Knowledge tracking method based on data enhanced convolutional memory network

Through the method of enhancing convolutional memory network based on data, adversarial problem sequences are generated and knowledge status is dynamically updated, which solves the problem of not being able to truly reflect students' actual learning situation and neglecting the connection between knowledge points in the existing technology, and achieves more accurate knowledge tracking and prediction of answering results improves the accuracy and robustness of the system.

CN119005307AInactive Publication Date: 2024-11-22GUANGDONG UNIV OF TECH

Patent Information

Application Number
CN202411107221.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-13
Publication Date
2024-11-22
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The existing knowledge tracking methods are trained and tested on clean learning data, which cannot truly reflect students' actual learning situation, and they independently deal with each knowledge point, ignoring the potential connection between knowledge points.

Method used

Adopting the method of data-enhanced convolutional memory network, generate adversarial problem sequences, integrate storage problem sequences and dynamically update students' knowledge status, use adversarial data-enhanced convolutional memory network model to predict students' answer results, and use convolutional key-value storage method Capture the relationship between knowledge points and identify the connection between adjacent knowledge points.

Benefits of technology

It improves the accuracy and robustness of knowledge tracking, can more accurately capture the complex relationship between students' knowledge state and potential concepts, and enhances the performance of the knowledge tracking system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119005307A_ABST
    Figure CN119005307A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of knowledge tracking methods, in particular to a knowledge tracking method based on a data enhanced convolutional memory network. The technical problems that when a knowledge tracking method is used, the actual learning condition of students cannot be truly reflected, each knowledge point is usually relatively independently processed, and the potential relation between the knowledge points is neglected are solved; according to the technical scheme, the knowledge tracking method based on the data enhanced convolutional memory network comprises the steps of generating a antagonism problem sequence; compared with a traditional knowledge tracking method, the method can truly reflect the actual learning condition of students, and solves the problem that the traditional method relatively independently processes each knowledge point and neglects the potential relation between the knowledge points. According to the knowledge tracking method, a method for calculating the similarity between an original question and an unanswered question is used for generating an antagonistic sample, the knowledge state of a student is dynamically represented by designing a key value storage network, and association between questions answered by the student is captured.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of knowledge tracking methods, and in particular to a knowledge tracking method based on a data enhanced convolutional memory network. Background Art

[0002] The purpose of knowledge tracing is to observe, characterize and quantify students' knowledge status, such as the degree of mastery of the skills contained in the textbooks. Traditional knowledge tracing research initially used Bayesian theorem for model design. The representative model is standard Bayesian knowledge tracing. BKT uses probabilistic graphical models such as hidden Markov models and Bayesian networks to track the changes in students' knowledge status when practicing skills. However, when using knowledge tracing methods, most models are only trained and tested on clean learning data, which cannot truly reflect students' actual learning situation, and often process each knowledge point relatively independently, while ignoring the potential connections between knowledge points. Summary of the invention

[0003] In order to overcome the problem that when using knowledge tracking methods, most models are only trained and tested on clean learning data, which cannot truly reflect students' actual learning situation, and often process each knowledge point relatively independently, while ignoring the potential connections between knowledge points.

[0004] The technical solution of the present invention is: a knowledge tracking method based on data-enhanced convolutional memory network, including the following methods:

[0005] S11: Generate a sequence of adversarial questions corresponding to the original questions;

[0006] S12: Integrate and store question sequences and dynamically update students’ knowledge status;

[0007] S13: Use adversarial data to enhance the convolutional memory network model to predict students’ answer results.

[0008] Preferably, the convolutional key-value storage method uses an attention-based method to map the association between answered questions and latent concepts to store their knowledge states in the learning task; and by using one-dimensional convolution to capture local patterns in the knowledge point vector, thereby identifying the connection between adjacent knowledge points, which helps to predict students' answers.

[0009] Preferably, when generating the adversarial question sequence, the following steps are included:

[0010] S21: Convert question text into embedding vector;

[0011] S22: Calculate the similarity scores between questions;

[0012] S23: Generate adversarial question sequences.

[0013] Preferably, when dynamically updating the knowledge state of students, the following steps are included:

[0014] S31: Calculate the embedding vectors between the question and each potential concept;

[0015] S32: Store the relevant information of the question sequence;

[0016] S33: Update and store the knowledge state of students;

[0017] S34: Extract the local features between potential question concepts.

[0018] Preferably, when extracting the local features between question concepts, the following steps are included:

[0019] S41: Extract the concept information included in the question;

[0020] S42: Add residual connections and generate a hierarchical convolutional layer H;

[0021] S43: Introduce non-linear factors.

[0022] Preferably, when converting the question text into an embedding vector, the following steps are included:

[0023] S51: Describe the text of the question as a sentence L containing m words i ={w1, w2, w3 ……, w m}, where w i represents a word, 1 < i ≤ m;

[0024] S52: Use the Word2vec method to convert each word into a vector with a dimension of d0, that is, each question text L i corresponds to a matrix with a dimension of d0 × m;

[0025] S53: Calculate the average value of the word vectors included in each question text in turn to obtain the embedding vector of the question text , where represents the embedding vector converted from a question text, 1 < i ≤ n, and n is the serial number of the question.

[0026] Preferably, when calculating the similarity score between questions, the following steps are included:

[0027] S61: Store the embedding vectors of the question text in Q emb one by one, where Q emb represents the question matrix for storing question embedding vectors;

[0028] S62: Through # Calculate the cosine similarity score ω between each question and all other questions, where for q e1 and The cosine similarity score ω ranges from [-1,1].

[0029] Preferably, when generating the adversarial question sequence, the following steps are included:

[0030] S71: After obtaining the ω of each question text vector and all other questions, sort all ω in descending order;

[0031] S72: Replace each question in the question sequence with the most similar unused question, thereby generating an adversarial question sequence;

[0032] S73: Merge the adversarial question sequence with the original dataset to obtain the input data q t , used for model training.

[0033] Preferably, when computing the embedding vector between the question and each potential concept, the following steps are included:

[0034] S81: First, input the question q t With the embedding matrix Multiply them together to get a continuous embedding vector

[0035] S82: By calculating the formula K t With each keyway M k The inner product of (i) is provided to the Softmax layer to obtain the vector w t Used to represent the connection between the problem and each potential concept, where M k (i) is the embedding vector of the i-th latent concept.

[0036] Preferably, when storing the relevant information of the question sequence, there are the following steps:

[0037] S91: Matrix Used to save the student's knowledge state at the current time t, where N is the number of potential concepts, d k is the state dimension. Before, the model needs to use w t right The weighted sum of all memory slots in the , we get the reading vector r representing the student's mastery of the knowledge point t :

[0038]

[0039] Here, r t Indicates from The knowledge state information related to the input question is included in the algorithm, but it does not include information about the question itself.

[0040] S92: r t Embedding vector with question Linear combination, and r t With k t Feed it to the Tanh layer to calculate the concept summary vector f t ∈R N :

[0041]

[0042] in W1 is a weight matrix and d1 is a bias vector, which enables the model to generate non-zero activation values ​​when the input is zero, so as to better learn and represent complex relationships in the data. t is a dimension d k A vector representing the problem q t Embedded expression in the model. t Reflects the input problem q t and in Information about the status of related concepts in .

[0043] S93: Finally, f t The combined memory vector is obtained by predicting the student's performance through a fully connected layer and a sigmoid activation function.

[0044]

[0045] It is the student performance predicted by the model based on the problem information and the student's knowledge status. W2 is a weight matrix and b2 is a bias vector, which enables the model to generate non-zero activation values ​​when the input is zero, so as to better learn and represent the complex relationships in the data. Local concept feature information (which represents the correlation between latent concepts) to predict the correct answer q t probability.

[0046] Preferably, when introducing the nonlinear factor, the following steps are included:

[0047] S101: The output of the hierarchical convolutional layer H Pass it to the ReLU function to introduce nonlinear factors;

[0048] S102: Pass The computation regulates the transmission and transformation of information in the learning process and controls the information flow in the learning process. is the output matrix of the hierarchical convolutional layer, which represents the correlation between latent concepts. t Represents the problem q at time t t The associated feature vector.

[0049] Preferably, when using the adversarial data enhanced convolutional memory network model to predict the student's answer results, the following steps are included:

[0050] S201: Generate a combined memory vector using the adversarial data enhancement method And the convolution summary vector generated in step 3 They are concatenated to form a new comprehensive feature vector, which can comprehensively consider the disturbance information of the problem and the local features of the relevant knowledge points to provide a more comprehensive representation;

[0051] S202:Use formula The concatenated comprehensive feature vector is input into a fully connected layer with a Sigmoid activation function, where W∈R |Q|×2N is the learnable weight matrix, b is the bias vector, p t To predict q t A scalar probability of a correct answer, indicating the model's probability that the student will correctly answer the input question q at the current time t t probability prediction.

[0052] Preferably, the advantages of adversarial data enhancement and convolutional memory networks are combined to more accurately capture the complex relationship between students' knowledge status and underlying concepts, thereby improving the performance of knowledge tracking. At the same time, the adversarial data enhanced convolutional memory network model can better predict students' answer results and improve the accuracy and robustness of the knowledge tracking system.

[0053] Beneficial effects of the present invention:

[0054] 1. Compared with traditional knowledge tracking methods, most models are only trained and tested on clean learning data, which cannot truly reflect students' actual learning situation. They often process each knowledge point relatively independently, while ignoring the potential connections between knowledge points. This knowledge tracking method uses the method of calculating the similarity between the original question and the unanswered question to generate adversarial samples, uses adversarial perturbations to generate challenging question sequences, and dynamically represents the students' knowledge status by designing a key-value storage network to capture the connection between students' answers to questions.

[0055] 2. The convolutional key-value storage method uses an attention-based approach to map the association between answered questions and latent concepts to store their knowledge status in the learning task; and by using one-dimensional convolution to capture local patterns in the knowledge point vector, it identifies the connection between adjacent knowledge points, which helps predict students' answers;

[0056] 3. Combining the advantages of adversarial data enhancement and convolutional memory networks, it can more accurately capture the complex relationship between students' knowledge status and underlying concepts, thereby improving the performance of knowledge tracking. At the same time, using adversarial data to enhance the convolutional memory network model can better predict students' answer results and improve the accuracy and robustness of the knowledge tracking system. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 What is shown is a schematic diagram of the workflow of a knowledge tracking method based on a data-enhanced convolutional memory network of the present invention;

[0058] Figure 2 Shown is a schematic diagram of the workflow of generating adversarial question sequences using a knowledge tracking method based on a data-enhanced convolutional memory network of the present invention. DETAILED DESCRIPTION

[0059] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0060] See also Figure 1-2 The present invention provides an embodiment: a knowledge tracking method based on a data-enhanced convolutional memory network, comprising the following method:

[0061] S11: Generate a sequence of adversarial questions corresponding to the original questions;

[0062] S12: Integrate and store question sequences and dynamically update students’ knowledge status;

[0063] S13: Use adversarial data to enhance the convolutional memory network model to predict students’ answer results.

[0064] Preferably, the convolutional key-value storage method uses an attention-based method to map the association between answered questions and latent concepts to store their knowledge states in the learning task; and by using one-dimensional convolution to capture local patterns in the knowledge point vector, thereby identifying the connection between adjacent knowledge points, which helps to predict students' answers.

[0065] Preferably, when generating the adversarial question sequence, the following steps are included:

[0066] S21: Convert question text into embedding vector;

[0067] S22: Calculate the similarity scores between questions;

[0068] S23: Generate an adversarial question sequence.

[0069] Preferably, when dynamically updating the knowledge state of the student, the following steps are included:

[0070] S31: Calculate the embedding vectors between the question and each potential concept;

[0071] S32: Store the relevant information of the question sequence;

[0072] S33: Update and store the knowledge state of the student;

[0073] S34: Extract the local features between potential question concepts.

[0074] Preferably, when extracting the local features between question concepts, the following steps are included:

[0075] S41: Extract the concept information contained in the question;

[0076] S42: Add a residual connection and generate a hierarchical convolutional layer H;

[0077] S43: Introduce a non-linear factor.

[0078] Preferably, when converting the question text into an embedding vector, the following steps are included:

[0079] S51: Describe the text of the question as a sentence L containing m words i ={w1, w2, w3……, w m}, where w i represents a word, 1 < i ≤ m;

[0080] S52: Use the Word2vec method to convert each word into a vector with a dimension of d0, that is, each question text L i corresponds to a matrix with a dimension of d0 × m;

[0081] S53: Calculate the average value of the word vectors contained in each question text in turn to obtain the embedding vector of the question text where represents the embedding vector converted from a question text, 1 < i ≤ n, and n is the serial number of the question.

[0082] Preferably, when calculating the similarity score between questions, the following steps are included:

[0083] S61: Store the embedding vectors of the question texts in Q emb in turn, where Q emb represents the question matrix for storing the question embedding vectors;

[0084] S62: Pass Compute the cosine similarity score ω between each question and all other questions, where for and The cosine similarity score ω ranges from [-1,1].

[0085] Preferably, when generating the adversarial question sequence, the following steps are included:

[0086] S71: After obtaining the ω of each question text vector and all other questions, sort all ω in descending order;

[0087] S72: Replace each question in the question sequence with the most similar unused question, thereby generating an adversarial question sequence;

[0088] S73: Merge the adversarial question sequence with the original dataset to obtain the input data q t , used for model training.

[0089] Preferably, when computing the embedding vector between the question and each potential concept, the following steps are included:

[0090] S81: First, input the question q t With the embedding matrix Multiply them together to get a continuous embedding vector

[0091] S82: By calculating the formula K t With each keyway M k The inner product of (i) is provided to the Softmax layer to obtain the vector w t Used to represent the connection between the problem and each potential concept, where M k (i) is the embedding vector of the i-th latent concept.

[0092] Preferably, when storing the relevant information of the question sequence, there are the following steps:

[0093] S91: Matrix Used to save the student's knowledge state at the current time t, where N is the number of potential concepts, d k is the state dimension. Before, the model needs to use w t right The weighted sum of all memory slots in the , we get the reading vector r representing the student's mastery of the knowledge point t :

[0094]

[0095] Here, r t Indicates from The knowledge state information related to the input question is included in the algorithm, but it does not include information about the question itself.

[0096] S92: r t Embedding vector with question Linear combination, and r t With k t Feed it to the Tanh layer to calculate the concept summary vector f t ∈R N :

[0097]

[0098] in W1 is a weight matrix and b1 is a bias vector, which enables the model to generate non-zero activation values ​​when the input is zero, so as to better learn and represent complex relationships in the data. t is a dimension d k A vector representing the problem q t Embedded expression in the model. t Reflects the input problem q t and in Information about the status of related concepts in .

[0099] S93: Finally, f t The combined memory vector is obtained by predicting the student's performance through a fully connected layer and a sigmoid activation function.

[0100]

[0101] It is the student performance predicted by the model based on the problem information and the student's knowledge status. W2 is a weight matrix and b2 is a bias vector, which enables the model to generate non-zero activation values ​​when the input is zero, so as to better learn and represent the complex relationships in the data. Local concept feature information (which represents the correlation between latent concepts) to predict the correct answer q t probability.

[0102] Preferably, when updating and storing the student's knowledge status, there are the following steps:

[0103] S101: The update of is controlled by two gating mechanisms: adding a gate h tand delete gate z t When the question and the student’s answer are input at time t, these two gating mechanisms can determine the proportion of matrix information update. Specifically, h t Write the new information into the matrix, and z t Delete old information from the matrix.

[0104] Assume (δ(q t ), y t ) is the question answer tuple obtained at time t. In order to obtain the embedding vector Embed it into the matrix Then, With the concept summary vector f t Connect to form updates Used to combine the latest question and answer information with the corresponding knowledge status summary to update Memory, delete gate vector It is obtained by the following formula:

[0105] z t =Sigmod(W e ·v t +b e )#

[0106] Among them, W e is a learnable weight matrix, b e is a bias vector.

[0107] S102: Given a deletion vector z t , the knowledge state is stored from Updated to

[0108]

[0109] Here t (i) represents the relevant weight, and Respectively and The i-th row vector in .

[0110] S103: Add the vector generated by the gate It is obtained by the following formula:

[0111] h t =Tanh(W a ·v t +b a )#

[0112] Among them, W a is a learnable weight matrix, ba is a bias vector.

[0113] S104: Finally, concept state memory Calculated as:

[0114]

[0115] This step will come from adding gate h t Controls the writing of new information to The corresponding i-th row in .

[0116] Preferably, when extracting the conceptual information contained in the title, the following steps are included:

[0117] S201: store the matrix M of the question information k Converted to a one-dimensional matrix vector and calculated through a one-dimensional sliding convolution window to get the output O t-1 To extract the potential conceptual information contained in the title:

[0118] O t-1 =f(Wx t-1 T +b)#

[0119] The input is represented as The weight corresponding to the sliding window is parameterized as W, b is the bias vector, and f represents the nonlinear transformation function. By learning the hierarchical convolution structure, the complex relationship between potential concepts can be more accurately captured, improving the performance and expression ability of the model.

[0120] Preferably, when adding a residual connection and generating a hierarchical convolutional layer H, the following steps are included:

[0121] S301: In order to prevent the subsequent learning interaction convolution from affecting the previous convolution module, the second half of the sliding window is shielded (i.e., set to 0). Therefore, the sliding window takes d consecutive learning interactions as input and maps them to a single output element. The number of feature maps is set to k, which means that there will be k convolution results. In addition, in order to speed up the training process, the output O of the convolution layer is t-1 Add a residual connection x t-1 As new output Right now:

[0122]

[0123] This means that the new output from the output of the convolutional layer is added to the original input to get the final output. This residual connection helps prevent the gradient vanishing problem while maintaining the stability and performance of the model.

[0124] S302: Next, k convolutional layers are stacked together to form a hierarchical convolutional layer H, where the lower layers are mainly used to capture the concepts contained in the questions in the recent period of time, while the higher layers can monitor a longer time range.

[0125] Preferably, when introducing the nonlinear factor, the following steps are included:

[0126] S401: The output of the hierarchical convolutional layer H Pass it to the ReLU function to introduce nonlinear factors;

[0127] S402: Pass The computation regulates the transmission and transformation of information in the learning process and controls the information flow in the learning process. is the output matrix of the hierarchical convolutional layer, which represents the correlation between potential concepts, y t Represents the problem q at time t t The associated feature vector.

[0128] Preferably, when using the adversarial data enhanced convolutional memory network model to predict the student's answer results, the following steps are included:

[0129] S501: Generate a combined memory vector using the adversarial data enhancement method And the convolution summary vector generated in step 3 They are concatenated to form a new comprehensive feature vector, which can comprehensively consider the disturbance information of the problem and the local features of related knowledge points to provide a more comprehensive representation;

[0130] S502: Utilize formula The concatenated comprehensive feature vector is input into a fully connected layer with a Sigmoid activation function, where W∈R |Q|×2N is the learnable weight matrix, b is the bias vector, p t To predict q t A scalar probability of a correct answer, indicating the model's probability that the student will correctly answer the input question q at the current time t t probability prediction.

[0131] Preferably, the advantages of adversarial data enhancement and convolutional memory networks are combined to more accurately capture the complex relationship between students' knowledge status and underlying concepts, thereby improving the performance of knowledge tracking. At the same time, the adversarial data enhanced convolutional memory network model can better predict students' answer results and improve the accuracy and robustness of the knowledge tracking system.

[0132] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the above embodiments, and various changes can be made within the knowledge scope of those skilled in the art without departing from the spirit of the present invention.

Claims

1. A knowledge tracking method based on data-enhanced convolutional memory network, including the following methods: S11: Generate a sequence of adversarial questions corresponding to the original questions; S12: Integrate and store question sequences and dynamically update students’ knowledge status; S13: Use adversarial data to enhance the convolutional memory network model to predict students’ answer results.

2. The knowledge tracking method based on data-enhanced convolutional memory network according to claim 1, characterized in that: When generating adversarial question sequences, the following steps are involved: S21: Convert question text into embedding vector; S22: Calculate the similarity scores between questions; S23: Generate adversarial question sequences.

3. The knowledge tracking method based on data-enhanced convolutional memory network according to claim 2 is characterized in that: When dynamically updating the student's knowledge status, the following steps are included: S31: Calculate the embedding vector between the question and each latent concept; S32: storing relevant information of the question sequence; S33: Update and store the student's knowledge status; S34: Extract local features between potential problem concepts.

4. The knowledge tracking method based on data-enhanced convolutional memory network according to claim 3 is characterized in that: When extracting local features between problem concepts, the following steps are included: S41: extracting conceptual information contained in the question; S42: Add residual connection and generate hierarchical convolution layer H; S43: Introducing nonlinear factors.

5. The knowledge tracking method based on data-enhanced convolutional memory network according to claim 4 is characterized in that: When converting the question text into an embedding vector, the following steps are involved: S51: Describe the text of the question as a sentence L containing m words i ={w1,w2,w3……,w m }, where w i Represents a word, 1 <i≤m; S52: Use the Word2vec method to convert each word into a vector of dimension d0, that is, each question text L i Corresponding to a matrix of dimension d0×m; S53: Calculate the average value of the word vectors contained in each question text in sequence to obtain the embedding vector of the question text where represents the embedding vector converted from a question text, 1 < i ≤ n, and n is the serial number of the question.

6. The knowledge tracking method based on data-enhanced convolutional memory network according to claim 5, characterized in that: When calculating the similarity score of a question, the following steps are involved: S61: Store the embedding vectors of the question text one by one in Q emb In, Q emb Represents the question matrix storing the question embedding vector; S62: Pass Compute the cosine similarity score ω between each question and all other questions, where for and The cosine similarity score ω ranges from [-1,1].

7. The knowledge tracking method based on data-enhanced convolutional memory network according to claim 6, characterized in that: When generating adversarial question sequences, the following steps are involved: S71: After obtaining each question text vector and the ω of all other questions, sort all ω in descending order; S72: Replace each question in the question sequence with the most similar unused question, thereby generating an adversarial question sequence; S73: Merge the adversarial question sequence with the original dataset to obtain the input data q t , used for model training.

8. The knowledge tracking method based on data-enhanced convolutional memory network according to claim 7, characterized in that: When computing the embedding vector between the question and each latent concept, the following steps are involved: S81: First, input the question q t With the embedding matrix Multiply them together to get a continuous embedding vector S82: By calculating the formula K t With each keyway M k The inner product of (i) is provided to the Softmax layer to obtain the vector w t Used to represent the connection between the problem and each potential concept, where M k (i) is the embedding vector of the i-th latent concept.

9. The knowledge tracking method based on data-enhanced convolutional memory network according to claim 8, characterized in that: Storing information about a sequence of questions involves the following steps: S91: Matrix Used to save the student's knowledge state at the current time t, where N is the number of potential concepts, d k is the state dimension, when generating the memory vector Before, the model needs to use w t right The weighted sum of all memory slots in the , we get the reading vector r representing the student's mastery of the knowledge point t : where r t Indicates from The knowledge state information related to the input question is included in the algorithm, but it does not include information about the question itself. S92: r t Embedding vector with question Linear combination, and r t With k t Feed it to the Tanh layer to calculate the concept summary vector f t ∈R N : in W1 is a weight matrix, b1 is a bias vector, which enables the model to generate non-zero activation values ​​when the input is zero, so as to better learn and represent complex relationships in the data. t is a dimension d k A vector representing the problem q t The embedded expression in the model, f t Reflects the input problem q t and in Information about the status of related concepts in . S93: Finally, f t The combined memory vector is obtained by predicting the student's performance through a fully connected layer and a sigmoid activation function. It is the student performance predicted by the model based on the question information and the student's knowledge status. W2 is a weight matrix and b2 is a bias vector, which enables the model to generate non-zero activation values ​​when the input is zero, so as to better learn and represent complex relationships in the data. Local concept feature information (which represents the correlation between latent concepts) to predict the correct answer q t probability.

10. The knowledge tracking method based on data-enhanced convolutional memory network according to claim 9, characterized in that: When introducing nonlinear factors, the following steps are included: S101: The output of the hierarchical convolutional layer H Pass it to the ReLU function to introduce nonlinear factors; S102: Pass The computation regulates the transmission and transformation of information in the learning process and controls the information flow in the learning process. is the output matrix of the hierarchical convolutional layer, which represents the correlation between potential concepts, y t Represents the problem q at time t t The associated feature vector.

11. The knowledge tracking method based on data-enhanced convolutional memory network according to claim 10, characterized in that: When using the adversarial data enhanced convolutional memory network model to predict student answer results, the following steps are included: S201: Generate a combined memory vector using the adversarial data enhancement method And the convolution summary vector generated in step 3 They are concatenated to form a new comprehensive feature vector, which can comprehensively consider the disturbance information of the problem and the local features of the relevant knowledge points to provide a more comprehensive representation; S202:Use formula The concatenated comprehensive feature vector is input into a fully connected layer with a Sigmoid activation function, where W∈R |Q|×2N is the learnable weight matrix, b is the bias vector, p t To predict q t A scalar probability of a correct answer, indicating the model's probability that the student will correctly answer the input question q at the current time t t probability prediction.

Citation Information

Patent Citations

  • Negative sample determination method, related equipment and readable storage medium

    CN113849603A

  • MKVMN model-based student learning ability prediction method

    CN117011098A

  • Knowledge tracking method based on one-dimensional residual convolution self-coding and related device

    CN117131930A

  • Knowledge tracking method and system based on dynamic neural network

    CN118014794A

  • Attention knowledge tracking method based on multi-scale space-time convolution

    CN118095531A

Cited By

  • Dynamic knowledge tracking method and device based on problem skill correlation

    CN120316617A