics, as well as multiple lesson types such as review lessons and problem-solving lessons. By adjusting the model parameters and evaluation standard weights, different scenarios can be adapted, which has wide applicability.

[0186] Those skilled in the art will readily understand that the above description is merely an embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A large-model-driven multi-agent method for instructional evaluation, characterized in that, include: Multimodal data acquisition and preprocessing in the classroom; Extract teaching behavior characteristics and teacher-student emotional characteristics to construct a classroom feature vector dataset; Enhanced knowledge retrieval is achieved by constructing a prior knowledge base; The teaching evaluation multi-agent construction and collaborative assessment were carried out to obtain the evaluation results of each agent. Generation and feedback of multi-dimensional teaching evaluation reports; Integrating continuous learning and adaptive optimization with emotional feedback.

2. The multi-agent method for teaching evaluation according to claim 1, characterized in that, The classroom multimodal data acquisition and preprocessing includes: Data collection: Video data is collected through classroom recording equipment, audio data is collected through audio collection equipment, and text data and interactive data are collected through the teaching interaction platform; Preprocessing: Frame extraction, noise reduction, and format standardization are performed on the acquired video data; noise reduction, frame segmentation, and feature alignment are performed on the acquired audio data; word segmentation, stop word removal, and semantic standardization are performed on the acquired text data.

3. The multi-agent method for teaching evaluation according to claim 1, characterized in that, The teaching behavior characteristics include teacher lecturing behavior characteristics, student participation behavior characteristics, and teacher-student interaction behavior characteristics. The emotional characteristics of teachers and students include facial expression emotional characteristics, voice emotional characteristics, and semantic emotional characteristics. We construct a classroom feature vector dataset by weighted integration of teaching behavior characteristics and teacher and student emotional characteristics.

4. The multi-agent method for teaching evaluation according to claim 1, characterized in that, The construction of the prior knowledge base specifically involves: Prior knowledge base construction: A structured knowledge base is constructed according to five knowledge types. The data sources for the five knowledge types include the corresponding subject curriculum standards, the corresponding lesson type teaching specifications, teacher and student emotion evaluation standards, typical teaching scenario emotion cases, and teaching evaluation cases; a MySQL database is used for storage, and a knowledge graph is constructed. Feature and Knowledge Encoding: The Sentence-BERT model is used to encode the classroom feature vectors in the classroom feature vector dataset and each knowledge vector in the knowledge base, generating a 768-dimensional vector; Knowledge retrieval and matching involves calculating the cosine similarity between the classroom feature vector and the 768-dimensional vector, and selecting the top five knowledge items with the highest similarity as augmented knowledge.

5. The multi-agent method for teaching evaluation according to claim 1, characterized in that, The construction and collaborative assessment of multi-agent teaching evaluation are as follows: The teaching evaluation system is constructed using a multi-agent hierarchical structure, comprising three layers: a strategy and coordination layer, a teaching analysis and collaboration layer, and a data perception layer. The teaching analysis and collaboration layer includes a lesson type analysis agent, a teaching behavior analysis agent, and a teaching content evaluation agent. The strategy and coordination layer includes a comprehensive decision-making agent. Clearly define the responsibilities of each intelligent agent, including its inputs, outputs, and core duties; Establish a collaborative mechanism process of "data sharing, task allocation, parallel evaluation, result fusion, and conflict coordination"; The integrated decision-making agent assigns evaluation tasks based on the lesson type analysis structure, and each agent conducts evaluations in parallel and synchronizes intermediate results; a weighted voting method is used to merge the intermediate results to obtain the evaluation results of each agent.

6. The multi-agent method for teaching evaluation according to claim 1, characterized in that, The multi-dimensional teaching evaluation report includes four core contents: (1) Teaching behavior evaluation (40%), including teaching behavior, student participation, teacher-student interaction scores and problem analysis; (2) Teaching content evaluation (30%), including content suitability, coverage of key points and difficulties, courseware design scores and optimization suggestions; (3) Emotional state evaluation (20%), including the trajectory of teacher and student emotional changes and the analysis of the impact of emotions on teaching effectiveness; and (4) Comprehensive suggestions (10%), including targeted teaching improvement measures.

7. The multi-agent method for teaching evaluation according to claim 1, characterized in that, The continuous learning and adaptive optimization that integrates emotional feedback specifically includes: Feedback data collection: Collect feedback from teachers and students on the evaluation report, the trajectory of changes in classroom emotions, and data related to teaching effectiveness; Learning sample set construction: Concatenate classroom feature vector dataset, agent evaluation results, and feedback data to construct a continuous learning sample set; Model parameter update: The model parameters of each agent are updated by combining federated learning and meta-learning. Evaluation criteria iteration: The evaluation criteria and weight allocation of each agent are dynamically adjusted based on continuous learning results and feedback data; Knowledge update: Incorporate new assessment experiences and feedback knowledge into the prior knowledge base to achieve knowledge iteration.

8. The multi-agent method for teaching evaluation according to claim 7, characterized in that, A method combining federated learning and meta-learning is used to update the model parameters of each agent, specifically; Federated learning initialization: Each agent acts as a local node and initializes the model parameters θ. local The integrated decision-making agent acts as the central node, initializing the global parameter θ. global ; Local training: Each node trains the model based on the local sample set S and updates the local parameters using the stochastic gradient descent algorithm; Parameter aggregation: The central node collects parameters from each local node and aggregates global parameters using a weighted average algorithm; Meta-learning optimization: The MAML algorithm is used to realize knowledge transfer and optimize the model's adaptive ability.

9. The multi-agent method for teaching evaluation according to claim 7, characterized in that, The feedback from teachers and students on the evaluation report was collected through a post-class questionnaire, which included the teachers' approval of the evaluation report and the students' ratings of their emotional experience in the classroom. The teaching effectiveness data includes data such as the accuracy rate of subsequent assignments and unit test scores.

10. The multi-agent method for teaching evaluation according to any one of claims 1-9, characterized in that, When the evaluation results of each agent deviate by more than 10%, the conflict coordination mechanism is triggered, and relevant knowledge from the knowledge enhancement module is invoked for a second evaluation to correct the deviation.

Images