Student cognitive state three-stage labeling method and system based on graph neural network and application

Abstract

The invention discloses a student cognitive state three-stage labeling system based on a graph neural network. The cognitive state of the student is the mastering degree of the student on different knowledge points in the learning process, and the marking system aims at marking the accurate cognitive state for the student through the historical answer information of the student. The core of the method is that a more reasonable student cognitive state labeling system is constructed through a graph neural network method and a three-stage learning process modeling mode. The system is helpful to more accurately evaluate the learning condition of the student, and provides powerful support for personalized teaching.

Description

Technical Field

[0001] The present invention belongs to the field of computer science and technology, and relates to a three-stage labeling method, system and application of student cognitive states based on graph neural network. Background Art

[0002] The task of labeling students' cognitive states is a hot topic in the field of computer intelligent education. Students' cognitive states refer to the degree of students' mastery of the corresponding knowledge points. The task of labeling students' cognitive states refers to labeling the dynamic cognitive states of different knowledge points in the learning process based on the students' past answers to questions related to different knowledge points. For example, a student first gets two addition questions right and wrong, and then gets a subtraction question wrong. The cognitive state labeling system can label the student's changing cognitive states of the two knowledge points of addition and subtraction during the learning process.

[0003] Traditional methods of annotating students' cognitive states mostly focus on predicting whether students' answers are right or wrong, while ignoring the rationality of the cognitive states generated by the annotations. This rationality includes but is not limited to: 1) when students improve their mastery of a certain knowledge point, their mastery of related knowledge points should also improve, and vice versa; 2) when students answer a question correctly, their mastery of the knowledge point tested should improve, and vice versa. Summary of the invention

[0004] In order to solve the deficiencies in the prior art, the purpose of the present invention is to provide a three-stage annotation method, system and application of student cognitive state based on graph neural network. The three-stage annotation method of student cognitive state based on graph neural network is used to model the correlation between different knowledge points in the process of knowledge mastery change, and to capture the reasonable knowledge mastery changes before and after answering questions through three-stage learning process modeling. The core of the method is to use graph neural network technology and three-stage learning process modeling to enhance the rationality of student cognitive state annotation.

[0005] The three-stage labeling method of student cognitive status based on graph neural network proposed in the present invention includes the following steps:

[0006] Step 1: Select a data set covering the answer records of students in different grades, and divide the data into training set, validation set and test set after data preprocessing;

[0007] Step 2: Calculate the correlation between different knowledge points through the preprocessed data set to form a relationship diagram of knowledge points that have been revised first and last and a relationship diagram of similar knowledge points;

[0008] Step 3: Initialize the student knowledge memory matrix and use the matrix elements to represent the memory code of each knowledge point;

[0009] Step 4: Represent the students’ history answers and knowledge points with learnable embedding vectors;

[0010] Step 5: Divide each student's answer to a question into three stages for modeling: the first stage is modeling knowledge retrieval, which retrieves and aggregates the relevant knowledge memory encoding of the question from the student's knowledge memory matrix to model the student's answer result and obtain the correct answer prediction score; the second stage is modeling memory reinforcement, which models the positive reinforcement and negative reinforcement of the student's knowledge memory according to the actual correct or incorrect answer situation, and updates the student's knowledge memory matrix; the third stage is modeling knowledge learning or forgetting, which models the learning and forgetting behavior of knowledge points according to the student's active learning behavior after answering the question, and updates the student's knowledge memory matrix; the knowledge point relationship graph neural network is integrated into these three steps to capture the changes in the cognitive state of related knowledge points;

[0011] Step 6: Train the system method model parameters, use the training samples to optimize the parameters of the network model according to the objective function, and use the validation set to tune the model parameters;

[0012] Step 7: Map students’ cognitive status through the student knowledge memory matrix.

[0013] In step 1, we selected public data sets such as ASSISTments2009-2010, ASSISTments2012-2013, and Junyi, covering the answer records of mathematics students in several stages of primary and secondary schools. We divided the answer sequences of students with more than 100 answers into one or more answer sequences with a length of 100, and filtered out the remaining sequences with less than 10 answers. We filled the sequences with a length greater than or equal to 10 but less than 100 with placeholder identifiers to 100, and then divided all the sequences into training set, validation set, and test set in a ratio of 8:1:1.

[0014] In the step 2, the knowledge point relationship graph and the similar knowledge point relationship graph are constructed in sequence. The specific method of constructing the graph is to first calculate the similarity scores of any two knowledge points. Prerequisites Then, based on the calculated similarity scores and prerequisite scores between any two knowledge points, a threshold is set to determine whether to establish a relationship edge between the knowledge point pairs.

[0015] The similarity score and the prerequisite score are calculated as follows:

[0016]

[0017]

[0018] Among them, U is the student set, u is a student in the set, A collection of history answers for students u, is the answer of student u at time t, is the answer of student u at time t'. for The correctness of the answer, 1 means correct answer, 0 means wrong answer; express Related knowledge points: I(·) is an indicator function. If the value of the expression input to the function is true, the output is 1, otherwise it outputs 0.

[0019] After obtaining the similarity score and the prerequisite score, a threshold η=0.7 is set for both the similarity and prerequisite relationships to distinguish between similar / non-similar knowledge point pairs and / or prerequisite / non-prerequisite knowledge point pairs. Then mark c i c j Similar knowledge points of Otherwise, it is not marked. Then mark c i c j Prerequisites for j c i The subsequent knowledge point of Otherwise, no mark is given.

[0020] The student knowledge memory matrix H in step 3 includes one or more knowledge points and knowledge memory encoding dimensions. The number of rows of the knowledge memory matrix H represents the number of knowledge points, and the number of columns of the knowledge memory matrix represents the knowledge memory encoding dimensions. The knowledge memory matrix H is initialized as a learnable parameter matrix H0, and the learnable parameter matrix H0 is continuously updated in the subsequent three-stage modeling of the modeling knowledge retrieval stage, the modeling memory reinforcement stage, and the modeling knowledge learning or forgetting stage: H0, H1-, H1, H2-, H2, ..., H t -,H t ,….H t - represents the knowledge memory matrix of the student before answering the question at time t (and only H1-=H0), H t Represents the knowledge memory matrix after answering the question at time t.

[0021] In step 4, the embedding vector representation of the student's history answer questions and knowledge points is in the form of dimension d e A learnable real number vector, wherein the specific value of each dimension in the learnable real number vector is learned during the model training process as an implicit representation of the student's historical answer questions and knowledge points.

[0022] In step 5, the knowledge point relationship graph neural network in the three stages is extended from the prototype knowledge point relationship graph neural network c iis a certain knowledge point. For knowledge point c i The input node prototype features, For knowledge point c i The output node prototype features of d0,d L are the dimensions of the input and output features, d1,…,d L-1 is the feature dimension of the middle layer. The process of modeling node features of the knowledge point relationship graph neural network is as follows:

[0023] a1. For each sub-layer of the prototype knowledge point relationship graph neural network (assuming it is the lth sub-layer), each sub-layer aggregates the knowledge point node features of the previous layer on each knowledge point relationship graph, and obtains the aggregated node features through feature transformation through a nonlinear feedforward neural network. Knowledge point relationship graph The aggregated node features of the previous layer of knowledge points and the aggregated node features after feature transformation are shown in the following formulas:

[0024] Aggregated previous layer knowledge point node features

[0025] Node features after feature transformation and aggregation

[0026] in, Indicates c i In the figure The set of neighbor nodes on , For the knowledge point relationship diagram The learnable matrix parameters of the sub-layer are used to model the message passing between the nodes of the graph neural network. ReLU(x)=max(0,x) is the activation function. For knowledge point c i ,c j In the knowledge point relationship diagram The correlation coefficient on .

[0027] a2. A sub-layer fuses the aggregated node features of the three knowledge point relationship graphs to obtain the overall features of the knowledge point nodes of the sub-layer:

[0028]

[0029] Among them, when the feature dimension of the previous sub-layer network is the same as the feature dimension of the current layer network, the overall features of the knowledge point nodes output by the previous layer need to be added after adding the node features of the three knowledge point relationship graphs. Otherwise, only the node features of the three knowledge point relationship graphs need to be added.

[0030] a3. After passing through L sublayers of the prototype knowledge point graph neural network, the output node prototype feature of each knowledge point of the prototype knowledge point graph neural network can be obtained

[0031] If the prototype feature of a single knowledge node The feature matrix composed of multiple knowledge point node prototype features is stacked. The number of rows is the number of knowledge points, and the number of columns is the node prototype feature dimension. The knowledge point node feature matrix can also be obtained. And the prototype knowledge point relationship graph neural network is represented as That is, the prototype feature of the input knowledge point node Stack in the same way.

[0032] The correlation coefficient of the knowledge point on the knowledge point relationship diagram indicates the strength of the relationship between the knowledge points. The calculation method is (taking knowledge point c as i ,c j In the knowledge point relationship diagram The correlation coefficient on is taken as an example):

[0033]

[0034] in, To calculate the knowledge point relationship diagram The learnable matrix parameters of the correlation coefficient between the above two knowledge points. is the activation function. For knowledge point c i ,c j The embedded vector representation of .

[0035] In the modeling knowledge retrieval stage described in step 5, taking the student answering the question at time t as an example, the specific steps are as follows:

[0036] Step 5.1.1. Use the knowledge retrieval extended from the prototype knowledge point relationship graph neural network to aggregate students' knowledge memory of different knowledge points:

[0037]

[0038] in, Initialized to the student's knowledge point c for the current answer question t Knowledge memory {d k} L+1 Represents the input feature dimension and the feature dimension of L sub-layers. Different from the prototype knowledge point relationship graph neural network, the knowledge retrieval knowledge point relationship graph neural network increases the demand coefficient of the current question for knowledge points when aggregating the previous layer of knowledge point memory, and modifies the aggregation process as follows:

[0039]

[0040] in, For the topic qt For knowledge point c i The demand coefficient, that is, whether the question requires the use of this knowledge point to answer. is the aggregation matrix parameter which is restricted to be positive. t ,c t The knowledge points tested in the question that the students are currently answering.

[0041] Step 5.1.2. Use the aggregated knowledge point c of the current answer question t The knowledge memory map is the mastery level (cognitive state) of the knowledge point and the difficulty of the current question. By comparing, we can get the probability of predicting whether the student can answer the question correctly:

[0042]

[0043] Among them, w h In order to map the knowledge memory of a knowledge point to the mapping vector of the degree of mastery of the knowledge point, it is also restricted to a positive number.

[0044] The requirement coefficient of the question for the knowledge point is calculated as follows (with question q i For knowledge point c j The demand coefficient is taken as an example):

[0045]

[0046] Among them, W req is the learnable parameter matrix. For the topic q i The embedding vector representation of For knowledge point c j The embedded vector representation of .

[0047] Difficulty level of the question It consists of a multi-layer perceptron MLP diff (·) is calculated, and the specific process is:

[0048]

[0049] in are learnable parameter matrices and vectors. For the topic q t The knowledge points it examines t The joint embedding vector representation of . is a vector concatenation operation. The multilayer perceptron process can be symbolized as Among them, d e is the embedding vector dimension, d h The middle dimension.

[0050] and w h Use the softmax function to restrict to positive numbers:

[0051]

[0052]

[0053] in, Representation Matrix The i-th row and j-th column of w h | i Represents vector w h The i-th dimension of . is the learnable parameter matrix and vector before restriction, d k Master the vector w for mapping knowledge points h Dimension.

[0054] In the modeling and memory strengthening stage described in step 5, taking the student finishing the question at time t as an example, the specific steps are as follows:

[0055] Step 5.2.1. Use another multilayer perceptron to calculate the correct answer q t The knowledge points tested by the following questions c t The initial knowledge memory positive reinforcement vector:

[0056]

[0057] in, is the vector concatenation operation, For the topic q t The knowledge points it examines t The joint embedding vector representation of Indicates that the student is answering knowledge point c t The knowledge memory vector of the previous knowledge point, d e is the embedding vector dimension, d h is the middle layer dimension, d k is the knowledge memory vector dimension.

[0058] Step 5.2.2. Set the initial knowledge memory positive reinforcement vectors of the remaining knowledge points to vector 0, and use them as the input vectors of the knowledge memory positive reinforcement knowledge point relationship graph neural network:

[0059]

[0060] in, That is knowledge point c i The initial knowledge memory positive reinforcement vector.

[0061] The initial knowledge memory positive reinforcement vectors of all knowledge points are stacked into an initial knowledge memory positive reinforcement matrix with the number of rows equal to the number of knowledge points and the number of columns equal to the vector dimension. And through the knowledge memory positive reinforcement knowledge point relationship graph neural network, the knowledge memory positive reinforcement matrix of the knowledge point after network propagation is obtained.

[0062]

[0063] Among them, {d k} L+1 The feature dimensions of the L layer of the neural network representing the relationship between the input layer and the knowledge memory positive reinforcement knowledge point are both d k , which is consistent with the dimension of the knowledge memory vector.

[0064] Step 5.2.3. Similar to steps 5.2.1 and 5.2.2, obtain the knowledge point c under examination t The initial knowledge memory positive reinforcement vector Then initialize the initial knowledge memory reverse reinforcement vector of each knowledge point Initial knowledge memory reverse reinforcement matrix stacked as knowledge points And reversely strengthen the knowledge point relationship graph neural network GNN through knowledge memory loss Get the knowledge memory reverse reinforcement matrix of knowledge points after network dissemination

[0065]

[0066]

[0067] Step 5.2.4. Determine whether the knowledge memory of the knowledge point is positively reinforced or negatively reinforced based on whether the answer to the current question is correct or not, and update the knowledge memory matrix:

[0068]

[0069] Among them, a t It is the correctness of the student's answer at time t, 1 indicates a correct answer and 0 indicates an incorrect answer.

[0070] In the knowledge learning or forgetting behavior modeling stage described in step 5, taking the student preparing to do the t+1 time question as an example, the specific steps are as follows:

[0071] Step 5.3.1. For the questions you just answered t and the questions to be answered t+1 , according to students' relevant knowledge points c t ,c t+1 Knowledge memory And the joint embedding vector representation of the topic Calculate whether students have learned knowledge point c t ,c t+1 Decision distribution: for learning knowledge point c i ∈{c t ,c t+1}, its decision distribution is:

[0072]

[0073] in, is the vector concatenation operation, d k is the dimension of knowledge memory, d e is the dimension of the embedding vector, MLP dcs The multilayer perceptron used in this generation process, d h is the dimension of the middle layer of the multi-layer perceptron.

[0074] The first dimension represents the knowledge points c that students learn i The probability of not learning knowledge point c i If the value of the first dimension is greater than the value of the second dimension, it means that the student decides to learn knowledge point c. i .

[0075] Step 5.3.2. For the knowledge points to be learned, students i The knowledge memory and the joint embedding vector representation of the question are expressed through a multi-layer perceptron MLP. prg Computational learning knowledge points c i The initial knowledge memory learning vector of:

[0076]

[0077] in, is the vector concatenation operation, d k is the dimension of knowledge memory, d e is the dimension of the embedding vector, d h is the dimension of the middle layer of the multi-layer perceptron.

[0078] Step 5.3.3. Set the initial knowledge memory learning vector of the knowledge point to be The initial knowledge memory learning vectors of the remaining knowledge points are set to vector 0 and used as the input vector of the knowledge learning knowledge point relationship graph neural network:

[0079]

[0080] The input vectors of all knowledge points are stacked as the initial knowledge memory learning matrix of the knowledge points Learning knowledge point relationship graph neural network GNN through knowledge prg Get the knowledge memory learning matrix of knowledge points after network dissemination

[0081]

[0082] Among them, the activation function ReLU is used to limit the output knowledge memory learning to be a positive number.

[0083] Step 5.3.4. Based on the time interval ΔT until the next answer t+1 =T t+1 -T t (T t ,T t+1 is the time point between the two answers), and updates the knowledge memory of each knowledge point after knowledge learning (in terms of c i For example):

[0084]

[0085] in, To answer the current question and the next time to answer the question, the student i Knowledge memory of knowledge points, through c i Knowledge memory learning vector of knowledge points to update.

[0086] in, To learn knowledge points c i The time-aware kernel function is used to model the process of knowledge point memory changing over time, and its growth rate decreases as the time interval increases:

[0087]

[0088] Among them, ⊙ is the Hadamard product. The student has learned the knowledge point c before answering the question at time t i The number of times. To learn knowledge points c i The parameters of the time-aware kernel function are learned through knowledge point relationship graph neural network GNN lrn The calculation process is as follows:

[0089] For knowledge point c i , learn knowledge point c i The calculation process of the time-aware kernel function parameters is:

[0090]

[0091] Among them, the input vector of the network is initialized as knowledge point ci The embedding vector representation of is: softplus(x)=log(1+e x ) is the activation function.

[0092] Step 5.3.5. Based on the time interval ΔT until the next answer t+1 , update the knowledge memory of each knowledge point after knowledge forgetting (in c i For example):

[0093]

[0094] in, To answer the current question and the next time to answer the question, the student i Knowledge memory of knowledge points. is the acquired knowledge point c i knowledge memory. To target forgotten knowledge points c i The time-aware kernel function of

[0095]

[0096] in, To target forgotten knowledge points c i The parameters of the time-aware kernel function are learned through knowledge point relationship graph neural network GNN frt The calculation process is as follows:

[0097] For knowledge point c i , forget knowledge point c i The calculation process of the time-aware kernel function parameters is:

[0098]

[0099] Among them, the input vector of the network is initialized as knowledge point c i The embedding vector representation of is:

[0100] The student cognitive state or knowledge mastery model in step 6 of the present invention uses a binary cross entropy loss function to train parameters. For the t-th answer of student u, the loss function is:

[0101]

[0102] in, is the correctness of student u’s answer at time t, 1 indicates a correct answer and 0 indicates an incorrect answer. The knowledge retrieval stage model predicts that student u will correctly answer question q at time t t probability.

[0103] The overall loss function is the average of the losses over all answers of all students in the training set:

[0104]

[0105] in, Gathering for students, A collection of answers from student u.

[0106] According to the loss training parameters, hyperparameters are adjusted on the validation set. When the model loss on the validation set does not decrease in 10 consecutive training cycles, training is stopped. The hyperparameter combination that can achieve the lowest validation set loss is selected.

[0107] The acquisition of the knowledge point cognitive state (knowledge mastery) described in step 7 uses the mapping vector w h Mapping students' understanding of knowledge point c at time t i Knowledge memory To the knowledge point c i cognitive state get:

[0108]

[0109] The present invention also provides a labeling system for implementing the above-mentioned student cognitive state labeling method, the system comprising: a student answer input encoding module, a three-stage student cognitive state modeling module, and a knowledge point relationship graph neural network module;

[0110] The student answer input encoding module represents the student's historical answer questions and knowledge points with a learnable embedding vector for use by subsequent modules;

[0111] The three-stage student cognitive state modeling module uses a continuously updated knowledge memory matrix to model the knowledge retrieval of students when answering questions based on the questions being examined, model the memory reinforcement of students when answering questions based on the correct or incorrect answers to the questions being examined, and model the knowledge learning or forgetting of students during the intervals between answers based on the questions being examined and the questions to be examined;

[0112] The knowledge point relationship graph neural network module models the process of knowledge retrieval, memory reinforcement, and knowledge learning or forgetting transfer between knowledge points during the student's answering process based on the similar knowledge point relationship graph and the sequential knowledge point relationship graph based on data statistics.

[0113] The method of the present invention also provides the application of the above-mentioned student cognitive status labeling method or the above-mentioned labeling system in online education platforms, actual teaching scenarios, etc. to label students' cognitive status and knowledge point mastery.

[0114] Compared with the prior art, the beneficial effects of the present invention include:

[0115] The knowledge point relationship graph neural network is used to model the changes in students' knowledge mastery of knowledge points during the answering process, making the cognitive state labeling results more reasonable. Its rationality is reflected in the aggregation and transmission of knowledge point features in the knowledge point relationship graph neural network, which simulates the knowledge transfer of students' knowledge mastery of knowledge points during the learning process: students transfer the knowledge mastery of the knowledge points that are taken before or after or similar knowledge points to answer the relevant questions of the knowledge point, and when learning or forgetting the relevant knowledge of the knowledge point, they transfer the changes in mastery to the knowledge points that are taken before or after or similar knowledge points.

[0116] It is more reasonable to use the three-stage modeling of knowledge retrieval, memory reinforcement, knowledge learning or forgetting to model the changes in students' cognitive states. Its rationality is reflected in the fact that knowledge retrieval models the psychological cognitive process of students retrieving knowledge stored in memory during the learning process. Memory reinforcement models the practice effect of the questions produced by students at the end of the test, that is, the increase or decrease in knowledge mastery caused by correct answers reinforcing the path of students' correct retrieval of relevant knowledge, and incorrect answers reinforcing the path of students' incorrect retrieval of relevant knowledge. Knowledge learning or forgetting models the active learning or passive forgetting process of students after the answering, which is reflected in the learning curve and forgetting curve of psychological cognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0117] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without paying any creative work.

[0118] Figure 1 It is a process schematic diagram of the three-stage labeling method of students' cognitive states based on graph neural network of the present invention.

[0119] Figure 2 It is a framework diagram of the entire network model of the present invention.

[0120] Figure 3 This is a model building process of an embodiment of the present invention. DETAILED DESCRIPTION

[0121] The present invention is further described in detail with reference to the following specific examples and drawings. The process, conditions, experimental methods, etc. for implementing the present invention, except for the contents specifically mentioned below, are all common knowledge and common common sense in the art and are not particularly limited by the present invention.

[0122] The cognitive state of students refers to the degree of mastery of different knowledge points by students during the learning process. The labeling system aims to label the students' accurate cognitive state through their historical answer information. The core of the method of the present invention is to construct a more reasonable student cognitive state labeling system through a graph neural network method and a three-stage learning process modeling method. The system helps to more accurately evaluate the learning status of students and provide strong support for personalized teaching.

[0123] The three-stage labeling method of student cognitive status based on graph neural network proposed in the present invention includes the following steps:

[0124] Step 1: Select ASSISTments2009-2010, ASSISTments2012-2013, and Junyi public data sets, covering the answer records of mathematics teaching students in several stages of primary and secondary schools, divide the student answer sequences with more than 100 answers into several answer sequences with a length of 100, and filter out the remaining sequences with less than 10 answers. After filling the sequences with a length greater than or equal to 10 but less than 100 with placeholder identifiers to 100, all sequences are divided into training set, validation set, and test set in a ratio of 8:1:1;

[0125] Step 2: Calculate the correlation between different knowledge points through the preprocessed data set to form a relationship diagram of knowledge points that have been revised first and last and a relationship diagram of similar knowledge points;

[0126] Step 3: Initialize the student knowledge memory matrix and use the matrix elements to represent the memory code of each knowledge point;

[0127] Step 4: Represent the students’ history answers and knowledge points with learnable embedding vectors;

[0128] Step 5: Divide each student's answer to a question into three stages for modeling: the first stage is modeling knowledge retrieval, which retrieves and aggregates the relevant knowledge memory encoding of the question from the student's knowledge memory matrix to model the student's answer results and obtain the predicted score for the correct answer; the second stage is modeling memory reinforcement, which models the positive reinforcement and negative reinforcement of the student's knowledge memory according to the actual correct or incorrect answer situation, and updates the student's knowledge memory matrix; the third stage is modeling knowledge learning or forgetting, which models the learning and forgetting behavior of knowledge points according to the student's active learning behavior after answering the question, and updates the student's knowledge memory matrix. The knowledge point relationship graph neural network is integrated into these three steps to capture the changes in the cognitive state of related knowledge points;

[0129] Step 6: Train the system method model parameters, use the training samples to optimize the parameters of the network model according to the objective function, and use the validation set to tune the model parameters;

[0130] Step 7: Map students’ cognitive status through the student knowledge memory matrix.

[0131] In the step 2, the knowledge point relationship graph and the similar knowledge point relationship graph are constructed in sequence. The specific method of constructing the graph is to first calculate the similarity scores of any two knowledge points. Prerequisites Then, based on the calculated similarity scores and prerequisite scores between any two knowledge points, a threshold is set to determine whether to establish a relationship edge between the knowledge point pairs.

[0132] The similarity score and the prerequisite score are calculated as follows:

[0133]

[0134]

[0135] Among them, U is the student set, u is a student in the set, A collection of history answers for students u, is the answer of student u at time t, is the answer of student u at time t'. for The correctness of the answer, 1 means correct answer, 0 means wrong answer; express I(·) is an indicator function. If the value of the expression input to the function is true, the output is 1, otherwise it outputs 0. After obtaining the similarity score and the prerequisite score, a threshold η = 0.7 is set for both relationships to distinguish similar / non-similar and prerequisite / non-prerequisite knowledge point pairs. Then mark c i c j Similar knowledge points of Otherwise, it is not marked. Then mark c i c j Prerequisites for j c i The subsequent knowledge point of Otherwise, no mark is given.

[0136] The student knowledge memory matrix H in step 3 includes multiple knowledge points and knowledge memory encoding dimensions. The number of matrix rows represents the number of knowledge points, and the number of matrix columns represents the knowledge memory encoding dimensions. It is initialized as a learnable parameter matrix H0 and is continuously updated in the three-stage modeling: H0, H1-, H1, H2-, H2, …, H t -,H t ,….H t - represents the knowledge memory matrix of the student before answering the question at time t (and only H1-=H0), H tRepresents the knowledge memory matrix after answering the question at time t.

[0137] In step 4, the embedding vector representation of the learnable questions and knowledge points is of dimension d e The specific value of each dimension is learned during the model training process as an implicit representation of the questions and knowledge points.

[0138] The knowledge point relationship graph neural network described in step 5 is extended from the prototype knowledge point relationship graph neural network c i is a certain knowledge point. For knowledge point c i The input node prototype features, For knowledge point c i The output node prototype features of d0,d L are the dimensions of the input and output features, d1,…,d L-1 is the feature dimension of the middle layer. The internal structure of the graph neural network is specifically described as:

[0139] a1. For each sub-layer of the prototype knowledge point relationship graph neural network (assuming it is the lth sub-layer), each sub-layer aggregates the knowledge point node features of the previous layer on each knowledge point relationship graph, and obtains the aggregated node features through feature transformation through a nonlinear feedforward neural network. For example:

[0140]

[0141]

[0142] in, Indicates c i In the figure The set of neighbor nodes on , For the knowledge point relationship diagram The learnable matrix parameters of the sub-layer are used to model the message passing between the nodes of the graph neural network. ReLU(x)=max(0,x) is the activation function. For knowledge point c i ,c j In the knowledge point relationship diagram The correlation coefficient on .

[0143] a2. A sub-layer fuses the aggregated node features of the three knowledge point relationship graphs to obtain the overall features of the knowledge point nodes of the sub-layer:

[0144]

[0145] Among them, when the feature dimension of the previous sub-layer network is the same as the feature dimension of the current layer network, the overall features of the knowledge point nodes output by the previous layer need to be added after adding the node features of the three knowledge point relationship graphs. Otherwise, only the node features of the three knowledge point relationship graphs need to be added.

[0146] a3. After passing through L sublayers of the prototype knowledge point graph neural network, the output node prototype feature of each knowledge point of the prototype knowledge point graph neural network can be obtained

[0147] If the prototype feature of a single knowledge node The feature matrix composed of multiple knowledge point node prototype features is stacked. The number of rows is the number of knowledge points, and the number of columns is the node prototype feature dimension. The knowledge point node feature matrix can also be obtained. And the prototype knowledge point relationship graph neural network is represented as That is, the prototype feature of the input knowledge point node Stack in the same way.

[0148] The correlation coefficient of the knowledge point on the knowledge point relationship diagram indicates the strength of the relationship between the knowledge points. The calculation method is (taking knowledge point c as i ,c j In the knowledge point relationship diagram The correlation coefficient on is taken as an example):

[0149]

[0150] in, To calculate the knowledge point relationship diagram The learnable matrix parameters of the correlation coefficient between the above two knowledge points. is the activation function. For knowledge point c i ,c j The embedded vector representation of .

[0151] In the knowledge retrieval modeling stage described in step 5, taking the student answering the question at time t as an example, the specific steps are as follows:

[0152] b1. Using knowledge retrieval extended from the prototype knowledge point relationship graph neural network to aggregate students’ knowledge memory of different knowledge points:

[0153]

[0154] in, Initialized to the student's knowledge point c for the current answer question t Knowledge memory {d k} L+1Represents the input feature dimension and the feature dimension of L sub-layers. Different from the prototype knowledge point relationship graph neural network, the knowledge retrieval knowledge point relationship graph neural network increases the demand coefficient of the current question for knowledge points when aggregating the previous layer of knowledge point memory, and modifies the aggregation process as follows:

[0155]

[0156] in, For the topic q t For knowledge point c i The demand coefficient, that is, whether the question requires the use of this knowledge point to answer. is the aggregation matrix parameter which is restricted to be positive. t ,c t The knowledge points tested in the question that the students are currently answering.

[0157] b2. Use the aggregated knowledge points tested by the current answer question c t The knowledge memory map is the mastery level (cognitive state) of the knowledge point and the difficulty of the current question. By comparing, we can get the probability of predicting whether the student can answer the question correctly:

[0158]

[0159] Among them, w h In order to map the knowledge memory of a knowledge point to the mapping vector of the degree of mastery of the knowledge point, it is also restricted to a positive number.

[0160] The coefficient of knowledge point requirement for the question is calculated as follows (with question q i For knowledge point c j The demand coefficient is taken as an example):

[0161]

[0162] Among them, W req is the learnable parameter matrix. For the topic q i The embedding vector representation of For knowledge point c j The embedded vector representation of .

[0163] Difficulty level of the question It consists of a multi-layer perceptron MLP diff (·) is calculated, and the specific process is:

[0164]

[0165] in are learnable parameter matrices and vectors. For the topic q t The knowledge points it examines t The joint embedding vector representation of . is a vector concatenation operation. The multilayer perceptron process can be symbolized as Among them, d e is the embedding vector dimension, d h The middle dimension.

[0166] and w h Use the softmax function to restrict to positive numbers:

[0167]

[0168]

[0169] in, Representation Matrix The i-th row and j-th column of w h | i Represents vector w h The i-th dimension of . is the learnable parameter matrix and vector before restriction, d k Master the vector w for mapping knowledge points h Dimension.

[0170] In the memory reinforcement modeling stage described in step 5, taking the example of a student finishing the question at time t, the specific steps are as follows:

[0171] c1. Use another multi-layer perceptron to calculate the correct answer q t The knowledge points examined later t The initial knowledge memory positive reinforcement vector:

[0172]

[0173] in, is the vector concatenation operation, For the topic q t The knowledge points it examines t The joint embedding vector representation of Indicates that the student is answering knowledge point c t The knowledge memory vector of the previous knowledge point, d e is the embedding vector dimension, d h is the middle layer dimension, d k is the knowledge memory vector dimension.

[0174] c2. Set the initial knowledge memory positive reinforcement vectors of the remaining knowledge points to vector 0, and use them as the input vectors of the knowledge memory positive reinforcement knowledge point relationship graph neural network:

[0175]

[0176] in, That is knowledge point c i The initial knowledge memory positive reinforcement vector of all knowledge points is stacked into an initial knowledge memory positive reinforcement matrix with the number of rows equal to the number of knowledge points and the number of columns equal to the vector dimension. Finally, the knowledge memory positive reinforcement matrix of the knowledge points after network propagation is obtained through the knowledge memory positive reinforcement knowledge point relationship graph neural network

[0177]

[0178] Among them, {d k} L+1 The feature dimensions of the L layer of the neural network representing the relationship between the input layer and the knowledge memory positive reinforcement knowledge point are both d k , which is consistent with the dimension of the knowledge memory vector.

[0179] c3. Similar to steps c1 and c2, get the knowledge point c under investigation t The initial knowledge memory positive reinforcement vector Then initialize the initial knowledge memory reverse reinforcement vector of each knowledge point Initial knowledge memory reverse reinforcement matrix stacked as knowledge points And reversely strengthen the knowledge point relationship graph neural network GNN through knowledge memory loss Get the knowledge memory reverse reinforcement matrix of knowledge points after network dissemination

[0180]

[0181]

[0182] c4. According to whether the answer to the current question is correct or not, determine whether the knowledge memory of the knowledge point is positive reinforcement or negative reinforcement, and update the knowledge memory matrix:

[0183]

[0184] Among them, a t It is the correctness of the student's answer at time t, 1 indicates a correct answer and 0 indicates an incorrect answer.

[0185] In the knowledge learning or forgetting behavior modeling stage described in step 5, taking the student preparing to do the t+1 time question as an example, the specific steps are as follows:

[0186] d1. For the questions you just answered qt and the questions to be answered t+1 , according to students' relevant knowledge points c t ,c t+1 Knowledge memory And the joint embedding vector representation of the topic Calculate whether students have learned knowledge point c t ,c t+1 Decision distribution: for learning knowledge point c i ∈{c t ,c t+1}, its decision distribution is:

[0187]

[0188] in, is the vector concatenation operation, d k is the dimension of knowledge memory, d e is the dimension of the embedding vector, MLP dcs The multilayer perceptron used in this generation process, d h is the dimension of the middle layer of the multi-layer perceptron.

[0189] Among them, if The first dimension represents the knowledge points c that students learn i The second digit indicates that knowledge point c is not learned. i If the value of the first dimension is greater than the value of the second dimension, it means that the student decides to learn knowledge point c. i .

[0190] d2. For the knowledge points that are decided to be studied, students should i The knowledge memory and the joint embedding vector representation of the question are expressed through a multi-layer perceptron MLP. prg Computational learning knowledge points c i The initial knowledge memory learning vector of:

[0191]

[0192] in, is the vector concatenation operation, d k is the dimension of knowledge memory, d e is the dimension of the embedding vector, d h is the dimension of the middle layer of the multi-layer perceptron.

[0193] d3. Set the initial knowledge memory learning vector of the knowledge point that determines the learning to The initial knowledge memory learning vectors of the remaining knowledge points are set to vector 0 and used as the input vector of the knowledge learning knowledge point relationship graph neural network:

[0194]

[0195] The input vectors of all knowledge points are stacked as the initial knowledge memory learning matrix of the knowledge points Learning knowledge point relationship graph neural network GNN through knowledge prg Get the knowledge memory learning matrix of knowledge points after network dissemination

[0196]

[0197] Among them, the activation function ReLU is used to limit the output knowledge memory learning to be a positive number.

[0198] d4. Based on the time interval ΔT until the next answer t+1 =T t+1 -T t (T t ,T t+1 is the time point between the two answers), and updates the knowledge memory of each knowledge point after knowledge learning (in terms of c i For example):

[0199]

[0200] in, To answer the current question and the next time to answer the question, the student i Knowledge memory of knowledge points, through c i Knowledge memory learning vector of knowledge points to update.

[0201] in, To learn knowledge points c i The time-aware kernel function is used to model the process of knowledge point memory changing over time, and its growth rate decreases as the time interval increases:

[0202]

[0203] Among them, ⊙ is the Hadamard product. The student has learned the knowledge point c before answering the question at time t i The number of times. To learn knowledge points c i The parameters of the time-aware kernel function are learned through knowledge point relationship graph neural network GNN lrn The calculation process is as follows:

[0204] For knowledge point c i , learn knowledge point c i The calculation process of the time-aware kernel function parameters is:

[0205]

[0206] Among them, the input vector of the network is initialized as knowledge point c i The embedding vector representation of is: is the activation function.

[0207] d5. Similar to step d4, according to the time interval ΔT until the next answer t+1 , update the knowledge memory of each knowledge point after knowledge forgetting (in c i For example):

[0208]

[0209] in, To answer the current question and the next time to answer the question, the student i Knowledge memory of knowledge points. is the acquired knowledge point c i knowledge memory. To target forgotten knowledge points c i The time-aware kernel function of

[0210]

[0211] in, To target forgotten knowledge points c i The parameters of the time-aware kernel function are learned through knowledge point relationship graph neural network GNN frt The calculation process is as follows:

[0212] For knowledge point c i , forget knowledge point c i The calculation process of the time-aware kernel function parameters is:

[0213]

[0214] Among them, the input vector of the network is initialized as knowledge point c i The embedding vector representation of is:

[0215] The acquisition of the knowledge point cognitive state (knowledge mastery) described in step 7 uses the mapping vector w h Mapping students' understanding of knowledge point c at time t i Knowledge memory To the knowledge point c i cognitive state get:

[0216]

[0217] The student cognitive state model in the present invention uses a binary cross entropy loss function to train parameters. For the rth answer of student u, the loss function is:

[0218]

[0219] in, is the correctness of student u’s answer at time t, 1 indicates a correct answer and 0 indicates an incorrect answer. The knowledge retrieval stage model predicts that student u will correctly answer question q at time t t probability.

[0220] The overall loss function is the average of the losses over all answers of all students in the training set:

[0221]

[0222] in, Gathering for students, A collection of answers from student u.

[0223] According to the loss training parameters, hyperparameters are adjusted on the validation set. When the model loss on the validation set does not decrease in 10 consecutive training cycles, training is stopped. The hyperparameter combination that can achieve the lowest validation set loss is selected.

[0224] The framework diagram of the entire network model in the present invention is as follows: Figure 2 As shown:

[0225] t - The knowledge memory matrix at the moment is retrieved by knowledge - Answer at any time to make predictions.

[0226] t - The knowledge memory matrix at time t receives the answer at time t, and obtains the knowledge memory matrix at time t through memory reinforcement.

[0227] The knowledge memory matrix at time t receives the answers at time t and time t+1, and obtains (t+1) through knowledge learning or forgetting. - Moment knowledge memory matrix.

[0228] t - Knowledge memory matrix at time t, knowledge memory matrix at time t, (t+1) - The knowledge memory matrix at time t is obtained by mapping - time, time t and (t+1) - The students’ cognitive state at each moment.

[0229] The knowledge point relationship graph neural network models the related changes in the degree of mastery between knowledge points, which runs through the three-stage student cognitive state labeling.

[0230] The model building flow chart in one embodiment of the present invention is as follows: Figure 3 As shown:

[0231] Take a student in a first-grade mathematics class in a primary school as an example. - Take the example of answering a question related to "subtraction within ten".

[0232] t - The moment model retrieves the students' knowledge memory of the knowledge point "subtraction within ten" and its similar or successively learned knowledge points, such as "addition within ten", from the knowledge memory matrix and aggregates them through the knowledge retrieval knowledge point relationship graph neural network to predict the answers to the target questions.

[0233] t - The moment model models the improvement or decline of students' mastery of the knowledge point "subtraction within ten" according to the answer result of the question, correct or incorrect, and transfers the improvement or decline of knowledge mastery to similar or successively learned knowledge points, such as "addition within ten" and other knowledge points according to the knowledge memory positive reinforcement knowledge point relationship graph neural network or the knowledge memory reverse reinforcement knowledge point relationship graph neural network, so as to obtain the knowledge memory matrix at moment t.

[0234] At time t, the model models whether the student continues to learn the knowledge point "subtraction within ten" or previews the knowledge point to be examined in the next time step, assuming it is "mixed operations of addition and subtraction within ten". The model models the knowledge growth of the knowledge point that the student decides to learn, and then propagates it to similar or previously learned knowledge points, such as "addition within ten", through the knowledge learning knowledge point relationship graph neural network, and uses the kernel function to simulate the learning curve. For the knowledge points that other students have learned but have not decided to learn at the current moment, the kernel function is used to simulate the forgetting curve. In summary, we can get (t+1) - Moment knowledge memory matrix.

[0235] The parameters in the above embodiments of the present invention are determined based on experimental results, that is, different parameter combinations are tested, a set of parameters with better evaluation indicators on the validation set is selected, and the results are evaluated on the test set. In actual testing, the above parameters can be appropriately adjusted according to needs to achieve the purpose of the present invention.

[0236] The protection content of the present invention is not limited to the above embodiments. Without departing from the spirit and scope of the present invention, changes and advantages that can be thought of by those skilled in the art are included in the present invention and are protected by the attached claims.

Claims

1. A three-stage labeling method for students' cognitive states based on graph neural networks, characterized in that: The steps include: Step 1: Select a data set covering the answer records of students in different grades, and divide the data into training set, validation set and test set after data preprocessing; Step 2: Calculate the correlation between different knowledge points through the preprocessed data set to form a relationship diagram of knowledge points that have been revised first and last and a relationship diagram of similar knowledge points; Step 3: Initialize the student knowledge memory matrix and use the matrix elements to represent the memory code of each knowledge point; Step 4: Represent the students’ history answers and knowledge points with learnable embedding vectors; Step 5: Divide each student's answer to a question into three stages: modeling knowledge retrieval stage, modeling memory reinforcement stage, and modeling knowledge learning or forgetting stage for modeling; Step 6: Train the system method model parameters, use the training samples to optimize the parameters of the network model according to the objective function, and use the validation set to tune the model parameters; Step 7: Map students’ cognitive status through the student knowledge memory matrix.

2. The marking method according to claim 1, characterized in that: In step 1, the data sets include ASSISTments2009-2010, ASSISTments2012-2013, and Junyi public data sets; and / or, The preprocessing refers to dividing the student answer sequence with more than 100 answers into one or more answer sequences with a length of 100, filtering out the remaining sequences with less than 10 answers, and filling the sequences with a length greater than or equal to 10 but less than 100 with placeholder identifiers to 100; the division ratio of the training set, validation set and test set is 8:1:

1.

3. The marking method according to claim 1, characterized in that: In step 2, first calculate the similarity score of any two knowledge points Prerequisites By setting a threshold, it is decided whether to establish a relationship edge between the knowledge point pairs; The similarity score and prerequisite scores The calculation of is as follows: Among them, U is the student set, u is a student in the set, A collection of history answers for students u, is the answer of student u at time t, is the answer of student u at time t', They are The correctness of the answer, 1 means correct answer, 0 means wrong answer; Respectively represent Related knowledge points: I(·) is an indicator function. If the value of the expression input to the function is true, the output is 1, otherwise it outputs 0; The thresholds of the similarity score and the prerequisite score are set to η = 0.7 to distinguish between similar / non-similar knowledge point pairs and / or prerequisite / non-prerequisite knowledge point pairs; if Then mark c i c j Similar knowledge points of Otherwise, no mark; if Then mark c i c j Prerequisites for j c i The subsequent knowledge point of Otherwise, no mark is given.

4. The marking method according to claim 1, characterized in that: In step three, the student knowledge memory matrix H includes one or more knowledge points and knowledge memory encoding dimensions. The number of rows of the knowledge memory matrix H represents the number of knowledge points, and the number of columns represents the knowledge memory encoding dimensions. The knowledge memory matrix H is initialized to a learnable parameter matrix H0.

5. The marking method according to claim 1, characterized in that: In step 4, the embedding vector representation of the student's history answer questions and knowledge points is in the form of dimension d e A learnable real number vector, wherein the specific value of each dimension in the learnable real number vector is learned during the model training process as an implicit representation of the student's historical answer questions and knowledge points; and / or, In step 5, the modeling knowledge retrieval stage retrieves and aggregates the relevant knowledge memory encoding of the question from the student knowledge memory matrix to model the student's answer result, and obtains the correct answer prediction score; the modeling memory reinforcement stage models the positive reinforcement of the student's knowledge memory and the reverse reinforcement of the knowledge memory according to the actual correct or incorrect answer situation, and updates the student's knowledge memory matrix; the modeling knowledge learning or forgetting stage models the learning and forgetting behavior of the knowledge point according to the student's active learning behavior after answering the question, and updates the student's knowledge memory matrix; In the three stages of the modeling knowledge retrieval stage, the modeling memory reinforcement stage, and the modeling knowledge learning or forgetting stage, the knowledge point relationship graph neural network is integrated to capture the changes in the cognitive state of the knowledge points; The knowledge point relationship graph neural network in the three stages of step 5 is extended from the prototype knowledge point relationship graph neural network c i For a certain knowledge point, For knowledge point c i The input node prototype features, For knowledge point c i Output node prototype features; d0,d L are the dimensions of the input and output features, d1,…,d L-1 is the feature dimension of the middle layer; the structure of the knowledge point relationship graph neural network is expressed as follows: a1. For each sub-layer of the prototype knowledge point relationship graph neural network, each of the sub-layers aggregates the knowledge point node features of the previous layer on each knowledge point relationship graph, and performs feature transformation through a nonlinear feedforward neural network to obtain aggregated node features; the knowledge point relationship graph The aggregated node features of the previous layer of knowledge points and the aggregated node features after feature transformation are shown in the following formulas: Aggregated previous layer knowledge point node features Node features after feature transformation and aggregation in, Indicates c i In the figure The set of neighbor nodes on , For the knowledge point relationship diagram The learnable matrix parameters of the sub-layer are used to model the message passing between the nodes of the graph neural network; ReLU(x)=max(0,x) is the activation function; For knowledge point c i ,c j In the knowledge point relationship diagram The correlation coefficient on ; a2. A sub-layer fuses the aggregated node features of the three knowledge point relationship graphs to obtain the overall features of the knowledge point nodes of the sub-layer: Among them, when the feature dimension of the previous sub-layer network is the same as the feature dimension of the current layer network, the overall features of the knowledge point nodes output by the previous layer are added after adding the node features of the three knowledge point relationship graphs; otherwise, only the node features of the three knowledge point relationship graphs are added; a3. After passing through L sublayers of the prototype knowledge point graph neural network, the output node prototype feature of each knowledge point of the prototype knowledge point graph neural network is obtained The prototype features of individual knowledge nodes The feature matrix is ​​composed of multiple knowledge point node prototype features. The number of rows is the number of knowledge points, and the number of columns is the node prototype feature dimension. Stacking The neural network representation of the prototype knowledge point relationship graph is is the knowledge point node feature matrix.

6. The marking method according to claim 5, characterized in that: The correlation coefficient of knowledge points on the knowledge point relationship diagram indicates the strength of the relationship between knowledge points. The calculation method is: in, To calculate the knowledge point relationship diagram The learnable matrix parameters of the correlation coefficient between the two knowledge points above; is the activation function; For knowledge point c i ,c j The embedded vector representation of .

7. The marking method according to claim 1, characterized in that: In the modeling knowledge retrieval stage, students answer questions at time t. The specific steps are as follows: Step 5.1.

1. Use the knowledge retrieval extended from the prototype knowledge point relationship graph neural network to aggregate students' knowledge memory of different knowledge points: in, Initialized to the student's knowledge point c for the current answer question t Knowledge memory {d k } L+1 Represents the input feature dimension and the feature dimension of L sub-layers. When the knowledge retrieval knowledge point relationship graph neural network aggregates the previous layer of knowledge point memory, it increases the demand coefficient of the current question for the knowledge point and modifies the aggregation process as follows: in, For the topic q t For knowledge point c i The demand coefficient, that is, whether the question requires the use of this knowledge point to answer; is the aggregation matrix parameter that is restricted to be positive; q t ,c t The knowledge points tested by the question that the student is currently answering; Step 5.1.

2. Use the aggregated knowledge point c of the current answer question t The knowledge memory map is the mastery level or cognitive state of the knowledge point, and the difficulty of the current question By comparing, we can get the probability of predicting whether the student can answer the question correctly: Among them, w h A mapping vector for mapping the knowledge memory of a knowledge point to the degree of mastery of the knowledge point restricted to positive numbers; Question q i For knowledge point c j The demand coefficient is calculated as: Among them, W req is the learnable parameter matrix, For the topic q i The embedding vector representation of For knowledge point c j Embedded vector representation of ; Difficulty level of the question By multi-layer perceptron MLP diff (·) is calculated, and the specific process is: in, are learnable parameter matrices and vectors, For the topic q t The knowledge points it examines t The joint embedding vector representation of , ⊕ is the vector concatenation operation, and the multi-layer perceptron process is symbolized as Among them, d e is the embedding vector dimension, d h It is the middle-level dimension; and w h Use the softmax function to restrict to positive numbers: in, Representation Matrix The i-th row and j-th column of w h | i Represents vector w h The i-th dimension of is the learnable parameter matrix and vector before restriction, d k Master the vector w for mapping knowledge points h Dimensions; and / or, In the modeling and memory reinforcement stage, students finish the questions at time t. The specific steps are as follows: Step 5.2.

1. Use another multilayer perceptron to calculate the correct answer q t The knowledge points tested by the following questions c t The initial knowledge memory positive reinforcement vector of: Among them, ⊕ is the vector concatenation operation, For the topic q t The knowledge points it examines t The joint embedding vector representation of Indicates that the student is answering knowledge point c t The knowledge memory vector of the previous knowledge point, d e is the embedding vector dimension, d h is the middle layer dimension, d k is the knowledge memory vector dimension; Step 5.2.

2. Set the initial knowledge memory positive reinforcement vectors of the remaining knowledge points to vector 0, and use them as the input vectors of the knowledge memory positive reinforcement knowledge point relationship graph neural network: in, For knowledge point c i The initial knowledge memory positive reinforcement vector; The initial knowledge memory positive reinforcement vectors of all knowledge points are stacked into an initial knowledge memory positive reinforcement matrix with the number of rows equal to the number of knowledge points and the number of columns equal to the vector dimension. Finally, the knowledge memory positive reinforcement matrix of the knowledge points after network propagation is obtained through the knowledge memory positive reinforcement knowledge point relationship graph neural network Among them, {d k } L+1 The feature dimensions of the L layer of the neural network representing the relationship between the input layer and the knowledge memory positive reinforcement knowledge point are both d k , which is consistent with the dimension of the knowledge memory vector; Step 5.2.

3. Get the knowledge point c under investigation t The initial knowledge memory positive reinforcement vector Then initialize the initial knowledge memory reverse reinforcement vector of each knowledge point Initial knowledge memory reverse reinforcement matrix stacked as knowledge points And reversely strengthen the knowledge point relationship graph neural network GNN through knowledge memory loss Get the knowledge memory reverse reinforcement matrix of knowledge points after network dissemination Step 5.2.

4. Determine whether the knowledge memory of the knowledge point is positively reinforced or negatively reinforced based on whether the answer to the current question is correct or not, and update the knowledge memory matrix: Among them, a t It is the correctness of the student’s answer at time t, 1 indicates a correct answer and 0 indicates an incorrect answer; and / or, In the stage of modeling knowledge learning or forgetting, students are ready to do the t+1 time problem. The specific steps are as follows: Step 5.3.

1. For the questions you just answered t and the questions to be answered t+1 , according to students' relevant knowledge points c t ,c t+1 Knowledge memory And the joint embedding vector representation of the topic Calculate whether students have learned knowledge point c t ,c t+1 Decision distribution: for learning knowledge point c i ∈{c t ,c t+1 }, its decision distribution is: Among them, ⊕ is the vector concatenation operation, d k is the dimension of knowledge memory, d e is the dimension of the embedding vector, MLP dcs The multilayer perceptron used in this generation process, d h is the dimension of the middle layer of the multi-layer perceptron; The first dimension represents the knowledge points c that students learn i The probability of not learning knowledge point c i If the value of the first dimension is greater than the value of the second dimension, it means that the student decides to learn knowledge point c i ; Step 5.3.

2. For the knowledge points to be learned, students i The knowledge memory and the joint embedding vector representation of the question are expressed through a multi-layer perceptron MLP. prg Computational learning knowledge points c i The initial knowledge memory learning vector of: Among them, ⊕ is the vector concatenation operation, d k is the dimension of knowledge memory, d e is the dimension of the embedding vector, d h is the dimension of the middle layer of the multi-layer perceptron; Step 5.3.

3. Set the initial knowledge memory learning vector of the knowledge point to be The initial knowledge memory learning vectors of the remaining knowledge points are set to vector 0 and used as the input vector of the knowledge learning knowledge point relationship graph neural network: The input vectors of all knowledge points are stacked as the initial knowledge memory learning matrix of the knowledge points Learning knowledge point relationship graph neural network GNN through knowledge prg Get the knowledge memory learning matrix of knowledge points after network dissemination Among them, the activation function ReLU is used to limit the output knowledge memory learning to be a positive number; Step 5.3.

4. Based on the time interval ΔT until the next answer t+1 =T t+1 -T t , update the knowledge memory of each knowledge point after knowledge learning: Among them, T t ,T t+1 For the two answering points, To answer the current question and the next time to answer the question, students should i knowledge memory, through c i Knowledge memory learning vector of knowledge points Make updates; in, To learn knowledge points c i The time-aware kernel function is used to model the process of knowledge point memory changing over time, and its growth rate decreases as the time interval increases: Among them, ⊙ is the Hadamard product, The student has learned the knowledge point c before answering the question at time t i The number of times, To learn knowledge points c i The parameters of the time-aware kernel function, Learning kernel function parameters through knowledge point relationship graph neural network GNN lrn The calculation process is as follows: For knowledge point c i , learn knowledge point c i The calculation process of the time-aware kernel function parameters is: Among them, the input vector of the knowledge point relationship graph neural network of the knowledge learning kernel function parameter is initialized to the embedded vector representation of the knowledge point ci: softplus(x)=log(1+e x ) is the activation function; Step 5.3.

5. Based on the time interval ΔT until the next answer t+1 , update the knowledge memory of each knowledge point after knowledge forgetting: in, To answer the current question and the next time to answer the question, the student i Knowledge memory of knowledge points, is the acquired knowledge point c i knowledge memory, To target forgotten knowledge points c i The time-aware kernel function of in, To target forgotten knowledge points c i The parameters of the time-aware kernel function, Learning kernel function parameters through knowledge point relationship graph neural network GNN frt The calculation process is as follows: For knowledge point c i , forget knowledge point c i The calculation process of the time-aware kernel function parameters is: The input vector of the knowledge point relationship graph neural network of the knowledge learning kernel function parameter is initialized as the knowledge point c i The embedding vector representation of is:

8. The marking method according to claim 1, characterized in that: In step 6, the student cognitive state or knowledge mastery model uses a binary cross entropy loss function to train parameters. For the t-th answer of student u, the loss function is: in, is the correctness of student u’s answer at time t, 1 indicates a correct answer and 0 indicates an incorrect answer; The knowledge retrieval stage model predicts that student u will correctly answer question q at time t t The probability of The overall loss function is the average of the losses over all answers of all students in the training set: in, Gathering for students, is the answer set of student u; According to the loss training parameters, hyperparameters are adjusted on the validation set. When the model loss on the validation set does not decrease in 10 consecutive training cycles, training is stopped and the hyperparameter combination that can achieve the lowest validation set loss is selected. and / or, In step 7, the knowledge point cognitive state or knowledge mastery is acquired using the mapping vector w h Mapping students' understanding of knowledge point c at time t i Knowledge memory To the knowledge point c i cognitive state get:

9. A labeling system for implementing the student cognitive status labeling method according to any one of claims 1 to 8, characterized in that: The system includes: a student answer input encoding module, a three-stage student cognitive state modeling module, and a knowledge point relationship graph neural network module; The student answer input encoding module represents the student's historical answer questions and knowledge points with a learnable embedding vector for use by subsequent modules; The three-stage student cognitive state modeling module uses a continuously updated knowledge memory matrix to model the knowledge retrieval of students when answering questions based on the questions being examined, model the memory reinforcement of students when answering questions based on the correct or incorrect answers to the questions being examined, and model the knowledge learning or forgetting of students during the intervals between answers based on the questions being examined and the questions to be examined; The knowledge point relationship graph neural network module models the process of knowledge retrieval, memory reinforcement, and knowledge learning or forgetting transfer between knowledge points during the student's answering process based on the similar knowledge point relationship graph and the sequential knowledge point relationship graph based on data statistics.

10. Application of the annotation method as described in any one of claims 1 to 8, or the annotation system as described in claim 9, in an online education platform or actual teaching scenario to annotate students' cognitive status and knowledge points.

Images