Knowledge tracking method, system and storage medium integrating learning behavior characteristics
Through the self-attention mechanism, the similarity between concepts and mastery states is calculated, combined with learning behavior characteristics, the concept mastery matrix is updated, which solves the problem of ignoring similarity and forgetful behavior in the existing technology, and improves the prediction accuracy of knowledge state tracking.
Patent Information
- Application Number
- CN202210746234.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-06-29
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2042-06-29
AI Technical Summary
The existing knowledge tracking method ignores the impact of similarity between concepts on the learning process, relies solely on exercise interaction data, and does not model forgetful behavior, resulting in insufficient prediction accuracy.
The self-attention mechanism is used to calculate the similarity between concepts and the similarity between mastery states, combine learning behavior characteristics, update the concept mastery matrix, and use correlation weights and behavioral attention vectors to predict the answering situation.
The prediction accuracy of knowledge state tracking is improved, similarity between concepts and forgetful behaviors are taken into account, and more learning behavior characteristics are input, which enhances the model's explanatory nature of the learning process.
Smart Images

Figure CN115062716B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of knowledge tracking technology, and more specifically, to a knowledge tracking method, system and storage medium integrating learning behavior characteristics. Background Art
[0002] At present, knowledge tracing methods mainly come from three types of models: Bayesian Knowledge Tracing (BKT), Deep Knowledge Tracing (DKT) and Dynamic Key-Value Memory Networks for Knowledge Tracing (DKVMN). Bayesian Knowledge Tracing uses the probability of going from an unlearned state to a learned state, the probability of forgetting previously learned concepts, the probability of correctly guessing without mastering concepts, and the probability of mastering concepts but making mistakes to track students' knowledge status, and updates the probability value in the students' exercise interaction data through the Hidden Markov Model (HMM). The BKT model not only assumes that students will not forget the knowledge they have learned, that is, the probability of forgetting previously learned concepts is 0, but also models each concept separately. But in fact, forgetting behavior accompanies the beginning and end of the learning process; concepts are not independent of each other, but are clearly hierarchical and interconnected. Deep knowledge tracking uses a recurrent neural network (RNN) or a long short-term memory network (LSTM) to track the student's knowledge state, represented by a hidden state vector. The dynamic key-value memory network uses a key matrix to store concepts and a value matrix to store the student's mastery of the concepts. Through these two matrices, the DKVMN model can accurately point out the student's specific knowledge state on each concept. Although the three models have achieved good results in predicting students' future answering, the data input to the model is only the student's exercise interaction data, and the parameter update process does not model forgetting behavior.
[0003] However, these methods still have defects. First, they ignore the impact of similarities between concepts on the learning process. Second, they use students' exercise interaction data to track their knowledge status, ignoring the fact that students have other rich learning behavior characteristics. Third, they do not model students' forgetting behavior during the learning process. Summary of the invention
[0004] In response to at least one defect or improvement need in the prior art, the present invention provides a knowledge tracking method, system and storage medium that integrate learning behavior characteristics, which improves prediction accuracy by using self-attention to explore similarities between concepts.
[0005] To achieve the above object, according to a first aspect of the present invention, a knowledge tracking method integrating learning behavior characteristics is provided, comprising:
[0006] The self-attention mechanism is used to calculate the similarity of concepts in the concept matrix. The concept matrix is used to represent the concepts related to the questions in the course. The calculation result is represented by the concept attention matrix. The similarity between the concepts contained in the predicted exercises is calculated based on the concept attention matrix, and the calculation result is represented by the association weight.
[0007] The self-attention mechanism is used to calculate the similarity between different concept mastery states in the concept mastery matrix. The concept mastery matrix is used to represent the learner's mastery state of the concept, and the calculation result is represented by the state attention matrix.
[0008] The self-attention mechanism is used to mine the connections between different features in the feature vector describing the user's behavior during the learning process, and the behavior attention vector is output;
[0009] Update the concept mastery matrix using the state attention matrix and behavior attention vector;
[0010] Multiply the association weight with the concept mastery matrix to obtain the vector of the student's concept mastery status. Input the vector of the student's concept mastery status and the embedding vector of the exercise to be predicted into the feedforward neural network to predict the user's answer.
[0011] Furthermore, the feature vector describing the user's behavior during learning is recorded as It includes exercise interaction data X and learning behavior feature data. Exercise interaction data X is used to describe the correctness of the user's answers at multiple moments before the prediction time. The learning behavior feature data includes the first type of behavior feature data and the second type of behavioral characteristic data Including several predefined behavioral feature data directly obtained from user behavior collection data, It includes several predefined behavior feature data related to the user's forgetting behavior calculated from the user behavior collection data.
[0012] Further, The three features of answer times, first response time, and request prompt times form a row vector with a dimension of 3. The number of times the same concept is repeatedly learned, the time interval between learning, and the time interval between learning the same concept form a row vector with a dimension of 3.
[0013] Furthermore, the feature vector of the user's behavior during learning is recorded as Mining with self-attention mechanism The connections between the different features in include:
[0014] A feedforward autoencoder with self-attention mechanism is used Perform dimensionality reduction to obtain the reduced dimensionality vector h t , where the feedforward autoencoder includes an encoder and a decoder. The encoder includes an input layer, a self-attention layer, a fully connected layer, and a hidden layer. The decoder includes a hidden layer, a fully connected layer, a self-attention layer, and an output layer. The encoder and the decoder are distributed in a mirror image.
[0015] Then use the self-attention mechanism to discover h t The connection between internal features is output as the behavior attention vector, denoted as o t .
[0016] Further,
[0017]
[0018] Softmax() represents the Softmax function, and T represents transpose.
[0019] Further, the state attention matrix is recorded as C t , the behavior attention vector is recorded as o t , using the state attention matrix C t and the behavior attention vector o t To model the students’ learning and forgetting behaviors, we can obtain the learning vector u t and the forget vector f t , using the learning vector u t and the forget vector f t Update the concept mastery matrix at time t-1 Get the concept mastery matrix at time t
[0020] Furthermore, C t and t After concatenation, it is input to a fully connected layer with a Tanh activation function. The weight matrix of the fully connected layer is w1, and the bias vector is b1, resulting in a dimension of d v The vector a t ,
[0021]
[0022] Use a fully connected layer with Tanh activation function to transform vector a t Convert to learning vector u t, to model the learning behavior and obtain the learning vector u t , the weight matrix of its fully connected layer is w u , the bias vector is b u ,
[0023]
[0024] The vector a t As the input of the key and value in the attention mechanism, the vector Projection to d v The output of the attention mechanism is input to the fully connected layer with Sigmoid activation function to obtain the forgetting vector f t ,
[0025]
[0026]
[0027]
[0028] Among them, W b For vector Projection to d v The weight term of the fully connected layer of dimension b b For vector Projection to d v The bias term of the fully connected layer of dimension W f is the weight term of the fully connected layer of the Sigmoid activation function, b f is the bias term of the fully connected layer of the Sigmoid activation function, Softmax() represents the Softmax function, and T represents transposition.
[0029] The updating formula of the concept mastery matrix is:
[0030]
[0031] Representation of concept mastery matrix The i-th column of Representation Concept Mastery Matrix The i-th column, w t (i) represents the association weight w t The i-th column of , where i ranges from 1 to N.
[0032] Furthermore, the concept matrix at time t-1 is recorded as the concept matrix Yes k ×N-dimensional matrix, d k is a matrix The dimension of each column, N is the number of concepts, calculated using the self-attention mechanism The similarity of the concepts in the concept is calculated using the concept attention matrix G t express,
[0033]
[0034] Exercise q t The embedding vector is denoted as k t , according to k t and the matrix G t Calculation Exercises t The similarity between the concepts contained in the calculation result is calculated using the association weight w t express,
[0035] w t =Softmax(k t ×G t )
[0036] The concept mastery matrix at time t-1 is recorded as Yes v ×N-dimensional matrix, d v is a matrix The dimension of each column is calculated using the self-attention mechanism The concept grasps the similarity between states, and the calculation result is the state attention matrix C t express,
[0037]
[0038] Among them, Softmax() represents the Softmax function, and T represents transpose.
[0039] According to a second aspect of the present invention, there is also provided a knowledge tracking system integrating learning behavior characteristics, comprising:
[0040] The concept attention module is used to calculate the similarity of concepts in the concept matrix using the self-attention mechanism. The concept matrix is used to represent the concepts related to the questions in the course. The calculation result is represented by the concept attention matrix. The similarity between the concepts contained in the predicted exercises is calculated based on the concept attention matrix, and the calculation result is represented by the association weight.
[0041] The state attention module is used to calculate the similarity between different concept mastery states in the concept mastery matrix using the self-attention mechanism. The concept mastery matrix is used to represent the learner's mastery state of the concept, and the calculation result is represented by the state attention matrix;
[0042] The behavior attention module uses the self-attention mechanism to mine the connections between different features in the feature vector describing the user's behavior during the learning process and outputs the behavior attention vector;
[0043] An updating module for updating the concept mastery matrix using the state attention matrix and the behavior attention vector;
[0044] The prediction module is used to multiply the association weight with the concept mastery matrix to obtain the vector of the student's concept mastery status, input the vector of the student's concept mastery status and the embedding vector of the exercise to be predicted into the feedforward neural network to predict the user's answer situation.
[0045] According to a third aspect of the present invention, there is also provided a storage medium storing a computer program executable by a processor, wherein when the computer program runs on the processor, the processor executes the steps of any one of the above methods.
[0046] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art:
[0047] (1) The present invention takes into account the impact of similarities between concepts on the student learning process when tracking the student's knowledge status. Existing models either do not consider this or use other methods. The present invention uses a self-attention mechanism to explore the similarities between concepts, which can have better explanatory power.
[0048] (2) The present invention inputs more information that affects the students' learning process into the model when tracking the students' knowledge status. Most existing models only use exercise interaction data to track the students' knowledge status, but the data contains limited information and cannot accurately track the students' knowledge status. The present invention combines the actual learning process and inputs rich learning behavior features into the model in addition to the exercise interaction data, so as to more accurately track the students' knowledge status.
[0049] (3) The present invention models forgetting behavior when tracking students' knowledge status. Most existing models do not model forgetting behavior, but forgetting behavior accompanies the entire learning process. It is obviously unreasonable to ignore modeling forgetting behavior. The present invention combines the actual situation and uses the attention mechanism to change the characteristics of the model's attention points on the input data, so that the model pays more attention to the features related to forgetting behavior when modeling forgetting behavior, thereby achieving the purpose of modeling forgetting behavior and having better interpretability. BRIEF DESCRIPTION OF THE DRAWINGS
[0050] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0051] Figure 1 A schematic diagram of a knowledge tracking system integrating learning behavior features provided in an embodiment of the present application;
[0052] Figure 2 A flowchart of a knowledge tracking method integrating learning behavior features provided in an embodiment of the present application;
[0053] Figure 3 Schematic diagram of a feedforward autoencoder with a self-attention mechanism provided in an embodiment of the present application. DETAILED DESCRIPTION
[0054] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.
[0055] The terms "first", "second", etc. in the specification and claims of this application and the above-mentioned drawings are used to distinguish different objects, rather than to describe a specific order. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or modules is not limited to the listed steps or modules, but optionally includes steps or modules that are not listed, or optionally includes other steps or modules inherent to these processes, methods, products or devices.
[0056] like Figure 1 As shown, the knowledge tracking system integrating learning behavior features of an embodiment of the present invention includes:
[0057] The concept attention module is used to calculate the similarity of concepts in the concept matrix using the self-attention mechanism. The concept matrix is used to represent the concepts related to the questions in the course. The calculation result is represented by the concept attention matrix. The similarity between the concepts contained in the predicted exercises is calculated based on the concept attention matrix, and the calculation result is represented by the association weight.
[0058] The state attention module is used to calculate the similarity between different concept mastery states in the concept mastery matrix using the self-attention mechanism. The concept mastery matrix is used to represent the learner's mastery state of the concept, and the calculation result is represented by the state attention matrix;
[0059] The behavior attention module uses the self-attention mechanism to mine the connections between different features in the feature vector describing the user's behavior during the learning process and outputs the behavior attention vector;
[0060] An updating module for updating the concept mastery matrix using the state attention matrix and the behavior attention vector;
[0061] The prediction module is used to multiply the association weight with the concept mastery matrix to obtain the vector of the student's concept mastery status, input the vector of the student's concept mastery status and the embedding vector of the exercise to be predicted into the feedforward neural network (fully connected neural network) to predict the user's answer situation.
[0062] like Figure 2 As shown, the knowledge tracking method integrating learning behavior features of an embodiment of the present invention comprises the steps of:
[0063] S101, using the self-attention mechanism to calculate the similarity of concepts in the concept matrix, the concept matrix is used to represent the concepts related to the questions in the course, the calculation result is represented by the concept attention matrix, and the similarity between the concepts contained in the exercises to be predicted is calculated based on the concept attention matrix, and the calculation result is represented by the association weight.
[0064] The concept matrix is used to represent the concepts related to the questions in the course. The concept matrix at time t-1 is recorded as It represents the vector representation of all concepts related to the questions in the course learned by the model at time t-1. The concept matrix is initialized before model training. During the model training process, the concept matrix continuously updates the vector representation of all concepts related to the questions in the course using the back propagation algorithm.
[0065] Calculating the Matrix The similarity between concepts in . The matrix that will store the concepts As the input of the self-attention mechanism in the module, d k is a matrix The dimension of each column is N, which is the number of different concepts in the course. Calculate the similarity between concepts and use the concept attention matrix G as the calculation result t Indicates that, then according to Exercise q t and the matrix G t Get the association weight w t .
[0066] The specific steps are as follows: Initialize the matrix storing concepts before model training matrix It is continuously updated during the model training process. It is used as the input of the self-attention mechanism to obtain the output concept attention matrix G t . t Convert to one-hot encoding and embedding matrix Multiply to get dimension d k The embedding vector k of the exercise t . k t With Gt Multiplication is converted into associated weight w through the Softmax function t , indicating that exercise q t Similarities between the concepts involved:
[0067]
[0068] w t =Softmax(k t ×G t )
[0069] S102, using a self-attention mechanism to calculate the similarity between different concept mastery states in a concept mastery matrix, where the concept mastery matrix is used to represent the learner's mastery state of the concept, and the calculation result is represented by a state attention matrix.
[0070] The concept mastery matrix is used to represent the learner's mastery of the concept. The concept mastery matrix at time t-1 is recorded as
[0071] Calculating the Matrix The matrix storing the concept mastery state As the input of the self-attention mechanism in the module, d v is a matrix The dimension of each column is N, which is the number of concepts in the course. Calculate the similarity of the user's mastery of different concepts, for example, the similarity between the user's mastery of concept 1 and the user's mastery of concept 2. The calculation result is expressed in the state attention matrix C. t express.
[0072] The specific steps are as follows: Initialize the matrix storing the concept mastery state before model training matrix It is continuously updated during the model training process. It is used as the input of the self-attention mechanism in the module to obtain the output state attention matrix C t :
[0073]
[0074] S103, using a self-attention mechanism to mine the connections between different features in a feature vector describing the user's behavior during the learning process, and outputting a behavior attention vector.
[0075] Receive the rich information generated by students during the learning process and input vector Contains exercise interaction information and learning behavior feature information, and the self-attention mechanism enables the model to capture vector Correlations between internal features.
[0076] The exercise interaction data X consists of students’ answers at multiple times, that is, X = (x1, x2, ..., x t ), x t =(q t ,r t ) is the student’s answer at time t, where q t represents the question answered by the student, r t 1 indicates that the student answered the question correctly, and 0 indicates that the student answered the question incorrectly. The learning behavior characteristic data represents some behavioral actions of the students in the learning process, such as the number of attempts to answer, the first response time, the number of prompt requests, etc.
[0077] Furthermore, the learning behavior characteristic data includes the first type of behavior characteristic data and the second type of behavioral characteristic data Including several predefined behavioral feature data directly obtained from user behavior collection data, It includes several predefined behavioral feature data reflecting the user's forgetting behavior calculated from the user's behavior collection data. It is the key behavioral feature data that reflects the user's forgetting behavior. The specific types of behavioral feature data can be pre-defined and calculated from the user behavior collection data. The methods can also be predefined.
[0078] Furthermore, the three features of the number of attempts to answer, the first response time, and the number of request prompts are combined into a row vector with a dimension of 3 Similarly, the number of times the same concept is repeatedly learned, the time interval between learning, and the time interval between learning the same concept form a row vector with a dimension of 3
[0079] The exercise interaction data x t Converted to one-hot encoding, To solve x t The sparse problem is to embed it into the matrix Multiply to get a dimension of d v vector Represents the student's exercise interaction data x at time t t . The vector vector Vector e t Composition Matrix matrix The first row represents the student's exercise interaction data x at time t t , each of the remaining rows represents a learning behavior feature, such as the second row represents the number of attempts to answer and the values in this row are the same. It can be seen that the matrix There are a lot of repeated features in , so the embodiment of the present invention preferably uses a feed forward autoencoder with self-attention mechanism (Feed Forward Autoencoder With Self-Attention, FFA-SA) to encode the matrix After dimensionality reduction, the self-attention mechanism is used to explore the connection between the internal features of the vector after dimensionality reduction, and the output is the behavioral attention vector.
[0080] Figure 3 The specific structure of FFA-SA is shown. Different from the traditional feedforward autoencoder, the self-attention layer is introduced in the embodiment of the present invention.
[0081] FFA-SA consists of two parts: encoder and decoder. The encoder consists of input layer, self-attention layer, fully connected layer and hidden layer; the decoder consists of hidden layer, fully connected layer, self-attention layer and output layer. The two parts are distributed in a mirror image.
[0082] In the encoder matrix As input, the self-attention mechanism is used to discover the connections between the internal features of the input data:
[0083]
[0084] The output matrix S e Containing the connections between features, it is input into the fully connected layer for the first dimensionality reduction and encoding:
[0085]
[0086] The weight matrix of the fully connected layer Bias Vector F e Input to the hidden layer to get the final dimension reduction and encoding results:
[0087]
[0088] The weight matrix of the hidden layer Bias Vector
[0089] In the decoder vector h t As input, it is used as a fully connected layer for the first dimension increase and decoding:
[0090]
[0091] The weight matrix of the fully connected layer Bias Vector F d Input to the self-attention layer to discover the internal connections of the features:
[0092]
[0093] The output matrix S d Containing the connections between features, the input and output layers are finally dimensionally upgraded and encoded:
[0094]
[0095] matrix With the matrix The closer it is, the better the autoencoder FFA-SA is trained. During the continuous training of FFA-SA, the output vector h of the hidden state t With the matrix The closer the characteristic information expressed is, the closer the vector h is. t The connection of internal features, the output of which is the behavior attention vector o t , o t It contains both students’ exercise interaction information and students’ learning behavior characteristics information:
[0096]
[0097] S104, using the state attention matrix and the behavior attention vector to update the concept mastery matrix.
[0098] The concept mastery matrix is initialized before model training. The concept mastery matrix is continuously updated during the model training process. The concept mastery matrix of the next moment in the adjacent moment is obtained by updating the concept mastery matrix of the previous moment in the adjacent moment. For example, the concept mastery matrix at moment t-2 is updated to obtain the concept mastery matrix at moment t-1, and the concept mastery matrix at moment t-1 is updated to obtain the concept mastery matrix at moment t.
[0099] The following is only updated get As an example, the updating principles at other times are the same and will not be described here.
[0100] Furthermore, using the state attention matrix C t and the behavior attention vector o t To model the students’ learning and forgetting behaviors, we can obtain the learning vector u t and the forget vector f t , using the learning vector u t and the forget vector f t Update the concept mastery matrix at time t-1 Get the concept mastery matrix at time t
[0101] The specific steps are: the output results C of the state attention module and the behavior attention module t and t Splice and input to a fully connected layer with Tanh activation function, whose weight matrix is w1 and bias vector is b1, to obtain a dimension d v The vector a t , is the vector o t and the matrix C t A comprehensive representation of .
[0102]
[0103] Learning behavior modeling: Use a fully connected layer with a Tanh activation function to transform the vector a t Convert to learning vector u t , the weight matrix of its fully connected layer is w u , the bias vector is b u ,
[0104]
[0105] Modeling forgetting behavior: transform vector a t As the input of the key and value in the attention mechanism, the vector Projection to d v The output of the attention mechanism is input to the fully connected layer with Sigmoid activation function to obtain the forgetting vector f t :
[0106]
[0107]
[0108]
[0109] Among them, W b For vector Projection to d v The weight term of the fully connected layer of dimension b b For vector Projection to d v The bias term of the fully connected layer of dimension W f is the weight term of the fully connected layer of the Sigmoid activation function, b f is the bias term of the fully connected layer of the Sigmoid activation function, Softmax() represents the Softmax function, and T represents transpose.
[0110] matrix Update:
[0111]
[0112] in, Representation Concept Mastery Matrix The i-th column of Representation Concept Mastery Matrix The i-th column, w t (i) represents the association weight w t The i-th column of , the value of i ranges from 1 to N, that is, each concept is updated one by one.
[0113] S105, multiply the association weight by the concept mastery matrix to obtain the vector of the student's concept mastery status, input the vector of the student's concept mastery status and the embedding vector of the exercise to be predicted into the feedforward neural network (fully connected neural network) to predict the user's answer.
[0114] The prediction module is based on the matrix To predict the student's answer at time t.
[0115] The specific steps are: t With the matrix Multiply them together to get vector n t , is the output of the student's concept mastery status:
[0116]
[0117] Considering that there are certain differences between exercises, such as different difficulty coefficients, the vector n t and the problem embedding vector k t The vector obtained by concatenating the concept contains both the student's mastery of the concept and the exercise information. This vector is input into the fully connected layer with the Tanh activation function. The weight matrix of the fully connected layer is w2, and the bias vector is b2. The vector i is obtained. t :
[0118]
[0119] Finally, an output layer with a Sigmoid activation function is used, whose weight matrix is w3 and bias vector is b3. t As input, it is used to predict the student's answer to the exercise q t Performance of:
[0120]
[0121] Simulation tests were conducted in different data sets. The knowledge tracking method of the present invention integrating learning behavior features was denoted as ILB-KT and compared with SAKT, DKT, and DKEMN. As shown in Table 1, it can be found that the prediction accuracy of the present invention is higher than that of the prior art.
[0122] Table 1
[0123]
[0124]
[0125] The implementation principle of the system is the same as the above method and will not be repeated here.
[0126] The present application also provides a storage medium storing a computer program executable by a processor, which, when the computer program is run on the processor, enables the processor to perform any of the steps of the above-mentioned knowledge tracking method integrating learning behavior characteristics. The computer-readable storage medium may include, but is not limited to, any type of disk, including a floppy disk, an optical disk, a DVD, a CD-ROM, a micro drive, and a magneto-optical disk, a ROM, a RAM, an EPROM, an EEPROM, a DRAM, a VRAM, a flash memory device, a magnetic card or an optical card, a nanosystem (including a molecular memory IC), or any type of medium or device suitable for storing instructions and / or data.
[0127] It should be noted that, for the aforementioned method embodiments, for the sake of simplicity, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the present application is not limited by the described order of actions, because according to the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present application.
[0128] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.
[0129] In the several embodiments provided in the present application, it should be understood that the disclosed system can be implemented in other ways. For example, the system embodiments described above are only schematic, such as the division of the modules, which is only a logical function division. There may be other division methods in actual implementation, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some service interfaces, indirect coupling or communication connection of systems or modules, which can be electrical or other forms.
[0130] The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical modules, that is, they may be located in one place or distributed on multiple network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.
[0131] In addition, each functional module in each embodiment of the present application can be integrated into a processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The above integrated modules can be implemented in the form of hardware or software functional modules.
[0132] If the integrated module is implemented in the form of a software function module and sold or used as an independent product, it can be stored in a computer-readable memory. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a memory, including a number of instructions to enable a computer device (which can be a personal computer, server or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned memory includes: U disk, read-only memory (ROM), random access memory (RAM), mobile hard disk, disk or optical disk and other media that can store program code.
[0133] A person skilled in the art may understand that all or part of the steps in the various methods of the above embodiments may be completed by entering a program to instruct the relevant hardware. The program may be stored in a computer-readable memory, and the memory may include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.
[0134] The above is only an exemplary embodiment of the present disclosure, and the scope of the present disclosure cannot be limited thereto. That is, any equivalent changes and modifications made according to the teachings of the present disclosure are still within the scope of the present disclosure. After considering the specification and practicing the disclosure here, those skilled in the art will easily think of the implementation scheme of the present disclosure. This application is intended to cover any modification, use or adaptation of the present disclosure, which follows the general principles of the present disclosure and includes common knowledge or customary technical means in the technical field not recorded in the present disclosure. The description and examples are regarded as exemplary only, and the scope and spirit of the present disclosure are defined by the claims.
[0135] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, 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.
[0136] It will be easily understood by those skilled in the art that the above description is only 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 in the protection scope of the present invention.
Claims
1. A knowledge tracking method integrating learning behavior characteristics, characterized in that: include: The self-attention mechanism is used to calculate the similarity of concepts in the concept matrix. The concept matrix is used to represent the concepts related to the questions in the course. The calculation result is represented by the concept attention matrix. The similarity between the concepts contained in the predicted exercises is calculated based on the concept attention matrix, and the calculation result is represented by the association weight. The self-attention mechanism is used to calculate the similarity between different concept mastery states in the concept mastery matrix. The concept mastery matrix is used to represent the learner's mastery state of the concept, and the calculation result is represented by the state attention matrix. The self-attention mechanism is used to mine the connections between different features in the feature vector describing the user's behavior during the learning process, and the behavior attention vector is output; Update the concept mastery matrix using the state attention matrix and behavior attention vector; Multiply the association weights with the concept mastery matrix to obtain the vector of the student's concept mastery status. Input the vector of the student's concept mastery status and the embedding vector of the exercise to be predicted into the feedforward neural network to predict the user's answer. in: The state attention matrix is denoted as C t , the behavior attention vector is recorded as o t , using the state attention matrix C t and the behavior attention vector o t To model the students’ learning and forgetting behaviors, we can obtain the learning vector u t and the forget vector f t , using the learning vector u t and the forget vector f t Update the concept mastery matrix at time t-1 Get the concept mastery matrix at time t C t and t After concatenation, it is input to a fully connected layer with a Tanh activation function. The weight matrix of the fully connected layer is w1, and the bias vector is b1, resulting in a dimension of d v The vector a t , Use a fully connected layer with Tanh activation function to transform vector a t Convert to learning vector u t , to model the learning behavior and obtain the learning vector u t , the weight matrix of its fully connected layer is w u , the bias vector is b u , The vector a t As the input of the key and value in the attention mechanism, the vector Projection to d v The output of the attention mechanism is input to the fully connected layer with Sigmoid activation function to obtain the forgetting vector f t , Among them, W b For vector Projection to d v The weight term of the fully connected layer of dimension b b For vector Projection to d v The bias term of the fully connected layer of dimension W f is the weight term of the fully connected layer of the Sigmoid activation function, b f is the bias term of the fully connected layer of the Sigmoid activation function, Softmax() represents the Softmax function, T represents the transpose, The updating formula of the concept mastery matrix is: Representation Concept Mastery Matrix The i-th column of Representation Concept Mastery Matrix The i-th column, w t (i) represents the association weight w t The i-th column of , where i ranges from 1 to N; The concept matrix at time t-1 is denoted as the concept matrix Yes k ×N-dimensional matrix, d k is a matrix The dimension of each column, N is the number of concepts, calculated using the self-attention mechanism The similarity of the concepts in the concept is calculated using the concept attention matrix G t express, Exercise q t The embedding vector is denoted as k t , according to k t and the matrix G t Calculation Exercises t The similarity between the concepts contained in the calculation result is calculated using the association weight w t express, w t =Softmax(k t ×G t ) The concept mastery matrix at time t-1 is recorded as Yes v ×N-dimensional matrix, d v is a matrix The dimension of each column is calculated using the self-attention mechanism The concept grasps the similarity between states, and the calculation result is the state attention matrix C t express, Among them, Softmax() represents the Softmax function, and T represents transpose.
2. The knowledge tracking method integrating learning behavior characteristics as claimed in claim 1, characterized in that: The feature vector describing the user's behavior during the learning process is recorded as It includes exercise interaction data X and learning behavior feature data. Exercise interaction data X is used to describe the correctness of the user's answers at multiple moments before the prediction time. The learning behavior feature data includes the first type of behavior feature data and the second type of behavioral characteristic data Including several predefined behavioral feature data directly obtained from user behavior collection data, It includes several predefined behavior feature data related to the user's forgetting behavior calculated from the user behavior collection data.
3. The knowledge tracking method integrating learning behavior characteristics as claimed in claim 2, characterized in that: The three features of answer times, first response time, and request prompt times form a row vector with a dimension of 3. The number of times the same concept is repeatedly learned, the time interval between learning, and the time interval between learning the same concept form a row vector with a dimension of 3.
4. The knowledge tracking method integrating learning behavior characteristics as claimed in claim 1, characterized in that: The characteristic vector of the user's behavior during the learning process is recorded as Mining with self-attention mechanism The connections between the different features in include: A feedforward autoencoder with self-attention mechanism is used Perform dimensionality reduction to obtain the reduced dimensionality vector h t , where the feedforward autoencoder includes an encoder and a decoder. The encoder includes an input layer, a self-attention layer, a fully connected layer, and a hidden layer. The decoder includes a hidden layer, a fully connected layer, a self-attention layer, and an output layer. The encoder and the decoder are distributed in a mirror image. Then use the self-attention mechanism to discover h t The connection between internal features is output as the behavior attention vector, denoted as o t .
5. The knowledge tracking method integrating learning behavior characteristics as claimed in claim 4, characterized in that: Softmax() represents the Softmax function, and T represents transpose.
6. A knowledge tracking system integrating learning behavior characteristics, characterized in that: include: The concept attention module is used to calculate the similarity of concepts in the concept matrix using the self-attention mechanism. The concept matrix is used to represent the concepts related to the questions in the course. The calculation result is represented by the concept attention matrix. The similarity between the concepts contained in the predicted exercises is calculated based on the concept attention matrix, and the calculation result is represented by the association weight. The state attention module is used to calculate the similarity between different concept mastery states in the concept mastery matrix using the self-attention mechanism. The concept mastery matrix is used to represent the learner's mastery state of the concept, and the calculation result is represented by the state attention matrix; The behavior attention module uses the self-attention mechanism to mine the connections between different features in the feature vector describing the user's behavior during the learning process and outputs the behavior attention vector; An updating module for updating the concept mastery matrix using the state attention matrix and the behavior attention vector; The prediction module is used to multiply the association weight with the concept mastery matrix to obtain the vector of the student's concept mastery state, input the vector of the student's concept mastery state and the embedding vector of the exercise to be predicted into the feedforward neural network to predict the user's answer; in: The state attention matrix is denoted as C t , the behavior attention vector is recorded as o t , using the state attention matrix C t and the behavior attention vector o t To model the students’ learning and forgetting behaviors, we can obtain the learning vector u t and the forget vector f t , using the learning vector u t and the forget vector f t Update the concept mastery matrix at time t-1 Get the concept mastery matrix at time t C t and t After concatenation, it is input to a fully connected layer with a Tanh activation function. The weight matrix of the fully connected layer is w1, and the bias vector is b1, resulting in a dimension of d v The vector a t , Use a fully connected layer with Tanh activation function to transform vector a t Convert to learning vector u t , to model the learning behavior and obtain the learning vector u t , the weight matrix of its fully connected layer is w u , the bias vector is b u , The vector a t As the input of the key and value in the attention mechanism, the vector Projection to d v The output of the attention mechanism is input to the fully connected layer with Sigmoid activation function to obtain the forgetting vector f t , Among them, W b For vector Projection to d v The weight term of the fully connected layer of dimension b b For vector Projection to d v The bias term of the fully connected layer of dimension W f is the weight term of the fully connected layer of the Sigmoid activation function, b f is the bias term of the fully connected layer of the Sigmoid activation function, Softmax() represents the Softmax function, T represents the transpose, The updating formula of the concept mastery matrix is: Representation Concept Mastery Matrix The i-th column of Representation Concept Mastery Matrix The i-th column, w t (i) represents the association weight w t The i-th column of , where i ranges from 1 to N; The concept matrix at time t-1 is denoted as the concept matrix Yes k ×N-dimensional matrix, d k is a matrix The dimension of each column, N is the number of concepts, calculated using the self-attention mechanism The similarity of the concepts in the concept is calculated using the concept attention matrix G t express, Exercise q t The embedding vector is denoted as k t , according to k t and the matrix G t Calculation Exercises t The similarity between the concepts contained in the calculation result is calculated using the association weight w t express, w t =Softmax(k t ×G t ) The concept mastery matrix at time t-1 is recorded as Yes v ×N-dimensional matrix, d v is a matrix The dimension of each column is calculated using the self-attention mechanism The concept grasps the similarity between states, and the calculation result is the state attention matrix C t express, Among them, Softmax() represents the Softmax function, and T represents transpose.
7. A storage medium, characterized in that: The device stores a computer program, and when the computer program is run on a processor, the processor is enabled to execute the steps of the method according to any one of claims 1 to 5.
