Knowledge tracking enhancement method and device based on adversarial training

Abstract

The application discloses a knowledge tracking enhancement method and device based on adversarial training, and relates to the field of intelligent education.The method comprises the following steps: constructing a training set, wherein the training set comprises a plurality of original knowledge point record sets and corresponding answer record sets; calculating interactive embedding representations representing the answering conditions of testers at each time step according to the original knowledge point record sets and the corresponding answer record sets; generating adversarial interactive embedding representations of the interactive embedding representations; training a deep neural network based on all the interactive embedding representations and the adversarial interactive embedding representations to obtain a knowledge tracking model; and obtaining a knowledge tracking result of the testers according to the knowledge tracking model.The application can more effectively enhance the knowledge tracking effect and improve the quantitative representation ability of the knowledge point mastery degree of students.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of smart education, and in particular to a knowledge tracking enhancement method and device based on adversarial training, which can more effectively enhance the knowledge tracking effect and improve the quantitative representation ability of students' knowledge point mastery. BACKGROUND

[0002] Education has always been a hot topic in society. Thanks to the rapid development of digital multimedia technology, smart education has also received more and more attention. Among them, knowledge tracking is considered a core task, and its purpose is to track students' knowledge point mastery during problem solving. Specifically, knowledge tracking predicts the probability of solving the next question through students' historical problem solving records, and regards the predicted probability as the students' mastery of the corresponding knowledge point. On this basis, the students' knowledge point mastery is dynamically tracked through time series prediction. Figure 1 A student practice process diagram is shown. The student has practiced the first four questions, of which the first, second and fourth questions are answered correctly, and the third question is answered incorrectly. Using knowledge tracking, the probability of answering the fifth question correctly can be predicted through the student's first four question answers, and the student's mastery of the corresponding knowledge point can be evaluated. This technology not only helps students better understand their knowledge point mastery, but also provides decision support for subsequent personalized question recommendation tasks.

[0003] As an important topic in the field of smart education, knowledge tracking has received continuous attention from researchers. In early work, researchers mainly followed the probabilistic model technology route to carry out research. This technology route assumes that the student learning process follows a Markov process, and estimates the student's knowledge point mastery by observing the student's learning performance. Bayesian Knowledge Tracing (BKT) is one of the representative methods. This method formalizes the student's knowledge point mastery as a set of binary variables, and uses a hidden Markov model to update the knowledge point mastery, achieving good tracking effect. Subsequent work has expanded on the basis of BKT, further improving the effect by incorporating more prior information, such as question difficulty, guessing probability, etc. Although these works have achieved good results, there are still two major defects. First, the premise assumption of these methods is often too simple, to some extent, it does not conform to the reality of learning, such as "there is no forgetting in the learning process". In addition, most of these methods rely on manual feature modeling methods, without considering more potential modeling factors, further limiting the algorithm performance.

[0004] In recent years, more and more researchers focus on using deep neural networks (DNN) to model knowledge tracing tasks and make some beneficial explorations due to the good representation ability of DNN. Generally, these methods are often based on recurrent neural networks and attention mechanisms, and representative methods include deep knowledge tracing (DKT), self-attentive knowledge tracing (SAKT), context-aware attentive knowledge tracing (AKT), etc. Compared with traditional methods based on hand-crafted feature modeling, these knowledge tracing methods based on DNN modeling in a data-driven manner have better performance. However, these methods still have the risk of overfitting, especially on small-scale datasets. In order to pursue better tracing effect, the generalization performance of the model needs to be further improved. Considering that adversarial training is an effective regularization method, the present application focuses on using adversarial training to improve the generalization ability of the knowledge tracing model based on DNN, so as to achieve better tracing effect. SUMMARY

[0005] In view of the potential overfitting risk of the knowledge tracing model based on deep neural network, the present application discloses a knowledge tracing enhancement method and device based on adversarial training, which uses adversarial training to train the knowledge tracing model, can enhance the generalization ability of the knowledge tracing model, and improve the knowledge tracing effect.

[0006] The technical content of the present application includes:

[0007] A knowledge tracing enhancement method based on adversarial training, the steps of which include:

[0008] Constructing a training set, the training set including a plurality of original knowledge point record sets and corresponding answer record sets;

[0009] According to the original knowledge point record set and the corresponding answer record set, the interactive embedding representation representing the testee's answer at each time step is calculated;

[0010] Generating an adversarial interactive embedding representation of the interactive embedding representation;

[0011] Based on all the interactive embedding representations and the adversarial interactive embedding representation, a deep neural network is trained to obtain a knowledge tracing model;

[0012] According to the knowledge tracing model, the knowledge tracing result of the testee is obtained.

[0013] Further, the computing the interactive embedding representation characterizing the testee's answering situation at each time step comprises:

[0014] 1) mapping each original knowledge point record s j and the corresponding answering record a j to an original knowledge point embedding representation and a student answering record embedding representation wherein j is a number, the original knowledge point record s j and the answering record a j are arranged in time sequence;

[0015] 2) concatenating the original knowledge point embedding representation and the student answering record embedding representation to obtain the interactive embedding representation e j .

[0016] Further, the interactive embedding representation represents a dimension concatenation operator, a j = 1 indicates that the question is answered correctly, and a j = 0 indicates that the question is answered incorrectly.

[0017] Further, the generating the adversarial interactive embedding representation of the interactive embedding representation comprises:

[0018] 1) generating an adversarial perturbation r' of the interactive embedding representation e j-1 ;

[0019] 2) adding the adversarial perturbation r' to the interactive embedding representation e j-1 to generate an adversarial interactive embedding representation e' j-1 .

[0020] Further, the adversarial perturbation wherein ε is a parameter controlling the amplitude of the adversarial perturbation, θ is a deep neural network parameter, and L is a loss function for training the deep neural network.

[0021] Further, the deep neural network comprises:

[0022] a long short-term memory network for capturing the relationship between the answering situations at different time steps based on the interactive embedding representation or the adversarial interactive embedding representation to obtain knowledge point hidden state information at different time steps;

[0023] an attention network for aggregating the knowledge point hidden state information at different time steps and emphasizing the knowledge point hidden state information at the current time step to obtain a comprehensive representation of the knowledge point hidden state;

[0024] A fully connected layer is configured to calculate a knowledge tracking result according to the comprehensive representation.

[0025] Further, the knowledge tracking result of the tester is obtained according to the knowledge tracking model, and the knowledge tracking result of the tester comprises:

[0026] 1) Obtain a set of original knowledge point records of the tester and a corresponding set of answer records;

[0027] 2) Calculate an interactive embedding representation representing the answering situation of the tester at each time step according to the set of original knowledge point records of the tester and the corresponding set of answer records;

[0028] 3) Input the interactive embedding representation representing the answering situation of the tester at each time step into the knowledge tracking model to obtain the knowledge tracking result of the tester.

[0029] Further, the knowledge tracking result comprises a probability value of answering the next original knowledge point record correctly.

[0030] A storage medium having a computer program stored therein, wherein the computer program is configured to execute any of the above methods when running.

[0031] An electronic device, characterized by comprising a memory and a processor, wherein the memory has a computer program stored therein, and the processor is configured to execute the computer program to execute any of the above methods.

[0032] Compared with the prior art, the present application has at least the following advantages:

[0033] 1. Considering that one-hot encoding cannot generate effective adversarial samples by adding slight perturbations, the present application changes the question and answer information into an embedding form and considers the different answering situations of students to establish an interactive embedding representation;

[0034] 2. Based on a long short-term memory (LSTM), an attention mechanism-based long short-term memory is constructed by utilizing the information of each historical time step while emphasizing the importance of the current time step, thereby further improving the representation ability of the model;

[0035] 3. Adversarial samples are introduced into the field of knowledge tracking, that is, the model generalization ability is improved and the model prediction performance is enhanced by collaborative training of original samples and adversarial samples. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 Student exercise schematic diagram.

[0037] Figure 2 Method flowchart of the present application.

[0038] Figure 3 Schematic diagram of backbone network of the present application. DETAILED DESCRIPTION

[0039] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings. Obviously, the described embodiments are only specific embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0040] The knowledge tracking enhancement method of the present application, as shown in Figure 2 , comprises:

[0041] constructing a training set, the training set comprising a plurality of original knowledge point record sets and corresponding answer record sets;

[0042] According to the original knowledge point record set and the corresponding answer record set, an interactive embedding representation representing the testee's answering situation at each time step is calculated.

[0043] generate an adversarial interactive embedding representation of the interactive embedding representation;

[0044] Based on all the interactive embedding representations and the adversarial interactive embedding representations, a deep neural network is trained to obtain a knowledge tracking model.

[0045] According to the knowledge tracking model, the knowledge tracking result of the testee is obtained.

[0046] As shown in Figure 3 , the backbone of the present application is composed of three parts, namely Interaction Projection, Attentive-LSTM based on attention mechanism and Response Prediction. They are introduced as follows.

[0047] 1)Interaction Projection: This part maps the original knowledge point record s j and the student answer record a j into original knowledge point embedding representation and student answer record embedding representation respectively, and further splices the two embedding representations according to the correctness of the student's answer to obtain the interactive embedding representation representing the student's answering situation:

[0048]

[0049] wherein, represents a dimension splicing operator, a j = 1 represents answering a question correctly, a j = 0 represents answering a question incorrectly.

[0050] 2) Attentive-LSTM: This part is composed of two modules, namely a knowledge hidden state modeling (KHS Modelling) module and a knowledge hidden state attention (KHS Attention) module. The KHS Modelling module is composed of an LSTM, which is used to capture the relationship between the student's answers at different time steps, and obtain the KHS at different time steps: {h1, h2, h3, … h t-1 On this basis, the KHS Attention module is designed based on the attention mechanism to adaptively aggregate the KHS information of different historical time steps while emphasizing the KHS information of the current time step, thereby obtaining a comprehensive representation of the KHS:

[0051]

[0052] wherein, represents the aggregated information obtained by implementing the attention mechanism on the KHS information of different historical time steps, h t-1 represents the KHS information of the current step.

[0053] 3) Response Prediction: This part is composed of a fully connected layer and a Sigmoid function, which is used to map the comprehensive representation obtained by the Attentive-LSTM to a probability value indicating the correct answer to the next question:

[0054]

[0055] wherein, a t represents the student's answer record, represents the model prediction result.

[0056] In addition, on the basis of establishing the backbone network, the invention uses adversarial training to enhance the generalization ability of the model. To this end, first, an adversarial sample needs to be constructed. The Fast Gradient Sign Method (FGSM) method is used to generate an adversarial perturbation r':

[0057]

[0058] wherein, ε is a parameter controlling the amplitude of the adversarial perturbation, where θ is the model parameter, and L is the loss function. Add the adversarial perturbation r' to e j-1 The above generates the adversarial interactive embedding representation e' j-1 :

[0059] e' j-1 = e j-1 + r'

[0060] On this basis, the original interactive embedding representation e j-1 and the adversarial interactive embedding representation e' j-1 are combined to update the model parameter θ:

[0061] θ←L(e j-1 , a j , θ) + βL(e' j-1 , a j , θ)

[0062] where β is a hyperparameter that controls the relative importance of adversarial training.

[0063] In the test process of the present application, the trained knowledge tracking model can be used to dynamically track the knowledge point mastery degree of the tester. According to all the original knowledge point records and answer records of the tester, the trained knowledge tracking model is used to update the probability values of the tester correctly answering each knowledge point in turn, and these probability values are taken as the mastery degrees of the tester for each knowledge point.

[0064] The above merely describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. An adversarial training based knowledge tracing enhancement method, comprising the steps of: constructing a training set comprising a plurality of original knowledge point record sets and corresponding answer record sets; calculating interactive embedding representations representing the testee's answering conditions at each time step based on the original knowledge point record sets and corresponding answer record sets; wherein the calculation of the interactive embedding representations representing the testee's answering conditions at each time step comprises: mapping each original knowledge point record s in the original knowledge point record set and the corresponding answer record set to an original knowledge point embedding representation j to an answer record embedding representation j to an answer record embedding representation to an answer record embedding representation wherein j is a number, the original knowledge point record s j to an answer record embedding representation j based on time sequence arrangement; concatenating the original knowledge point embedding representation with the student answer record embedding representation obtaining the interaction embedding representation e j generating an adversarial interaction embedding representation of the interaction embedding representation; wherein the generating an adversarial interaction embedding representation of the interaction embedding representation comprises: producing an interactive embedding representation e j-1 of an adversarial perturbation r' adding an adversarial perturbation r' to the interaction embedding representation e j-1 above, generating an adversarial interaction embedding representation e' j-1 ; training a deep neural network based on all interactive embedding representations and adversarial interactive embedding representations to obtain a knowledge tracing model; obtaining the knowledge tracing result of the testee according to the knowledge tracing model.

2. The method of claim 1, wherein, The interaction embedding represents a represents a dimension concatenation operator j = 1 represents a correct answer to a question, a j = 0 represents an incorrect answer to a question.

3. The method of claim 1, wherein, Robustness to adversarial perturbations where ε is a parameter controlling the magnitude of adversarial perturbations, θ are the parameters of the deep neural network, and L is a loss function for training the deep neural network.

4. The method of claim 1, wherein, The deep neural network comprises: a long short-term memory network for capturing the relationship between the answering conditions at different time steps based on the interactive embedding representations or the adversarial interactive embedding representations to obtain knowledge point hidden state information at different time steps; an attention network for aggregating the knowledge point hidden state information at different time steps and emphasizing the knowledge point hidden state information at the current time step to obtain a comprehensive representation of the knowledge point hidden state; a fully connected layer for calculating the knowledge tracing result according to the comprehensive representation.

5. The method of claim 1, wherein, The obtaining of the knowledge tracing result of the testee according to the knowledge tracing model comprises: 1) obtaining the original knowledge point record set and the corresponding answer record set of the testee; 2) calculating the interactive embedding representations representing the testee's answering conditions at each time step based on the original knowledge point record set and the corresponding answer record set of the testee; 3) inputting the interactive embedding representations representing the testee's answering conditions at each time step into the knowledge tracing model to obtain the knowledge tracing result of the testee.

6. The method of claim 1, wherein, The knowledge tracing result comprises a probability value of answering the next original knowledge point record correctly.

7. A storage medium having stored therein a computer program, wherein, The computer program is configured to execute the method of any one of claims 1-6 when running.

8. An electronic device, comprising: A computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to execute the computer program to perform the method of any one of claims 1-6.

Images