Neural network learning diagnosis method for step-by-step modeling of answering process based on cognitive load enhancement

Abstract

The invention belongs to the field of education data mining, and provides a neural network learning diagnosis method for step-by-step modeling of an answering process based on cognitive load enhancement, which comprises a cognitive attribute extraction stage, a cognitive attribute interaction stage, a cognitive load enhancement stage and a diagnosis analysis stage. According to the method, the cognitive load is introduced into the model as a new attribute influencing the answering result of the student for the first time, the LSTM is adopted to perform step-by-step modeling on the answering process of the student, and the attention mechanism is utilized to fuse the student attribute and the practice attribute, so that the transmission of time sequence information is realized. According to the method, step-by-step modeling is carried out on the answering steps of the students through the long-short-term memory network, the reaction process data in the answering records of the students are fully utilized to represent the cognitive load of the students, the cognitive load is fused into the cognitive diagnosis model, and the accuracy and effectiveness of model diagnosis are improved.

Description

Technical Field

[0001] The present invention belongs to the field of educational data mining, and provides a neural network learning diagnosis method based on step-by-step modeling of the answering process with enhanced cognitive load for the intelligent diagnosis task of the learner's knowledge and skill mastery level. Technical Background

[0002] In the context of the continuous development of intelligent education, cognitive diagnosis has gradually become an important means of personalized education. By modeling the cognitive processing process of learners and mining the potential abilities and skill states of learners, educators can break through the limitations of traditional assessment methods, more conveniently and accurately diagnose and analyze the knowledge mastery status of students in subject knowledge, and provide timely feedback on the weak knowledge and skills of learners. However, the answering responses of students are composed of a series of complex psychological activities, and most of the existing research directly uses the answering result data of students to mine the knowledge state, ignoring the data generated during the response process.

[0003] Cognitive diagnosis models are important means to achieve cognitive diagnosis. At present, more and more researchers are committed to the development of cognitive diagnosis models. Traditional cognitive diagnosis models perform probabilistic modeling on the answering process of students through different learning assumptions, and then diagnose the knowledge mastery status of learners. On the one hand, based on the knowledge state of learners, it can generally be divided into two situations: potential characteristic abilities and specific knowledge and skills. The cognitive diagnosis model based on the potential characteristic abilities of learners is represented by the item response theory. Its characteristic is to assume that the answering results of learners are affected by the learners' potential abilities and the difficulty of the test questions, and model the learners' potential cognitive abilities as continuous parameters. On the other hand, the cognitive diagnosis model based on the specific knowledge and skill state is represented by the connected deterministic input noise "AND" gate model. This model models the cognitive state of students as a binary discrete vector, and each dimension of the vector represents the mastery level of students' certain specific knowledge ability. It is assumed that learners can only answer the test questions correctly if they master all the knowledge points examined by the test questions. The emergence of neurocognitive diagnosis has greatly improved the computational efficiency and the ability to process multi-dimensional features. Therefore, some other existing technologies advocate using neural networks to learn the interaction function between the students' knowledge level and the knowledge characteristics of the test questions from the data, and constructing a neurocognitive diagnosis model with good generalization ability.

[0004] In recent years, research has delved deeply into the process data generated during students' answering processes. Among them, response speed, as one of the important characteristics in the learning process of students, has received relatively extensive attention. Modeling methods for response time or response speed include modeling based on the independent distribution of response time itself, response models that include response time variables, response time models that include response variables, and hierarchical framework models that incorporate both response time and response variables. However, most of these models are constructed based on the assumption that students' response speed is constant.

[0005] The response process of students is very complex and directly affects their answering results. The data generated during the response process can also reflect the knowledge state of students to a certain extent. Therefore, it is very necessary and reasonable to study the response process of students. However, it is not enough to model the response process of students only from the perspective of response speed, and the data generated during the answering process of students has not been fully explored. Summary of the Invention

[0006] The object of the present invention is to address the problems of single modeling of students' response process data and insufficient data utilization in traditional cognitive diagnosis models. A neural network learning diagnosis method for step-by-step modeling of the answering process based on enhanced cognitive load is proposed. This method performs step-by-step modeling of students' answering steps through a long short-term memory network, and fully utilizes the response process data in students' answer records to represent students' cognitive load and incorporate it into the cognitive diagnosis model, exploring the changes in the knowledge state during the students' response process and the impact of cognitive load on the students' knowledge state during this process, so as to improve the accuracy and effectiveness of model diagnosis.

[0007] To achieve the object of the invention, the following technical solutions are adopted.

[0008] A neural network learning diagnosis method for step-by-step modeling of the answering process based on enhanced cognitive load includes the following steps:

[0009] (1) Cognitive attribute extraction: Collect student numbers, question numbers, knowledge points corresponding to the questions, response time, and number of prompts from the student's answering sequence, and extract student attributes and question attributes from these data. Student attributes include knowledge attributes, ability attributes, prompt attributes, and speed attributes of students. Question attributes include difficulty attributes, knowledge attributes, and workload attributes.

[0010] (2) Cognitive attribute interaction: After obtaining the 7 cognitive attributes of students and questions, based on relevant cognitive diagnosis theories, use the attention mechanism and LSTM to achieve the interaction between these 7 cognitive attributes, thereby capturing the complex relationships between them and obtaining cognitive factors incorporating the interaction information of these attributes: knowledge factors, ability factors, and updated speed attributes.

[0011] (3)Enhanced cognitive load. Use LSTM to perform temporal modeling on the hint attributes obtained in step (1) and the updated speed attributes obtained in step (2) respectively, so as to achieve step-by-step modeling of the student's answering process. Further combine the reaction speed processed by LSTM with the reaction time and the workload of the questions to obtain the final representation of the speed attribute. Finally, use a fully connected layer to fuse the hint attributes processed by LSTM and the final speed attribute to obtain the cognitive load factor. Then, use the attention mechanism to further fuse the cognitive load factor with the knowledge factor and the ability factor obtained in step (2) to obtain the final representations of the knowledge factor and the ability factor.

[0012] (4)Diagnostic analysis. The diagnostic stage includes a main task and an auxiliary task: the main task is to predict the student's answering reaction, and the auxiliary task is to predict the number of hints of the student. Use the hint attributes extracted in step (1) as the input of the auxiliary task diagnostic module, and use the self-supervised learning algorithm to predict the number of student hints. At the same time, introduce the contrastive learning algorithm to divide the positive and negative samples of the hint attributes, which can enable the model to learn the differences in the student's knowledge mastery state under different numbers of hints. Concatenate the knowledge factor, the ability factor, and the cognitive load factor obtained in step (3) as the input of the main task, and output the prediction of the student's answering reaction. Finally, introduce the guessing and mistake parameters, model the guessing and mistake parameters with the number of hints predicted by the self-supervised learning algorithm, and combine the predicted student answering reaction of the main task. Use the DINA diagnostic formula to predict the student's final answering result.

[0013] (5)Collect the data set, set the optimizer, and iteratively train the model until convergence. Finally, predict the student's answering reaction and obtain the student's diagnostic report.

[0014] In the above technical solution, the specific method of step (1) includes:

[0015] (1-1)Cognitive attribute extraction includes extracting student attributes and question attributes. Student attributes include knowledge attributes, ability attributes, hint attributes, and speed attributes. Question attributes include difficulty attributes, workload attributes, and knowledge attributes, that is, the Q matrix (the Q matrix corresponds to the knowledge points examined by the questions, the columns represent knowledge points, the rows represent questions, and the elements only take binary values of 0 or 1. For example, if the first question examines knowledge point 1, then the first row and the first column are marked as 1, and the other columns in the first row are marked as 0). For student attributes, by multiplying the one-hot vector corresponding to the student id with the trainable matrix, the student's knowledge attributes, ability attributes, hint attributes, and speed attributes can be obtained. For question attributes, by multiplying the one-hot vector corresponding to the question id with the Q matrix and the trainable matrix, the knowledge attributes, difficulty attributes, and workload attributes of the questions can be obtained.

[0016] In the above technical solution, the specific method of step (2) includes:

[0017] (2-1) Multiply the knowledge attributes of the student by the knowledge attributes of the test question to obtain the state of the student on the knowledge points examined by the test question, that is, the knowledge factor.

[0018] (2-2) Use the attention mechanism to perform attention matching on the extracted hint attributes with the ability attributes of the student and the difficulty attributes of the test question respectively, so as to model the ability of the student under the current cognitive load level and the test question difficulty for the student's cognitive load level, and subtract the two to obtain the true ability of the student, that is, the ability factor.

[0019] (2-3) Similarly, use the attention mechanism to match the workload attribute of the test question and the speed attribute of the student to obtain an updated representation of the speed attribute.

[0020] In the above technical solution, the specific method of step (3) includes:

[0021] (3-1) Decompose the student's response process into several response steps, use a Long-Short-Term Memory Network (LSTM) to perform temporal modeling on the student's answering process, and use the hint attributes extracted in step (1) as the input of the LSTM. After multi-step transmission, the output of the hidden layer is obtained.

[0022] (3-2) Similarly, use LSTM to perform temporal modeling on the updated speed attribute in step (2) to obtain the output of the hidden layer after multi-step transmission, then multiply it by the reaction time to calculate the workload of the student within the reaction time, and finally subtract the workload attribute of the test question to obtain the final representation of the speed attribute.

[0023] (3-3) Fuse the results of steps (3-1) and (3-2) through a fully connected layer to obtain a representation of the cognitive load factor.

[0024] (3-4) Use the attention mechanism to capture the relationship between the ability factor and knowledge factor of the student and the cognitive load factor. Match the knowledge factor and ability factor obtained in step (2) with the cognitive load factor obtained in step (3-3), so that the knowledge factor and ability factor share the temporal information in the cognitive load factor, and output the final knowledge factor and ability factor.

[0025] In the above technical solution, the specific method of step (4) includes:

[0026] (4-1) Use the hint attribute as the input of the auxiliary task, and adopt a self-supervised learning algorithm to predict the number of student hints. At the same time, introduce a contrastive learning algorithm to divide the positive and negative samples of the hint attribute embedding, so that the model can learn the differences in the student's knowledge mastery state under different numbers of hints.

[0027] (4-2) Model the student's guessing and error probabilities based on the predicted number of hints in step (4-1), and predict the probabilities of guessing and making mistakes.

[0028] (4-3) Concatenate the knowledge factor, ability factor, and cognitive load factor obtained in step (3) as the input of the main task, output the prediction of the student's response to the question, and use the DINA formula to combine the prediction of the student's response obtained from the main task with the guessing and error parameters to obtain the final prediction of the response.

[0029] (4-4) Adopt the method of calculating the error gradient under each weight in real time, which is a classic method of training a neural network by combining optimization methods (such as gradient descent, etc.), perform gradient descent to reduce the loss function, and optimize the parameters.

[0030] In the above technical solution, the specific method in step (5) includes:

[0031] (5-1) Collect three datasets, namely Junyi, PISA2015, and Assistment2017.

[0032] (5-2) In the feedback neural network structure, use the cross-entropy loss function, mean squared error loss function, and contrastive loss as the joint loss function to measure the loss between the predicted value and the true value.

[0033] (5-3) Perform backpropagation, and select the method of calculating the error gradient under each weight in real time to update the parameters.

[0034] (5-4) Select the optimization algorithm optimizer.step() and the backpropagation algorithm backward() to minimize the loss function.

[0035] Compared with the existing technology, the present invention has the following obvious outstanding substantial features and remarkable technical progress:

[0036] 1. The present invention proposes a neural network learning diagnosis method based on step-by-step modeling of the answering process with enhanced cognitive load. Aiming at the problems of single modeling of the student's response process data and insufficient data utilization in the traditional cognitive diagnosis model, based on the cognitive diagnosis theory, the multi-dimensional features of students and test questions are characterized, and further, the method of cognitive diagnosis analysis is used to model and train the multi-dimensional features, diagnose the student's knowledge mastery state, and predict the student's future performance.

[0037] 2. The method of the present invention includes a cognitive attribute extraction stage, a cognitive attribute interaction stage, a cognitive load enhancement stage, and a diagnostic analysis stage. It makes full use of the reaction process data in the student answer record, introduces cognitive load into the model as a new attribute that affects the student's answer result for the first time, uses LSTM to model the student's answer process step by step, and uses the attention mechanism to merge the student attribute with the exercise attribute, thus realizing the transmission of temporal information. Different from previous studies, this method makes more full use of the data in the student's reaction process to achieve more accurate diagnosis.

[0038] 3. The method of the present invention uses the cognitive diagnosis method to build a neural network learning diagnosis model based on step-by-step modeling of the answering process with cognitive load enhancement, so that it has the ability to perform cognitive diagnosis and prediction on students' answer records. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a scene graph for implementing the method of the present invention.

[0040] Figure 2 It is a neural network learning diagnosis framework model diagram based on step-by-step modeling of the answering process with cognitive load enhancement according to the present invention. DETAILED DESCRIPTION

[0041] The method of the present invention studies the existing public data sets and believes that in addition to the reaction speed, the number of prompts has an important influence on the students' reaction process. Constructivist theory believes that knowledge is acquired by learners in a certain context, with the help of others (including teachers and learning partners), using necessary learning materials, and through the construction of meaning. Therefore, the number of prompts, as an important feature of the data of students' interaction with the external environment, has a certain influence on the students' reaction process and knowledge state. In addition, cognitive load theory believes that the brain's working memory capacity is limited, and excessive load will affect the students' response process. Prompts, as an external intervention, can help students reduce cognitive load by reducing the complexity of learning tasks. It can be seen that the degree of cognitive load will affect the students' answer results, and the number of prompts is an important indicator for measuring the degree of students' cognitive load. The more prompts students need, the higher their current cognitive load level is, and vice versa. Similarly, reaction speed and reaction time will also reflect the students' cognitive load level to a certain extent. The longer the student's reaction time and the slower the reaction speed, the greater the cognitive load the student faces. Therefore, it is extremely important to integrate cognitive load into cognitive diagnosis to evaluate the students' knowledge status.

[0042] The present invention discloses a neural network learning diagnosis framework for step - by - step modeling of the answering process based on enhanced cognitive load. It mainly performs step - by - step modeling of the student's answering process through a long - short - term memory network. At the same time, based on the cognitive load theory, it fully utilizes the process data in the student's answering record data, extracts the number of prompts, response time, response speed, and the workload of the test questions, jointly represents the student's cognitive load factors and introduces them into the model, thereby improving the effectiveness and accuracy of the model. Specifically, the cognitive factors affecting the student's answering results in the present invention are divided into knowledge factors, ability factors, and cognitive load factors. The knowledge factors reflect whether the student's knowledge mastery meets the test requirements of the questions; the ability factors are also important factors that affect the student's answering response. For example, in IRT, it is assumed that the student's answering result is affected by their latent ability; the cognitive load factors cannot be directly observed, and their external manifestations are in the reaction process data that is easy to collect during the student's reaction process. In order to represent the cognitive load factors, this method first extracts from the student's answering record. By analyzing the student's answering record, it extracts the process data therein, including the number of prompts, response speed, response time, and the workload of the test questions, and jointly calculates the cognitive load factors. Specifically, the method of the present invention first matches the number of prompts with attributes such as the student's response speed, ability, and question difficulty through an attention mechanism to capture the complex relationships between these attributes. In addition, LSTM is used to perform time - series modeling on the number of prompts and response speed of the student respectively, so as to achieve step - by - step modeling of the student's answering process, and fuse the output of the hidden layer obtained by processing through a fully - connected layer to obtain a representation of the cognitive load factors. Finally, the present invention uses a multi - layer neural network to fully mine information from the above three cognitive factors and fit the complex non - linear relationship between the student and the test questions, trains the model parameters through the task of predicting the student's answering results, and introduces self - supervised learning and contrast learning in the diagnosis module to achieve multi - task prediction, and optimizes the model through a joint loss function. Through experimental comparative analysis, the neural network learning diagnosis method based on step - by - step modeling of the answering process with enhanced cognitive load proposed by the present invention improves the prediction accuracy of the answering results.

[0043] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0044] An embodiment of the present invention provides a neural network learning diagnosis method for step - by - step modeling of the answering process based on enhanced cognitive load, including the following steps:

[0045] (1) Cognitive attribute extraction

[0046] According to the different sources of attributes, the attribute extraction module can divide the extracted attributes into student attributes and question attributes.

[0047] (1-1) For student attributes, this method uses the knowledge attribute k i , the ability attribute a i , the hint attribute d i and the speed attribute v i to represent a student. Specifically, by multiplying the one-hot vector corresponding to the student id with the trainable matrices K, A, D, and V, the knowledge attribute k i , the ability attribute a i , the hint attribute d i , and the speed attribute v i of student s i can be obtained:

[0048] k i = h i s × K a i = h i s × A d i = h i s × D v i = h i s × V (Equation 1.1)

[0049] where h i s ∈ {0, 1} N represents the one-hot vector corresponding to the student id, and K, A ∈ R N×L represent the trainable knowledge matrix and ability matrix of N students respectively. To achieve step-by-step modeling of the answering process, this method assumes that the reaction process of a student consists of Z different reaction steps, and the number of hint times at each step is independent of each other. Therefore, the trainable hint matrix is defined as a three-dimensional tensor, i.e., D ∈ R N×Z×L , d i ∈ R Z×L , d i = {d i1 , d i2 ,..., d iz} indicates that the number of hint times of student s i at step z is d iz ∈ R L , and the number of hint times in the student's answering record is used as the initial embedding of the trainable hint matrix. Similarly, the reaction speed at each step is also independent of each other. Therefore, the trainable speed matrix of a student is defined in the same way, i.e., V ∈ R N×Z×L , v i ∈ R Z×L . Since there is no relevant research indicating how many reaction steps the student's reaction process specifically contains, this method takes the number of reaction steps Z as a hyperparameter.

[0050] For the test item attributes, this method uses the knowledge attribute q j , the difficulty attribute b j , and the workload attribute g j to represent the test items. Specifically, by multiplying the one-hot vector corresponding to the test item id with the Q matrix and the trainable matrices D, G, the knowledge attribute q j of the test item e j , the difficulty attribute b j and the workload attribute g j can be obtained:

[0051] q j = h j e ×Q b j = h j e ×B g j = h j e ×G (Formula 1.2)

[0052] (2) Interaction of cognitive attributes

[0053] After obtaining the 7 cognitive attributes of students and test items, the attention mechanism and LSTM are used to realize the interaction between these 7 cognitive attributes, so as to capture the complex relationships between them and obtain the cognitive factors integrated with the interaction information of these attributes: knowledge factors, ability factors, and the updated speed attribute.

[0054] (2-1) The knowledge factor indicates whether the student has sufficient mastery of the knowledge points involved in the test item. The knowledge attribute k i of the student represents the mastery degree of the student on each knowledge point, and the knowledge attribute q j of the test item comes from the Q matrix marked by experts and represents the examination status of the test item on each knowledge point. The DINA model believes that the factors affecting the student's answering result are only related to whether the student masters the knowledge points examined by the test item and have nothing to do with the knowledge points not examined by the test item. In view of this, multiplying the knowledge attribute k i of the student s i and the knowledge attribute q j of the test item e j can retain the status of the student on the knowledge points examined by the test item, that is, the knowledge factor α ij :

[0055] α ij = k i ⊙q j (Formula 2.1)

[0056] (2-2) Ability factors affect students' answering results at the non-knowledge level. This means that in addition to mastering the specific knowledge points required by the test questions, students also need to have good reading comprehension skills to understand the true intention of the questions and then give effective answers. Item response theory (IRT) has an accurate description of ability, pointing out that students' potential ability is related to the difficulty level of the test questions: facing more difficult questions, students need to show stronger ability to have a greater chance of getting the correct answer. In addition, this method believes that students' ability is closely related to the number of prompts. According to constructivist theory, the student's answer response process is not just an interaction between students and exercises. The prompts obtained during the interaction between students and the external environment will also affect the student's answer results. Cognitive load theory believes that prompts, as an external cognitive resource, can help students reduce cognitive load by reducing the complexity of learning tasks. If a student needs more prompts, the cognitive load he faces in the answering process will be greater. Therefore, in real educational scenarios, students' actual ability should be based on their ability after receiving prompts. This method first uses the attention mechanism to match the students' prompt attributes and ability attributes. Specifically, by calculating the cosine similarity w between the two vectors of prompt attributes and ability attributes a To indicate the degree of matching, w a According to the formula w a =Softmax(cs(a i ,d i )) is calculated, cs represents the similarity function. Then the matching degree and the ability attribute are multiplied to obtain the intermediate representation of the ability factor Right now The number of hints a student receives is also closely related to the difficulty of the test questions. The more difficult the test questions are, the more hints a student will request. Therefore, the attention mechanism is also used to match the hint attributes of the student with the difficulty attributes of the test questions. The calculation process is the same as before, and the intermediate representation of the difficulty attributes of the test questions is obtained. Ability Factorθ ij Can be used by students i Intermediate representation of ability factors and test questions j The difficulty attribute is in the middle It means subtraction, that is:

[0057]

[0058] (2-3) The workload attribute of the test question corresponds to the student speed attribute. The workload attribute refers to the amount of work required to complete a test question. Each reaction step of the student can be understood as completing a part of the workload of the test question. Therefore, it is necessary to match the student's reaction speed with the workload of the test question. Specifically, first calculate the workload attribute g of the test question jand the speed v at each reaction step iz The matching degree between them. This method calculates the cosine similarity w between two vectors v to represent the matching degree, that is, w v = Softmax(cs(g i , v i ))). Then, multiply the matching degree by the corresponding reaction speed to obtain the updated speed attribute That is

[0059] (3) Enhanced cognitive load

[0060] (3-1) The reaction process of students can be decomposed into several reaction steps, and each reaction step has its own independent number of prompts. Since the number of prompts obtained by students is interspersed in the reaction steps of students, each reaction step has its corresponding prompt information. The degree of cognitive load faced by students at the current step can be reflected by the number of prompts in each step. Since the reaction steps of students are carried out in sequence, the previous steps will affect the subsequent steps. The prompt of the previous step will play a certain auxiliary role in the students' answers, which will reduce the current degree of cognitive load and also affect the cognitive load of the subsequent steps. Therefore, the Long-Short Term Memory Network (LSTM) is used to simulate this sequence feature. LSTM can not only process features with sequence relationships, but also reasonably model the influence between the front and back elements in the sequence feature through memory and forgetting units. This method uses the prompt attribute d i extracted in step (1) as the input of LSTM, and after Z steps, the output of the last unit of LSTM and the hidden layer output of each intermediate unit The specific calculation process is as follows:

[0061] i z = σ(W i × [h z-1 , d iz + b i ) (Formula 3.1)

[0062] f z = σ(W f × [h z-1 , d iz + b f ) (Formula 3.2)

[0063] o z = σ(W o × [h z-1 , d iz + bo ) (Formula 3.3)

[0064]

[0065] h z = o z ⊙tanh(c z )(Formula 3.6)

[0066] h z is the result of the number of hint times in each step being affected by the previous step, and the output h of the last unit Z includes all the information from the first step to the Z-th step, which is denoted as In the formula, σ represents the sigmoid activation function, and tanh represents the tanh activation function.

[0067] (3-2) Since each reaction step has its own independent reaction speed, which is the same as the student's cognitive load level, and the reaction speed also has temporal characteristics, an LSTM is also used to model it. Take as the input of the LSTM, and after Z steps, the output of the last unit of the LSTM is obtained and the hidden layer output of each intermediate unit The calculation process is the same as that of Formulas 2.3 to 2.8, and the reaction time t ij is multiplied by to calculate the workload that the student can complete within the reaction time, and then subtracted from the workload attribute g of the test question j to determine whether the workload requirement of the test question can be met, so as to obtain the final speed attribute representation That is

[0068] (3-3) The student's reaction speed is also affected by the number of hint times. The more hint times the student requests, the more difficult the question is, and the slower the answering speed will naturally be. On the contrary, the answering speed will increase. To achieve this interaction relationship, the obtained in step (3-1) and the obtained in step (3-2) are passed into the fully connected layer and fused through the non-linear activation function ReLu to obtain the cognitive load factor representation CL ij , that is

[0069] (3-4) In a series of steps of the student's answering reaction, some of the reaction steps are related to the knowledge mastery level, and some of the reaction steps are related to the student's potential ability. Therefore, this method uses an attention network to capture the complex correlation relationship. Specifically, the knowledge factor α ij obtained in step (2) and the ability factor θij are respectively matched with the cognitive load factor CL ij to transfer the rich temporal information in CL ij to the knowledge factor and the ability factor. Therefore, it is necessary to calculate the cosine similarity between CL ij and the knowledge factor α ij and the ability factor θ ij respectively, and then multiply by CL after normalization through the Softmax function to obtain the final representations of the knowledge factor and the ability factor ij and and

[0070] (4) Diagnostic analysis

[0071] (4-1) When a student answers a question, the number of hint requests is affected by both the potential ability and the question difficulty. When facing the same question, different students may request different numbers of hints, and when the same student faces questions of different difficulties, the number of hint requests may also be different. Therefore, this method introduces a self-supervised learning algorithm to predict the number of hint requests of students by inputting the ability factor a i of the student and the difficulty factor b j of the question into a fully connected layer That is:

[0072]

[0073] And the mean squared error (MSE) is selected as the loss function, that is:

[0074]

[0075] where represents the predicted value of the number of hint requests of the i-th student, and d i ' represents the actual number of hint requests of the student.

[0076] In addition, this method also optimizes the embedding of hint attributes by introducing contrastive learning, thereby improving the accuracy of model prediction. The core idea of contrastive learning is to optimize the embedding space so that the similarity between similar samples is greater and the similarity between dissimilar samples is smaller. As a feature, the number of hints can be associated and optimized with other features such as students' abilities and question difficulties through contrastive learning, which will help the model better capture the cognitive load level of students in different situations and improve the model's prediction ability. Specifically, this method uses the K-means clustering algorithm to divide the number of hints in the dataset into different clusters to achieve clustering of the number of hints for students. Here, the clusters are divided into 3 categories according to the level of the number of hints: low number of hints, medium number of hints, and high number of hints. Then, positive and negative sample pairs are generated according to the divided categories. If students belong to the same category, it means that the behaviors or performances of these students are similar, then they are positive sample pairs, otherwise they are negative sample pairs. The positive and negative sample pairs can help the model better distinguish the embedding representations of different types of students. After obtaining the positive and negative sample pairs, the contrastive loss is obtained by calculating the Euclidean distance between the positive and negative sample pairs. The calculation formula is as follows:

[0077]

[0078] where y i ∈{0,1} represents the label of the sample pair, 1 represents the positive sample pair, and 0 represents the negative sample pair. d i represents the Euclidean distance between two samples. The obtained contrastive loss is used in subsequent training.

[0079] (4-2) Inspired by the DINA model, this method also considers the influence of two parameters, namely slip s and guess g, on the student's answering process. That is, even if a student has mastered all the concepts included in the question, they may make a wrong answer due to a slip. In another case, even if a student has not fully mastered all the concepts included in the question, they may make a correct answer by guessing. The number of hints requested by the student has a strong correlation with these two parameters. When the number of hints requested by the student is more, it means that the question is more difficult for the student, and it is more likely to answer the question by guessing. When the number of hints is less, it means that the student is more likely to answer correctly, and then the probability of answering the question wrong due to a slip will be greater. Therefore, this method models the student's guessing and slipping probabilities according to the predicted number of hints in step (4-1). The process is as follows:

[0080]

[0081] where W g ,W s ∈R 1×L represent the trainable matrices for guessing and slipping respectively, and b g ,bs is the bias term. represents the probability that a student answers correctly due to guessing, represents the probability that a student answers incorrectly due to a mistake.

[0082] (4 - 3) To explore the complex relationship between students and test questions and predict students' answering results, this method designs a prediction layer composed of a three - layer fully - connected network. After concatenating the three final cognitive factors obtained, they are used as the input of the prediction layer, that is:

[0083]

[0084] y * = σ(f FC (σ(f FC (σ(f FC (x ij )))))) (Formula 4.7)

[0085] By combining and y * , the diagnostic formula in the DINA model is used to predict students' answering results That is:

[0086]

[0087] The binary cross - entropy function is used as the loss function for the part of predicting students' answers. The formula is as follows:

[0088]

[0089] where y is the true value of the students' answering results.

[0090] (4 - 4) The method of calculating the error gradient under each weight in real - time is adopted. It is a classic method for training neural networks by combining optimization methods (such as gradient descent, etc.), and it consists of two parts: excitation propagation and weight update.

[0091] (4 - 4 - 1) In the excitation propagation stage, each iteration is carried out in two steps:

[0092] Step 1: Input the training results into the network to obtain the excitation response;

[0093] Step 2: Take the difference between the excitation response and the corresponding output target to obtain the response errors of the output layer and the hidden layer.

[0094] (4 - 4 - 2) In the weight update stage, two steps are carried out for each weight:

[0095] Step 1: Multiply the input excitation and the response error to obtain the weight gradient;

[0096] Step 2: Multiply this gradient by the learning rate, then take its inverse, and add it to the weights.

[0097] (5) Collect the dataset and train the network structure

[0098] (5-1) Collect three real-world datasets, namely Junyi, PISA2015, and Assistment2017

[0099] Junyi: The Junyi dataset comes from "Junyi Academy". The original Junyi dataset contains more than 20 million response records, distributed in different learning sessions. To conform to the static assumption of cognitive diagnosis, this method analyzes the relationship between students' answer data and the exercises and concepts contained.

[0100] PISA2015 dataset: PISA is a globally authoritative online test with high-quality test items, and it records students' answer results and answering times. The PISA2015 dataset includes computer-based PISA mathematics data. In this study, 17 computer-scored dichotomous items were selected and used. The database for analysis contains dichotomous answer data and continuous answering time data of 6,000 randomly selected students.

[0101] Assistment2017: Assistment2017 originated from the "2017 ASSISTments Data Mining Competition" and provides the answering records of learners from 2014 to 2017 and the relationship between the exercises and concepts contained.

[0102] (5-2) Train using the combined loss function

[0103] This method combines formulas (4.2), (4.3), and (4.9) to design a combined loss function to measure the loss between the predicted value and the true value, and proves the effectiveness of the model by pursuing a lower loss value. The formula can be described as:

[0104] loss = L CDM + λL d +(1 - λ)L (Formula 5.1)

[0105] Among them, λ represents the weight parameter affecting the loss.

[0106] (5-3) Perform backpropagation and select the method of calculating the error gradient under each weight in real time to update the parameters

[0107] Fusing the difficulty, discrimination of the test question features and the mastery degree X of the student features of the test question in the front, after receiving the mixed input X, transfer X to the first fully connected layer (Linear layer). X undergoes a linear mapping in the first fully connected layer to obtain z 1 , and then process it through the sigmoid activation function to obtain X 1 . Then transmit X 1 into the second fully connected layer and repeat the above steps. After repeating the linear-sigmoid processing twice, the mapping product X 2 is obtained. The formula description is as follows:

[0108]

[0109] X i+1 = sigmoid(z i ) (Formula 5.3)

[0110] In this model, backpropagation plays a role in updating the parameters for fitting. ΔW ij is the update formula for the parameters, and the formula description is as follows:

[0111]

[0112] The variable W ij represents the neuron weight between i and j. Define ΔW ij as the weight update, η is the learning rate, represents the partial derivative of the mean squared error function. X i is the output of the current neuron, and δ j is the error generated by the j neuron of the current layer (i.e., the error between the actual value and the predicted value). The input part X i leading to neuron j is obtained by the weighted sum of the output X i of the upper-layer neuron I.

[0113] (5-4) Select the optimization algorithm optimizer.step() and the backpropagation algorithm backward() to minimize the loss function

[0114] W ij = W ij + ΔW ij , so W ij = W ij — ηX i δ j (Formula 5.6)

[0115] It should be noted that the description of the content of the embodiments of the present invention included above is for explaining in detail the technical features of the present invention. Without departing from the present invention, several improvements and modifications made are also protected by the present invention. Therefore, the protection scope of the present invention should be based on the content defined by the claims of this application.

[0116] The content not described in detail in this specification belongs to the prior art well known to those skilled in the art.

Claims

1. A neural network learning diagnosis method based on step-by-step modeling of the answering process with cognitive load enhancement, characterized by The method comprises the following steps: (1) Cognitive attribute extraction: collect the student number, test number, knowledge point corresponding to the test, reaction time and number of prompts from the students' answer sequence, and extract student attributes and test attribute from these data. Student attributes include students' knowledge attributes, ability attributes, prompt attributes and speed attributes. Test attribute includes difficulty attribute, knowledge attribute and workload attribute. (2) Cognitive attribute interaction: After obtaining the seven cognitive attributes of students and test questions, the attention mechanism and LSTM are used to realize the interaction between these seven cognitive attributes, capture the relationship between them, and obtain cognitive factors that incorporate the interactive information of these attributes: knowledge factor, ability factor, and updated speed attribute; (3) Cognitive load enhancement: LSTM is used to perform time series modeling on the prompt attribute obtained in step (1) and the updated speed attribute obtained in step (2), thereby realizing step-by-step modeling of the student's answering process, and the reaction speed processed by LSTM is further combined with the reaction time and the test workload to obtain the final speed attribute representation. Finally, the prompt attribute processed by LSTM and the final speed attribute are fused with the fully connected layer to obtain the cognitive load factor; the attention mechanism is used to further fuse the cognitive load factor with the knowledge factor and ability factor obtained in step (2) to obtain the final representation of the knowledge factor and ability factor; (4) Diagnostic analysis. The diagnostic stage includes a main task and an auxiliary task. The main task is to predict the student's response, and the auxiliary task is to predict the number of prompts given to the student. The prompt attributes extracted in step (1) are used as the input of the auxiliary task diagnosis module. The self-supervised learning algorithm is used to predict the number of student prompts. At the same time, a contrastive learning algorithm is introduced to divide the prompt attributes into positive and negative samples, so that the model can learn the difference in the student's knowledge mastery status under different prompt times. The knowledge factors, ability factors and cognitive load factors obtained in step (3) are spliced ​​as the input of the main task, and the prediction of the student's response is output. Finally, guessing and error parameters are introduced, and the guessing and error parameters are modeled using the number of prompts predicted by the self-supervised learning algorithm. Combined with the student's response predicted by the main task, the DINA diagnostic formula is used to predict the student's final answer result. (5) Collect the data set, set the optimizer, iteratively train the model until convergence, and finally predict the student’s response to obtain the student’s diagnostic report.

2. The neural network learning diagnosis method based on step-by-step modeling of the answering process with cognitive load enhancement according to claim 1 is characterized in that The cognitive attribute extraction in step (1) specifically includes: Cognitive attribute extraction includes extracting student attributes and test question attributes. Student attributes include knowledge attributes, ability attributes, prompt attributes and speed attributes. Test question attributes include difficulty attributes, workload attributes and knowledge attributes, namely the Q matrix. For student attributes, the student’s knowledge attributes, ability attributes, prompt attributes and speed attributes are obtained by multiplying the one-hot vector corresponding to the student id with the trainable matrix. For test question attributes, the knowledge attributes, difficulty attributes and workload attributes of the test question are obtained by multiplying the one-hot vector corresponding to the test question id with the Q matrix and the trainable matrix.

3. The neural network learning diagnosis method based on step-by-step modeling of the answering process with cognitive load enhancement according to claim 1 is characterized in that The cognitive attribute interaction in step (2) specifically includes: (2-1) Multiply the student's knowledge attribute by the test item's knowledge attribute to obtain the student's status on the knowledge point tested by the test item, i.e., the knowledge factor; (2-2) Using the attention mechanism, the extracted prompt attributes are matched with the student's ability attributes and the difficulty attributes of the test questions, so as to model the student's ability at the current cognitive load level and the difficulty of the test questions for the student's cognitive load level. The two are subtracted to obtain the student's true ability, i.e., the ability factor; (2-3) The attention mechanism is also used to match the workload attribute of the test questions with the speed attribute of the students to obtain an updated speed attribute representation.

4. The neural network learning diagnosis method based on step-by-step modeling of the answering process with cognitive load enhancement according to claim 1 is characterized in that The cognitive load enhancement in step (3) specifically includes: (3-1) Decompose the student's response process into several response steps, use the long short-term memory network to perform temporal modeling of the student's answer process, use the prompt attributes extracted in step (1) as the input of the LSTM, and obtain the output of the hidden layer after multiple steps of transmission; (3-2) LSTM is also used to perform time series modeling on the speed attribute updated in step (2) to obtain the hidden layer output after multi-step transmission, which is then multiplied by the reaction time to calculate the student's workload within the reaction time, and finally the workload attribute of the test question is subtracted to obtain the final speed attribute representation; (3-3) The results of step (3-1) and step (3-2) are fused through a fully connected layer to obtain a cognitive load factor representation; (3-4) The attention mechanism is used to capture the relationship between students' ability factors, knowledge factors and cognitive load factors, and the knowledge factors and ability factors obtained in step (2) are matched with the cognitive load factors obtained in step (3-3) so that the knowledge factors and ability factors share the temporal information in the cognitive load factors, and the final knowledge factors and ability factors are output.

5. The neural network learning diagnosis method based on step-by-step modeling of the answering process with cognitive load enhancement according to claim 1 is characterized in that The diagnostic analysis in step (4) specifically includes: (4-1) The prompt attribute is used as the input of the auxiliary task, and a self-supervised learning algorithm is used to predict the number of prompts given by students. At the same time, a contrastive learning algorithm is introduced to divide the prompt attribute embedding into positive and negative samples, so that the model can learn the differences in students' knowledge mastery status under different prompt times; (4-2) Modeling the probability of students’ guesses and errors based on the number of hints predicted in step (4-1) to predict the probability of guesses and errors; (4-3) The knowledge factors, ability factors and cognitive load factors obtained in step (3) are combined as the input of the main task, and the prediction of the student's answer response is output. The DINA formula is used to combine the student's answer response prediction obtained by the main task with the guess and error parameters to obtain the final answer response prediction; (4-4) The method of calculating the error gradient under each weight in real time is adopted, and gradient descent is performed to reduce the loss function to optimize the parameters.

6. The neural network learning diagnosis method based on step-by-step modeling of the answering process with cognitive load enhancement according to claim 1 is characterized in that In step (5), the data set is collected and the network structure training specifically includes: (5-1) Collect three datasets, namely Junyi, PISA2015, and Assistment2017; (5-2) In the feedback neural network structure, the cross entropy loss function, mean square error loss function and contrast loss are used as a joint loss function to measure the loss between the predicted value and the true value; (5-3) Perform back propagation and select a method to calculate the error gradient under each weight in real time to update the parameters; (5-4) Select the optimization algorithm optimizer.step() and the backpropagation algorithm backward() to minimize the loss function.

Images