Intelligent learning knowledge ability tracking method, electronic device and storage medium

Abstract

The present invention provides a method for tracking knowledge and ability of intelligent learning, an electronic device and a storage medium, the method comprising: inputting a user identification to be tested, a current question, and a priori correlation weight between the current question and each knowledge point into a knowledge and ability tracking model, and obtaining the probability of the user to be tested answering the current question correctly; the knowledge and ability tracking model includes the mastery status of the user to be tested for each knowledge point. The present invention introduces the priori relationship between the question and each knowledge point, and combines the relationship between the network learning question and each knowledge point, thereby establishing an explicit mapping correspondence between the current question and each knowledge point. Therefore, it is not only possible to accurately predict the probability of answering the current question correctly, but also to update the knowledge point mastery status matrix according to the answer result of the current question, so that the mastery status of the user to be tested for each knowledge point can be known.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to an intelligent learning knowledge ability tracking method, electronic equipment and storage medium. Background Art

[0002] The knowledge and ability tracking model is based on the student behavior sequence. Tracking the student's mastery of knowledge is the core and key to building an adaptive teaching system. In an adaptive teaching system, whether it is accurate question push or student learning path planning, the first step is to accurately estimate the student's mastery of knowledge points.

[0003] However, deep learning models such as recurrent neural networks that directly model student behavior time series data are currently unable to obtain students' mastery of knowledge points. For example, when the model inputs the answer to a question, the model can predict the probability of answering all questions (done and not done) correctly in the next step, but because the model represents the student's current mastery of all knowledge points with a hidden state, the model cannot output the student's mastery of specific knowledge points. For another example, when the model inputs the answer to the knowledge point (one or more) corresponding to the question, it can predict the probability of answering all knowledge points (done and not done) correctly in the next step, but this is not the student's mastery of the knowledge points, that is, it still cannot output the student's mastery of specific knowledge points. Summary of the invention

[0004] In view of the problems in the prior art, an embodiment of the present invention provides an intelligent learning knowledge ability tracking method, an electronic device and a storage medium.

[0005] Specifically, the embodiment of the present invention provides the following technical solutions:

[0006] In a first aspect, an embodiment of the present invention provides a method for tracking knowledge capabilities of intelligent learning, including:

[0007] Obtain the current question corresponding to the user to be tested, and the prior correlation weights between the current question and each knowledge point;

[0008] Input the identification of the user to be tested, the current question, and the prior correlation weights between the current question and each knowledge point into the knowledge ability tracking model to obtain the probability of the user to be tested answering the current question correctly; wherein the knowledge ability tracking model includes the mastery status of each knowledge point by the user to be tested, and the mastery status of each knowledge point by the user to be tested is determined based on the questions that the user to be tested has done in a historical time period, the answer results of the questions that the user to be tested has done, and the prior correlation weights between the questions that the user to be tested has done and each knowledge point;

[0009] Among them, the knowledge and ability tracking model is obtained by inputting the questions done by each user in the historical training data, the answer results of the questions done by each user, and the prior correlation weights between the questions done by each user and each knowledge point into the initial knowledge and ability tracking model, and training based on machine learning.

[0010] Furthermore, the intelligent learning knowledge ability tracking method further includes:

[0011] The answer result of the user to be tested for the current question is obtained, and the mastery status of each knowledge point of the user to be tested is updated according to the answer result.

[0012] Furthermore, the knowledge and ability tracking model includes a topic knowledge point mapping model;

[0013] Accordingly, the identification of the user to be tested, the current question, and the prior correlation weights between the current question and each knowledge point are input into the knowledge ability tracking model to obtain the probability of the user to be tested answering the current question correctly, specifically including:

[0014] Inputting the topic feature data of the current topic into the topic knowledge point mapping model to obtain the first correlation weight between the current topic and each knowledge point;

[0015] Determine the priori correlation weights between the current question and each knowledge point according to the knowledge point information marked on the current question;

[0016] Determine the final correlation weight between the current question and each knowledge point according to the first correlation weight between the current question and each knowledge point, and the prior correlation weight between the current question and each knowledge point, and determine the mapping correspondence between the current question and each knowledge point according to the final correlation weight between the current question and each knowledge point;

[0017] Obtaining the mastery status of each knowledge point of the user to be tested from the knowledge point mastery status matrix corresponding to the user to be tested; wherein the knowledge point mastery status matrix stores the mastery status of each knowledge point of the user to be tested; wherein the knowledge point mastery status is used to characterize the mastery degree of the knowledge point of the user to be tested;

[0018] Determine the knowledge point status characteristic data of the user to be tested for all the knowledge points in the current question according to the mastery status of the user to be tested for each knowledge point and the final correlation weight between the current question and each knowledge point;

[0019] Integrate the knowledge point state feature data and the question feature data to obtain a feature vector containing both question information and knowledge point state information;

[0020] Based on the feature vector that contains both question information and knowledge point status information, the probability of the user to be tested answering the current question correctly is predicted, and based on the answer result of the user to be tested to the current question and the mapping correspondence between the current question and each knowledge point, the mastery status of the user to be tested for each knowledge point in the knowledge point mastery status matrix is ​​updated.

[0021] Furthermore, according to the knowledge point information marked on the current question, determining the priori correlation weights between the current question and each knowledge point includes:

[0022] When there are n knowledge points marked on the current question, the prior relevance weights corresponding to the n knowledge points are determined to be 1 / n respectively (if the specific weight of each knowledge point has been marked, the relevance weight is the marked weight), and the prior relevance weights of other knowledge points are determined to be 0, n≥1.

[0023] Further, according to the first correlation weight between the current question and each knowledge point, and the prior correlation weight between the current question and each knowledge point, determining the final correlation weight between the current question and each knowledge point includes:

[0024] The final correlation weights between the current question and each knowledge point are determined according to the first relationship model, wherein the first relationship model is:

[0025] Including but not limited to weighted average, for example,

[0026] w ti =α*softmax(M i m t )+(1-α)*{c t}

[0027] Among them, α is an adjustable parameter in [0,1]; w ti represents the final correlation weight between the current question and knowledge point i; softmax(M i m t ) represents the first correlation weight between the current question and each knowledge point; {c t} represents the prior correlation weight between the current question and each knowledge point; M represents the question-knowledge point mapping matrix, which is used to store the mapping between the current question and each knowledge point, and its size is n*d m , n is the number of knowledge points, d m is the dimension of the feature data of the current question; M i represents the i-th row of M, m t Represents the topic feature data of the current topic; where the expression of the soft max function is Where i=1...C.

[0028] Furthermore, the knowledge and ability tracking model also includes a forgetting decay network model;

[0029] Accordingly, the method further comprises:

[0030] Obtaining the time interval between the current and the last update of the knowledge point mastering state matrix; wherein the time interval is used to input the forgetting decay network model to obtain the decay rate;

[0031] Accordingly, obtaining the mastery status of each knowledge point of the user to be tested from the knowledge point mastery status matrix corresponding to the user to be tested includes:

[0032] The decay rate is used to decay and update the knowledge point mastery states in the knowledge point mastery state matrix corresponding to the user to be tested, and the knowledge point mastery states after decay and update are obtained.

[0033] Furthermore, the forgetting decay network model is: t =sigmoid(W d t t +b d ), where d t represents the decay rate, t t represents the time interval, W d and b d represents the network parameters, W d and b d The value of is obtained through training; the expression of the sigmoid function is

[0034] Correspondingly, the decay rate is used to decay and update the knowledge point mastering state in the knowledge point mastering state matrix corresponding to the user to be tested, including:

[0035] Attenuation update of the knowledge point mastery state in the knowledge point mastery state matrix corresponding to the user to be tested is performed using a second relationship model according to the attenuation rate;

[0036] Among them, the second relationship model is:

[0037] Among them, N t represents the knowledge point mastering state matrix after attenuation update, which is used to store the mastering state of each knowledge point of the user to be tested; the size is n*d n , n is the number of knowledge points, d n The dimension representing the mastery status of knowledge points, N t-1 Represents the most recently updated knowledge point mastering state matrix, N ti N t The i-th row, N t-1,iN t-1 The i-th row of represents dot product, 1 represents all 1 vector, w ti Represents the final correlation weight between the current question and knowledge point i; () T Represents the transposed matrix of ().

[0038] Furthermore, according to the mastery status of each knowledge point by the user to be tested and the final correlation weight between the current question and each knowledge point, the knowledge point status characteristic data of the user to be tested for all knowledge points in the current question is determined, including:

[0039] The knowledge point state characteristic data of the user to be tested for all the knowledge points in the current question is determined according to the third relationship model; wherein the third relationship model is:

[0040]

[0041] Among them, k t Represents the knowledge point status feature data of the user to be tested for all knowledge points in the current question.

[0042] Furthermore, the knowledge point state feature data and the question feature data are integrated together to obtain a feature vector containing both question information and knowledge point state information, including:

[0043] The knowledge point state feature data k t and topic feature data m t Integrate them together and get the feature vector s through a fully connected neural network t :

[0044] s t =tanh(W s [k t ,m t ]+b s )

[0045] Among them, the expression of tanh function is: W f and b f Are two network parameters obtained through training.

[0046] Furthermore, the knowledge and ability tracking model also includes: an ability network model and a question difficulty network model;

[0047] Accordingly, according to the feature vector including both the question information and the knowledge point status information, predicting the probability of the user to be tested answering the current question correctly includes:

[0048] Obtaining the ability value of the user to be tested for the current question according to a feature vector including both question information and knowledge point status information, and an ability network model corresponding to the user to be tested;

[0049] Based on the question feature data and the question difficulty network model, obtaining a difficulty value of the current question;

[0050] Based on the ability value and the difficulty value, the probability of the user to be tested answering the current question correctly is predicted.

[0051] Further, according to the feature vector including both the question information and the knowledge point status information, and the ability network model corresponding to the user to be tested, the ability value of the user to be tested for the current question is obtained, including:

[0052] Using the capability network model θ tj =tanh(W θ s t +b θ ) obtain the ability value of the user to be tested for the current question; wherein, θ tj Indicates ability value, W θ and b θ Represents the parameters of the capability network model, obtained through training;

[0053] According to the question feature data and the question difficulty network model, the difficulty value of the current question is obtained, including:

[0054] Using the question difficulty network model β j =tanh(W β m t +b β ) to obtain the difficulty value of the current question; where β j Indicates the difficulty value, W β and b β The parameters of the network model representing the difficulty of the question are obtained through training;

[0055] Based on the ability value and the difficulty value, predicting the probability of the user to be tested answering the current question correctly includes: using a prediction model p t =sigmoid(a*θ tj -β j ) predicts the probability of the user to be tested answering the current question correctly; where p t It represents the probability of answering correctly. According to IRT theory, the value of a is 3.0.

[0056] Further, according to the answer result of the user to be tested for the current question and the mapping relationship between the current question and each knowledge point, the mastery status of each knowledge point of the user to be tested in the knowledge point mastery status matrix is ​​updated, including but not limited to:

[0057] According to the answer result vector v of the user to be tested for the current question t And the mapping relationship between the current question and each knowledge point, first erase the original memory and then write the new memory; wherein the erase vector and the write vector are e t and h t ;

[0058] e t =sigmoid(W e n t +b e )

[0059] h t =tanh(W h n t +b h )

[0060] Among them, W e and b e is the network parameter, obtained in advance through training; W h and b h are network parameters, obtained in advance through training;

[0061] The original memory is erased according to the fourth relational model, wherein the fourth relational model is:

[0062]

[0063] Write new memory according to the fifth relational model, where the fifth relational model is:

[0064]

[0065] Among them, N t+1,i represents the updated knowledge point mastering state matrix, N t+1,i N t+1 The i-th row of .

[0066] Furthermore, when the answer result is a binary 0 or 1 result, the loss function is a binary cross entropy; when the answer result is a score rate [0, 1], the loss function is a multivariate cross entropy;

[0067] In a second aspect, an embodiment of the present invention further provides a resource push method based on the intelligent learning knowledge capability tracking method as described in the first aspect, comprising:

[0068] According to the mastery status of each knowledge point by the user to be tested, test questions and / or learning resources are pushed to the user to be tested.

[0069] In a third aspect, an embodiment of the present invention further provides a test paper composition method based on the answer prediction method and knowledge ability tracking as described in the first aspect, comprising:

[0070] Automatically compile the test paper based on the mastery status of each knowledge point by each user to be tested within the specified assessment scope and the knowledge points marked on the test questions to be compiled; and / or, automatically compile the test paper based on the predicted scores of each test question for each user to be tested within the specified assessment scope; and / or, based on the mastery status of each knowledge point by users to be tested in each region and the predicted scores of users to be tested in each region for different questions, map student groups in different regions to the same question space to achieve comparability of scores in different regions.

[0071] In a fourth aspect, an embodiment of the present invention further provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of the intelligent learning knowledge ability tracking method as described in the first aspect are implemented, and / or, when the processor executes the program, the steps of the resource pushing method as described in the second aspect are implemented, and / or, when the processor executes the program, the steps of the test paper assembling method as described in the third aspect are implemented.

[0072] It can be seen from the above technical scheme that the intelligent learning knowledge ability tracking method, electronic device and storage medium provided by the embodiment of the present invention first obtain the current question corresponding to the user to be tested, the prior correlation weight between the current question and each knowledge point, and then input the user identification to be tested, the current question, and the prior correlation weight between the current question and each knowledge point into the knowledge ability tracking model to obtain the probability of the user to be tested answering the current question correctly; wherein, the knowledge ability tracking model contains the mastery status of the user to be tested for each knowledge point, and the mastery status of the user to be tested for each knowledge point is determined based on the questions done by the user to be tested in a historical time period, the answer results of the questions done by the user to be tested, and the prior correlation weights of the questions done by the user to be tested and each knowledge point; wherein, the knowledge ability tracking model is obtained after training based on machine learning by inputting the questions done by each user in the historical training data, the answer results of the questions done by each user, and the prior correlation weights of the questions done by each user and each knowledge point into the initial knowledge ability tracking model. The embodiment of the present invention introduces the prior relationship between questions and knowledge points, and combines the relationship between online learning questions and various knowledge points, thereby establishing an explicit mapping correspondence between the current question and each knowledge point. Therefore, it is not only possible to accurately predict the probability of answering the current question correctly, but also possible to update the knowledge point mastery status matrix according to the answer result of the current question, so that the mastery status of each knowledge point of the user to be tested can be known. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0074] Figure 1 A flow chart of a method for tracking intelligent learning knowledge capabilities provided by an embodiment of the present invention;

[0075] Figure 2 A schematic diagram of a knowledge capability tracking model provided by an embodiment of the present invention;

[0076] Figure 3 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0077] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0078] At present, deep learning models such as recurrent neural networks that directly model student behavior time series data cannot obtain students' mastery of knowledge points. For example, when the model inputs the answer to the question, the model can predict the probability of answering all questions (done and not done) correctly in the next step, but because the model represents the student's current mastery of all knowledge points with a hidden state, the model cannot output the student's mastery of specific knowledge points. For another example, when the model inputs the answer to the knowledge point (one or more) corresponding to the question, it can predict the probability of answering all knowledge points (done and not done) correctly in the next step, but this is not the student's mastery of the knowledge point, that is, it still cannot output the student's mastery of specific knowledge points. Moreover, in the time series, there will be discontinuity and fluctuations. In terms of prediction accuracy, the loss of question dimension information will lead to a decrease in the model's prediction accuracy.

[0079] It is understandable that the goal of the model is not only to predict the next answer, but also to calculate the knowledge point mastery after the current answer result is input. The deep recurrent neural network model cannot achieve this goal. The model provided by the embodiment of the present invention can use the current input result to update the knowledge point mastery state matrix, and then obtain the knowledge point state of the corresponding memory card slot position. It is understandable that an important goal of the model is to obtain the mastery state of the knowledge points marked on the question, so the model needs to input the prior relationship between the marked questions and the knowledge points as a constraint condition, that is, the correlation weight between the question and the knowledge point is composed of the marked prior relationship and the relationship learned by the network. The intelligent learning knowledge ability tracking method provided by the present invention will be explained and illustrated in detail below through specific embodiments.

[0080] Figure 1 FIG. 2 is a flow chart showing a method for tracking the knowledge and ability of intelligent learning provided by an embodiment of the present invention. Figure 1 As shown, the intelligent learning knowledge ability tracking method provided by the embodiment of the present invention includes the following steps:

[0081] Step 101: Obtain the current question corresponding to the user to be tested, and the priori correlation weights between the current question and each knowledge point;

[0082] Step 102: Input the ID of the user to be tested, the current question, and the prior correlation weights between the current question and each knowledge point into the knowledge and ability tracking model to obtain the probability of the user to be tested answering the current question correctly; wherein, the knowledge and ability tracking model includes the mastery status of the user to be tested on each knowledge point, and the mastery status of the user to be tested on each knowledge point is determined based on the questions done by the user to be tested in a historical time period, the answer results of the questions done by the user to be tested, and the prior correlation weights between the questions done by the user to be tested and each knowledge point; wherein, the knowledge and ability tracking model is obtained after training based on machine learning by inputting the questions done by each user in the historical training data, the answer results of the questions done by each user, and the prior correlation weights between the questions done by each user and each knowledge point into the initial knowledge and ability tracking model.

[0083] It is understandable that the intelligent learning knowledge ability tracking method further includes:

[0084] After the current question is answered, the answer result of the user to be tested for the current question is obtained, and the mastery status of each knowledge point of the user to be tested is updated according to the answer result.

[0085] In this embodiment, the priori relevance weights between the question and each knowledge point are determined based on the knowledge point information marked on the question. For example, assuming that there are n knowledge points marked on the current question, it can be determined that the priori relevance weights of the corresponding knowledge points are all 1 / n (if the specific weight of each knowledge point has been marked, the relevance weight is the marked weight), and the priori relevance weights of the remaining knowledge points are all 0.

[0086] In the present embodiment, it should be noted that an important goal of the present embodiment is to obtain the mastery status of the knowledge points marked on the questions, so the prior relationship between the marked questions and the knowledge points needs to be input as a strong signal. Since the prior relationship between the questions and the knowledge points is introduced, and the relationship between the online learning questions and the knowledge points is combined, an explicit mapping correspondence between the questions and the knowledge points can be established, which can not only achieve accurate prediction of the probability of answering the questions correctly, but also update the mastery status of the knowledge points according to the answer results of the questions, so that the mastery status of the user to be tested for each knowledge point can be known.

[0087] It can be understood that this embodiment first obtains the current question corresponding to the user to be tested, the prior correlation weights between the current question and each knowledge point, and then inputs the user identification to be tested, the current question, and the prior correlation weights between the current question and each knowledge point into the knowledge and ability tracking model to obtain the probability of the user to be tested answering the current question correctly; wherein, the knowledge and ability tracking model includes the mastery status of the user to be tested for each knowledge point, and the mastery status of the user to be tested for each knowledge point is determined based on the questions done by the user to be tested in a historical time period, the answer results of the questions done by the user to be tested, and the prior correlation weights between the questions done by the user to be tested and each knowledge point; wherein, the knowledge and ability tracking model is obtained after training based on machine learning by inputting the questions done by each user in the historical training data, the answer results of the questions done by each user, and the prior correlation weights between the questions done by each user and each knowledge point into the initial knowledge and ability tracking model. The embodiment of the present invention introduces the prior relationship between questions and knowledge points, and combines the relationship between online learning questions and various knowledge points, thereby establishing an explicit mapping correspondence between the current question and each knowledge point. Therefore, it is not only possible to accurately predict the probability of answering the current question correctly, but also possible to update the knowledge point mastery status matrix according to the answer result of the current question, so that the mastery status of each knowledge point of the user to be tested can be known.

[0088] The intelligent learning knowledge ability tracking method provided by the present invention is explained and illustrated in more detail below through a more specific scheme.

[0089] Based on the content of the above embodiment, in this embodiment, the knowledge ability tracking model includes a question knowledge point mapping model;

[0090] Accordingly, the user ID to be tested, the current question, and the prior correlation weights between the current question and each knowledge point are input into the knowledge ability tracking model to obtain the probability of the user to be tested answering the current question correctly, specifically including:

[0091] Inputting the topic feature data of the current topic into the topic knowledge point mapping model to obtain the first correlation weight between the current topic and each knowledge point;

[0092] Determine the priori correlation weights between the current question and each knowledge point according to the knowledge point information marked on the current question;

[0093] Determine the final correlation weight between the current question and each knowledge point according to the first correlation weight between the current question and each knowledge point, and the prior correlation weight between the current question and each knowledge point, and determine the mapping correspondence between the current question and each knowledge point according to the final correlation weight between the current question and each knowledge point;

[0094] From the knowledge point mastering state matrix corresponding to the user to be tested, obtaining the mastering state of each knowledge point of the user to be tested; wherein the knowledge point mastering state matrix stores the dynamic mastering state of each knowledge point of the user to be tested; wherein the knowledge point mastering state is used to characterize the mastering degree of the user to be tested on the knowledge point;

[0095] Determine the knowledge point status characteristic data of the user to be tested for all the knowledge points in the current question according to the mastery status of the user to be tested for each knowledge point and the final correlation weight between the current question and each knowledge point;

[0096] Integrate the knowledge point state feature data and the question feature data to obtain a feature vector containing both question information and knowledge point state information;

[0097] Based on the feature vector that contains both question information and knowledge point status information, the probability of the user to be tested answering the current question correctly is predicted, and based on the answer result of the user to be tested to the current question and the mapping correspondence between the current question and each knowledge point, the mastery status of the user to be tested for each knowledge point in the knowledge point mastery status matrix is ​​updated.

[0098] In this embodiment, the question-knowledge point mapping model is used to characterize the correlation between question feature data and the first relevance weight of each knowledge point. Therefore, the first relevance weight between the current question and each knowledge point can be obtained through the question-knowledge point mapping model.

[0099] In this embodiment, the prior correlation weights between the current question and each knowledge point can be determined based on the knowledge point information marked on the current question. For example, assuming that there are n knowledge points marked on the current question, it can be determined that the prior correlation weights of the corresponding knowledge points are all 1 / n (if the specific weight of each knowledge point has been marked, the correlation weight is the marked weight), and the prior correlation weights of the remaining knowledge points are all 0.

[0100] In this embodiment, it should be noted that an important goal of this embodiment is to obtain the mastery status of the knowledge points marked on the questions, so the prior relationship between the marked questions and the knowledge points needs to be input as a constraint condition, that is, the correlation weight between the questions and the knowledge points is composed of the marked prior relationship and the relationship learned through the network.

[0101] In the present embodiment, it should be noted that, since an explicit mapping correspondence between the current question and each knowledge point is established based on the first correlation weight between the current question and each knowledge point, as well as the prior correlation weight between the current question and each knowledge point, the knowledge point mastery status matrix can be subsequently updated according to the answer result of the current question, thereby making it possible to know the mastery status of each knowledge point of the user to be tested in real time.

[0102] In the present embodiment, it should be noted that the knowledge point status of the user to be tested for each knowledge point stored in the knowledge point mastery status matrix is ​​updated in real time according to the answering situation of the user to be tested, that is, the status of the user to be tested for each knowledge point stored in the knowledge point mastery status matrix is ​​obtained by updating the status of each knowledge point in real time according to the historical answering situation of the user to be tested for each question with marked knowledge points; when the user to be tested has no historical answering record, the status of each knowledge point initially stored in the knowledge point mastery status matrix is ​​the average knowledge point status of all users of the training model data.

[0103] In this embodiment, since the current question includes at least one knowledge point, the knowledge point status characteristic data of the user to be tested for all knowledge points in the current question can be finally determined based on the user to be tested's mastery of each knowledge point and the final correlation weight between the current question and each knowledge point.

[0104] In this embodiment, the knowledge point state feature data and the question feature data are integrated together to obtain a feature vector containing both question information and knowledge point state information, so that the probability of the user to be tested answering the current question correctly can be predicted based on the feature vector containing both question information and knowledge point state information. Since this embodiment further considers the question dimension information in terms of prediction accuracy, the prediction accuracy of the probability of answering correctly will be improved.

[0105] In this embodiment, after the test is completed, the mastery status of each knowledge point of the user to be tested in the knowledge point mastery status matrix is ​​updated according to the answer result of the user to be tested for the current question and the mapping correspondence between the current question and each knowledge point, so that the mastery status of each knowledge point of the user to be tested can be known, and it is also convenient to accurately predict the success rate of the next test according to the mastery status of each knowledge point of the user to be tested.

[0106] It can be seen from the above technical scheme that the intelligent learning knowledge ability tracking method provided by the embodiment of the present invention first obtains the first correlation weight between the current question and each knowledge point, and at the same time determines the prior correlation weight between the current question and each knowledge point according to the knowledge point information marked on the current question, so that the final correlation weight between the current question and each knowledge point can be determined according to the first correlation weight between the current question and each knowledge point and the prior correlation weight between the current question and each knowledge point. At the same time, the embodiment of the present invention also obtains the mastery status of the user to be tested for each knowledge point from the knowledge point mastery status matrix corresponding to the user to be tested; thereby, the mastery status of the user to be tested for each knowledge point and the final correlation weight between the current question and each knowledge point can be determined. The present invention further comprises the following steps: first, the step of: obtaining a characteristic vector which includes both the characteristic vector of the question information and the characteristic vector of the knowledge point status information, and then predicting the probability of the user to be tested to answer the current question correctly according to the characteristic vector which includes both the characteristic vector of the question information and the characteristic vector of the knowledge point status information. Then, according to the answer result of the user to be tested to the current question and the mapping relationship between the current question and each knowledge point, the mastery status of each knowledge point of the user to be tested in the knowledge point mastery status matrix is ​​updated, so that the mastery status of each knowledge point of the user to be tested can be known, and at the same time, it is also convenient to accurately predict the success rate of answering the next question according to the mastery status of each knowledge point of the user to be tested.

[0107] Based on the content of the above embodiment, in this embodiment, according to the knowledge point information marked on the current question, the priori correlation weights between the current question and each knowledge point are determined, including:

[0108] When there are n knowledge points marked on the current question, the prior relevance weights corresponding to the n knowledge points are determined to be 1 / n respectively (if the specific weight of each knowledge point has been marked, the relevance weight is the marked weight), and the prior relevance weights of other knowledge points are determined to be 0, n≥1.

[0109] In this embodiment, the prior relationship information between the current question and each knowledge point is introduced, so that on the one hand, an explicit mapping correspondence between the current question and each knowledge point can be established, and on the other hand, the final weights of the question and each knowledge point can be determined based on the prior weights of the question and each knowledge point and the weights learned through the network. This allows the determined final weights to more accurately reflect the correlation between the question and each knowledge point, and at the same time, the potential relationship between the knowledge points can be explored.

[0110] Based on the content of the above embodiment, in this embodiment, according to the first correlation weight between the current question and each knowledge point, and the prior correlation weight between the current question and each knowledge point, the final correlation weight between the current question and each knowledge point is determined, including:

[0111] The final correlation weights between the current question and each knowledge point are determined according to the first relationship model, wherein the first relationship model is:

[0112] Including but not limited to weighted average, for example,

[0113] w ti =α*soft max(M i m t )+(1-α)*{c t}

[0114] Among them, α is an adjustable parameter in [0,1], for example, it can take the value of 0.5; w ti represents the final correlation weight between the current question and knowledge point i; softmax(M i m t ) represents the first correlation weight between the current question and each knowledge point; {c t} represents the prior correlation weight between the current question and each knowledge point; M represents the question-knowledge point mapping matrix, which is used to store the mapping between the current question and each knowledge point, and its size is n*d m , n is the number of knowledge point slots, d m is the dimension of the feature data of the current question; M i represents the i-th row of M, m t Represents the topic feature data of the current topic; where the expression of the softmax function is Where i=1...C.

[0115] In this embodiment, the final correlation weight between the current question and each knowledge point is determined according to the first relationship model. It can be seen from the first relationship model that the correlation weight w between the question and each knowledge point is t By input {c t The final weights of the question and each knowledge point are determined based on the prior weights of the question and each knowledge point and the initial weights learned through the network, so that the final weights determined can more accurately reflect the relationship between the question and each knowledge point.

[0116] Based on the content of the above embodiment, in this embodiment, the knowledge ability tracking model also includes a forgetting decay network model; accordingly, the intelligent learning knowledge ability tracking method also includes:

[0117] Obtaining the time interval between the current and the last update of the knowledge point mastering state matrix; wherein the time interval is used to input the forgetting decay network model to obtain the decay rate;

[0118] Accordingly, obtaining the mastery status of each knowledge point of the user to be tested from the knowledge point mastery status matrix corresponding to the user to be tested includes:

[0119] The decay rate is used to decay and update the knowledge point mastery status in the knowledge point mastery status matrix corresponding to the user to be tested, and the knowledge point mastery status after decay and update is obtained.

[0120] This embodiment adds the factor of time decay on the basis of the above embodiments. Since students' mastery of knowledge points will decrease as time goes by, the pure right or wrong answer sequence loses the time factor in the answer sequence, which will inevitably affect the accuracy of prediction and the calculation accuracy of knowledge status. The same question done one day ago and one month ago will have different effects on the current situation. Adding the input of the forgetting decay network and the time dimension can not only improve the prediction accuracy of the model, but also the forgetting decay network can learn different decay rates for each knowledge point. The Ebbinghaus forgetting curve is only applicable to certain specific scenarios. It is more reasonable to apply the forgetting decay rate learned from a large amount of historical data to related scenarios.

[0121] Based on the content of the above embodiment, in this embodiment, the forgetting decay network model is: t =sigmoid(W d t t +b d ), where d t represents the decay rate, t t represents the time interval, W d and b d represents the network parameters, W d and b d The value of is obtained through training; the expression of the sigmoid function is

[0122] Accordingly, the decay rate is used to decay and update the knowledge point mastering state in the knowledge point mastering state matrix corresponding to the user to be tested, including but not limited to:

[0123] Attenuation update of the knowledge point mastery state in the knowledge point mastery state matrix corresponding to the user to be tested is performed using a second relationship model according to the attenuation rate;

[0124] Among them, the second relationship model is:

[0125] Among them, N t represents the knowledge point mastering state matrix after attenuation update, which is used to store the mastering state of each knowledge point of the user to be tested; the size is n*d n , n is the number of knowledge points, d n To grasp the dimension of the state matrix for knowledge points, N t-1 Represents the most recently updated knowledge point mastering state matrix, N ti N t The i-th row, N t-1,i N t-1 The i-th row of represents dot product, 1 represents all 1 vector, w ti Represents the final correlation weight between the current question and knowledge point i; () T Represents the transposed matrix of ().

[0126] In this embodiment, a forgetting decay network d is used according to the time interval t =sigmoid(W d t t +b d ) to obtain the decay rate, and then adopt the second relationship model according to the acquired decay rate The knowledge point mastery states in the knowledge point mastery state matrix corresponding to the user to be tested are updated with decay, so that the knowledge point mastery state matrix after decay update can be obtained.

[0127] Based on the content of the above embodiment, in this embodiment, according to the mastery status of each knowledge point by the user to be tested and the final correlation weight between the current question and each knowledge point, the knowledge point status feature data of the user to be tested for all knowledge points in the current question is determined, including:

[0128] The knowledge point state characteristic data of the user to be tested for all the knowledge points in the current question is determined according to the third relationship model; wherein the third relationship model is:

[0129]

[0130] Among them, k t Represents the knowledge point status feature data of the user to be tested for all knowledge points in the current question.

[0131] In this embodiment, based on the mastery status of each knowledge point by the user to be tested and the final correlation weight between the current question and each knowledge point, the knowledge point status characteristic data of the user to be tested for all knowledge points in the current question is determined according to the third relationship model, so that the user's comprehensive mastery results of all knowledge points contained in the current question can be accurately obtained.

[0132] Based on the content of the above embodiment, in this embodiment, the knowledge point state feature data and the question feature data are integrated together to obtain a feature vector containing both question information and knowledge point state information, including:

[0133] The knowledge point state feature data k t and topic feature data m t Integrate them together and get the feature vector s through a fully connected neural network t :

[0134] s t =tanh(W s [k t ,m t ]+b s )

[0135] Among them, the expression of tanh function is: W s and b s are two neural network parameters obtained through training.

[0136] In this embodiment, the knowledge point state feature data r t and topic feature data k t Integrate them together and get the feature vector f through the fully connected neural network t , so that the accuracy of answering questions can be predicted based on the characteristic vector that contains both the question feature data and the knowledge point status feature data.

[0137] Based on the content of the above embodiment, in this embodiment, the knowledge ability tracking model also includes: an ability network model and a question difficulty network model;

[0138] Accordingly, according to the feature vector including both the question information and the knowledge point status information, predicting the probability of the user to be tested answering the current question correctly includes:

[0139] Obtaining the ability value of the user to be tested for the current question according to a feature vector including both question information and knowledge point status information, and an ability network model corresponding to the user to be tested;

[0140] Based on the question feature data and the question difficulty network model, obtaining a difficulty value of the current question;

[0141] Based on the ability value and the difficulty value, the probability of the user to be tested answering the current question correctly is predicted.

[0142] In this embodiment, it can be understood that the ability values ​​of the same user for different questions are different. Therefore, the ability value of the user to be tested for the current question can be obtained based on the feature vector that contains both question information and knowledge point status information, and the ability network model corresponding to the user to be tested. In addition, even if two questions contain the same knowledge points, the difficulty of the two questions is not necessarily the same. For example, some questions are difficult and some are simple. Therefore, a question difficulty network model is added to obtain the difficulty value of the current question, and then the ability value and the difficulty value are combined to predict the probability of answering the question correctly.

[0143] In this embodiment, it can be understood that the ability network model can output the student's ability value for the question and the ability value for the knowledge point (i.e., the knowledge point mastery status), and the question difficulty network model can output the question difficulty value, and then combine with the IRT theory to predict the probability of answering the question correctly, thereby enhancing the interpretability of the model.

[0144] Based on the content of the above embodiment, in this embodiment, according to the feature vector including both the question information and the knowledge point state information, and the ability network model corresponding to the user to be tested, the ability value of the user to be tested for the current question is obtained, including:

[0145] Using the capability network model θ tj =tanh(W θ s t +b θ ) obtain the ability value of the user to be tested for the current question; wherein, θ tj Indicates ability value, W θ and b θ Represents the parameters of the capability network model, obtained through training;

[0146] Based on the question feature data and the question difficulty network model, the difficulty value of the current question is obtained, including:

[0147] Using the question difficulty network model β j =tanh(W β m t +b β ) to obtain the difficulty value of the current question; where β j Indicates the difficulty value, W β and b β The parameters of the network model representing the difficulty of the question are obtained through training;

[0148] Based on the ability value and the difficulty value, predicting the probability of the user to be tested answering the current question correctly includes: using a prediction model p t =sigmoid(a*θ tj -βj ) predicts the probability of the user to be tested answering the current question correctly; where p t It represents the probability of answering correctly. According to IRT theory, the value of a is 3.

[0149] Based on the content of the above embodiment, in this embodiment, according to the answer result of the user to be tested for the current question and the mapping correspondence between the current question and each knowledge point, the mastery status of each knowledge point of the user to be tested in the knowledge point mastery status matrix is ​​updated, including but not limited to:

[0150] According to the answer result vector v of the user to be tested for the current question t And the mapping relationship between the current question and each knowledge point, first erase the original memory and then write the new memory; wherein the erase vector and the write vector are e t and h t ;

[0151] e t =sigmoid(W e n t +b e )

[0152] h t =tanh(W h n t +b h )

[0153] Among them, W e and b e is the network parameter, obtained in advance through training; W h and b h are network parameters, obtained in advance through training;

[0154] The original memory is erased according to the fourth relational model, wherein the fourth relational model is:

[0155]

[0156] Write new memory according to the fifth relational model, where the fifth relational model is:

[0157]

[0158] Among them, N t+1,i represents the updated knowledge point mastering state matrix, N t+1,i N t+1 The i-th row of .

[0159] In this embodiment, the question answer vector n is obtained. t (The dimension is d n), then erase the original memory and write the new memory, the erase vector and write vector are e t 、h t , e t =sigmoid(W e n t +b e ), h t =tanh(W h n t +b h ), erase the original memory: Writing new memory: Thereby, the mastering status of each knowledge point of the user to be tested in the knowledge point mastering status matrix is ​​updated.

[0160] Based on the content of the above embodiment, in this embodiment, when the answer result is a binary 0 or 1 result, the loss function is a binary cross entropy; when the answer result is a score rate, the loss function is a multivariate cross entropy.

[0161] In this embodiment, it should be noted that the answer results of many subjective questions are not just right or wrong (1 and 0), and the answer results of multiple-choice questions are not just right or wrong (1 and 0). Usually, the answer results can output a score rate (between 0-1). This embodiment supports the input of score rate. The current answer result input is not just 1 and 0, but can also include score levels such as 0.1, 0.2...

[0162] It is understandable that if a student has a considerable number of answer records, the model provided in this embodiment can output the student's mastery of all questions in the question space, that is, obtain the student's predicted scores on all questions; the model can also automatically mine the knowledge concepts implicit in the questions, and then obtain the student's mastery of all knowledge in the knowledge space. In particular, at the first moment, if a student does not have any historical record of answering questions, then his first knowledge state is based on a non-personalized state, which is generated based on large-scale training data. After the first knowledge state, the subsequent knowledge states will become more and more personalized.

[0163] This embodiment uses a memory network to implement a knowledge ability tracking model. The network model uses a matrix (question knowledge point mapping matrix) to store all knowledge points, and another matrix (knowledge point mastery status matrix) to store and update students' mastery of knowledge points. While tracking the mastery status of different knowledge points, the model can also capture the relationship between different knowledge points. The model maintains a state for each knowledge point, and automatically learns the correlation between the current question input and each knowledge point. When a new question is input, the model will first select the knowledge point related to the current question, and then update the status of the relevant knowledge point according to the correctness of the student's answer. In order to obtain the mastery status of each knowledge point, the model of this embodiment maps the knowledge point status stored in the knowledge point mastery status matrix card slot with the knowledge points marked on the question, and uses the prior relationship between the marked question and the knowledge point as a constraint. In a preferred embodiment, the model of this embodiment also adds the time interval input of the question and the forgetting decay network. The knowledge point status in the knowledge point mastery status matrix is ​​first forgotten and decayed, and then the knowledge point status is used to predict the correct probability of the student's answer, and finally the knowledge point status in the matrix is ​​updated according to the correctness of the question. For details of the model network structure, see Figure 2 As shown, the specific description is as follows:

[0164] 1. Model input:

[0165] q t is the topic id, q t ∈{1,2,...,Q}, where Q is the number of questions; a t For the answer result, a t ∈{0,1}, the score rate input is multiple score levels, qa t =q t +a t *Q,qa t ∈{1,2,...,2Q}; {c t} is the softmax input of the prior relationship between the question and the knowledge point (i.e., if the question contains n knowledge points, the value of the position where the knowledge point is located is 1 / n, and the others are 0); t t The time interval between the current question time and the last question time involving the same knowledge point. For example, if it is the first time to do the question, the time interval is entered as 0, and if it is not the first time, the time interval is entered as log 10 (1+interval hours).

[0166] 2. Get the mapping weight of question knowledge points:

[0167] q t Through the topic embedding matrix (shape is q*d m , q is the number of questions) to get the embedding vector m of the questiont (The dimension is d m ), M (shape is n*d m , n is the number of knowledge points) is the question-knowledge-point mapping matrix, which stores the mapping between questions and knowledge points; N t (shape is n*d v , n is the number of knowledge points) is the knowledge point mastery status matrix, which stores the mastery status of the knowledge points. The relevance weight w between the question and each knowledge point t By input {c t} is combined with the softmax vector of the output of the knowledge point mapping matrix of the question, including but not limited to the weighted average method, for example,

[0168] w ti =α*soft max(M i m t )+(1-α)*{c t}

[0169] Among them, α is an adjustable parameter in [0,1], for example, it can be set to 0.5 by default; w ti is the correlation weight vector w t The i-th element of M i is the i-th row vector of M, Where i=1...C.

[0170] 3. Forget decay:

[0171] Input t t Connect a forget decay fully connected neural network and output the decay rate d of each knowledge point t ,

[0172] d t =sigmoid(W d t t +b d )

[0173] Attenuate memory of the knowledge point status in the knowledge point mastery status matrix, including but not limited to,

[0174]

[0175] Among them, 1 is a vector of all 1s, is the dot product,

[0176] 4. Prediction process:

[0177] First, from the knowledge point to grasp the state matrix N t Read the knowledge point state k t ,

[0178]

[0179] Among them, N ti N t The i-th row vector of .

[0180] Then, k t and m t Connect them together and then connect a fully connected neural network to get the feature vector s t ,

[0181] s t =tanh(W s [k t ,m t ]+b s )

[0182] Then connect the student ability network θ tj and the question difficulty network β j ,

[0183] θ tj =tanh(W θ s t +b θ )

[0184] β j =tanh(W β m t +b β )

[0185]

[0186] Finally, according to IRT theory (IRT theory is Item Response Theory, also known as Question Response Theory, is a mathematical model used to analyze test scores or questionnaire survey data.

[0187] The model expression is as follows:

[0188]

[0189] Among them, P represents the probability of answering the question correctly, a represents the discrimination of the question, b represents the difficulty parameter of the question, c represents the guessing parameter of the question, θ represents the student ability value, and c = 0, a = 1), the predicted probability of answering the question correctly p t :

[0190] p t =sigmoid(3.0*θ tj -β j )

[0191] 5. Knowledge point mastery status matrix update:

[0192] qa t After answering the question, the embedding matrix (shape is 2q*d n , q is the number of questions) to get the question answer vector n t (The dimension is d n ), then erase the original memory and then write the new memory, including but not limited to, the erase vector and the write vector are e respectively t 、h t :

[0193] e t =sigmoid(W e n t +b e )

[0194] h t =tanh(W h n t +b h )

[0195] Erase the original memory:

[0196]

[0197] Writing new memory:

[0198]

[0199] Specifically, the model training and usage process is as follows:

[0200] 1. Training Model

[0201] The input of the model is the sequence of question answering records, the correspondence data between questions and knowledge points, the direct output of the model is the predicted probability of answering the question correctly, and the indirect output is the student's ability value for the question, the ability value for the knowledge point, and the difficulty value of the question.

[0202] In this embodiment, for the knowledge ability tracking model that only adds the prior relationship input of the question knowledge points, when training the model, the data format of the input model (taking a certain subject a as an example, the format of other subjects is the same) is as shown in (1) of the following 1), and for the knowledge ability tracking model that simultaneously adds the prior relationship input of the question knowledge points and the time interval input, when training the model, the data format of the input model (taking a certain subject a as an example, the format of other subjects is the same) is as shown in (2) of the following 1):

[0203] 1) Question sequence data

[0204] (1) No time input for answering questions

[0205] (student id, [[[question id], question result], [[question id], question result], [[question id], question result], ...])

[0206] For example, (101893639,[[[2],1],[[5],0],[[3],1]])

[0207] The question id is a converted continuous id starting from 0, and the answer result is 1 for a correct answer and 0 for an incorrect answer.

[0208] (2) Add time input for answering questions

[0209] (student id, [[[question id], question result, question time], [[question id], question result, question time], [[question id], question result, question time], ...])

[0210] For example, (101893639,[[[2],1,'2020-08-02 16:07:10'],[[5],0,'2020-08-02 16:08:11'],[[3],1,'2020-08-02 16:08:56']])

[0211] The question id is a converted continuous id starting from 0, and the answer result is 1 for correct answer and 0 for wrong answer.

[0212] 2) The prior relationship data between questions and knowledge points are shown in Table 1:

[0213] Table 1

[0214] Topic ID Knowledge point id 0 0 1 1 2 2,3 3 4,5,6 4 7 5 8 ... ...

[0215] Among them, the question id is a continuous id starting from 0 after conversion; the knowledge point id is a continuous id starting from 0 after conversion.

[0216] 2. Use the model

[0217] This embodiment calculates the students' mastery of knowledge points and predicts the probability of answering questions based on the students' real-time answering records. After the answering results update the knowledge point status of the knowledge point mastering status matrix, the corresponding knowledge point status is obtained. If you want to obtain the knowledge point mastering status in the i-th memory card slot, the process is as follows:

[0218] Assume w t is [0,..,w i ,..0],w i The position is set to 1, and then the knowledge point state k in the knowledge point mastery state matrix is ​​read t ,

[0219]

[0220] And the characteristic vector s t No connection to the topic related information,

[0221] s t =tanh(W s [k t ,0]+b s )

[0222] Then, through the student ability network, the student's ability value for the knowledge point, that is, the degree of mastery of the knowledge point, is calculated.

[0223] θ tj =tanh(W θ s t +b θ )

[0224] According to the above description, the intelligent learning knowledge ability tracking method provided by the embodiment of the present invention has the following advantages:

[0225] I. The model provided in this embodiment takes the question as input, automatically learns the correlation between the question and each knowledge point, and then combines the prior relationship between the question and the knowledge point to dynamically output the mastery status of each knowledge point. For the time series model represented by the deep recurrent neural network, if the model inputs the answer to the question, it can predict the probability of answering all the questions (done and not done) correctly in the next step, but the model represents the student's current mastery of all knowledge points with a hidden state, resulting in the model being unable to output the student's mastery of specific knowledge points. If the model inputs the answer to the question corresponding to the knowledge point (one or more), it can predict the probability of answering all the knowledge points (done and not done) correctly in the next step, but this is not the student's mastery of the knowledge point, and in the time series, there will be discontinuity and fluctuations. In terms of prediction accuracy, due to the loss of the question dimension information, there will be a decrease in prediction accuracy.

[0226] In addition, the main goal of the model is to be able to calculate the knowledge point mastery status after the current input answer result, but the deep recurrent neural network model cannot achieve this goal. The model of this embodiment can use the current input result to update the knowledge point mastery status matrix, and then obtain the knowledge point status of the corresponding memory card slot position.

[0227] II. The model structure of this embodiment includes not only a student ability network and a question difficulty network, but also a forgetting decay network.

[0228] The student ability network can output the student's ability value for the question and the ability value for the knowledge point (i.e. the mastery status of the knowledge point), and the question difficulty network can output the question difficulty value, and then combine the IRT theory to predict the probability of answering the question correctly, which enhances the interpretability of the model. The student's mastery of the knowledge point will decrease with time. The pure right and wrong answer sequence loses the time factor in the answer sequence, which will inevitably affect the accuracy of the prediction and the calculation accuracy of the knowledge status. The same question done a day ago and a month ago will have different effects on the current situation. Adding the input of the forgetting decay network and the time dimension can not only improve the prediction accuracy of the model, but also the forgetting decay network can learn different decay rates for each knowledge point. The Ebbinghaus forgetting curve is only applicable to certain specific scenarios. It is more reasonable to apply the forgetting decay rate learned from a large amount of historical data to related scenarios.

[0229] III. The model of this embodiment supports the input of score rate

[0230] The answer results of many subjective questions are not just right or wrong (1 and 0), and the answer results of multiple-choice questions are not just right or wrong (1 and 0). Usually, the answer results can output a score rate (between 0-1). The model of this embodiment supports the input of score rate. The current answer result input is not just 1 and 0, but can also include score levels such as 0.1, 0.2...

[0231] Taking subjects a, b, c, d, e, and f shown in Table 2 below as examples, the model evaluation indicators are auc (area under curve) and acc (accuracy, threshold 0.5), with a value range of [0,1]. The larger the auc and acc indicators, the better.

[0232] Table 2

[0233]

[0234] For the knowledge ability tracking model that does not add the prior relationship input of question knowledge points and time interval input, its model indicators are shown in Table 3 below:

[0235] Table 3

[0236] discipline auc acc a 0.8281 0.7934 b 0.8531 0.8038 c 0.8414 0.7958 d 0.9028 0.8630 e 0.9086 0.8682 f 0.9405 0.9118

[0237] For the knowledge ability tracking model that adds the prior relationship input of question knowledge points and the time interval input, its model indicators are shown in Table 4 below:

[0238] Table 4

[0239] discipline auc acc a 0.8632 0.8295 b 0.8734 0.8224 c 0.8674 0.8164 d 0.9212 0.8823 e 0.9283 0.8864 f 0.9589 0.9246

[0240] From the above, we can see that the performance of each subject model has been greatly improved.

[0241] Based on the same inventive concept, another embodiment of the present invention further provides a resource push method based on the intelligent learning knowledge capability tracking method described in the above embodiment, comprising:

[0242] According to the mastery status of each knowledge point by the user to be tested and / or the predicted answer results of the user to be tested to each question in the question bank, test questions and / or learning resources are pushed to the user to be tested.

[0243] In this embodiment, it should be noted that the knowledge and ability tracking model is modeled based on the student behavior sequence, and tracking the student's mastery of knowledge is the core and key to building an adaptive teaching system. In the adaptive teaching system, whether it is to push accurate questions or plan the path of student learning, the first step is to accurately estimate the student's mastery of knowledge. It can be seen that this embodiment can be used for personalized teaching process. Specifically, the student's academic behavior can be predicted. The excellent student knowledge and ability tracking model needs to have the ability to predict scores. An important evaluation indicator of the model is the accuracy of score prediction. The score prediction is defined as: obtaining a student's learning history and a question representation, the model can predict the student's score on the question. The personalized teaching system can accurately predict the student's answering situation, which is conducive to improving the student's autonomous learning efficiency, because the system can gradually push moderately difficult exercises to students, avoiding recommending too difficult or too simple questions, wasting students' time and energy. In addition, this embodiment can also perform personalized diagnosis and learning on the student's academic level. The goal of student ability diagnosis is to model the student's academic ability level at the knowledge level and help students discover their own weak knowledge points so that students can learn autonomously. For example, questions that students have not mastered well can be automatically pushed to strengthen the students' practice and reinforcement of these knowledge points. At the same time, based on the prediction results of knowledge points, it can also combine the teaching and research experience of education experts to help students learn independently and assist teachers in targeted teaching; personalized learning goes a step further and directly recommends appropriate learning resources to students based on their knowledge status and prior knowledge in the field of education, such as learning videos and test questions. Using the historical learning data of individual students and related groups to provide accurate learning plans, breaking the one-size-fits-all learning method in traditional teaching, can effectively improve students' learning efficiency.

[0244] Based on the same inventive concept, another embodiment of the present invention further provides a test paper composition method based on the intelligent learning knowledge ability tracking method described in the above embodiment, including one or more of the following methods:

[0245] ① Automatically generate the test paper based on the mastery status of each knowledge point of each user within the specified assessment scope and the knowledge points marked on the test paper to be generated;

[0246] ② Automatically generate the test paper based on the predicted scores of each test question for each user within the specified evaluation range;

[0247] ③ Based on the dynamic mastery status of each knowledge point by the users in each region, as well as the predicted scores of different questions by the users in each region, the student groups in different regions are mapped to the same question space to achieve comparability of scores in different regions.

[0248] In this embodiment, different test papers can be set for different application scenarios. For example, for classroom consolidation exercises, the mastery of each knowledge point by each user to be tested within the specified evaluation range can be obtained according to the mastery status of each knowledge point by each user to be tested within the specified evaluation range and the knowledge points marked on the test questions to be set. Then, according to the mastery of each knowledge point by each user to be tested within the specified evaluation range, suitable questions are selected for test paper setting, so that each user to be tested can effectively consolidate the knowledge points that they do not understand or are not proficient in during classroom consolidation exercises. For another example, for exam application scenarios, purposeful and targeted test papers can be set according to the predicted scores of each test question to be set by each user to be tested within the specified evaluation range, according to the normal distribution requirements of the exam scores or other requirements.

[0249] In addition, it should be noted that this embodiment is also conducive to the scientific assessment of the ability differences between students. This embodiment can map student groups in different regions to the same question space based on the mastery status of each knowledge point by the users to be tested in each region, as well as the predicted scores of the users to be tested in each region for different questions, so as to achieve comparability of scores in different regions, making it convenient for teaching and research personnel to understand the normalized academic ability distribution of different student groups, thereby formulating the next teaching plan.

[0250] In this embodiment, it should be noted that the knowledge ability tracking model is modeled based on the student behavior sequence. Tracking the student's knowledge mastery status is the core and key to building an adaptive teaching system. In an adaptive teaching system, whether it is accurate question push or student learning path planning, the first step is to accurately estimate the student's knowledge mastery.

[0251] The importance of knowledge and ability tracking model for personalized teaching:

[0252] First, the student knowledge and ability tracking model can predict students' academic behavior. An excellent student knowledge and ability tracking model needs to have the ability to predict scores. An important evaluation indicator of the model is the accuracy of score prediction. Score prediction is defined as: obtaining a student's learning history and the representation of a question, the model can predict the student's score on the question. The personalized teaching system can accurately predict the student's answering situation. First, it is conducive to improving students' autonomous learning efficiency, because the system can gradually push moderately difficult exercises to students, avoiding recommending too difficult or too simple questions, wasting students' time and energy; secondly, it is conducive to improving teachers' teaching efficiency. For example, when teachers are compiling papers, the system can tell teachers the predicted scores of students in this class on each test question, which is convenient for teachers to compile papers more purposefully; finally, it is conducive to scientifically evaluating the ability differences between students. Through score prediction, student groups from different schools and regions can be mapped to the same question space to achieve comparable scores, which is convenient for teaching and research personnel to understand the normalized academic ability distribution of different student groups, so as to formulate the next teaching plan.

[0253] Second, the student knowledge and ability tracking model can provide personalized diagnosis of students' academic level. The goal of student ability diagnosis is to model students' academic ability level at the knowledge level. Combined with the teaching and research experience of education experts, it can help students discover their own weak knowledge points so that students can learn independently and assist teachers in targeted teaching; personalized learning goes a step further and directly recommends appropriate learning resources to students based on their knowledge status and prior knowledge in the field of education, such as learning videos and test questions. Using the historical learning data of individual students and related groups to provide accurate learning plans, breaking the one-size-fits-all learning method in traditional teaching, can effectively improve students' learning efficiency.

[0254] Based on the same inventive concept, another embodiment of the present invention provides an intelligent learning knowledge ability tracking device, including: an acquisition module and a prediction module; wherein:

[0255] An acquisition module is used to acquire the current question corresponding to the user to be tested and the prior correlation weights between the current question and each knowledge point;

[0256] A prediction module is used to input the identification of the user to be tested, the current question, and the prior correlation weights between the current question and each knowledge point into the knowledge ability tracking model to obtain the probability of the user to be tested answering the current question correctly; wherein the knowledge ability tracking model contains the mastery status of the user to be tested for each knowledge point, and the mastery status of the user to be tested for each knowledge point is determined based on the questions that the user to be tested has done in the historical time period, the answer results of the questions that the user to be tested has done, and the prior correlation weights between the questions that the user to be tested has done and each knowledge point;

[0257] Among them, the knowledge and ability tracking model is obtained by inputting the questions done by each user in the historical training data, the answer results of the questions done by each user, and the prior correlation weights between the questions done by each user and each knowledge point into the initial knowledge and ability tracking model, and training based on machine learning.

[0258] Since the intelligent learning knowledge ability tracking device provided in the embodiment of the present invention can be used to execute the intelligent learning knowledge ability tracking described in the above embodiment, its working principle and beneficial effects are similar, so they will not be described in detail here. For specific contents, please refer to the introduction of the above embodiment.

[0259] In this embodiment, it should be noted that the various modules in the device of the embodiment of the present invention can be integrated into one or deployed separately. The above modules can be combined into one module or further divided into multiple sub-modules.

[0260] Based on the same inventive concept, another embodiment of the present invention provides an electronic device, see Figure 3 , the electronic device specifically includes the following contents: a processor 301, a memory 302, a communication interface 303 and a communication bus 304;

[0261] The processor 301, the memory 302, and the communication interface 303 communicate with each other via the communication bus 304;

[0262] The processor 301 is used to call the computer program in the memory 302. When the processor executes the computer program, all steps of the above-mentioned intelligent learning knowledge and ability tracking method are implemented. For example, when the processor executes the computer program, the following process is implemented: obtaining the current question corresponding to the user to be tested and the prior correlation weight between the current question and each knowledge point; inputting the user identification to be tested, the current question, and the prior correlation weight between the current question and each knowledge point into the knowledge and ability tracking model to obtain the probability of the user to be tested answering the current question correctly; wherein, the knowledge and ability tracking model contains the mastery status of the user to be tested for each knowledge point, and the mastery status of the user to be tested for each knowledge point is determined based on the questions done by the user to be tested in the historical time period, the answer results of the questions done by the user to be tested, and the prior correlation weights of the questions done by the user to be tested and each knowledge point; wherein, the knowledge and ability tracking model is obtained after training based on machine learning by inputting the questions done by each user in the historical training data, the answer results of the questions done by each user, and the prior correlation weights of the questions done by each user and each knowledge point into the initial knowledge and ability tracking model.

[0263] It can be understood that the detailed functions and extended functions that can be executed by the computer program can refer to the description of the above embodiments.

[0264] In addition, the processor may also implement the steps of the resource pushing method described in the above embodiment when executing the computer program, and / or the processor may implement the steps of the test paper assembly method described in the above embodiment when executing the program.

[0265] Based on the same inventive concept, another embodiment of the present invention provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, all the steps of the above-mentioned intelligent learning knowledge ability tracking method are implemented. For example, when the processor executes the computer program, the following process is implemented: obtaining the current question corresponding to the user to be tested, and the prior correlation weights between the current question and each knowledge point; inputting the user identification to be tested, the current question, and the prior correlation weights between the current question and each knowledge point into the knowledge ability tracking model, and obtaining the correct answer of the user to be tested to the current question. probability; wherein the knowledge and ability tracking model includes the mastery status of each knowledge point by the user to be tested, and the mastery status of each knowledge point by the user to be tested is determined based on the questions done by the user to be tested in a historical time period, the answer results of the questions done by the user to be tested, and the prior correlation weights between the questions done by the user to be tested and each knowledge point; wherein the knowledge and ability tracking model is obtained after training based on machine learning by inputting the questions done by each user in the historical training data, the answer results of the questions done by each user, and the prior correlation weights between the questions done by each user and each knowledge point into the initial knowledge and ability tracking model.

[0266] It can be understood that the detailed functions and extended functions that can be executed by the computer program can refer to the description of the above embodiments.

[0267] In addition, the processor may also implement the steps of the resource pushing method described in the above embodiment when executing the computer program, and / or the processor may implement the steps of the test paper assembly method described in the above embodiment when executing the program.

[0268] In addition, the logic instructions in the above-mentioned memory can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0269] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the embodiments of the present invention. Those of ordinary skill in the art may understand and implement them without creative effort.

[0270] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the intelligent learning knowledge ability tracking method described in each embodiment or some parts of the embodiment.

[0271] In addition, in the present invention, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0272] In addition, in the present invention, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0273] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for tracking intelligent learning knowledge and ability, characterized in that: include: Obtain the current question corresponding to the user to be tested, and the prior correlation weights between the current question and each knowledge point; Input the identification of the user to be tested, the current question, and the prior correlation weights between the current question and each knowledge point into the knowledge ability tracking model to obtain the probability of the user to be tested answering the current question correctly; wherein the knowledge ability tracking model includes the mastery status of each knowledge point by the user to be tested, and the mastery status of each knowledge point by the user to be tested is determined based on the questions that the user to be tested has done in a historical time period, the answer results of the questions that the user to be tested has done, and the prior correlation weights between the questions that the user to be tested has done and each knowledge point; The knowledge and ability tracking model is obtained by inputting the questions each user has done in the historical training data, the answer results of the questions each user has done, and the prior correlation weights between the questions each user has done and each knowledge point into the initial knowledge and ability tracking model, and training based on machine learning; The knowledge ability tracking model includes a topic knowledge point mapping model; Accordingly, the user ID to be tested, the current question, and the prior correlation weights between the current question and each knowledge point are input into the knowledge ability tracking model to obtain the probability of the user to be tested answering the current question correctly, specifically including: Inputting the topic feature data of the current topic into the topic knowledge point mapping model to obtain the first correlation weight between the current topic and each knowledge point; Determine the priori correlation weights between the current question and each knowledge point according to the knowledge point information marked on the current question; Determine the final correlation weight between the current question and each knowledge point according to the first correlation weight between the current question and each knowledge point and the prior correlation weight between the current question and each knowledge point; Determine the knowledge point status characteristic data of the user to be tested for all the knowledge points in the current question according to the mastery status of the user to be tested for each knowledge point and the final correlation weight between the current question and each knowledge point; Integrate the knowledge point state feature data and the question feature data to obtain a feature vector containing both question information and knowledge point state information; The probability of the user to be tested answering the current question correctly is predicted based on the feature vector containing both the question information and the knowledge point status information.

2. The intelligent learning knowledge ability tracking method according to claim 1 is characterized in that: Also includes: The answer result of the user to be tested for the current question is obtained, and the mastery status of each knowledge point of the user to be tested is updated according to the answer result.

3. The intelligent learning knowledge ability tracking method according to claim 1 is characterized in that: The step of inputting the identification of the user to be tested, the current question, and the prior correlation weights between the current question and each knowledge point into the knowledge ability tracking model to obtain the probability of the user to be tested answering the current question correctly specifically includes: Determine the mapping relationship between the current question and each knowledge point according to the final correlation weight between the current question and each knowledge point; Obtaining the mastery status of each knowledge point of the user to be tested from the knowledge point mastery status matrix corresponding to the user to be tested; wherein the knowledge point mastery status matrix stores the mastery status of each knowledge point of the user to be tested; wherein the knowledge point mastery status is used to characterize the mastery degree of the knowledge point of the user to be tested; According to the answer result of the user to be tested to the current question and the mapping correspondence between the current question and each knowledge point, the mastery status of each knowledge point of the user to be tested in the knowledge point mastery status matrix is ​​updated.

4. The intelligent learning knowledge ability tracking method according to claim 3 is characterized in that: Determining the priori relevance weights between the current question and each knowledge point according to the knowledge point information marked on the current question includes: When n knowledge points are marked on the current question, the priori relevance weights corresponding to the n knowledge points are determined to be 1 / n respectively, and the priori relevance weights of other knowledge points are determined to be 0, n≥1.

5. The intelligent learning knowledge ability tracking method according to claim 3 is characterized in that: Determining the final correlation weight between the current question and each knowledge point according to the first correlation weight between the current question and each knowledge point and the prior correlation weight between the current question and each knowledge point includes: The final correlation weights between the current question and each knowledge point are determined according to the first relationship model, wherein the first relationship model is: Using the weighted average method, ； in, is an adjustable parameter in the range [0, 1]; Indicates the current question and knowledge point i The final relevance weight of Indicates the first relevance weight between the current question and the knowledge point; represents the prior correlation weight between the current question and the knowledge point; M represents the question-knowledge point mapping matrix, which is used to store the mapping between the current question and the knowledge point, and its size is , n is the number of knowledge points, d m is the dimension of the feature data of the current question; M i Indicates the first i Line, m t Represents the topic feature data of the current topic; wherein, The expression of the function is , where i=1...C.

6. The intelligent learning knowledge ability tracking method according to any one of claims 3 to 5, characterized in that: The knowledge ability tracking model also includes a forgetting decay network model; Accordingly, the method further comprises: Obtaining the time interval between the current and the last update of the knowledge point mastering state matrix; wherein the time interval is used to input the forgetting decay network model to obtain the decay rate; Accordingly, obtaining the mastery status of each knowledge point of the user to be tested from the knowledge point mastery status matrix corresponding to the user to be tested includes: The decay rate is used to decay and update the knowledge point mastery status in the knowledge point mastery status matrix corresponding to the user to be tested, and the knowledge point mastery status after decay and update is obtained.

7. The intelligent learning knowledge ability tracking method according to claim 6 is characterized in that: The forgetting decay network model is: ; where d t represents the decay rate, t t Indicates the time interval, W d and b d represents the network parameters, W d and b d The value of is obtained through training; sigmoid The expression of the function is ; Correspondingly, the decay rate is used to decay and update the knowledge point mastering state in the knowledge point mastering state matrix corresponding to the user to be tested, including: Attenuation update of the knowledge point mastery state in the knowledge point mastery state matrix corresponding to the user to be tested is performed using a second relationship model according to the attenuation rate; Among them, the second relationship model is: ; Among them, N t represents the knowledge point mastering state matrix after attenuation update, which is used to store the mastering state of each knowledge point by the user to be tested; the size is , n is the number of knowledge points, d n The dimension representing the mastery status of knowledge points, N t-1 Represents the most recently updated knowledge point mastering state matrix, N ti N t No. i OK, N t-1,i N t-1 No. i OK, represents dot product, 1 represents all-1 vector, Indicates the current question and knowledge point i The final relevance weight of express The transposed matrix of .

8. The intelligent learning knowledge ability tracking method according to claim 7, characterized in that: According to the mastery status of each knowledge point by the user to be tested and the final correlation weight between the current question and each knowledge point, the knowledge point status characteristic data of the user to be tested for all knowledge points in the current question is determined, including: The knowledge point state characteristic data of the user to be tested for all the knowledge points in the current question is determined according to the third relationship model; wherein the third relationship model is: ； Among them, k t Represents the knowledge point status feature data of the user to be tested for all knowledge points in the current question.

9. The intelligent learning knowledge ability tracking method according to claim 8, characterized in that: The knowledge point state feature data and the question feature data are integrated together to obtain a feature vector containing both question information and knowledge point state information, including: The knowledge point state feature data k t and topic feature data m t Integrate them together and get the feature vector s through a fully connected neural network t : ； Among them, the expression of tanh function is: , W s and b s Are two network parameters obtained through training.

10. The intelligent learning knowledge ability tracking method according to claim 9, characterized in that: The knowledge and ability tracking model also includes: an ability network model and a question difficulty network model; Accordingly, according to the feature vector including both the question information and the knowledge point status information, predicting the probability of the user to be tested answering the current question correctly includes: Obtaining the ability value of the user to be tested for the current question according to a feature vector including both question information and knowledge point status information, and an ability network model corresponding to the user to be tested; Based on the question feature data and the question difficulty network model, obtaining a difficulty value of the current question; Based on the ability value and the difficulty value, the probability of the user to be tested answering the current question correctly is predicted.

11. The intelligent learning knowledge ability tracking method according to claim 10, characterized in that: According to the feature vector including both the question information and the knowledge point status information, and the ability network model corresponding to the user to be tested, the ability value of the user to be tested for the current question is obtained, including: Adopting the Capability Network Model Obtaining the ability value of the user to be tested for the current question; wherein, Indicates the ability value. and Represents the parameters of the capability network model, obtained through training; According to the question feature data and the question difficulty network model, the difficulty value of the current question is obtained, including: Using the Question Difficulty Network Model Get the difficulty value of the current question; where, Indicates the difficulty value, and The parameters of the network model representing the difficulty of the question are obtained through training; Based on the ability value and the difficulty value, predict the probability of the user to be tested answering the current question correctly, including: using a prediction model Predict the probability of the user to be tested answering the current question correctly; wherein, It represents the probability of answering correctly. According to IRT theory, the value of a is 3.

0.

12. The intelligent learning knowledge ability tracking method according to claim 10, characterized in that: According to the answer result of the user to be tested for the current question and the mapping relationship between the current question and each knowledge point, the mastery status of each knowledge point of the user to be tested in the knowledge point mastery status matrix is ​​updated, including: According to the answer result vector of the user to be tested for the current question v t And the mapping relationship between the current question and each knowledge point, first erase the original memory and then write the new memory; wherein the erase vector and the write vector are respectively e t and h t ; ； in, and is the network parameter, obtained in advance through training; W h and b h are network parameters, obtained in advance through training; The original memory is erased according to the fourth relational model, wherein the fourth relational model is: ； Write new memory according to the fifth relational model, where the fifth relational model is: ； Among them, N t+1,i represents the updated knowledge point mastering state matrix, N t+1,i N t+1 No. i OK.

13. The intelligent learning knowledge ability tracking method according to claim 11, characterized in that: When the answer result is a binary result of 0 and 1, the loss function is the binary cross entropy; when the answer result is a score rate [0,1], the loss function is the multivariate cross entropy.

14. A resource push method based on the intelligent learning knowledge ability tracking method according to any one of claims 1 to 13, characterized in that: include: According to the mastery status of each knowledge point by the user to be tested and / or the predicted answer results of the user to be tested to each question in the question bank, test questions and / or learning resources are pushed to the user to be tested.

15. A test paper composition method based on the intelligent learning knowledge ability tracking method according to any one of claims 1 to 13, characterized in that: include: Automatically generate the test paper based on the mastery status of each knowledge point of each user within the specified evaluation range and the knowledge points marked on the test paper to be generated; and / or, Automatically generate the test paper based on the predicted scores of each test question for each user within the specified evaluation range; and / or, According to the mastery status of each knowledge point by the users in each region, as well as the predicted scores of the users in each region for different questions, the student groups in different regions are mapped to the same question space to achieve comparability of scores in different regions.

16. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the intelligent learning knowledge ability tracking method as described in any one of claims 1 to 15 are implemented, and / or, when the processor executes the program, the steps of the resource pushing method as described in claim 14 are implemented, and / or, when the processor executes the program, the steps of the test paper assembling method as described in claim 15 are implemented.

Images