Multi-knowledge point dynamic knowledge tracking method applied to intelligent education system

Abstract

The application discloses a kind of multi-knowledge point dynamic knowledge tracking methods applied to intelligent education system.The method includes the following steps:1, design Res-embedding method pre-training exercise embedding vector;2, construct historical interaction, position coding, the embedding vector of real answer condition;3, calculate the weight of student current answer exercise and student historical answer interaction;4, maintain student knowledge mastery matrix;5, aggregate the mapping relationship between student historical interaction data and current answer exercise;6, predict student current exercise answer correct probability;7, update student knowledge state.According to the consistency measure, the best discriminant result is calculated.The advantage of the application lies in that the exercise representation is pre-trained by Res-embedding method.The concept of student learning absorption index is proposed, which is used for the update operation of student knowledge state.The application greatly improves the method effect, improves the robustness of knowledge tracking algorithm, and can be applied to different student exercise answer scenes.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of intelligent education data mining, and relates to a dynamic knowledge tracking method based on multiple knowledge points. According to student answering interaction data, the Res-embedding learning mechanism is used to mine the corresponding relationship between each exercise and the corresponding knowledge point, and the mastery of each knowledge point by the student is predicted. The method can be applied to an online answering system to assist teachers in assigning personalized after-school exercises for each student, thereby improving the learning efficiency of each student. TECHNICAL BACKGROUND

[0002] With the development of the times, modern education deeply influenced by informatization reform has derived a novel teaching mode of online and offline integration. Through mining and analysis of the massive teaching data generated, the data should be used to improve teaching quality and efficiency. Some domestic education platforms mine student learning data and provide intelligent education services such as adaptive exercises and learning path recommendations. In order to provide uninterrupted continuous student learning services, knowledge tracking is proposed and serves as an extremely important research topic in the field of intelligent education. Knowledge tracking dynamically mines the mastery of knowledge points by each student at different times using the past exercise answering interaction records of each student. Existing knowledge tracking methods are divided into traditional learning methods and deep learning methods. Generally, traditional learning methods cannot fully utilize the past answering interaction records of students, resulting in biased prediction of the mastery of knowledge points by students and the inability to provide optimal online learning guidance. Therefore, deep learning methods have more extensive utilization value in the field of intelligent education. At the same time, knowledge tracking based on deep learning methods is an important way for personalized exercises and path recommendations for students. The more accurate the mastery of student learning is, the more beneficial it is for students to fill in weak knowledge points.

[0003] For traditional knowledge tracking methods, the existing traditional methods mainly include item response theory, factor analysis-based method and knowledge tracking method based on hidden Markov characteristics. The traditional methods are relatively simple, among which the item response theory and the factor analysis method analyze the student level under the assumption that the learning process is unchanged in a period of time. The knowledge tracking method based on hidden Markov characteristics uses the characteristics of Markov model to represent the student knowledge state by binary attribute (master and not master), and the change of student knowledge state only depends on the knowledge state at the last time. The learning process of students is a gradual process, and although the traditional method takes good effect, it does not conform to the transfer of student knowledge state. Therefore, knowledge tracking based on deep learning is proposed. First, the knowledge tracking model based on recurrent neural network is proposed, which gives the knowledge tracking a stronger problem representation ability, establishes the implicit association between knowledge points, and establishes the interactive multi-dimensional knowledge state of students in the hidden space. Secondly, the knowledge tracking model based on recurrent neural network breaks away from the Markov characteristics, and can consider the mastery state of the past knowledge, so that the modeling of student learning behavior is more reasonable. On the basis of the knowledge tracking based on recurrent neural network, the method considering the forgetting characteristics of students is proposed, which makes the transfer process of student knowledge state more reasonable. In order to further model the relationship between knowledge points and improve the transfer process of student knowledge state, the dynamic key-value pair storage network uses memory to enhance the network influence on the change of student mastery of knowledge points. The knowledge tracking model based on graph neural network models the relationship between problems with knowledge points according to the interactive information of student problem answering, effectively simulates the knowledge transmission process in the learning process of students. However, the above models only rely on the knowledge points of students to model the students, thereby ignoring the association between the problems and the knowledge points, and some problems may correspond to the same knowledge points, but the problem description, difficulty and other ways are completely different, which will affect the prediction of the model on the performance of student behavior. In addition, the current knowledge tracking theory is based on the monotonicity assumption, which considers that the student's wrong answer indicates that the knowledge point is not mastered, and vice versa. SUMMARY

[0004] The application discloses a multi-knowledge point dynamic knowledge tracking method applied to an intelligent education system, pre-trains input exercises through a Res-embedding method, mines the relationship between the exercises and corresponding multi-skills, projects the input exercises, position coding and historical answer interaction information into a high-dimensional linear space to be commonly used as the input of the method, establishes the weight relationship between the current answer exercises and the historical answer interaction records by using a self-attention mechanism, establishes the representation between the past knowledge points mastered by students and the current exercises by reading the method, the representation is an aggregation operation on the past answer interaction, and the knowledge performance of students is updated by using the real answer conditions when the exercises are answered.

[0005] In the first step, the Res-embedding method is used to pre-train the input exercises.

[0006] In the second step, the historical answer interaction of students, position coding and real answer conditions are projected into a high-dimensional linear space through an embedding layer.

[0007] In the third step, the self-attention mechanism method is used to establish the weight relationship between the current answer exercises and the historical answer exercises of students.

[0008] In the fourth step, a knowledge point mastering matrix of each student is maintained, the current exercises are represented by using the knowledge mastering conditions of students according to the weight relationship obtained in the third step, and the historical answer conditions are summarized to the current exercises.

[0009] In the fifth step, the mastering conditions of students under the current exercises are predicted by aggregating the historical interaction information and the information of the current exercises.

[0010] In the sixth step, the knowledge point mastering matrix of each student is updated for dynamically pursuing the knowledge mastering conditions of students, and the update content depends on the real answer conditions of students.

[0011] Compared with the prior art, the present application has the following significant features: (1) the present application links each question to multiple knowledge points. In order to alleviate the "sparsity disaster" caused by the large amount of questions and the sparseness of student answer data, the prior art method only considers the skill corresponding to each question. The present application uses the Res-embedding method to integrate the characteristics of multiple skills and questions, thereby alleviating the "data disaster" problem and considering the unique characteristics of each question in the method; (2) the method proposes a novel state updating method for simulating the learning process of students. The prior art method assumes that if a student answers a question correctly, the proficiency of the corresponding knowledge point increases, and vice versa. The present method believes that even if a student makes a mistake in a question, the proficiency of the knowledge point should not decrease significantly, and in some cases, it can even increase.

[0012] The present application uses student answer history interaction data to mine and predict the knowledge point mastery of students, thereby overcoming the data sparsity problem caused by the lack of student answer interaction information. The present application also proposes a novel student knowledge point proficiency updating method to improve the accuracy of the knowledge point mastery of students. The following knowledge tracking examples and drawings further describe the present application in detail. BRIEF DESCRIPTION OF DRAWINGS

[0013] Figure 1 is a flowchart of the present application.

[0014] Figure 2 is a visualization effect diagram of the pre-trained question embedding vector. The left diagram is a visualization effect diagram of the 2009-2010 ASSISTment dataset, and the right diagram is a visualization effect diagram of the ASSISTment Challenge dataset.

[0015] Figure 3 is a visualization diagram of the knowledge mastery state of two students DETAILED DESCRIPTION

[0016] In combination with Figure 1 , the implementation process of the present application is described in detail below, and the steps are as follows:

[0017] Step 1: using the Res-embedding method to pre-train the input question. That is, input a set of questions and their corresponding knowledge points, and obtain the appropriate question representation according to the self-learning weight. Taking a question with C corresponding knowledge points as an example, the specific process is as follows:

[0018] (1) input the question q and the corresponding C knowledge points {s1, s2, s3,... s C}, according to the self-attention mechanism method, obtain the weight relationship between each knowledge point and the question q then the question embedding vector Can be expressed as:

[0019]

[0020] (2) To consider the weight of the exercise embedding vector representation more reasonable, the exercise feature matrix is introduced To distinguish the unique characteristics between each exercise. Where N represents the total number of exercises in the exercise library, d k Indicates the dimension of the linear space of the embedding vector. Finally, the exercise embedding vector q e Is expressed as:

[0021]

[0022] Where q feature Indicates the exercise feature embedding vector corresponding to the exercise q in the exercise feature matrix Concat[,,] represents the concatenation operation.

[0023] Second, the student's historical answer interaction, position coding and true answer situation are projected into a high-dimensional linear space through the Embedding layer. After projection through the embedding layer, the historical exercise answer interaction matrix is represented as I e , the position coding matrix is represented as pos e , the currently answering exercise matrix is represented as Q e , the corresponding knowledge point matrix of Q e Is represented as S e .

[0024] Third, the weight relationship between the current answering exercise and the historical interaction answer is considered using the self-attention mechanism method. If the self-attention mechanism is used, the input query, key, and value need to be defined in advance. Then it is specifically expressed as:

[0025] query=F(concat(Q e , pos e , I e )),

[0026] key=F(I e ),

[0027] value=F(S e ),

[0028] Where F(·) is a full connection layer. Next, the weight relationship between the current answering exercise and the historical answer record needs to be defined. Assume that the input at time t is query t , key t-1 , value t, the weight between the current answer of the student i at time t and the history answer is defined as:

[0029]

[0030] where z i can represent any scalar or function. Finally, a mapping relationship x t between the current answer of the student i at time t and the history answer record is established , which is specifically represented as:

[0031] x t = w t value t

[0032] The fourth step is to maintain a knowledge point mastery matrix M v for each student. We assume that the answer of each student to each question will be projected into a vector space R t , and any interaction information should be represented as a vector r i ∈ R t , then the interaction vector r t at the current time t is represented as:

[0033]

[0034] The weight between the current answer of the student i at time t and the history answer is obtained. The knowledge point mastery matrix of the student i at time t is represented as:

[0035] The fifth step is to predict the mastery of the student under the current question f t by aggregating the history interaction information r t and the current question information x t . Specifically, it is represented as:

[0036] f t = F tanh (concat[r t , x t ]),

[0037] F tanh (·) represents the combination of a fully connected layer and a tanh function, where Finally, the probability p t that the student answers correctly for the current answer question q is obtained through a fully connected layer and a sigmoid function. The specific process is:

[0038] p t = F sigmoid (f t ),

[0039] F sigmoid (·) represents the combination of full connection layer and sigmoid function, where

[0040] The sixth step, after predicting the latest performance of the students, to dynamically track the knowledge mastery of the students, the knowledge point mastery matrix M of each student needs to be updated, and the update content depends on the input of the student's real answer qa v . The specific process is: t

[0041] (1) According to step one, q e and student answer record a t , the student answer vector at time t is obtained The specific expression is:

[0042]

[0043] And the interaction vector r t at the current time t in step four is considered together to update the student knowledge point mastery matrix M v The specific expression is:

[0044]

[0045] (2) To update the module to build a more effective knowledge tracking model, after obtaining the student's real answer qa t , directly through a layer of full connection and tanh layer, for the update operation of the knowledge point mastery matrix M of each student V , the specific process is:

[0046] c t = F tanh (qa t )

[0047]

[0048]

[0049] Indicates the temporary transition knowledge point mastery matrix. Indicates the knowledge point mastery matrix of student i at time t, which is used for predicting whether the student can answer correctly at time t+1. The effect of the present application can be further illustrated by the following simulation experiment:

[0050] Simulation conditions

[0051] ​To verify the effectiveness of our proposed invention, we conduct simulation experiments using three real student answering interaction datasets, namely the 2009-2010 ASSISTment dataset, the ASSISTment Challenge dataset, and the KDD Cup algebra_2005_2006 dataset. The 2009-2010 ASSISTment dataset is collected by the ASSISTment online education platform and contains 282,790 student answering records, 17,717 exercises, and 123 knowledge points. The dataset has a repeated submission problem, and we need to stack the multi-knowledge point exercises to generate an answering interaction. The ASSISTment Challenge dataset is from the 2017 ASSISTments data mining competition and has more student answering interactions than the 2009-2010 ASSISTment dataset. The dataset has 942,816 answering records, 3,162 exercises, and 102 knowledge points. The ASSISTment Challenge dataset has the same questions as the 2009-2010 ASSISTment dataset, so it needs to be processed in the same way as the 2009-2010 ASSISTment dataset. The last dataset, KDD Cup algebra_2005_2006, is from the 2010 KDD Cup competition and is different from the above two datasets. Each interaction in the dataset represents a step, so we integrate the exercises and steps into a question. The dataset includes 606,401 interaction information, 172,994 exercises, and 112 knowledge points. Before the experiment, we preprocessed the dataset, and the simulation experiment and related comparison experiments were conducted on the Ubuntu 20.04 operating system, with Python 3.7 and Pytorch 1.7 as the programming language environment, and two GeForce 1080Ti GPUs for hardware support.

[0052] The evaluation index used in the invention is the area size under the ROC curve (AUC). To demonstrate the effectiveness of our invention, we also implemented several widely used comparison methods, namely:

[0053] (1) DKT: This model implements a recurrent neural network to build a student knowledge point mastery model and summarizes the student's learning situation.

[0054] (2) DKVMN: This model uses key-value pairs to store the knowledge point and student knowledge point mastery level embedding vectors and updates the knowledge point mastery level embedding vectors through a forgetting gate mechanism.

[0055] (3)DKT+: This model is based on the improvement of the DKT model, which solves the problem of input reconstruction failure and unsustainable prediction of the DKT model.

[0056] (4)GKT: This model introduces the knowledge point distribution of the graph structure, and uses the change of the state of the adjacent nodes to update the learning state of the current student.

[0057] (5)SAKT: This model uses a self-attention mechanism to capture the relationship between the current answer knowledge point and the past record.

[0058] (6)HawkestKT: This method is the first time to introduce the Hawkest process into the knowledge tracking field

[0059] Simulation experiment result analysis

[0060] Table 1 is the simulation experiment result of the method of the application and other comparative methods on three data sets. From the experimental results, our results have the best effect in three different data sets, which embodies the universality of the method of the application. There are several phenomena worth paying attention to: (1) There are more interaction information in each question in the KDD Cup algebra_2005_2006 data set, so all methods have achieved good results in this data set. (2) The method based on question input (HawestKT, SAKT, DMKT) is usually better than the method based on knowledge point input (DKT, DKVMN, GKT, and DKT+), which leads to the phenomenon that using different information as input may lead to performance stratification. The original exercise input also leads to the problem of sparsity. This further proves the superiority of the embedded exercise input concept of DMKT. In addition, the DKT and DKT+ models are based on recurrent neural networks, and their overall performance is better than the model based on memory enhanced network. Especially DKT+ performs well in KDD algebra_2005_2006 data set, reaching 82.5%. The reason may be that DKT+ uses a reasonable method to update the knowledge proficiency of students, and the fact that the interaction sequence of each student in KDD algebra_2005_2006 is long. Therefore, the learning absorption index concept proposed by our method is considered in the updating process of the learning state of the student. Therefore, the performance in the three data sets is at least 1.0% higher than that of other methods.

[0061] Table 2 is the ablation simulation result of the method of the application. In order to prove the role of each module of the method, we have carried out some ablation research. We design three kinds of method variants.

[0062] (1)RQE: This variant method removes the question embedding vector.

[0063] (2) RFQE: This variant removes the problem feature vector.

[0064] (3) RLAI: This variant removes the learning absorption index concept.

[0065] First, RQE performs worse than the original method. The reason is that the change of RQE is the lack of multi-skill information after the corresponding module is removed. This leads to insufficient information embedded in the practice, resulting in performance degradation. In addition, RQE still achieves better performance compared to the practice input model (Hawest KT, SAKT). This proves that the remaining part of the original method has superior fitting performance. Then, we remove the problem feature embedding and only consider the multi-knowledge point embedding. The performance of RFQE decreases significantly. The reason is that when the exercises have similar skills, they share the same representation. This leads to the fact that the single input of the model cannot distinguish the problems. Finally, RLAI deletes the student's interaction vector r t when updating the interaction state. The results of RLAI are worse than the original method, and this variant does not consider the learning ability of students. Whenever the student answers a practice, it is assumed that he will obtain all the knowledge. This unreasonable behavior will cause the tracking of the student's knowledge state to be biased.

[0066] To further demonstrate the effect of Res-embedding and the learning absorption index concept, we visualize the changes in the problem embedding vector and the student's knowledge point state, which are shown by Figure 2 and Figure 3 respectively.

[0067] Table 1

[0068]

[0069] Table 2

[0070]

Claims

1. A multi-knowledge point dynamic knowledge tracking method applied to an intelligent education system, comprising the following steps Step 1: Using the Res-embedding method, pre-training the input exercises, fusing the multi-knowledge point embedding vector and the exercise vector, and through continuous training, representing the input exercises appropriately, Step 2: Student historical answer interaction, position encoding and real answer situation are projected into a high-dimensional linear space through an embedding layer, Step 3: Use the self-attention mechanism method to establish the weight relationship between the student's current answer exercise and the historical answer exercise interaction, Step 4: Maintain a knowledge point mastery matrix for each student, and use the student's knowledge mastery to represent the current exercise according to the weight relationship obtained in Step 3, which contains the summary of the historical answer situation to the current exercise, Step 5: Predict the student's current exercise answer situation by aggregating historical interaction information and current exercise information, Step 6: To dynamically track the student's knowledge mastery, the knowledge point mastery matrix of each student needs to be updated, and the update content depends on the input student's real answer situation.

2. The method of claim 1, wherein the method further comprises: Calculate the exercise embedding vector, the specific process is: (1) input the exercise q and the corresponding C knowledge points {s1, s2, s3,...s C}, according to the self-attention mechanism method, the weight relationship between each knowledge point and the exercise q is obtained i is the 0th to the Cth, and the exercise weight embedding vector can be represented as: (2) To consider the more reasonable representation of the embedded vectors of exercises, the exercise feature matrix is introduced to distinguish the unique features between each exercise, where N represents the total number of exercises in the exercise library, d k represents the dimension of the linear space of the embedded vector, and finally the exercise embedded vector q e is represented as: where q feature denotes the problem feature matrix denotes the problem feature embedding vector corresponding to problem q, and concat[, ] denotes the concatenation operation.

3. The method of claim 1, wherein the method further comprises: The embedding vector of the student historical answer interaction, the position coding embedding vector and the embedding vector of the true answer condition are calculated, and the specific steps are that the student historical answer interaction, the position coding and the true answer condition are projected to a high-dimensional linear space through an embedding layer. After the projection through the embedding layer, the historical exercise answer interaction matrix is represented as I e , the position coding vector is represented as pos e , the current answer exercise matrix is represented as Q e , and the corresponding knowledge point matrix is represented as S e . e .

4. The method of claim 1, wherein the method further comprises: Use the self-attention mechanism method to consider the weight relationship W between the current answer exercise and the historical interaction answer, if the self-attention mechanism is used, the input query, key, value need to be defined in advance, then the specific representation is: query = F(concat(Q e , pos e , I e )), key=F(I e ), value = F(S e ), Where F(·) refers to the full connection layer, next, the weight relationship between the current answer to the question and the historical answer interaction needs to be defined, assuming that the input at time t is query t , key t-1 , value t The corresponding weight of student i between the answer to the question at time t and the historical answer is defined as: ​ wherein z i may represent any scalar or function; ultimately, a mapping between the current answer to the history of answers is established x t , which is represented in detail as: x t = w t value t .

5. The method of claim 1, wherein the method further comprises: Maintain a matrix M of knowledge mastery for each student v We assume that each student's answer to each problem is projected into the vector space R t Any interaction information should be represented as a vector r i ∈R t Then the interaction vector r t at the current time t is represented as: weight between the answer of the exercise and the historical answer for student i at time t in step three, represents the knowledge point mastery matrix of student i at time t.

6. The method of claim 1, wherein the method further comprises: By aggregating historical interaction information r t and current exercise information x t to predict the mastery of the student under the current exercise f t , which is specifically represented as: f t = F tanh (concat[r t , x t ]), F tanh (·) denotes the combination of a fully connected layer with a tanh function, where Finally, the student's probability p of answering the current question q correctly t will be obtained after passing through a layer of fully connected and sigmoid function, the specific process is: p t = F sigmoid (f t ), F sigmoid (·) denotes the combination of a fully connected layer and a sigmoid function, where 7. The method of claim 1, wherein the method further comprises: When the student's latest performance is predicted, the knowledge matrix M of each student needs to be updated to dynamically track the student's knowledge mastery v The update content depends on the input of the student's real answer qa t The specific process is: (1) Get the exercise embedding vector q according to step one e and student answer record a t , define the student answer vector at time t Specifically represented as: And the current t moment of step four interaction vector r t For updating the student knowledge point mastery matrix M v In the process, the student knowledge state is updated by referring to the learning absorption index, which is specifically represented as: (2) To update the module for constructing more effective knowledge tracking model, get the real answer of students qa t Then, directly through a fully connected and tanh layer, for the final each student's knowledge point mastery matrix M V update operation, the specific process is: c t = F tanh (qa t ), represents a temporary transition knowledge point mastery matrix, represents a knowledge point mastery matrix of student i at time t, used to predict whether student i can answer the exercise at time t+1.

Images