Method for classifying teacher teaching behavior based on classroom speech text analysis
By constructing a dataset of teachers' teaching behaviors and using graph convolution and graph attention neural networks for interpolation and fusion, the problem of existing methods being affected by subjective factors is solved, realizing the automated classification and quantitative evaluation of teachers' teaching behaviors and improving teaching quality.
Patent Information
- Application Number
- CN202311059140.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-08-22
- Publication Date
- 2026-01-02
- Estimated Expiration
- 2043-08-22
AI Technical Summary
Existing methods for classifying teachers' teaching behaviors mainly rely on questionnaires and expert evaluations, which are easily influenced by subjective factors, time-consuming and labor-intensive, lack standardized coding methods and datasets, and cannot meet the needs of automated classification.
A dataset of teacher teaching behaviors was constructed and interpolated and fused using graph convolutional neural networks and graph attention neural networks. By combining pointwise mutual information and word frequency-inverse document frequency algorithms, a teacher teaching behavior classification network was constructed, and speech text was extracted from videos for automatic classification through speech recognition.
It enables accurate and automatic classification of teachers' teaching behaviors, provides objective quantitative evaluation criteria, and improves teaching quality and self-reflection effectiveness.
Smart Images
Figure CN117093898B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application belongs to the cross field of artificial intelligence and pedagogy, and particularly relates to a teacher teaching behavior classification method based on classroom speech text analysis. BACKGROUND
[0002] The classroom is the first main position for teachers to teach and educate, and the classroom teaching behavior of teachers directly affects the learning effect and physical and mental development of students. Data shows that the teacher speech behavior accounts for about 80% of the entire classroom behavior, and efficient speech behavior can effectively mobilize the enthusiasm and initiative of students and improve the teaching effect. Therefore, scientific and reasonable analysis of the teacher behavior in the classroom is not only conducive to promoting the reform of modern teaching evaluation, but also can improve the self-reflection and teaching analysis effect of teachers. The main problems existing in the current teaching behavior classification method are: (1) the current research on teacher classroom speech behavior mostly adopts questionnaire survey method, expert evaluation method, etc., and the research results are easily affected by subjective factors; (2) the teacher behavior evaluation needs the cooperation of students and experts, which is time-consuming and laborious, and the cost of self-reflection is high.
[0003] With the rapid development of information technology and artificial intelligence, the education industry is gradually moving towards informatization, and it has important practical significance and practical value to use advanced computer technology to analyze the teacher classroom speech text generated in the teaching process and feed back the teaching behavior to the teachers. However, at present, there is a lack of standard coding method and data set for teacher teaching behavior classification, and there are few studies on automatic classification of teaching behavior, which cannot meet the actual needs. SUMMARY
[0004] The technical problem to be solved by the present application is to overcome the shortcomings of the prior art and provide a teacher teaching behavior classification method based on classroom speech text analysis with accurate classification results.
[0005] The technical scheme adopted to solve the above technical problems is: a teacher teaching behavior classification method based on classroom speech text analysis, comprising the following steps:
[0006] Step 1. Constructing a teacher teaching behavior data set according to classroom speech text
[0007] Teacher teaching video is collected in a natural classroom environment, audio information is stripped from the original video in units of sentences, teacher speech text data is obtained through speech recognition, each sentence in the teacher speech text data is labeled with a corresponding class label according to the teacher teaching behavior, a teacher teaching behavior data set is formed, the class label is composed of a classroom teaching class, a classroom interaction class, a teacher evaluation class, and a classroom management class, the teacher teaching behavior data set is randomly divided into a training set and a test set according to 9:1, in each running process, 90% of the data in the training set is randomly extracted for training the model, and the remaining data constitutes a validation set;
[0008] Step 2. Constructing the teacher teaching behavior classification network
[0009] The teacher teaching behavior classification network is composed of two layers of graph convolutional neural network and two layers of graph attention neural network in parallel and then interpolated, the graph convolutional neural network is composed of a first layer of graph convolutional layer, a dropout layer and a second layer of graph convolutional layer in series, the graph attention neural network is composed of a first layer of dropout layer, a first layer of attention layer, a second layer of dropout layer and a second layer of attention layer in series, the first layer of attention layer has 8 attention modules, the second layer of attention layer has 1 attention module, the two networks share parameters, and m is the interpolation coefficient, m∈[0, 1];
[0010] Step 3. Constructing a text graph
[0011] The sentences D in the teacher teaching behavior data set are represented by a graph G, which contains document nodes D n , word nodes W n , edges W e between word nodes, and edges D e between word nodes and document nodes, the weight of edge W e between word nodes is calculated using the pointwise mutual information algorithm, the weight of edge D e between word nodes and document nodes is calculated using the term frequency-inverse document frequency algorithm, and the confidence of the word is calculated at each iteration Words with a confidence higher than 0.9 are added to the graph G as keywords to form a text graph as input to the teacher teaching behavior classification network;
[0012] Step 4. Training the teacher teaching behavior classification network
[0013] Step 4.1. Initialization of the teacher teaching behavior classification network
[0014] The parameters of the teacher teaching behavior classification network are initialized in the Xavier manner;
[0015] Step 4.2. Setting the hyperparameters of the teacher teaching behavior classification network
[0016] In the training process, the batch size is 64, the learning rate is 0.01, the pointwise mutual information window is 10, and the dropout rate is 0.5;
[0017] Step 4.3. Training the teacher teaching behavior classification network
[0018] All the texts in the training set are input into the teacher teaching behavior classification network, forward propagation is performed, the loss function L is determined, the loss is minimized using the Adam optimizer, and all parameters are updated until the loss function converges, the training is completed, and the trained teacher teaching behavior classification network is obtained;
[0019] Step 5. Detecting the teacher teaching behavior classification network with the test set
[0020] The verification set is first input into the trained teacher teaching behavior classification network, the parameters of the teacher teaching behavior classification network are adjusted, and thus the optimal teacher teaching behavior classification network is obtained, and then the test set is input into the teacher teaching behavior classification network, and the teacher teaching behavior classification result is output.
[0021] As a preferred technical solution, the classroom lecture category in step 1 includes the behavior of the teacher providing the teacher's personal insights, authoritative insights and objective facts to the students on the course content;
[0022] The classroom interaction category includes the behavior of the teacher and the students interacting with each other during the teaching process, specifically including asking questions to a single student, asking questions to the whole class and participating in student discussions;
[0023] The teacher evaluation category includes the behavior of the teacher evaluating the performance of the students during the teaching process, specifically including affirming or denying the answers of the students and praising the behavior of the students;
[0024] The classroom management category includes all behaviors unrelated to the course content, specifically the behaviors of maintaining discipline and arranging student grouping.
[0025] As a preferred technical solution, the confidence of the word in step 3 is determined according to the following formula:
[0026]
[0027] In the formula, c is the category label, q c is the number of the cth category of documents containing the word w i in the marked document, and n is the number of categories.
[0028] As a preferred technical solution, the method for interpolating the two-layer graph convolutional neural network and the two-layer graph attention neural network in step 2 is as follows:
[0029] The text feature X and the self-loop adjacency matrix are input into the graph convolutional neural network and the graph attention neural network respectively, and the final prediction result P of the model is determined according to the following formula:
[0030] P = mP1 + (1-m)P2
[0031]
[0032]
[0033]
[0034] In the formula, m is the interpolation coefficient, m∈[0,1], P1 is the prediction result of the graph convolutional neural network, P2 is the prediction result of the graph attention neural network, log_softmax(·) is the logarithm of the normalized exponential function, and GC (1) (·) represents the first layer of the graph convolutional neural network, GC (2) (·) represents the second layer of the graphical convolutional neural network, and ReLU(·) is the linear rectified activation function; GA (1) (·) represents the first layer of the graphical attention neural network, GA (2) (·) represents the second layer of the graph attention neural network, where I is the identity matrix and A is the adjacency matrix of the graph.
[0035] The beneficial effects of this invention are as follows:
[0036] This invention utilizes a self-training method based on classroom speech text information to add calculated text keywords as pseudo-labeled data to the dataset. Linear interpolation is then performed on a graph convolutional network and a graph attention network to obtain a teacher teaching behavior classification network. By constructing a teacher teaching behavior dataset and a teacher teaching behavior classification network, automatic classification of teacher teaching behaviors based on classroom speech text analysis is achieved. This provides an objective basis for quantitatively evaluating teachers' classroom abilities and helps guide teachers to optimize teaching methods and improve teaching quality. Attached Figure Description
[0037] Figure 1 This is a flowchart of the present invention.
[0038] Figure 2 This is a schematic diagram of the self-built teacher teaching behavior dataset of this invention.
[0039] Figure 3 This is a schematic diagram of the teacher teaching behavior classification network in Embodiment 1 of the present invention. Detailed Implementation
[0040] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments, but the present invention is not limited to the following embodiments.
[0041] exist Figure 1 In this embodiment, a method for classifying teacher teaching behaviors based on classroom verbal text analysis includes the following steps:
[0042] Step 1. Construct a dataset of teacher teaching behaviors based on classroom speech texts.
[0043] Videos of teachers lecturing in natural classroom settings are collected. Audio information is extracted from the original videos, sentence by sentence, and speech text data is obtained through speech recognition. Each sentence in the teacher's speech text data is then labeled with a corresponding category based on the teacher's teaching behavior, forming a teacher teaching behavior dataset. Figure 2 The teacher teaching behavior dataset was randomly divided into a training set and a test set in a 9:1 ratio. During each run, 90% of the data was randomly extracted from the training set to train the model, and the remaining data formed the validation set.
[0044] The category labels consist of classroom teaching category, classroom interaction category, teacher evaluation category, and classroom management category.
[0045] Classroom instruction includes the act of teachers providing students with their personal insights, the opinions of authority figures, and objective facts regarding classroom content.
[0046] Classroom interaction categories include teacher-student and student-student interactions during the teaching process, specifically including asking questions to individual students, asking questions to the whole class, and participating in student discussions.
[0047] Teacher evaluation categories include the behaviors by which teachers evaluate students’ performance during the teaching process, specifically including affirming or negating students’ answers and praising students.
[0048] The classroom management category includes all behaviors unrelated to classroom content, specifically maintaining discipline and arranging students into groups.
[0049] Step 2. Construct a teacher teaching behavior classification network
[0050] The teacher instruction behavior classification network consists of a two-layer graph convolutional neural network and a two-layer graph attention neural network connected in parallel, followed by interpolation. Figure 3 The graph convolutional neural network is composed of a first graph convolutional layer, a dropout layer, and a second graph convolutional layer connected in series. The graph attention neural network is composed of a first dropout layer, a first attention layer, a second dropout layer, and a second attention layer connected in series. The first attention layer has 8 attention modules, and the second attention layer has 1 attention module. The two networks share parameters, and the interpolation coefficient m∈[0,1]. In this embodiment, m is 0.5.
[0051] The method for interpolating two-layer graph convolutional neural networks and two-layer graph attention neural networks is as follows:
[0052] The text features X and the self-circular adjacency matrix are used. The inputs are fed into the graph convolutional neural network and the graph attention neural network respectively, and the final prediction result P of the model is determined by the following formula:
[0053] P = mP1 + (1 - m)P2
[0054]
[0055]
[0056]
[0057] where m is an interpolation coefficient, m e [0, 1], P1 is the prediction result of the graph convolutional neural network, P2 is the prediction result of the graph attention neural network, log softmax(·) is a logarithmic operation on a normalized exponential function, GC (1) (·) is a first layer graph convolutional neural network, GC (2) (·) is a second layer graph convolutional neural network, ReLU(·) is a linear rectifier activation function; GA (1) (·) is a first layer graph attention neural network, GA (2) (·) is a second layer graph attention neural network, I is an identity matrix, and A is an adjacency matrix of the graph.
[0058] Step 3. Constructing a text graph
[0059] The sentences D in the teacher teaching behavior dataset are represented by a graph G, which contains document nodes D n , word nodes W n , edges W e between word nodes, and edges D e between word nodes and document nodes, the weights of edges W e between word nodes are calculated using the pointwise mutual information algorithm, the weights of edges D e between word nodes and document nodes are calculated using the term frequency-inverse document frequency algorithm, and the confidence of the word is calculated at each iteration. Words with a confidence higher than 0.9 are added to the graph G as keywords to form a text graph, which is used as the input of the teacher teaching behavior classification network.
[0060] The confidence of the word is determined according to the following formula:
[0061]
[0062] where c is the class label, q c is the number of the cth class of documents containing the word w i in the marked document, and n is the number of classes.
[0063] Step 4. Training the teacher teaching behavior classification network
[0064] Step 4.1. Initialization of the teacher teaching behavior classification network
[0065] The parameters of the teacher teaching behavior classification network are initialized in the Xavier manner;
[0066] Step 4.2. Set the hyperparameters of the teacher teaching behavior classification network
[0067] During the training process, the batch size is 64, the learning rate is 0.01, the point-wise mutual information window is 10, and the dropout rate is 0.5;
[0068] Step 4.3. Train the teacher teaching behavior classification network
[0069] All the texts in the training set are input into the teacher teaching behavior classification network for forward propagation to determine the loss function L, the loss is minimized using the Adam optimizer, and all parameters are updated until the loss function converges, the training is completed, and the trained teacher teaching behavior classification network is obtained;
[0070] Step 5. Test the teacher teaching behavior classification network with the test set
[0071] First, input the validation set into the trained teacher teaching behavior classification network to adjust the parameters of the teacher teaching behavior classification network, thereby obtaining the optimal teacher teaching behavior classification network, and then input the test set into the teacher teaching behavior classification network to output the teacher teaching behavior classification result.
[0072] To verify the beneficial effects of the present application, the inventors used the teacher teaching behavior classification method based on classroom speech text analysis of Example 1 of the present application (referred to as Example 1) and "Kim Y. Convolutional neural networks for sentence classification [J]. arXiv preprint arXiv: 1408.5882, 2014." (referred to as Comparative Experiment 1), "Joulin A, Grave E, Bojanowski P, et al. Bag of tricks for efficient text classification [J]. arXiv preprint arXiv: 1607.01759, 2016." (referred to as Comparative Experiment 2), "Johnson R. Zhang T. Deep pyramid convolutional neural networks for text categorization [C] / / Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2017: 562-570." (referred to as Comparative Experiment 3) to conduct comparative experiments, and the classification results were evaluated by testing the precision and the average Macro-F1 of macro-F1. The experimental results are shown in Table 1.
[0073] Table 1 Experimental results of the method of Example 1 and comparative experiments
[0074]
[0075]
[0076] As can be seen from Table 1, compared with Comparative Experiments 1-3, Example 1 has the highest test accuracy and Macro-F1 value, and the scores of each index are obviously improved. The test accuracy and Macro-F1 value of Example 1 are increased by 8.3% and 8.3% respectively compared with Comparative Experiment 1, by 9.1% and 9.5% respectively compared with Comparative Experiment 2, and by 19.2% and 20.0% respectively compared with Comparative Experiment 3.
[0077] The above experiments show that the scores of each index of the present application are better than those of the comparative experiments, and the optimal effect is obtained on the self-built teacher teaching behavior data set, which can effectively classify the teacher teaching behavior and can be used for automatically classifying the teacher's teaching behavior based on classroom speech text analysis in the natural environment of the classroom.
Claims
1. A teacher teaching behavior classification method based on classroom speech text analysis, characterized in that, Comprising the following steps: Step 1. Constructing a teacher teaching behavior dataset according to classroom speech text Teacher teaching video is collected in a natural classroom environment, audio information is stripped from the original video in units of sentences, teacher speech text data is obtained through speech recognition, each sentence in the teacher speech text data is labeled with a corresponding class label according to the teacher teaching behavior, a teacher teaching behavior dataset is formed, the class label is composed of classroom teaching class, classroom interaction class, teacher evaluation class and classroom management class, the teacher teaching behavior dataset is randomly divided into training set and test set according to 9:1, in each running process, 90% of the data in the training set is randomly extracted for training model, and the remaining data constitutes a validation set; Step 2. Constructing a teacher teaching behavior classification network The teacher teaching behavior classification network is composed of two layers of graph convolutional neural network and two layers of graph attention neural network in parallel and then interpolated, the graph convolutional neural network is composed of a first layer of graph convolutional layer, a dropout layer and a second layer of graph convolutional layer connected in turn, the graph attention neural network is composed of a first layer of dropout layer, a first layer of attention layer, a second layer of dropout layer and a second layer of attention layer connected in turn, the first layer of attention layer has 8 attention modules, the second layer of attention layer has 1 attention module, the two networks share parameters, and m is an interpolation coefficient, m∈[0,1]; Step 3. Constructing a text graph The sentences D in the teacher teaching behavior data set are represented by a graph G, which contains document nodes D n , word nodes W n , edges W e between word nodes, and edges D e between word nodes and document nodes, the weights of edges W e between word nodes are calculated using the pointwise mutual information algorithm, the weights of edges D e between word nodes and document nodes are calculated using the term frequency-inverse document frequency algorithm, and the confidence of the word is calculated at each iteration Words with a confidence higher than 0.9 are added to the graph G as keywords to form a text graph, which is used as the input of the teacher teaching behavior classification network; Step 4. Training the teacher teaching behavior classification network Step 4.
1. Initialization of the teacher teaching behavior classification network The parameters of the teacher teaching behavior classification network are initialized in the Xavier manner; Step 4.
2. Setting the super parameters of the teacher teaching behavior classification network In the training process, the batch size is 64, the learning rate is 0.01, the point-wise mutual information window is 10, and the dropout rate is 0.5; Step 4.
3. Training the teacher teaching behavior classification network All the texts in the training set are input into the teacher teaching behavior classification network, forward propagation is performed, the loss function L is determined, the Adam optimizer is used to minimize the loss, and all the parameters are updated until the loss function converges, the training is completed, and the trained teacher teaching behavior classification network is obtained; Step 5. Detecting the teacher teaching behavior classification network with the test set First, input the validation set into the trained teacher teaching behavior classification network, adjust the parameters of the teacher teaching behavior classification network, and then input the test set into the teacher teaching behavior classification network, and output the teacher teaching behavior classification result.
2. The method of claim 1, wherein the method of classifying teacher teaching behaviors based on classroom discourse text analysis is characterized by, The classroom teaching class in step 1 includes the behavior of the teacher providing the teacher's personal insights, authoritative insights and objective facts to the students on the course content; The classroom interaction class includes the behavior of the teacher and the students interacting with each other during the teaching process, specifically including asking questions to a single student, asking questions to the whole class and participating in student discussion; The teacher evaluation class includes the behavior of the teacher evaluating the performance of the students during the teaching process, specifically including affirming or denying the answers of the students and praising the students. The classroom management category includes all behaviors that are not related to the content of the lesson, specifically behaviors that maintain discipline and arrange student grouping. 3.The method of claim 1, wherein, the confidence of the word in step 3 is determined according to the following formula: where c is a class label, q c is the number of documents in class c that contain the word w i in the document, and n is the number of classes.
4. The method of claim 1, wherein the teacher teaching behavior classification based on classroom speech text analysis is characterized by, The interpolation method of the two-layer graph convolutional neural network and the two-layer graph attention neural network in step 2 is: The text features X and the self-circular adjacency matrix are used. The inputs are fed into the graph convolutional neural network and the graph attention neural network respectively, and the final prediction result P of the model is determined by the following formula: P = mP1 + (1-m)P2 wherein m is an interpolation coefficient, m∈[0, 1], P1 is a prediction result of the graph convolutional neural network, P2 is a prediction result of the graph attention neural network, log_softmax(·) is a logarithmic operation on a normalized exponential function, GC (1) (·) is a first layer graph convolutional neural network, GC (2) (·) is a second layer graph convolutional neural network, ReLU(·) is a linear rectifier activation function; GA (1) (·) is a first layer graph attention neural network, GA (2) (·) is a second layer graph attention neural network, I is an identity matrix, and A is an adjacency matrix of the graph.