A deep knowledge tracing method and system based on cognitive theory

Abstract

The application discloses a kind of fusion cognitive theory's deep knowledge tracking method and system, the method includes the following steps: obtaining answer record set, obtains recent attempt time interval and past attempt times, obtains forgetting behavior embedding by embedding layer and is spliced, obtains problem embedding vector by splicing problem index embedding, knowledge concept index embedding and forgetting behavior embedding, obtains problem-response embedding vector by splicing problem embedding vector and corresponding answer response, generates problem embedding vector, problem-response embedding vector and knowledge mastery state based on question encoder, knowledge encoder and knowledge retriever respectively, generates student ability value, question discrimination, exercise difficulty based on embedding layer respectively, and calculates answer guess rate, constructs question response theory model based on deep knowledge tracking and carries out iterative training, and makes answer performance prediction based on optimized deep knowledge tracking model.The application can make accurate evaluation to the knowledge mastery state of student.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of knowledge tracing, and particularly relates to a deep knowledge tracing method and system fusing cognitive theory. BACKGROUND

[0002] Knowledge tracing is a key link to realize personalized teaching, which evaluates the cognitive level of learners and tracks the changes of the cognitive level according to the learning behaviors of learners, and further predicts the learning performance. At present, the mainstream knowledge tracing methods mainly include two categories: knowledge tracing based on long short-term memory network and knowledge tracing based on attention mechanism. However, due to the lack of guidance of domain knowledge, the output results have volatility, and lack of interpretability.

[0003] In the direction of knowledge tracing based on long short-term memory network, some researchers apply deep learning to the knowledge tracing task, and use recurrent neural network or long short-term memory recurrent network modeling, which obtains good results. Some researchers make a jump connection between the correlation of hidden units in long short-term memory recurrent network to expand the sequence learning capacity of long short-term memory recurrent network. However, these existing knowledge tracing models based on recurrent neural network or long short-term memory recurrent network can only extend the length of the learning sequence of recurrent or long short-term memory recurrent network to a certain extent, and cannot fundamentally solve the problem, still have the problem of long-term dependence, and cannot really use long sequence input data.

[0004] In the direction of knowledge tracing based on attention, some researchers use self-attention mechanism instead of recurrent neural network to build model framework, which solves the problem of long-term dependence because it does not depend on the framework of recurrent neural network. Some researchers use encoder and decoder to embed exercises and answers respectively, and obtain better prediction performance through multiple attention layers, and solve the problems of too shallow attention layer and lack of experience in embedding calculation. Some researchers add two time attributes in the decoder; or use the context information of answering questions, or construct exercise embedding and interaction embedding, and use monotonic attention to model the forgetting behavior of students, which has a certain effect. However, these existing knowledge tracing models based on attention mechanism, although they get rid of the framework of recurrent neural network or long short-term memory recurrent network, and do not have the problem of long-term dependence, but also lose the ability of recurrent neural network to model sequence information, and the influence of position embedding on sequence information needs further research. SUMMARY

[0005] In order to overcome the defects and deficiencies existing in the prior art, the present application provides a deep knowledge tracking method and system fused with cognitive theory, the present application obtains a student's answer history data set, obtains student answer records from the data set, including exercise index embedding, knowledge concept index embedding and answer response information, obtains question embedding vectors with forgetting information and question-response embedding vectors by using embedding operation; the question embedding vectors and the question-response embedding vectors are encoded to obtain context-aware question embedding vectors and question-response embedding vectors; the context-aware question embedding vectors and the question-response embedding vectors are input into a knowledge retriever to obtain a student's knowledge mastery state for the current question; student ability value, question discrimination, exercise difficulty and answer guessing rate are obtained by embedding operation; a question response theory model based on deep knowledge tracking is constructed to predict the student's subsequent answer performance; the deep knowledge tracking model is fed back by gradient based on the objective function, the network parameters are updated, and the optimized deep knowledge tracking model is obtained by iterative training, and the probability of the student's correct answer to the next exercise is predicted according to the model, the present application combines deep knowledge tracking and cognitive theory, accurately evaluates the student's knowledge mastery state according to the student's answer history, can better avoid the student's blind exercise, and further improves the level of individualized teaching and learning.

[0006] In order to achieve the above purpose, the present application adopts the following technical scheme:

[0007] The present application provides a deep knowledge tracking method fused with cognitive theory, comprising the following steps:

[0008] Obtain a student's answer record set {(q1, c1, r1),..., (qt, ct, rt)} at continuous t time points, wherein q represents an exercise index, c represents a knowledge concept index, and r represents an answer response; t t t t t t

[0009] Obtain the latest attempt time interval and the past attempt times according to the knowledge concept index sequence, the latest attempt point time interval is the time interval between the current time knowledge concept index and the last time answer same knowledge concept index, and the past attempt times are the number of times that the current time knowledge concept index appears in the knowledge concept index sequence;

[0010] The latest attempt time interval is input into an embedding layer G1 to obtain an embedding representation of the latest attempt time interval, the past attempt times are input into an embedding layer G2 to obtain an embedding representation of the past attempt times, and the embedding representation of the latest attempt time interval and the embedding representation of the past attempt times are spliced to obtain a forgetting behavior embedding;

[0011] ​​​​​​The question index q t Input into embedding layer E1 to obtain a question index embedding, the knowledge concept index c t Input into embedding layer E2 to obtain a knowledge concept index embedding, the answer response r t Input into embedding layer E3 to obtain an answer response embedding;

[0012] Splice the question index embedding, the knowledge concept index embedding, and the forgetting behavior embedding to obtain a question embedding vector x t , and construct a question embedding vector sequence (x1, …, xt) t );

[0013] Splice the question embedding vector x i and the corresponding answer response r i to obtain a question-response embedding vector y i , and construct a question-response embedding vector sequence (y1, …, yt) t-1 , wherein 1≤i≤t-1;

[0014] Input the question embedding vector sequence (x1, …, xt) t into a question encoder to generate context-aware question embedding vectors

[0015] Input the question-response embedding vector sequence (y1, …, yt) t-1 into a knowledge encoder to generate context-aware question-response embedding vectors

[0016] Input the question embedding vector sequence (x1, …, xt) t and the question-response embedding vector sequence (y1, …, yt) t-1 into a knowledge retriever to output a student's knowledge mastery state h t ;

[0017] Splice the student's knowledge mastery state h t and the question embedding vector x t , and input into embedding layer B1 to generate a student ability value θ t ;

[0018] Input the question embedding vector x t into embedding layer B2 to generate a question discrimination degree a t ;

[0019] Input the question embedding vector x t into embedding layer B3 to generate a question difficulty b t ;

[0020] Calculate the answer guessing rate g according to the student ability value and the student's answer record sett ;

[0021] According to the student ability value θ t , the question discrimination a t , the exercise difficulty b t and the answer guess rate g t , a question response theory model based on deep knowledge tracking is constructed to predict the probability of the student correctly answering the current question

[0022] The loss function value is calculated according to the probability of the student correctly answering the question predicted by the question response theory model and the real answer response of the student, and the model is updated and adjusted in reverse to obtain an optimized deep knowledge tracking model through iterative training;

[0023] The answer record set of the student at the t continuous time is input into the optimized deep knowledge tracking model for answer performance prediction, and the answer response r t at the t time is output.

[0024] As a preferred technical solution, the question embedding vector sequence (x1, …, x t ) is input into the question encoder to generate a context-aware question embedding vector Specifically represented as:

[0025]

[0026] Wherein, W a1 , W a2 and W a3 are weight matrices of full connection layers, represents the transpose of W a1 , represents the transpose of x t .

[0027] As a preferred technical solution, the question-response embedding vector sequence (y1, …, y t-1 ) is input into the knowledge encoder to generate a context-aware question-response embedding vector Specifically represented as:

[0028]

[0029] Wherein, W b1 , W b2 and W b3 are weight matrices of full connection layers, represents the transpose of W b1 , represents the transpose of y t-1 .

[0030] As a preferred technical solution, the question embedding vector sequence (x1, …, x t ) and the question-response embedding vector sequence (y1, …, y t-1 ) are input into the knowledge retriever, and the knowledge mastery state h t of the student is output, which is specifically represented as:

[0031]

[0032] Wherein, W c1 , W c2 and W c3 are weight matrices of full connection layers, represents the transpose of W c1 , represents the transpose of x t .

[0033] As a preferred technical solution, the answer guessing rate g t is calculated according to the student ability value and the student's answer record set, which is specifically represented as:

[0034]

[0035] Wherein, t-1 represents t-1 moment.

[0036] As a preferred technical solution, the probability of the student correctly answering the current question is predicted , which is specifically represented as:

[0037]

[0038] Wherein, d is a constant.

[0039] As a preferred technical solution, the loss function value is calculated according to the probability of the student correctly answering the question predicted by the item response theory model and the real answer response of the student, and the loss function is represented as:

[0040]

[0041] Wherein, represents the real answer response of the i-th student at t moment, represents the probability of the i-th student correctly answering the current question at t moment predicted by the model, and l represents the loss function.

[0042] This invention also provides a deep knowledge tracking system that integrates cognitive theory, including: a question record set acquisition module, a recent attempt time interval calculation module, a past attempt count calculation module, an embedding module, a splicing module, a sequence construction module, a question encoder, a knowledge encoder, a knowledge retrieval device, a question guessing rate calculation module, a question response theory model construction module, a loss function value calculation module, an iterative training module, and a prediction module;

[0043] The answer record set acquisition module is used to acquire the student's answer record set {(q1, c1, r1), ..., (q...} at t consecutive time points. t c t r t )}, where q t c represents the exercise index. t Represents a knowledge concept index, r t Indicates a response to the question;

[0044] The recent attempt time interval calculation module is used to calculate the recent attempt time interval based on the knowledge concept index sequence. The recent attempt time interval is the time interval between the knowledge concept index at the current moment and the same knowledge concept index in the most recent answer.

[0045] The past attempt count calculation module is used to calculate the past attempt count based on the knowledge concept index sequence. The past attempt count is the number of times the knowledge concept index has appeared in the knowledge concept index sequence at the current moment.

[0046] The embedding module is used to input the most recent attempt time interval into the embedding layer G1 to obtain the embedded representation of the most recent attempt time interval, input the past attempt count into the embedding layer G2 to obtain the embedded representation of the past attempt count, and input the exercise index q t The input is fed into the embedding layer E1 to obtain the exercise index embedding, and the knowledge concept index c is then used. t The input is fed into the embedding layer E2 to obtain the knowledge concept index embedding, and the answer response r is then used. t The input is fed into the embedding layer E3 to obtain the answer response embedding, which incorporates the student's knowledge mastery status h. t and the problem embedding vector x t The input is concatenated into the embedding layer B1 to generate the student's ability value θ. t Embed the problem into vector x t The input is fed into the embedding layer B2 to generate the question discrimination score a. t Embed the problem into vector x t Input into the embedding layer B3 generates exercise difficulty b t ;

[0047] The concatenation module is used to concatenate the embedding representation of the most recent attempt time interval and the embedding representation of the past attempt count to obtain the forgetting behavior embedding, and to concatenate the exercise index embedding, the knowledge concept index embedding, and the forgetting behavior embedding to obtain the question embedding vector x. t Embed the problem into vector x i With the corresponding answer response r i The concatenation yields the problem-response embedding vector y. i ;

[0048] The sequence construction module is used to construct based on the question embedding vector x. t Construct the problem embedding vector sequence (x1, ..., x) t Based on the question-response embedding vector y i Constructing the problem-response embedding vector sequence (y1, ..., y2) yields the following sequence: t-1 ), where 1≤i≤t-1;

[0049] The question encoder is used to embed questions into a vector sequence (x1, ..., x). t Generate context-aware question embedding vectors

[0050] The knowledge encoder is used to embed the question-response vector sequence (y1, ..., y2) into a question-response embedding vector sequence (y1, ..., y3). t-1 Generating context-aware problem-response embedding vectors

[0051] The knowledge retrieval unit is used to retrieve information based on the question embedding vector sequence (x1, ..., x). t ) and the question-response embedding vector sequence (y1, ..., y t-1 Generate students' knowledge mastery status h t ;

[0052] The answer guessing rate calculation module is used to calculate the answer guessing rate g based on the student's ability value and the student's answer record set. t ;

[0053] The question response theory model construction module is used to construct the student's ability value θ. t Question discrimination a t Exercise difficulty b t and the guessing rate g t Construct a question response theory model based on deep knowledge tracing to predict the probability of a student answering the current question correctly.

[0054] The loss function value calculation module is used to calculate the loss function value based on the probability of the student answering the question correctly predicted by the question response theory model and the student's actual answer response.

[0055] The iterative training module is configured to update the adjustment model in reverse using the loss function value, and the iterative training obtains an optimized deep knowledge tracking model.

[0056] The prediction module is configured to input a set of student answering records at consecutive t time points into the optimized deep knowledge tracking model to predict the answering performance, and output a response r corresponding to the response at the t time point. t .

[0057] As a preferred technical solution, the prediction module is configured to predict the probability of the student correctly answering the current question. Specifically, the prediction module is configured to predict the probability of the student correctly answering the current question according to the following formula:

[0058]

[0059] wherein d is a constant.

[0060] As a preferred technical solution, the loss function value calculation module is configured to calculate the loss function value according to the probability of the student correctly answering the question predicted by the question response theoretical model and the real answering response of the student, and specifically, the loss function value calculation module is configured to calculate the loss function value according to the following formula:

[0061]

[0062] wherein, represents the real answering response of the i-th student at the t time point, represents the probability of the i-th student correctly answering the current question at the t time point predicted by the model, and l represents the loss function.

[0063] Compared with the prior art, the present application has the following advantages and beneficial effects:

[0064] (1) The present application adopts a question response theoretical model based on deep knowledge tracking to predict the answering performance of the student, solves the limitation of the traditional question response theoretical model that uses shallow features of the student answering data for prediction, achieves the effect of improving the prediction accuracy and interpretability of the knowledge tracking system, and meets the demand of the intelligent education field for personalized teaching.

[0065] (2) The present application adopts a technical solution of integrating a forgetting mechanism into student behavior modeling, solves the defect that the user features used by the traditional student behavior model are relatively single, and achieves the technical effect of more accurately evaluating and predicting the answering performance of the student. BRIEF DESCRIPTION OF DRAWINGS

[0066] Figure 1 FIG. 1 is a flowchart of a deep knowledge tracking method according to the present application. DETAILED DESCRIPTION

[0067] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0068] Example 1

[0069] The knowledge tracking task accurately assesses students' knowledge mastery based on their answer history. The knowledge tracking task is described as follows:

[0070] set up Given a dataset containing the answer information of n students, where each student's answer information... It consists of answer records at mi time points. For student i's answer record at time t, a triplet is used. It means that among them It is an index of exercises. Represents a knowledge concept index. This indicates the response to the question, that is, the performance in answering the question. For example, This indicates that student i correctly answered the question about the knowledge concept at time t. Exercises

[0071] In this embodiment, for simplicity, the superscript i is omitted. The input is the student's past historical answer record information {(q1, c1, r1), ..., (q t-1 c t-1 r t-1 The output is a prediction of the student's understanding of knowledge concept c at time t. t Question q t response r t Therefore, the knowledge tracing task can be interpreted as predicting the student's answer to question q at the next moment, i.e., time t, based on the student's past answer records at the deadline t-1. t The response performance, namely: r t The probability that = 1 is:

[0072] P(r t =1|(q1,c1,r1),…,(q t-1 c t-1 r t-1 ), (q t c t )).

[0073] like Figure 1As shown, this embodiment provides a deep knowledge tracking method that integrates cognitive theory. Based on students' answer history, it accurately assesses students' knowledge mastery and predicts their subsequent answer performance. The method includes the following steps:

[0074] S1: Obtain the student's answer records at consecutive time points t {(q1, c1, r1), ..., (q t c t r t )};

[0075] S2: Obtain the knowledge concept index sequence (c1, ..., c) from the above answer record set. t The most recent attempt time interval and the number of past attempts are obtained based on the knowledge concept index sequence.

[0076] Here, the most recent attempt time interval is the time interval between the current knowledge concept index and the most recent answer with the same knowledge concept index, and the past attempt count is the number of times the current knowledge concept index has appeared in the knowledge concept index sequence (including the current time). Then, the most recent attempt time interval is input into the embedding layer G1 to obtain the embedding representation of the most recent attempt time interval. Similarly, the past attempt count is input into the embedding layer G2 to obtain the embedding representation of the past attempt count. The embedding representations of the most recent attempt time interval and the past attempt time interval are concatenated to obtain the forgetting behavior embedding f. t .

[0077] In this embodiment, both the embedding layer G1 and the embedding layer G2 are implemented using fully connected layer networks. The weight coefficients of the embedding layers G1 and G2 are learnable and can be optimized through training iterations.

[0078] S3: Index the exercises q t The input is fed into the embedding layer E1 to obtain the exercise index embedding, and the knowledge concept index c is then used. t The input is fed into the embedding layer E2 to obtain the knowledge concept index embedding, and the answer response r is then used. t The input is fed into the embedding layer E3 to obtain the answer response embedding. The question index embedding, knowledge concept index embedding, and forgetting behavior embedding are then incorporated into the embedding. t The three are concatenated to obtain the problem embedding vector x. t Similarly, the problem embedding vector sequence (x1, ..., x) can be obtained. t Embed the problem into vector x. i (1≤i≤t-1) and the response r i (1≤i≤t-1) concatenation yields the problem-response embedding vector y. i (1≤i≤t-1), thus obtaining the problem-response embedding vector sequence (y1, ..., y t-1); since the above-mentioned problem embedding vectors and problem-response embedding vectors contain forgetting behavior information, the behavior characteristics of students can be better modeled.

[0079] In the embodiment, the embedding layer E1, the embedding layer E2 and the embedding layer E3 are all implemented by using a fully connected layer network, and the weight coefficients of the embedding layer E1, the embedding layer E2 and the embedding layer E3 are learnable and are optimized through training iteration.

[0080] S4: input the problem embedding vector sequence (x1, …, x t to the question encoder to generate context-aware problem embedding vectors Specifically, it is represented as:

[0081]

[0082] wherein, W a1 , W a2 and W a3 are weight matrices of the fully connected layer, represents the transpose of W a1 , represents the transpose of x t ;

[0083] input the problem-response embedding vector sequence (y1, …, y t-1 to the knowledge encoder to generate context-aware problem-response embedding vectors Specifically, it is represented as:

[0084]

[0085] wherein, W b1 , W b2 and W b3 are weight matrices of the fully connected layer, represents the transpose of W b1 , represents the transpose of y t-1 ;

[0086] input the problem embedding vector sequence (x1, …, x t ) and the problem-response embedding vector sequence (y1, …, y t-1 ) to the knowledge retriever to output the knowledge mastery state h t of the student, which is specifically represented as:

[0087]

[0088] wherein, W c1 , W c2 and W c3 are weight matrices of the fully connected layer, denotes the transpose of W c1 ;

[0089] The above is one implementation of the problem encoder, the knowledge encoder and the knowledge retriever using the attention network, and of course, other attention networks can also be implemented.

[0090] S5: Concatenate the student's knowledge mastery state h t and the problem embedding vector x t , input into the embedding layer B1 to generate the student ability value θ t ;

[0091] Input the problem embedding vector x t into the embedding layer B2 to generate the question discrimination degree a t ;

[0092] Input the problem embedding vector x t into the embedding layer B3 to generate the exercise difficulty b t ;

[0093] Calculate the answer guess rate g t according to the student ability value and the student's answer record set, and the answer guess rate g t is the ratio of the number of correct answers to the total number of historical answers of the student without mastering the knowledge point, and the calculation formula is as follows:

[0094]

[0095] In the embodiment, the embedding layer B1, the embedding layer B2 and the embedding layer B3 are all implemented by using a fully connected layer network, and the weight coefficients of the embedding layer B1, the embedding layer B2 and the embedding layer B3 are learnable and optimized through training iteration.

[0096] S6: According to the student ability value θ t , the question discrimination degree a t , the exercise difficulty b t and the answer guess rate g t obtained in step S5, a question response theory model based on deep knowledge tracking is constructed to predict the subsequent answer performance of the student;

[0097] In the embodiment, the question response theory model based on deep knowledge tracking predicts the probability of the student correctly answering the current question , which is specifically expressed as:

[0098]

[0099] Wherein, d is a constant, and the embodiment is preferably 1.702;

[0100] S7: Calculate the loss function value according to the probability of the student correctly answering the question predicted by the item response theory model and the student's real answer response, and use the loss function value to update and adjust the model in reverse, if the training process does not converge, execute steps S1 to S7 cyclically until the training is completed, and obtain the optimized deep knowledge tracking model;

[0101] In this embodiment, the calculation formula of the loss function is as follows:

[0102]

[0103] Wherein, represents the real answer response of the ith student at time t, represents the probability of the model predicting the correct answer of the ith student at time t.

[0104] In this embodiment, the loss function value is used to guide the update of the model parameters, and the model is optimized by gradient descent method, and the updated model parameters include: embedding layer G1 and embedding layer G2 in step S2, embedding layer E1, embedding layer E2 and embedding layer E3 in step S3, question encoder (weight matrix W a1 , W a2 and W a3 ), knowledge encoder (weight matrix W b1 , W b2 and W b3 ) and knowledge retriever (weight matrix W c1 , W c2 and W c3 ) in step S4, and embedding layer (B1, B2 and B3) in step S5.

[0105] S8: Input a student's continuous t-time answer record set (the student does not appear in the above training process, and the answer response at time t is not available), predict the student's answer performance through the above trained deep knowledge tracking model, and output the response r t .

[0106] Embodiment 2

[0107] The embodiment provides a deep knowledge tracking system integrating cognitive theory, which is used to realize the deep knowledge tracking method integrating cognitive theory in the above embodiment 1. The system comprises: an answer record set acquisition module, a recent attempt time interval solving module, a past attempt times solving module, an embedding module, a splicing module, a sequence construction module, a question encoder, a knowledge encoder, a knowledge retriever, an answer guessing rate calculation module, an item response theory model construction module, a loss function value calculation module, an iterative training module and a prediction module.

[0108] In the embodiment, the answer record set obtaining module is configured to obtain a set of answer records {(q1, c1, r1),..., (qt, ct, rt)} of the student at consecutive t time points, where q represents an exercise index, c represents a knowledge concept index, and r represents an answer response. t t t t t t

[0109] In the embodiment, the recent attempt time interval solving module is configured to obtain a recent attempt time interval according to the sequence of knowledge concept indexes, the recent attempt point time interval being a time interval between the knowledge concept index at the current time point and the same knowledge concept index of the last answer.

[0110] In the embodiment, the past attempt times solving module is configured to obtain a past attempt times according to the sequence of knowledge concept indexes, the past attempt times being a number of times that the knowledge concept index at the current time point has appeared in the sequence of knowledge concept indexes.

[0111] In the embodiment, the embedding module is configured to input the recent attempt time interval into an embedding layer G1 to obtain an embedding representation of the recent attempt time interval, input the past attempt times into an embedding layer G2 to obtain an embedding representation of the past attempt times, input the exercise index q into an embedding layer E1 to obtain an exercise index embedding, input the knowledge concept index ct into an embedding layer E2 to obtain a knowledge concept index embedding, input the answer response r into an embedding layer E3 to obtain an answer response embedding, concatenate the knowledge mastery state h of the student and the problem embedding vector x to input into an embedding layer B1 to generate a student ability value θ, input the problem embedding vector x into an embedding layer B2 to generate a question discriminability a, and input the problem embedding vector x into an embedding layer B3 to generate an exercise difficulty b. t t t t t t t t t

[0112] In the embodiment, the concatenation module is configured to concatenate the embedding representation of the recent attempt time interval and the embedding representation of the past attempt times to obtain a forgetting behavior embedding, concatenate the exercise index embedding, the knowledge concept index embedding, and the forgetting behavior embedding to obtain a problem embedding vector x, concatenate the problem embedding vector x and the corresponding answer response r to obtain a problem-response embedding vector y. t i i i

[0113] In the embodiment, the sequence construction module is configured to construct a sequence based on the problem embedding vector x​​​​​​​​​​​​​​​​​​​t A sequence of question embedding vectors (x1, …, xt) is constructed t A question-response embedding vector y is constructed based on the sequence of question embedding vectors (x1, …, xt) and the sequence of response embedding vectors (r1, …, rt) i A sequence of question-response embedding vectors (y1, …, yt-1) is constructed t-1 wherein 1≤i≤t-1

[0114] In this embodiment, the question encoder is configured to generate a context-aware question embedding vector t based on the sequence of question embedding vectors (x1, …, xt) Specifically, the question encoder is configured to generate the context-aware question embedding vector

[0115]

[0116] wherein W a1 , W a2 and W a3 are weight matrices of full connection layers, represents the transpose of W a1 , represents the transpose of x t .

[0117] In this embodiment, the knowledge encoder is configured to generate a context-aware question-response embedding vector t-1 based on the sequence of question-response embedding vectors (y1, …, yt-1) Specifically, the knowledge encoder is configured to generate the context-aware question-response embedding vector

[0118]

[0119] wherein W b1 , W b2 and W b3 are weight matrices of full connection layers, represents the transpose of W b1 , represents the transpose of y t-1 .

[0120] In this embodiment, the knowledge retriever is configured to generate a knowledge mastery state h t of a student based on the sequence of question embedding vectors (x1, …, xt) and the sequence of question-response embedding vectors (y1, …, yt-1) t-1 Specifically, the knowledge retriever is configured to generate the knowledge mastery state h t based on the context-aware question embedding vector

[0121]

[0122] wherein W c1 , W c2 and W c3 are weight matrices of full connection layers, represents the transpose of W c1transpose of represents x t transpose of

[0123] In this embodiment, the answer guessing rate calculation module is configured to calculate the answer guessing rate g t , and specifically expressed as:

[0124]

[0125] wherein t-1 represents t-1 moment;

[0126] In this embodiment, the item response theory model construction module is configured to construct an item response theory model based on deep knowledge tracking according to the student ability value θ t , the item discrimination a t , the exercise difficulty b t and the answer guessing rate g t , and predict the probability of the student correctly answering the current question , and specifically expressed as:

[0127]

[0128] wherein d is a constant;

[0129] In this embodiment, the loss function value calculation module is configured to calculate the loss function value according to the probability of the student correctly answering the question predicted by the item response theory model and the real answer of the student, and specifically expressed as:

[0130]

[0131] wherein represents the real answer of the i-th student at t moment, represents the probability of the i-th student correctly answering the current question at t moment predicted by the model, and l represents the loss function;

[0132] In this embodiment, the iterative training module is configured to update and adjust the model in reverse by using the loss function value, and iteratively train to obtain an optimized deep knowledge tracking model;

[0133] In this embodiment, the prediction module is configured to input the answer record set of the student at the continuous t moments into the optimized deep knowledge tracking model to predict the answer performance, and output the response r t at the t-th moment.

[0134] The above embodiments are the preferred embodiments of the present application, but the embodiments of the present application are not limited to the above embodiments, and any changes, modifications, substitutions, combinations, simplifications, etc. made without departing from the spirit and principles of the present application should be equivalent replacement manners and should be included in the protection scope of the present application.

Claims

1. A deep knowledge tracking method integrating cognitive theory, characterized in that, Includes the following steps: Obtain the set of students' answer records at consecutive time points t. ,in, This represents the exercise index. Represents a knowledge concept index. Indicates a response to the question; The most recent attempt time interval and the number of past attempts are obtained from the knowledge concept index sequence. The most recent attempt time interval is the time interval between the knowledge concept index at the current moment and the most recent answer with the same knowledge concept index. The number of past attempts is the number of times the knowledge concept index at the current moment has appeared in the knowledge concept index sequence. The most recent attempt time interval is input into the embedding layer. The embedded representation of the most recent attempt time interval is obtained, and the number of past attempts is input into the embedding layer. Obtain the embedding representation of the number of past attempts, and concatenate the embedding representation of the most recent attempt time interval with the embedding representation of the number of past attempts to obtain the forgetting behavior embedding; The exercise index Input to the embedding layer The exercise index embedding is obtained, and the knowledge concept index is obtained. Input to the embedding layer The knowledge concept index embedding is obtained, and the answer response is... Input to the embedding layer Get the embedded answer response; The question embedding vector is obtained by concatenating the question index embedding, the knowledge concept index embedding, and the forgetting behavior embedding. The problem embedding vector sequence is constructed. ; Embedding the problem into a vector and corresponding answer response Concatenating the results yields the problem-response embedding vector. The question-response embedding vector sequence is constructed. ,in, ; Embed the problem into a vector sequence The input is fed into the question encoder to generate a context-aware question embedding vector. ; Embedding the question-response vector sequence The input is fed into a knowledge encoder to generate a context-aware question-response embedding vector. ; Embed the problem into a vector sequence Problem-response embedding vector sequence The information is input into the knowledge retrieval system, which then outputs the student's knowledge mastery status. ; Assess students' knowledge mastery and question embedding vector The data is then concatenated and input into the embedding layer. Generate student ability values ; Embedding the problem into a vector Input to the embedding layer Generate question discrimination ; Embedding the problem into a vector Input to the embedding layer Generating Exercise Difficulty ; The guessing rate is calculated based on students' ability scores and their answer records. ; Based on student ability value Question discrimination Difficulty of the exercises and guessing rate Construct a question response theory model based on deep knowledge tracing to predict the probability of a student answering the current question correctly. ; The loss function value is calculated based on the probability of students answering questions correctly predicted by the question response theory model and the students' actual answer responses. The loss function value is then used to update and adjust the model in reverse, and the optimized deep knowledge tracking model is obtained through iterative training. The optimized deep knowledge tracing model is input with the student's answer records over t consecutive time periods to predict answer performance, and the corresponding answer response at time t is output. .

2. The deep knowledge tracking method based on fusion cognitive theory according to claim 1, characterized in that, The embedding of the problem into a vector sequence The input is fed into the question encoder to generate a context-aware question embedding vector. Specifically, it is expressed as: ； in, , and These are all weight matrices for fully connected layers. express transpose, express The transpose of .

3. The deep knowledge tracking method based on fusion cognitive theory according to claim 1, characterized in that, The embedding of the question-response vector sequence The input is fed into a knowledge encoder to generate a context-aware question-response embedding vector. Specifically, it is expressed as: ； in, , and These are all weight matrices for fully connected layers. express transpose, express .

4. The deep knowledge tracking method based on fusion cognitive theory according to claim 1, characterized in that, The embedding of the problem into a vector sequence Problem-response embedding vector sequence The information is input into the knowledge retrieval system, which then outputs the student's knowledge mastery status. Specifically, it is expressed as: ； in, , and These are all weight matrices for fully connected layers. express transpose, express The transpose of .

5. The deep knowledge tracking method based on fusion cognitive theory according to claim 1, characterized in that, The guessing rate is calculated based on students' ability scores and their answer records. Specifically, it is expressed as: ； in, express time.

6. The deep knowledge tracking method based on fusion cognitive theory according to claim 1, characterized in that, The probability of predicting the student's correct answer to the current question. Specifically, it is expressed as: ； in, It is a constant.

7. The deep knowledge tracking method based on fusion cognitive theory according to claim 1, characterized in that, The loss function is calculated based on the probability of a student answering a question correctly predicted by the question-response theory model and the student's actual response. The loss function is expressed as follows: ； in, This represents the actual response of the k-th student at time t. This represents the probability predicted by the model that the k-th student will correctly answer the current question at time t. This represents the loss function.

8. A deep knowledge tracking system integrating cognitive theory, characterized in that, include: The module includes modules for obtaining answer records, calculating the most recent attempt time interval, calculating the number of past attempts, embedding, splicing, constructing sequences, question encoders, knowledge encoders, knowledge retrieval, calculating the answer guessing rate, constructing a question response theoretical model, calculating the loss function value, iterative training, and prediction. The answer record acquisition module is used to acquire the student's answer record set at t consecutive time points. ,in, This represents the exercise index. Represents a knowledge concept index. Indicates a response to the question; The recent attempt time interval calculation module is used to calculate the recent attempt time interval based on the knowledge concept index sequence. The recent attempt time interval is the time interval between the knowledge concept index at the current moment and the same knowledge concept index in the most recent answer. The past attempt count calculation module is used to calculate the past attempt count based on the knowledge concept index sequence. The past attempt count is the number of times the knowledge concept index has appeared in the knowledge concept index sequence at the current moment. The embedding module is used to input the most recent attempt time interval into the embedding layer. Obtain the embedded representation of the most recent attempt time interval, and input the past attempt count into the embedding layer. Obtain the embedding representation of the number of past attempts, and index the exercises. Input to the embedding layer Obtain the exercise index embedding and index the knowledge concepts. Input to the embedding layer Obtain the knowledge concept index embedding and then respond with the answer. Input to the embedding layer By embedding the answer response, the student's knowledge mastery status can be obtained. and question embedding vector Concatenate input to the embedding layer Generate student ability values Embedding the problem into a vector Input to the embedding layer Generate question discrimination Embedding the problem into a vector Input to the embedding layer Generating Exercise Difficulty ; The concatenation module is used to concatenate the embedding representation of the most recent attempt time interval and the embedding representation of the past attempt count to obtain the forgetting behavior embedding, and to concatenate the exercise index embedding, knowledge concept index embedding, and forgetting behavior embedding to obtain the question embedding vector. Embedding the problem into a vector and corresponding answer response Concatenating the results yields the problem-response embedding vector. ; The sequence construction module is used to construct based on the question embedding vector. Construct the problem embedding vector sequence Based on question-response embedding vectors Constructing the question-response embedding vector sequence ,in, ; The question encoder is used to embed questions into a vector sequence. Generate context-aware question embedding vectors ; The knowledge encoder is used to embed question-response vector sequences. Generating context-aware problem-response embedding vectors ; The knowledge retrieval device is used to retrieve information based on a question embedding vector sequence. Problem-response embedding vector sequence Generate students' knowledge mastery status ; The answer guessing rate calculation module is used to calculate the answer guessing rate based on the student's ability value and the student's answer record set. ; The question response theory model construction module is used to construct a model based on students' ability values. Question discrimination Difficulty of the exercises and guessing rate Construct a question response theory model based on deep knowledge tracing to predict the probability of a student answering the current question correctly. ; The loss function value calculation module is used to calculate the loss function value based on the probability of the student answering the question correctly predicted by the question response theory model and the student's actual answer response. The iterative training module is used to back-update and adjust the model using the loss function value, and iterative training yields an optimized deep knowledge tracking model. The prediction module is used to input the student's answer records for t consecutive time periods into the optimized deep knowledge tracing model to predict the student's answer performance and output the corresponding response at time t. .

9. The deep knowledge tracking system based on fusion cognitive theory according to claim 8, characterized in that, The probability of predicting the student's correct answer to the current question. Specifically, it is expressed as: ； in, It is a constant.

10. The deep knowledge tracking system based on fusion cognitive theory according to claim 8, characterized in that, The loss function value calculation module is used to calculate the loss function value based on the probability of a student correctly answering a question predicted by the question response theory model and the student's actual answer response. Specifically, it is expressed as follows: ； in, This represents the actual response of the k-th student at time t. This represents the probability predicted by the model that the k-th student will correctly answer the current question at time t. This represents the loss function.

Images