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

Abstract

The present application belongs to the field of educational data mining, and provides a neural network learning diagnosis method based on step-by-step modeling of answering process enhanced by cognitive load, which comprises a cognitive attribute extraction stage, a cognitive attribute interaction stage, a cognitive load enhancement stage and a diagnosis analysis stage, fully utilizes the reaction process data in the student answering record, firstly introduces cognitive load as a new attribute affecting the student answering result into the model, adopts LSTM to model the student answering process step by step, and utilizes the attention mechanism to fuse the student attributes and exercise attributes, and realizes the transmission of time sequence information. The method of the present application models the student answering steps step by step through the long short-term memory network, fully utilizes the reaction process data in the student answering record to represent the cognitive load of the student and integrates it into the cognitive diagnosis model, and improves the accuracy and effectiveness of the model diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of educational data mining. For the intelligent diagnostic task of learners' knowledge and skills mastery, it provides a neural network learning diagnostic method based on step-by-step modeling of the answering process with enhanced cognitive load. Technical Background

[0002] In the context of the continuous development of intelligent education, cognitive diagnosis has gradually become an important tool for personalized education. By modeling learners' cognitive processing and exploring their potential abilities and skills, educators can break through the limitations of traditional assessment methods and more conveniently and accurately diagnose and analyze students' knowledge mastery in subject areas, providing timely feedback on learners' weak knowledge and skills. However, students' responses are composed of a series of complex psychological activities, and most existing studies directly use students' response data to explore their knowledge status, ignoring the data generated by students during the response process.

[0003] Cognitive diagnostic models are an important means of achieving cognitive diagnosis, and an increasing number of researchers are dedicated to their development. Traditional cognitive diagnostic models use different learning hypotheses to probabilistically model students' answering processes, thereby diagnosing learners' knowledge mastery. On one hand, based on learners' knowledge states, they can generally be divided into two categories: latent characteristic abilities and specific knowledge skills. Cognitive diagnostic models based on learners' latent characteristic abilities are represented by item response theory, which assumes that learners' answers are influenced by their latent abilities and the difficulty of the test questions, and models students' latent cognitive abilities as continuous parameters. On the other hand, cognitive diagnostic models based on specific knowledge skill states are represented by the connected deterministic input noise AND gate model. This model models students' cognitive states as a binary discrete vector, where each dimension represents the student's mastery of a specific knowledge ability. It assumes that learners can only answer questions correctly if they master all the knowledge points tested in the test. The emergence of neurocognitive diagnosis has greatly improved computational efficiency and the ability to handle multidimensional features. Therefore, some existing technologies advocate using neural networks to learn the interaction function between students' knowledge level and test item knowledge characteristics from data, and to build a neurocognitive diagnostic model with good generalization ability.

[0004] In recent years, research has delved into the procedural data generated during student responses, with reaction speed, as a key characteristic of student learning, receiving considerable attention. Modeling methods for reaction time or reaction speed include models based on the independent distribution of reaction time, response models incorporating reaction time variables, reaction time models incorporating response variables, and hierarchical framework models that combine both reaction time and response variables. However, most of these models are built upon the assumption that student reaction speed is constant.

[0005] Students' response process is highly complex and directly impacts their answers. The data generated during this process can also reflect students' knowledge levels to some extent. Therefore, research on students' response process is both necessary and reasonable. However, modeling the student response process solely from the perspective of reaction speed is insufficient; the data generated during the answering process has not yet been fully explored. Summary of the Invention

[0006] The purpose of this invention is to address the problems of traditional cognitive diagnostic models, such as limited modeling of students' response process data and insufficient data utilization. It proposes a neural network learning diagnostic method based on step-by-step modeling of the response process with enhanced cognitive load. This method uses a long short-term memory network to model students' response steps step by step and fully utilizes the response process data in students' answer records to represent their cognitive load and integrate it into the cognitive diagnostic model. It explores the changes in students' knowledge state during the response process and the impact of cognitive load on their knowledge state, thereby improving the accuracy and effectiveness of the model's diagnostic capabilities.

[0007] To achieve the purpose of the invention, the present invention adopts the following technical solution.

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

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

[0010] (2) Cognitive attribute interaction: After obtaining the seven cognitive attributes of students and test questions, based on relevant cognitive diagnostic theories, attention mechanism and LSTM are used to realize the interaction between these seven cognitive attributes, thereby capturing the complex relationship between them and obtaining cognitive factors that incorporate the interactive information of these attributes: knowledge factors, ability factors and updated speed attributes.

[0011] (3) Enhanced cognitive load: LSTM is used to perform temporal modeling on the cue attributes obtained in step (1) and the updated speed attributes obtained in step (2), thereby realizing step-by-step modeling of the student's answering process. The reaction speed after LSTM processing is further combined with the reaction time and the workload of the test questions to obtain the final speed attribute representation. Finally, a fully connected layer is used to fuse the cue attributes after LSTM processing and the final speed attribute to obtain the cognitive load factor. Then, an 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.

[0012] (4) Diagnostic Analysis: The diagnostic phase 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. The prompt attributes extracted in step (1) are used as input to the auxiliary task diagnostic module. A self-supervised learning algorithm is used to predict the number of prompts. At the same time, a contrastive learning algorithm is introduced to divide the prompt attributes into positive and negative samples, which allows the model to learn the differences in students' knowledge mastery under different numbers of prompts. The knowledge factors, ability factors, and cognitive load factors obtained in step (3) are concatenated and used as input to the main task, outputting a prediction of the student's response. Finally, guessing and error parameters are introduced. The number of prompts predicted by the self-supervised learning algorithm is used to model the guessing and error parameters. Combined with the student's response predicted by the main task, the DINA diagnostic formula is used to predict the student's final response.

[0013] (5) Collect the dataset, set the optimizer, iteratively train the model until convergence, and finally predict the student's response to 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 test item attributes. Student attributes include knowledge attributes, ability attributes, hint attributes, and speed attributes. Test item attributes include difficulty attributes, workload attributes, and knowledge attributes, i.e., the Q matrix (the Q matrix corresponds to the knowledge points tested in the test item; columns represent knowledge points, rows represent test items, and elements are a binary matrix with only 0 or 1 values. For example, if the first test item 1 is tested, then the first column of the first row is labeled 1, and the other columns of the first row are labeled 0). For student attributes, the knowledge attributes, ability attributes, hint attributes, and speed attributes of the student can be obtained by multiplying the one-hot vector corresponding to the student ID with the trainable matrix. For test item attributes, the knowledge attributes, difficulty attributes, and workload attributes of the test item can be obtained by multiplying the one-hot vector corresponding to the test item ID with the Q matrix and the trainable matrix.

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

[0017] (2-1) Multiply the student’s knowledge attributes by the knowledge attributes of the test questions to obtain the student’s status on the knowledge points tested by the test questions, i.e., knowledge factors.

[0018] (2-2) An attention mechanism is used to match the extracted cue attributes with the student’s ability attributes and the difficulty attributes of the test questions, thereby modeling the student’s ability under the current cognitive load level and the difficulty of the test questions for the student’s cognitive load level. The difference between the two is the student’s true ability, i.e., the ability factor.

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

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

[0021] (3-1) The student's response process is decomposed into several response steps. A Long Short-Term Memory Network (LSTM) is used to perform temporal modeling of the student's response process. The prompt attributes extracted in step (1) are used as the input of the LSTM. After multiple transmission steps, the output of the hidden layer is obtained.

[0022] (3-2) Similarly, LSTM is used to perform time-series modeling on the updated velocity attribute in step (2) to obtain the hidden layer output after multiple transmission steps. Then, it is multiplied by the reaction time to calculate the workload of the student within the reaction time. Finally, the workload attribute of the test question is subtracted to obtain the final velocity attribute representation.

[0023] (3-3) The results of steps (3-1) and (3-2) are fused through a fully connected layer to obtain a representation of cognitive load factors.

[0024] (3-4) Use attention mechanisms to capture the relationship between students' ability factors, knowledge factors and cognitive load factors. Match the knowledge factors and ability factors obtained in step (2) 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 output the final knowledge factors and ability factors.

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

[0026] (4-1) The prompt attributes are used as input for the auxiliary task. A self-supervised learning algorithm is used to predict the number of prompts for students. At the same time, a contrastive learning algorithm is introduced to divide the prompt attribute embeddings into positive and negative samples, so that the model can learn the differences in students' knowledge mastery status under different prompt counts.

[0027] (4-2) Based on the number of prompts predicted in step (4-1), model the probability of students guessing and making mistakes, and predict the probability of guessing and making mistakes.

[0028] (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 output is the prediction of the student's response. The DINA formula is used to combine the student's response prediction obtained from the main task with the guess and error parameters to obtain the final response prediction.

[0029] (4-4) The method of calculating the error gradient under each weight in real time is a classic method of training neural networks by combining optimization methods (such as gradient descent). Gradient descent is used to reduce the loss function in order to optimize the parameters.

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

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

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

[0033] (5-3) Perform backpropagation and select a method to calculate 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 existing technologies, this invention has the following obvious and prominent substantive features and significant technological advancements:

[0036] 1. This invention proposes a neural network learning diagnostic method based on step-by-step modeling of the response process with enhanced cognitive load. It addresses the problems of traditional cognitive diagnostic models, such as the single modeling of students' response process data and insufficient data utilization. Based on cognitive diagnostic theory, it represents the multidimensional features of students and test questions, and further uses cognitive diagnostic analysis methods to model and train the multidimensional features, diagnose students' knowledge mastery status, and predict students' future performance.

[0037] 2. The method of this invention includes a cognitive attribute extraction stage, a cognitive attribute interaction stage, a cognitive load enhancement stage, and a diagnostic analysis stage. It fully utilizes the process data of student responses, and for the first time introduces cognitive load as a new attribute influencing student responses into the model. It employs LSTM to model the student's response process step by step and uses an attention mechanism to fuse student attributes with practice attributes, achieving the transmission of temporal information. Unlike previous studies, this method makes fuller use of data from the student's response process to achieve more accurate diagnosis.

[0038] 3. The method of the present invention builds a neural network learning diagnostic model based on the step-by-step modeling of the answering process with enhanced cognitive load through the cognitive diagnostic method, so that it has the ability to perform cognitive diagnosis and prediction on students' answer records. Attached Figure Description

[0039] Figure 1 This is a scenario diagram illustrating the implementation of the method of this invention.

[0040] Figure 2 This is a model diagram of the neural network learning diagnostic framework based on step-by-step modeling of the response process with enhanced cognitive load, as described in this invention. Detailed Implementation

[0041] This invention, through research on existing public datasets, argues that besides reaction speed, the number of prompts significantly impacts students' response process. Constructivist theory posits that knowledge is acquired by learners in specific contexts, with the help of prompts from others (including teachers and learning partners), and by utilizing necessary learning materials, through meaning construction. Therefore, the number of prompts, as a crucial feature of student interaction data, influences students' response process and knowledge status. Furthermore, cognitive load theory suggests that the brain's working memory capacity is limited; excessive load can affect students' response process. Prompts, as an external intervention, can help alleviate cognitive load by reducing the complexity of learning tasks. Thus, the level of cognitive load affects students' answers, and the number of prompts is a key indicator of cognitive load. More prompts indicate a higher level of cognitive load, and vice versa. Similarly, reaction speed and reaction time also reflect cognitive load levels to some extent; longer reaction times and slower reaction speeds indicate a greater cognitive load. Therefore, integrating cognitive load into cognitive diagnosis to assess students' knowledge status is extremely important.

[0042] This invention discloses a neural network learning diagnostic framework based on step-by-step modeling of the answering process using cognitive load enhancement. It primarily models the student's answering process step-by-step using a long short-term memory network. Based on cognitive load theory, it fully utilizes procedural data from student answer records to extract prompt counts, reaction time, reaction speed, and task workload, collectively representing the student's cognitive load factors and incorporating them into the model, thereby improving the model's effectiveness and accuracy. Specifically, this invention categorizes cognitive factors influencing student answering results into knowledge factors, ability factors, and cognitive load factors. Knowledge factors reflect whether the student's knowledge mastery meets the requirements of the test questions; ability factors are also important factors affecting student responses, for example, in IRT, it is assumed that student answering results are influenced by their potential abilities; cognitive load factors are not directly observable but are externally manifested in easily collected procedural data during the student's response process. To characterize cognitive load factors, this method first extracts them from student answer records. By analyzing these records, procedural data is extracted, including prompt counts, reaction speed, reaction time, and task workload, which are then used to calculate the cognitive load factors. Specifically, the method of this invention first matches the number of prompts with attributes such as students' reaction speed, ability, and question difficulty through an attention mechanism, thereby capturing the complex relationships between these attributes. Furthermore, LSTM is used to perform temporal modeling of the number of prompts and students' reaction speed, thus achieving step-by-step modeling of the student's answering process. The processed hidden layer outputs are then fused through a fully connected layer to obtain a representation of cognitive load factors. Finally, this invention utilizes a multi-layer neural network to fully mine information from the above three cognitive factors and fit the complex nonlinear relationship between students and questions. The model parameters are trained by predicting student answer results, and self-supervised learning and contrastive learning are introduced into the diagnostic module to achieve multi-task prediction. A joint loss function is used to optimize the model. Experimental comparative analysis shows that the neural network learning diagnostic method based on step-by-step modeling of the answering process with enhanced cognitive load proposed in this invention improves the accuracy of answer prediction.

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

[0044] This invention provides a neural network learning diagnostic method based on step-by-step modeling of the response process with enhanced cognitive load, comprising the following steps:

[0045] (1) Cognitive attribute extraction

[0046] The attribute extraction module can categorize extracted attributes into student attributes and test question attributes based on their different sources.

[0047] (1-1) For student attributes, this method uses knowledge attribute k. i Ability attribute a i , prompt attribute d i and speed attribute v i To represent students. Specifically, by multiplying the one-hot vector corresponding to the student's ID with the trainable matrices K, A, D, and V, we can obtain the student's ID. i Knowledge attribute k i Ability attribute a i , prompt attribute d i and speed attribute v i :

[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 (Formula 1.1)

[0049] Among them, h i s ∈{0,1} N Let K, A ∈ R be the one-hot vector corresponding to a student ID. N×L Let D represent the trainable knowledge matrix and ability matrix of N students, respectively. To achieve step-by-step modeling of the response process, this method assumes that the student's response process consists of Z distinct response steps, with the number of prompts at each step being independent. Therefore, the trainable prompt 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} represents student s i The number of prompts in step z is d. iz ∈R L Furthermore, the number of prompts in the student's response record is used as the initial embedding of the trainable prompt matrix. Similarly, the reaction speed at each step is independent, so the student's trainable speed matrix is ​​defined in the same way, i.e., V∈R N×Z×L v i ∈R Z×L Since there is currently no relevant research on the specific number of reaction steps involved in a student's reaction process, this method uses the number of reaction steps Z as a hyperparameter.

[0050] For question attributes, this method uses the knowledge attribute q. j Difficulty attribute b j and workload attribute g j This is used to represent test questions. Specifically, by multiplying the one-hot vector corresponding to the test question id with the Q matrix and the trainable matrices D and G, we can obtain the test question e. j Knowledge attribute q j Difficulty attribute b j and workload attribute g j :

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

[0052] (2) Cognitive attribute interaction

[0053] After obtaining the seven cognitive attributes of students and test questions, attention mechanisms and LSTM are used to realize the interaction between these seven cognitive attributes, thereby capturing the complex relationship between them and obtaining cognitive factors that incorporate the interaction information of these attributes: knowledge factors, ability factors, and updated speed attributes.

[0054] (2-1) Knowledge factors indicate whether students have a sufficient grasp of the knowledge points involved in the test questions; students' knowledge attribute k i This indicates the student's level of mastery over various knowledge points; the knowledge attribute q of the test question. j Derived from the expert-annotated Q-matrix, it represents the state of examination of each knowledge point by the test questions. The DINA model posits that the factors influencing students' answers are only related to whether students have mastered the knowledge points tested by the test questions, and are unrelated to the knowledge points not tested by the test questions. Therefore, the student's s... i Knowledge attribute k i and test question e j Knowledge attribute q j Multiplying these components preserves the student's state regarding the knowledge points tested in the test, i.e., knowledge factor α. ij :

[0055] α ij =k i ⊙q j (Formula 2.1)

[0056] (2-2) Ability factors influence students' performance at the non-knowledge level. This means that in addition to mastering the specific knowledge points required by the questions, students also need good reading comprehension skills to understand the true intent of the questions and thus provide effective answers. Item Response Theory (IRT) provides a precise description of ability, pointing out that students' potential abilities are related to the difficulty level of the questions: facing more difficult questions, students need to demonstrate stronger abilities to have a greater chance of arriving at the correct answer. Furthermore, this method argues that students' abilities are closely related to the number of prompts. According to constructivist theory, the student's response process is not merely an interaction between the student and the exercise; prompts received during the interaction with the external environment also affect the student's answer. Cognitive load theory suggests that prompts, as an external cognitive resource, can help reduce students' cognitive load by lowering the complexity of the learning task. The more prompts a student needs, the greater the cognitive load they face during the answering process. Therefore, in real educational situations, students' actual abilities should be based on their abilities after receiving prompts. This method first employs an attention mechanism to match students' cue attributes and ability attributes. Specifically, it calculates the cosine similarity w between the two vectors of cue attributes and ability attributes. a To indicate the degree of matching, w a From formula w a =Softmax(cs(a i ,d i The calculation shows that cs represents the similarity function. Then, the matching degree and the ability attribute are multiplied to obtain an intermediate representation of the ability factor. Right now The number of times students request hints is closely related to the difficulty of the questions; the more difficult the questions, the more hints students request. Therefore, we also use an attention mechanism to match the student's hint attributes with the question's difficulty attributes, following the same calculation process as before, to obtain an intermediate representation of the question's difficulty attribute. Ability Factor θ ij Can be made by students i Intermediate representation of ability factors and test question e j Difficulty attribute in the middle This means that the result is obtained by subtraction, that is:

[0057]

[0058] (2-3) Corresponding to the student's speed attribute is the workload attribute of the test question. 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, the workload attribute g of the test question needs to be calculated. jand the rate v at each reaction step iz The degree of matching between the two vectors is determined by calculating the cosine similarity w between them. v To represent the degree of matching, i.e., w v =Softmax(cs(g i ,v i Then, the matching degree and the corresponding reaction speed are multiplied to obtain the updated speed attribute. Right now

[0059] (3) Increased cognitive load

[0060] (3-1) The student's response process can be decomposed into several response steps, each with its own independent number of prompts. Since the number of prompts received by the student is interspersed throughout the student's response steps, each response step has its corresponding prompt information. The number of prompts in each step can reflect the cognitive load faced by the student in the current step. Since the student's response steps are performed sequentially, the previous steps will affect the subsequent steps. The prompts in the previous step will play a certain auxiliary role in the student's answer, reduce the current cognitive load, and also affect the cognitive load of the subsequent steps. Therefore, a Long-Short Term Memory Network (LSTM) is used to simulate this sequence feature. LSTM can not only handle features with sequence relationships, but also reasonably model the influence between elements before and after in the sequence feature through memory and forgetting units. This method extracts the prompt attribute d from step (1). i As the input to the LSTM, the output of the last unit of the LSTM is obtained after Z steps. and the hidden layer output of each unit in the middle. 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 The number of prompts for each step is influenced by the results of previous steps, and the output h of the last unit is... Z It includes all the information from step 1 to step Z, and 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 rate, which corresponds to the student's cognitive load, and the reaction rate also exhibits temporal characteristics, LSTM is also used to model it. As the input to the LSTM, the output of the final unit of the LSTM is obtained after Z steps. and the hidden layer output of each unit in the middle. The calculation process is the same as that of formulas 2.3 to 2.8, and the reaction time t is used. ij and Multiply the results to calculate the amount of work the student can complete within the reaction time, and then multiply them by the workload attribute g of the test question. j Subtraction is used to determine whether the workload requirement of the question is met, thus obtaining the final representation of the speed attribute. Right now

[0068] (3-3) Students' response speed is also affected by the number of prompts. The more prompts a student requests, the more difficult the question, and the slower the response speed will be; conversely, the fewer prompts a student requests, the faster the response speed will be. To achieve this interaction, the information obtained in step (3-1) will be used... And obtained in step (3-2) The input is fed into a fully connected layer and fused using the non-linear activation function ReLU to obtain the cognitive load factor representation (CL). ij ,Right now

[0069] (3-4) In the series of steps in a student's response, some steps are related to the degree of knowledge mastery, while others are related to the student's potential ability. Therefore, this method uses attention networks to capture the complex correlations. Specifically, it uses the knowledge factor α obtained in step (2) to... ij And ability factor θij Respectively related to cognitive load factors CL ij Perform matching to obtain CL ij The rich temporal information is transmitted to knowledge and ability factors. Therefore, it is necessary to calculate CL separately. ij With knowledge factor α ij And ability factor θ ij The cosine similarity between them was then normalized using the Softmax function and compared with CL. ij Multiplying them yields the final representation of the knowledge and ability factors. and

[0070] (4) Diagnostic Analysis

[0071] (4-1) The number of hints a student requests when answering questions is influenced by both their potential ability and the difficulty of the question. Different students may request different numbers of hints for the same question, and the same student may request different numbers of hints for questions of varying difficulty. Therefore, this method introduces a self-supervised learning algorithm to address the student's ability factor a i and the difficulty factor of the test questions b j The input is fed into a fully connected layer to predict the number of prompts for the student. Right now:

[0072]

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

[0074]

[0075] in, d represents the predicted number of prompts for the i-th student. i 'Indicates the actual number of prompts given by the student.

[0076] Furthermore, this method also improves the accuracy of model prediction by introducing contrastive learning to optimize the embedding of cue attributes. The core idea of ​​contrastive learning is to optimize the embedding space so that similar samples have greater similarity and dissimilar samples have less similarity. The number of cue points, as a feature, can be correlated and optimized with other features such as students' abilities and test difficulty through contrastive learning. This will help the model better capture the cognitive load of students in different contexts and improve the model's predictive ability. Specifically, this method uses the K-means clustering algorithm to cluster the number of cue points in the dataset according to different clusters. Here, the clusters are divided into three categories based on the number of cue points: low cue points, medium cue points, and high cue points. Then, positive and negative sample pairs are generated according to the categories. If students belong to the same category, it means that their behaviors or performances are similar, and they are positive sample pairs; otherwise, they are negative sample pairs. 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] Among them, y i ∈{0,1} represents the label of a sample pair, where 1 represents a positive sample pair and 0 represents a negative sample pair. d i This represents the Euclidean distance between two samples. The resulting contrastive loss is used in subsequent training.

[0079] (4-2) Inspired by the DINA model, this method also considers the impact of two parameters, error (s) and guess (g), on the student's answering process. That is, even if a student has mastered all the concepts in the question, they may still make an incorrect answer due to error. Conversely, even if a student has not fully mastered all the concepts in the question, they may still make a correct answer through guessing. The number of hints requested by the student has a strong correlation with these two parameters. The more hints a student requests, the more difficult the question is for that student, and the more likely they are to answer by guessing. Conversely, fewer hints indicate a higher probability of correct answering, thus increasing the probability of incorrect answering due to error. Therefore, this method models the student's guessing and error probabilities based on the number of hints predicted in step (4-1), as follows:

[0080]

[0081] Among them, W g W s ∈R 1×L Let b represent the trainable matrices for guessing and error, respectively. g ,bs It is a deviation item. This represents the probability that a student answers correctly by guessing. This indicates the probability that a student answered incorrectly due to a mistake.

[0082] (4-3) To uncover the complex relationship between students and test questions and predict students' answers, this method designs a prediction layer consisting of a three-layer fully connected network. The three final cognitive factors obtained are concatenated and used as the input to the prediction layer, namely:

[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' answer results. Right now:

[0086]

[0087] The binary cross-entropy function is used as the loss function for predicting student responses, as shown in the following formula:

[0088]

[0089] Where y is the actual value of the student's answer.

[0090] (4-4) The method of real-time calculation of error gradient under each weight is a classic method of training neural networks by combining optimization methods (such as gradient descent). It consists of two parts: activation propagation and weight update.

[0091] (4-4-1) In the incentive propagation phase, each iteration is performed in two steps:

[0092] Step 1: Input the training results into the network to obtain stimulus responses;

[0093] Step 2: Difference the stimulus response with the corresponding output target to obtain the response error of the output layer and the hidden layer.

[0094] (4-4-2) In the weight update phase, two steps are performed for each weight:

[0095] Step 1: Multiply the input excitation and response errors 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: Junyi, PISA 2015, and Assistment 2017.

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

[0100] The PISA 2015 dataset: PISA is a globally recognized authoritative online testing system with high-quality test items, recording students' answers and response times. The PISA 2015 dataset includes computer-based PISA mathematical data. This study selected and used 17 computer-scored dichotomous items. The database used for analysis contains dichotomous response data and continuous response time data from a randomly selected group of 6000 students.

[0101] Assistment2017: Assistment2017 originated from the "2017 ASSISTments Data Mining Competition" and provides learners' answer records from 2014 to 2017, as well as the relationships between the exercises and concepts included.

[0102] (5-2) Training with joint loss function

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

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

[0105] Where λ represents the weighting parameter that affects the loss.

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

[0107] The previous layer integrates the difficulty, discrimination, and student's level of mastery of the test questions (X). Upon receiving the mixed input X, X is transferred to the first fully connected layer (Linear layer). X undergoes linear mapping in the first fully connected layer to obtain z1, which is then processed by the sigmoid activation function to obtain X1. X1 is then passed to the second fully connected layer, and the above steps are repeated. After repeating the linear-sigmoid process twice, the mapping product X2 is obtained. The formula is described as follows:

[0108]

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

[0110] In this model, backpropagation serves to update the parameters for fitting, ΔW ij The formula for updating the parameters is described below:

[0111]

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

[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 Therefore, W ij =W ij —ηX i δ j (Formula 5.6)

[0115] It should be noted that the above description of the examples of the present invention is for the purpose of explaining the technical features of the present invention in detail. Various improvements and modifications made without departing from the present invention are also protected by the present invention; therefore, the scope of protection of the present invention should be determined by the content defined in the claims of this application.

[0116] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

Claims

1. A neural network learning diagnostic method based on step-by-step modeling of the answering process enhanced by cognitive load, characterized in that The method comprises the following steps: (1) cognitive attribute extraction, collecting student number, test number, knowledge point corresponding to the test, reaction time and prompt number from the student's answer sequence, and extracting student attributes and test attributes from the data, the student attributes including student knowledge attributes, ability attributes, prompt attributes and speed attributes, and the test attributes including difficulty attributes, knowledge attributes and workload attributes; (2) cognitive attribute interaction, after obtaining the seven cognitive attributes of the student and the test, the interaction between the seven cognitive attributes is realized by using the attention mechanism and LSTM to capture the relationship between them, and the cognitive factors of the interaction information are obtained, including the knowledge factor, the ability factor and the updated speed attribute; (3) cognitive load enhancement, the prompt attribute obtained in step (1) and the updated speed attribute obtained in step (2) are respectively modeled by using LSTM, so as to realize 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 by using a fully connected layer to obtain the cognitive load factor; the cognitive load factor is further fused with the knowledge factor and the ability factor obtained in step (2) by using the attention mechanism, to obtain the final representation of the knowledge factor and the ability factor; (4) diagnostic analysis, the diagnosis 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 student's prompt number; the prompt attribute extracted in step (1) is used as the input of the auxiliary task diagnosis module, a self-supervised learning algorithm is used to predict the student's prompt number, and a contrast learning algorithm is introduced to divide the prompt attribute into positive and negative samples, so that the model learns the difference in the student's knowledge mastery state under different prompt numbers, the knowledge factor, the ability factor and the cognitive load factor obtained in step (3) are spliced and used as the input of the main task, and the prediction of the student's answering reaction is output, finally, the guessing and error parameters are modeled by using the self-supervised learning algorithm to predict the prompt number, combined with the student's answering reaction predicted by the main task, the DINA diagnosis formula is used to predict the student's final answering result; (5) collecting data set, setting optimizer, iteratively training model until convergence, finally predicting student's answering reaction and obtaining student's diagnosis report.

2. The neural network learning diagnosis method based on the step-by-step modeling of the answering process according to the cognitive load enhancement of claim 1, wherein The cognitive attribute extraction in step (1) specifically comprises: The cognitive attribute extraction includes extracting student attributes and test attributes, the student attributes including knowledge attributes, ability attributes, prompt attributes and speed attributes, and the test attributes including difficulty attributes, workload attributes and knowledge attributes, i.e. Q matrix; for the student attributes, the knowledge attributes, the ability attributes, the prompt attributes and the speed attributes of the student are obtained by multiplying the one-hot vector corresponding to the student id with the trainable matrix; for the test attributes, the knowledge attributes, the difficulty attributes and the workload attributes of the test are obtained by multiplying the one-hot vector corresponding to the test id with the Q matrix and the trainable matrix.

3. The neural network learning diagnosis method based on the step-by-step modeling of the answering process according to the cognitive load enhancement of claim 1, characterized in that The cognitive attribute interaction in step (2) specifically includes: (2-1) multiplying the knowledge attribute of the student and the knowledge attribute of the test question to obtain the state of the student on the knowledge point examined by the test question, i.e., the knowledge factor; (2-2) using an attention mechanism to match the extracted prompt attribute with the ability attribute of the student and the difficulty attribute of the test question, thereby modeling the ability of the student under the current cognitive load level and the difficulty of the test question for the student's cognitive load level, and subtracting the two to obtain the true ability of the student, i.e., the ability factor; (2-3) using an attention mechanism to match the work load attribute of the test question and the speed attribute of the student to obtain an updated speed attribute representation.

4. The neural network learning diagnosis method based on the step-by-step modeling of the answering process according to the cognitive load enhancement of claim 1, wherein The cognitive load enhancement in step (3) specifically includes: (3-1) decomposing the reaction process of the student into several reaction steps, using a long short-term memory network to model the time sequence of the student's answering process, and using the prompt attribute extracted in step (1) as the input of the LSTM, and after multiple transmissions, obtaining the output of the hidden layer; (3-2) using LSTM to model the time sequence of the updated speed attribute in step (2), obtaining the hidden layer output after multiple transmissions, and then multiplying the reaction time to calculate the work load of the student within the reaction time, and finally subtracting the work load attribute of the test question to obtain the final speed attribute representation; (3-3) fusing the results of step (3-1) and step (3-2) through a fully connected layer to obtain a cognitive load factor representation; (3-4) using an attention mechanism to capture the relationship between the ability factor and the knowledge factor of the student and the cognitive load factor, matching the knowledge factor and the ability factor obtained in step (2) with the cognitive load factor obtained in step (3-3), so that the knowledge factor and the ability factor share the time sequence information in the cognitive load factor, and output the final knowledge factor and ability factor.

5. The neural network learning diagnostic method based on step-by-step modeling of the answering process with enhanced cognitive load according to claim 1, characterized in that The diagnostic analysis in step (4) specifically includes: (4-1) using the prompt attribute as the input of the auxiliary task, using a self-supervised learning algorithm to predict the number of student prompts, and introducing a contrastive learning algorithm to divide the positive and negative samples of the prompt attribute embedding, so that the model learns the difference in the student's knowledge mastery state under different prompt frequencies; (4-2) modeling the guessing and error probabilities of the student according to the predicted number of prompts in step (4-1), and predicting the probabilities of guessing and errors; (4-3) concatenating the knowledge factor, the ability factor, and the cognitive load factor obtained in step (3) as the input of the main task, outputting the prediction of the student's answering reaction, and combining the student's answering reaction prediction obtained by the main task with the guessing and error parameters using the DINA formula to obtain the final answering reaction prediction; (4-4) using the method of real-time calculation of error gradient under each weight to make gradient descent to reduce the loss function, in order to optimize the parameters.

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

Images