A cross-domain based adaptive exercise adaptation method and application

Abstract

The application discloses a kind of based on cross-domain adaptive exercise adaptation method, comprising: at least two different subjects learning data in student online learning system is respectively obtained, one is as source data domain, another is as target data domain;Select the similar problem text in the source data domain and the target data domain;According to the problem text in the selected source data domain and the problem text in the target data domain, construct knowledge state evaluation model;According to the knowledge state evaluation model, the knowledge state of corresponding student is evaluated, and the corresponding adaptive exercise is obtained according to the knowledge state;According to the exercise result obtained by the adaptive exercise, the above process is repeated to evaluate the knowledge state again, to realize the individualized exercise adaptation of adaptive learning.It can be based on a small amount of learning data accurately obtain student knowledge state, and realize intelligent exercise adaptation according to the knowledge state of student, to improve the efficiency of student learning and teacher teaching.

Description

Technical Field

[0001] This invention relates to the field of intelligent education technology, and in particular to a cross-domain adaptive exercise adaptation method, a cross-domain adaptive exercise adaptation device, an electronic device, and a computer-readable storage medium. Background Technology

[0002] In the era of big data, the resources in online learning systems have exploded. Students struggle to accurately and quickly select the learning materials they need from this vast amount of information, easily falling into an "information maze." Teachers, too, face challenges due to the large number of students and the characteristics of online learning systems, making it difficult to accurately understand each student's learning progress and level. This often leads to a "sea of ​​questions" approach, resulting in low efficiency for both student learning and teacher instruction.

[0003] To achieve adaptive exercise matching, it is essential to accurately obtain students' current knowledge status in order to match exercises accordingly. However, in current intelligent education scenarios, students' online learning data is often limited, resulting in inaccurate perception of students' knowledge status. Therefore, proposing a method that can accurately obtain students' knowledge status based on limited learning data to achieve intelligent exercise matching, thereby improving the efficiency of student learning and teacher instruction, is a pressing issue that needs to be addressed. Summary of the Invention

[0004] The technical problem to be solved by this invention is to address the shortcomings of existing methods by providing a cross-domain adaptive exercise matching method. This method can accurately obtain students' knowledge status based on a small amount of learning data and achieve intelligent exercise matching based on students' knowledge status, thereby improving the efficiency of student learning and teacher teaching.

[0005] On one hand, this invention proposes a cross-domain adaptive exercise matching method, comprising: acquiring learning data from at least two different subjects in a student's online learning system, one as a source data domain and the other as a target data domain; selecting question texts in the source data domain that are similar to those in the target data domain; constructing a knowledge state assessment model based on the selected question texts in the source data domain and the target data domain; assessing the knowledge state of the corresponding student based on the knowledge state assessment model, and obtaining corresponding matching exercises based on the knowledge state; repeating the above process again to assess the knowledge state based on the practice results obtained from the matching exercises, so as to achieve personalized exercise matching for adaptive learning.

[0006] In one embodiment of the present invention, selecting question texts in the source data domain that are similar to those in the target data domain includes: encoding the question in the source data domain through a bidirectional long short-term memory network layer and fusing the contextual features of the question; connecting the hidden layer feature vectors in two directions via a connection layer; combining the semantic representation vectors of all time steps into a single vector with max pooling via a max pooling layer; and decoding the single vector through the long short-term memory network layer to obtain a reconstruction of the question in the source data domain, thereby selecting question texts that achieve the target similarity.

[0007] In one embodiment of the present invention, constructing the knowledge state evaluation model includes: concatenating the question text in the source data domain and the question text in the target data domain and representing them as a semantic representation containing answer information; calculating the hidden state of the student's answer at each time step as the student's knowledge state based on the semantic representation; obtaining the probability of the student's correct answer to the current question based on the knowledge state and calculating the training loss function to construct the knowledge state evaluation model.

[0008] In one embodiment of the present invention, the construction of the knowledge state evaluation model further includes: minimizing the distribution difference of the evaluation results of the knowledge state evaluation model between the source data domain and the target data domain using the maximum average difference method to obtain the joint distribution loss of the knowledge state evaluation model; and training the knowledge state evaluation model by minimizing the joint distribution loss using the backpropagation algorithm.

[0009] In one embodiment of the present invention, the step of evaluating the knowledge status of the corresponding student according to the knowledge status evaluation model includes: inputting the historical learning data of all students in the target data domain into the knowledge status evaluation model to obtain the probability of each student answering all questions correctly; obtaining the knowledge level vector matrix of all students according to the probability of each student answering all questions correctly; determining the similarity vectors of the knowledge level vector matrix using cosine similarity, and obtaining the individual students with similar knowledge status according to the similarity vectors.

[0010] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:

[0011] (1) The cross-domain adaptive exercise matching method proposed in this invention utilizes the principle of transfer learning to evaluate and train students’ knowledge status based on their learning data from different subjects. The proposed knowledge status evaluation model can accurately obtain students’ current knowledge status and perform intelligent exercise matching even when students’ historical practice answer data is insufficient, by using historical answer data from other subjects to assist in training. This can effectively improve the efficiency of students’ learning and teachers’ teaching.

[0012] (2) By inputting the historical learning data of all students in the target data domain into the knowledge status assessment model for training, students with similar knowledge status are obtained. When assessing the knowledge status of individual students, information of other students with similar knowledge status is added to the reference, which can propose an accurate relative difficulty index of exercises, making the intelligent exercise adaptation more in line with the real education scenario. Attached Figure Description

[0013] Figure 1 This is a flowchart of a cross-domain adaptive exercise adaptation method proposed in an embodiment of the present invention;

[0014] Figure 2 This is a schematic diagram of the functional module for selecting problem text as proposed in an embodiment of the present invention;

[0015] Figure 3 This is a schematic diagram of the overall functional modules of an adaptive exercise matching method based on cross-domain proposed in an embodiment of the present invention;

[0016] Figure 4 This is a schematic diagram of the structure of a cross-domain adaptive exercise adaptation device proposed in an embodiment of the present invention;

[0017] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention;

[0018] Figure 6 This is a schematic diagram of the structure of a computer-readable storage medium provided in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0020] like Figure 1As shown, the first embodiment of the present invention proposes a cross-domain adaptive exercise adaptation method, which includes, for example: step S1, acquiring learning data from at least two different subjects in a student's online learning system, one as the source data domain and the other as the target data domain; step S2, selecting question texts in the source data domain that are similar to those in the target data domain; step S3, constructing a knowledge state assessment model based on the selected question texts in the source data domain and the target data domain; step S4, assessing the knowledge state of the corresponding student based on the knowledge state assessment model, and obtaining corresponding adaptive exercises based on the knowledge state; step S5, repeating the above process again to assess the knowledge state based on the practice results obtained from the adaptive exercises, so as to achieve personalized exercise adaptation for adaptive learning.

[0021] In step S1, for example, the source data field x is obtained. s and target data domain x T Source data field x S =(s1,s2,s3s4,.....,s t ),s t =(g t ,r t ); where g t r represents the question answered by the student at time step t. t This represents the corresponding score. If the student answers question g correctly... t Then r t =1, otherwise r t =0. Target data domain x T =(d1,d2,d3,....,d t ), d t =(z t ,a t ); where z t a represents the question answered by the student at time step t. t This represents the corresponding score. If the student answers question z correctly... t Then a t =1, otherwise a t =0. For the source data domain x S The complete representation of the student's answer to the question, g′ t = (w1, w2, w3, ..., w3); w is a single word embedded by a pre-trained Word2Vec, and L is the number of words.

[0022] In step S2, combined Figure 2 As shown, to fully consider the contextual semantic relationships of a problem, Bi-LSTM is used to reconstruct the problem. The pre-trained problem representation g... t′ The computation process of the forward long short-term memory network fed into Bi-LSTM is as follows:

[0023]

[0024] Where, w L This represents the Lth pre-trained word; h t-1 w represents the hidden state value at time t-1; f ,w i ,w o ,w a ,u f ,u i ,u o ,u a b represents the weighting coefficient; f b i ,b o ,b a represents the bias value; tanh represents the tangent hyperbolic function; σ represents the activation function Sigmoid; ⊙ represents the Hadamard product.

[0025] The computation process for the inverse long short-term memory network is similar, except for the weight coefficients and bias values. To fully utilize contextual information, the hidden state at each time step of the Bi-LSTM is used as the output. Since Bi-LSTM is used, the output of the hidden layer is... and t represents time.

[0026] Since the hidden state in each direction only contains context information for one direction, the hidden states in both directions... and It is beneficial to combine them into a vector to capture the contextual semantic information of each word. Therefore, the semantic representation at each time step is:

[0027]

[0028] Then, the semantic representations of all time steps are combined into a single vector with max pooling:

[0029] q j =max(η 1j ,η 2j ,η 3j ,η 4j ,... ,η Lj (3)

[0030] Here, the j-th dimension of the embedding vector q can be formalized as q j η ij It is the semantic representation η of the i-th step. i The j-th dimension.

[0031] The obtained problem representation q is used as the initial hidden state and fed into a Long Short-Term Memory network to obtain the reconstruction of each problem in the source domain. Regarding the problem of the target data domain z t ′ =(e1,e2,e3,...,e L ); e is a single word embedded by a pre-trained Word2Vec algorithm; L is the number of words. The reconstruction for each problem in the target domain is obtained using the same method. If a model has relatively small errors for two data domains simultaneously, then the data in those two domains have relatively high similarity. Therefore, the reconstruction error is constructed as follows:

[0032]

[0033] To select a problem, for example, a variable 'u' is added for selection. S In the loss function, the total loss function is as follows:

[0034]

[0035] Where, n s ,n T Indicates the number of questions in the domain and the target domain; These are the reconstructions of the source domain and the target domain problems, respectively. It is a selection indicator. 1 indicates selection; 0 indicates no selection. The loss function Γ(u) S ) is u S Regular terms on, to avoid u S The case where all values ​​are zero is defined as follows:

[0036]

[0037] Where λ∈{0,1} represents the importance of the regularization term. The parameter of the automatic selector for the alternating update problem is θ. auto and u S When u is fixed S When, θ auto The backpropagation algorithm can be used to update when θ auto When fixed, u S Update using the following method:

[0038]

[0039] Through the above training process, the problems selected from the source domain are denoted as follows: The corresponding reply is recorded as The problem of the target data domain is denoted as The corresponding reply is recorded as And then embed them as described above.

[0040] In step S3, combined Figure 3 As shown, the training process of the knowledge state assessment model is as follows:

[0041] The problem automatic selection module obtained and Place it in the connector layer, as shown in the following formula:

[0042]

[0043] Obtain a semantic representation containing the answer information. N represents the number of questions.

[0044] The obtained semantic representation The hidden state h at each time step is obtained by feeding it into the feature computation layer (LSTM). t This is considered as the student's knowledge status.

[0045] The obtained student status h t The values ​​are then classified using a Softmax function in the output layer. The Softmax function maps values ​​to a range of [0,1], thus yielding the probability y of a student correctly answering the current question. t :

[0046] y t = sigmod(W out *h t +b out (9)

[0047] Among them, W out With b out These are the output layer parameters. The training loss function is as follows:

[0048]

[0049] Where n is the number of students, i and t represent the i-th student and the t-th time step; ι(.,.) represents the binary cross-entropy loss; This indicates that student i answers the question at time t+1. Independent hot coding. The true label indicates whether the question was answered correctly.

[0050] To minimize the difference between the two domain distributions, the output layer of the trained knowledge state evaluation model is discarded and replaced with an adaptive layer (fully connected layer), while the other layers remain unchanged. The data selected in step S2 is then used: And the data in the target domain: The data is fed into the knowledge state assessment model for training. The model parameters are shared during training with data from both data domains. Specifically, the process is as follows:

[0051] The output of the feature computation layer in the knowledge state assessment model: Place it into the adaptive layer, which can be viewed as a function f:

[0052] α t =adaptation(h t ;θ adp )=f(h t (11)

[0053] Among them, h t The output of the feature computation layer represents the student's knowledge state. θ adp This represents the parameters of the adaptive layer.

[0054] The output of the adaptive layer in the knowledge state assessment model: To minimize domain differences, the maximum average difference method is used for calculation:

[0055]

[0056] The joint distributed loss of the entire knowledge state assessment model is:

[0057]

[0058] Here, γ controls the weights of the two terms. The model is trained by minimizing the joint distribution loss using the backpropagation algorithm.

[0059] A well-trained knowledge state evaluation model is obtained through the above process. However, since the number of questions differs between the source and target domains, the dimensions of the output layer also differ between the two domains. Therefore, before applying the model to the target domain, it is necessary to abandon the output layer trained on the source data domain and train a new output layer with the target domain data after the adaptive layer. That is, by minimizing formula (10), only the parameters of the new output layer are updated.

[0060] Furthermore, in step S4, the implementation process of personalized exercise adaptation is as follows:

[0061] The historical target domain learning data of all students are fed into the knowledge state assessment model to obtain the probability y of each student's correct answer to all questions. t The knowledge mastery level of all students is represented by the vector matrix U. m×N Where m represents the number of students and N represents the number of questions.

[0062] Because matching exercises based on individual knowledge levels does not take into account information about similar students, it ignores the group nature of learning. Therefore, in this embodiment, the knowledge mastery level matrix U... m×N For example, using cosine similarity to judge vector U i and U j The degree of similarity. Assume U i ={a1,a2,a 3, ...,a n},U j ={b1,b2,b3,...,b n},have:

[0063]

[0064] Cosine similarity is used to determine the degree of similarity between two vectors. Furthermore, students with cosine similarity scores exceeding a certain threshold can be selected as similar users. For example, for target user i, their final exercise score vector f is obtained. i :

[0065] f i =ρ·U i +(1-ρ)·average(N i (15)

[0066] Among them, U i Let N be the student's knowledge level vector, and average(N) be the average(N) of the student's knowledge level vector. i ) represents the average of the knowledge level vectors of similar students, and ρ is the control ratio.

[0067] When generating exercise recommendations, a suitable difficulty range [β1, β2] (β2 > β1) is determined as the difficulty interval of the exercises. Then, the exercise score vector f of student i can be used as a reference. i The system recommends exercises to students with a correct answer probability between β2 and β1. For example, if β1 is set to 0.6 and β2 is set to 0.8, then exercises with a correct answer probability between 0.6 and 0.8 can be recommended to students based on the predicted exercise score vector.

[0068] In step S5, after students practice the personalized exercises, the system reassesses their knowledge level and adjusts their learning methods and the difficulty parameters of the exercise matching process accordingly.

[0069] In summary, the cross-domain adaptive exercise matching method proposed in the first embodiment of this invention utilizes the principle of transfer learning to evaluate and train students' knowledge status based on their learning data from different subjects. The proposed knowledge status evaluation model can accurately obtain students' current knowledge status even when historical practice answer data is insufficient, by using historical answer data from other subjects to assist in training. This enables intelligent exercise matching, effectively improving the efficiency of student learning and teacher instruction. By inputting the historical learning data of all students in the target data domain into the knowledge status evaluation model for training, individual students with similar knowledge statuses are obtained. When evaluating the knowledge status of individual students, information from other students with similar knowledge statuses is added for reference, resulting in accurate relative difficulty indicators for exercises. This makes the intelligent exercise matching more consistent with real-world educational scenarios.

[0070] In addition, such as Figure 4 As shown, the second embodiment of the present invention proposes a cross-domain adaptive exercise adaptation device 20, which includes, for example, a learning data acquisition module 201, a question text selection module 202, a knowledge state evaluation model construction module 203, an exercise adaptation module 204, and an adaptive learning module 205.

[0071] The learning data acquisition module 201 acquires learning data from at least two different subjects in the student's online learning system, one as the source data domain and the other as the target data domain. The question text selection module 202 selects question texts from the source data domain that are similar to those from the target data domain. The knowledge state assessment model construction module 203 constructs a knowledge state assessment model based on the selected question texts from the source and target data domains. The exercise adaptation module 204 assesses the knowledge state of the corresponding student based on the knowledge state assessment model and obtains corresponding adapted exercises based on the knowledge state. The adaptive learning module 205 repeats the above process to reassess the knowledge state based on the practice results obtained from the adapted exercises, thereby achieving personalized exercise adaptation for adaptive learning.

[0072] It is worth mentioning that the cross-domain adaptive exercise adaptation method implemented by the cross-domain adaptive exercise adaptation device disclosed in the second embodiment of the present invention is as described in the first embodiment above, and therefore will not be described in detail here. Optionally, each component and the other operations or functions described above are for implementing the method described in the first embodiment, and the beneficial effects of this embodiment are the same as those of the first embodiment above. For the sake of brevity, they will not be repeated here.

[0073] like Figure 5As shown, the third embodiment of the present invention also provides an electronic device, which includes: a memory and one or more processors connected to the memory, the memory storing a computer program, and the processor executing the computer program to implement the steps of the cross-domain adaptive exercise adaptation method described in the first embodiment above, and the electronic device provided in this embodiment has the same beneficial effects as the cross-domain adaptive exercise adaptation method provided in the first embodiment.

[0074] like Figure 6 As shown, the fourth embodiment of the present invention also provides a computer-readable storage medium storing a computer program thereon. When the program is executed by a processor, it implements the steps of the cross-domain adaptive exercise adaptation method described in the first embodiment above. The computer-readable storage medium provided in this embodiment has the same beneficial effects as the cross-domain adaptive exercise adaptation method provided in the first embodiment.

[0075] The computer-readable storage medium may include, but is not limited to, any type of disk, including floppy disks, optical disks, DVDs, CD-ROMs, microdrives, as well as magneto-optical disks, ROMs, RAMs, EPROMs, EEPROMs, DRAMs, VRAMs, flash memory devices, magnetic cards or optical cards, nanosystems (including molecular memory ICs), or any type of medium or device suitable for storing instructions and / or data.

[0076] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0077] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0078] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage device, which may include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0079] The foregoing description is merely an exemplary embodiment of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure. Those skilled in the art will readily conceive of embodiments of this disclosure upon considering the specification and practicing the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described herein. The specification and embodiments are to be considered exemplary only, and the scope and spirit of this disclosure are defined by the claims.

[0080] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0081] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A cross-domain adaptive exercise adaptation method, characterized in that, include: Acquire learning data from at least two different subjects in the student's online learning system, with one subject as the source data domain and the other as the target data domain; Select question text from the source data domain that is similar to that from the target data domain; A knowledge state assessment model is constructed based on the question texts in the selected source data domain and the question texts in the target data domain. The knowledge status of the corresponding student is assessed according to the knowledge status assessment model, and the corresponding suitable exercises are obtained according to the knowledge status. Based on the practice results obtained from the adapted exercises, repeat the above process to re-evaluate the knowledge state, so as to achieve personalized exercise adaptation for adaptive learning; The step of selecting question texts in the source data domain that are similar to those in the target data domain includes: The problem in the source data domain is encoded through a bidirectional long short-term memory network layer, and the contextual features of the problem are fused together. The output of the bidirectional long short-term memory network layer connects the hidden layer feature vectors in two directions via a connection layer. The output of the connection layer is combined into a single vector with max pooling by a max pooling layer, which combines the semantic representation vectors of all time steps. The single vector is decoded by a long short-term memory network layer to reconstruct the problem in the source data domain, so as to select the problem text that achieves the target similarity. The construction of the knowledge state assessment model includes: The question text in the source data domain and the question text in the target data domain are concatenated and represented as a semantic representation containing answer information; The hidden state of the student's answer at each time step is calculated based on the semantic representation as the student's knowledge state; Based on the knowledge state, the probability of the student's correct answer to the current question is obtained, and the training loss function is calculated to construct the knowledge state evaluation model.

2. The cross-domain adaptive exercise adaptation method according to claim 1, characterized in that, The construction of the knowledge state assessment model also includes: The maximum average difference method is used to minimize the distribution difference of the evaluation results of the knowledge state evaluation model between the source data domain and the target data domain, thereby obtaining the joint distribution loss of the knowledge state evaluation model; The knowledge state evaluation model is trained by minimizing the joint distribution loss using the backpropagation algorithm.

3. The cross-domain adaptive exercise adaptation method according to claim 1, characterized in that, The step of assessing the knowledge status of the corresponding student according to the knowledge status assessment model includes: Input the historical learning data of all students in the target data domain into the knowledge state assessment model to obtain the probability of each student answering all questions correctly. The knowledge level vector matrix of all students is obtained based on the probability of each student answering all questions correctly. The knowledge level vector matrix is ​​used to determine similar vectors using cosine similarity, and student individuals with similar knowledge states are obtained based on the similar vectors.

4. A cross-domain adaptive exercise adaptation device, characterized in that, include: The learning data acquisition module is used to acquire learning data from at least two different subjects in the student's online learning system, with one subject as the source data domain and the other as the target data domain. The question text selection module is used to select question texts in the source data domain that are similar to those in the target data domain; The knowledge state assessment model construction module is used to construct a knowledge state assessment model based on the question text in the selected source data domain and the question text in the target data domain. The exercise matching module is used to assess the knowledge status of the corresponding student according to the knowledge status assessment model, and obtain the corresponding matching exercises according to the knowledge status. The adaptive learning module is used to repeat the above process and re-evaluate the knowledge state based on the practice results obtained from the adapted exercises, so as to achieve personalized exercise adaptation for adaptive learning. The question text selection module is specifically used for: The problem in the source data domain is encoded through a bidirectional long short-term memory network layer, and the contextual features of the problem are fused together. The output of the bidirectional long short-term memory network layer connects the hidden layer feature vectors in two directions via a connection layer. The output of the connection layer is combined into a single vector with max pooling by a max pooling layer, which combines the semantic representation vectors of all time steps. The single vector is decoded by a long short-term memory network layer to reconstruct the problem in the source data domain, so as to select the problem text that achieves the target similarity. The knowledge state assessment model construction module is specifically used for: The question text in the source data domain and the question text in the target data domain are concatenated and represented as a semantic representation containing answer information; The hidden state of the student's answer at each time step is calculated based on the semantic representation as the student's knowledge state; Based on the knowledge state, the probability of the student's correct answer to the current question is obtained, and the training loss function is calculated to construct the knowledge state evaluation model.

5. An electronic device, characterized in that, include: A memory and one or more processors connected to the memory, the memory storing a computer program, the processors being configured to execute the computer program to implement the steps of the method according to any one of claims 1-3.

6. A computer-readable storage medium, characterized in that, It stores a computer program executable by an access authentication device, which, when run on the access authentication device, causes the access authentication device to perform the steps of the method described in any one of claims 1-3.

Images