A learning goal recommendation method based on a double-flow knowledge embedding network

Abstract

The application provides a learning target recommendation method based on a double-flow knowledge embedding network, mainly relates to the technical field of intelligent human-computer interaction, and comprises the following steps: obtaining a learner's answer sequence and corresponding knowledge concepts, and outputting a Q matrix; obtaining the weight of the knowledge concepts in the answer sequence through a trained knowledge tracking model, predicting the correct probability of the learner's next answer sequence, outputting the recommendation probability value of the nearest neighbor associated knowledge concept, generating at least one recommended learning target according to the recommendation probability value; predicting the learner's mastery of the recommended learning target based on the trained knowledge demand model, outputting the predicted correct probability of the answer, arranging the corresponding recommended learning target from high to low according to the predicted correct probability, and generating a recommended learning path; through the method provided by the application, the adaptation of the nearest neighbor learning target to individualization can be improved while reducing the required cost, thereby improving the learning efficiency of the learner.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent human-computer interaction, and in particular to a learning goal recommendation method based on a double-flow knowledge embedding network. BACKGROUND

[0002] With the deepening of the process of education informatization, more and more information technology is applied to education and teaching, and emerging online learning mode gradually develops to become an indispensable part of learning. In traditional classroom education, the learning activities of learners completely depend on the content customized for the classroom, and the customization of teaching content is limited by the class hours and the teaching site. For further learning goals, it is necessary to rely on the teaching experience of teachers to make unified customization for all learners in the same classroom. Therefore, in the current education and learning mode, it is difficult to develop a personalized learning plan for each learner. How to track the knowledge state of learners according to their learning data on various learning systems, and then develop a suitable personalized learning plan for learners, has great significance for improving the learning efficiency of learners.

[0003] Knowledge tracking can effectively estimate the learning situation of learners according to their test data, and can also obtain the performance state of learners on specific knowledge points according to the data corresponding to the knowledge points and the questions. However, the learning training plan of learners is generally the same, and there is little adaptive improvement for the learning situation of specific learners. If the learning plan is developed for each learner by relying on teachers, it will consume huge time and labor cost, and such a method has high requirements for the teaching experience of teachers, and cannot efficiently, quickly and accurately develop a learning goal for learners. SUMMARY

[0004] The present application provides a learning goal recommendation method based on a double-flow knowledge embedding network, which is used to solve the defects in the prior art and realize intelligent customization of personalized learning goals of learners. The purpose is to improve the adaptation of the nearest neighbor learning goal to individualization while reducing the required cost, so as to improve the learning efficiency of learners.

[0005] The present application provides a learning goal recommendation method based on a double-flow knowledge embedding network, which comprises:

[0006] S1 obtains the answer sequence of a learner and the knowledge concept corresponding to each question, and outputs each question and the corresponding knowledge concept as a Q matrix;

[0007] S2 inputs the answer sequence and the corresponding Q matrix into the trained knowledge tracking model, obtains the weight of a single knowledge concept in the knowledge concept of the answer sequence, outputs the mastery state of each learned knowledge concept of the learner, and predicts the answer correct probability of the learner for the next answer sequence through the knowledge tracking model;

[0008] S3 selects a preset nearest related knowledge concept based on the weight of a single learned knowledge concept based on the trained knowledge demand model, outputs a recommendation probability value of the nearest related knowledge concept, and outputs at least one recommended learning target corresponding to the related knowledge concept to the learner in descending order of the recommendation probability value;

[0009] S4 predicts the mastery of the learner for multiple recommended learning targets based on the trained knowledge demand model, outputs a predicted answer correct probability, arranges the corresponding recommended learning targets in descending order of the predicted correct probability, and generates a recommended learning path.

[0010] According to the learning target recommendation method based on the double-flow knowledge embedding network provided by the application, in step S1, the answer sequence of the learner is obtained as:

[0011] S i ={(e1,r1),(e2,r2),(e3,r3),...,(e t ,r t ),};

[0012] Wherein, e t is the tth question in the answer sequence, r t is the answer of the student S i to the question e t ;

[0013] If the answer is correct, r t is 1, otherwise 0.

[0014] According to the learning target recommendation method based on the double-flow knowledge embedding network provided by the application, in step S1, each of the questions and the corresponding knowledge concept is output as a Q matrix, for the element q ij in the Q matrix, if it is 1, the question e i of the ith row includes the corresponding knowledge concept k j ; if it is 0, the question e i of the ith row does not include the corresponding knowledge concept k j ;

[0015] Wherein, the question e i includes at least one knowledge concept k j .

[0016] According to the learning target recommendation method based on the double-flow knowledge embedding network provided by the application, after the answer sequence and the corresponding Q matrix are obtained in step S2, student embedding and knowledge concept embedding are generated, including:

[0017] The student embedding includes learner information and question text information;

[0018] The knowledge concept embedding includes the question text information and corresponding knowledge concept information;

[0019] The knowledge concept embedding s k，t is concatenated with the student embedding c k，t , and the formula is applied:

[0020]

[0021] Further, the answer result y k，t of the learner to the question is obtained, the embedding a k，t fusing the answer of the student is generated, and a k，t is input into the knowledge tracking model:

[0022]

[0023] When y k，t is 1, the answer result of the learner to the question is correct; and when y k，t is 0, the answer result of the learner to the question is incorrect.

[0024] Further, in step S2, the knowledge concept k t of the i-th knowledge concept in the knowledge concept of the question e t in the answer sequence is obtained, including:

[0025] The knowledge concept corresponding to each question e t is converted into one-hot encoding k t ∈{0,1}, and the dimension of the one-hot encoding is the total number K of all knowledge concepts.

[0026] Based on the embedding matrix W k , the one-hot encoding k t is converted into a low-dimensional vector , and the continuous value of the low-dimensional vector is v t =W k T k t .

[0027] Based on the inner product M i of v t and the knowledge memory vector, the weight w is calculated by performing softmax calculation.

[0028]

[0029] where 1≤i≤K.

[0030] According to the learning target recommendation method based on the double-flow knowledge embedding network provided by the application, in step S2, the LSTM network is used to update a learner practice question answering sequence e t after knowledge state The application formula is:

[0031]

[0032] The output learning k current knowledge state h k，t , the knowledge state h k，t and the embedding c k，t of the knowledge concept are input into the Sigmoid function, and the probability of the learner k answering the question e t+1 is obtained, and the application formula is:

[0033]

[0034] p k，t+1 =σ(W s s k，t +b s );

[0035] where h is the dimension of the hidden state vector, g k，t , f k，t , o k，t , r k，t , h k，t are all vector parameters, W i , W f , W fc , W s , W o , W c and U i , U f , U o , U c are all weight matrices, b i , b f , b o , b c , b fc , b s are bias vectors, represents the multiplication operation, and σ(.) and tanh(.) represent the sigmoid and tanh activation functions, respectively.

[0036] p k，t+1 is the probability of the learner k correctly answering the question e t+1 .

[0037] According to the learning target recommendation method based on the double-flow knowledge embedding network provided by the application, before step S3, a loss function is calculated.

[0038]

[0039] wherein r t is a real answer result of the learner to the question, represents a predicted answer result of the learner to the question predicted by the knowledge tracking model.

[0040] Further, the difference in the mastery of the knowledge concept after the learner learns the adjacent two recommended learning targets is obtained, and the formula is applied.

[0041]

[0042] wherein d k，t is the correct answer probability of the question related to the knowledge concept k at time t, and m k，t is the number of times of recommending the learning target including the knowledge concept k.

[0043] On the other hand, the application further provides a learning target recommendation system based on the double-flow knowledge embedding network, comprising:

[0044] A data preprocessing module is configured to obtain a question answer sequence of a learner and a knowledge concept corresponding to each question, and output each question and the corresponding knowledge concept as a Q matrix.

[0045] A question answer prediction module is configured to input the question answer sequence and the corresponding Q matrix into a trained knowledge tracking model, obtain the weight of a single knowledge concept in the knowledge concept of the question answer sequence, output the mastery state of each knowledge concept learned by the learner, and predict the correct answer probability of the next question answer sequence of the learner through the knowledge tracking model.

[0046] A learning target recommendation module is configured to select a preset nearest related knowledge concept based on the weight of a single learned knowledge concept, output the recommendation probability value of the nearest related knowledge concept, and output at least one recommended learning target corresponding to the related knowledge concept to the learner in descending order of the recommendation probability value.

[0047] A learning path generation module is configured to predict the mastery of multiple recommended learning targets of the learner based on a trained knowledge demand model, output the predicted correct answer probability, arrange the corresponding recommended learning target in descending order of the predicted correct answer probability, and generate a recommended learning path.

[0048] The application further provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method according to any one of the above when executing the program.

[0049] The application further provides a non-transitory computer-readable storage medium, which stores a computer program, wherein the computer program is executable by a processor to implement the steps of the method according to any one of the above.

[0050] Overall, compared with the prior art, the above technical solutions conceived by the application can achieve the following beneficial effects:

[0051] (1) The double-flow knowledge embedding network of student modeling and knowledge modeling is adopted to simulate the knowledge state process of students, so that the knowledge tracking model is more in line with the knowledge mastery of students in the learning process, and thus the network trained is more interpretable.

[0052] (2) The nearest neighbor learning goal of students is intelligently adapted based on the framework of reinforcement learning, the intelligent adaptation made according to the knowledge state of students can take into account long-term returns, realize an automatic and optimized design strategy, and effectively reduce the human cost of teaching and learning. BRIEF DESCRIPTION OF DRAWINGS

[0053] In order to more clearly illustrate the technical solutions in the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description are some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0054] Figure 1 is one of the flowcharts of the learning goal recommendation method provided by the application

[0055] Figure 2 is another flowchart of the learning goal recommendation method provided by the application. DETAILED DESCRIPTION

[0056] In order to make the purpose, technical solutions and advantages of the application clearer, the technical solutions in the application will be described clearly and completely in the following with reference to the drawings in the application. Obviously, the described embodiments are some of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0057] The terms "comprise", "comprising", "have", "having", "include", "including", "contain", "containing", "provide", "providing", "offer", "offering", "specify", "specifying", "preserve", "preserving" and any variations thereof in the Specification and in the Claims shall not be construed as referring to the whole post- granted claim elements and / or method steps only as they can also refer to processing the whole originally-enumerated claim elements and / or method steps. For example, a process, method, object, or apparatus that "comprises" or "specifies" several steps can not necessarily be limited to the listed steps or elements but can include additional or fewer steps or elements.

[0058] In one embodiment, as shown in FIG. 1, the present application provides a learning target recommendation method based on a double-flow knowledge embedding network, which specifically comprises the following steps: Figures 1-2

[0059] S1: obtaining a learner's answer sequence and a knowledge concept corresponding to each question, and outputting each question and the corresponding knowledge concept as a Q matrix;

[0060] S2: inputting the answer sequence and the corresponding Q matrix into a trained knowledge tracking model, obtaining the weight of a single knowledge concept in the knowledge concept of the answer sequence, outputting the mastery state of each learned knowledge concept by the learner, and predicting the answer correct probability of the next answer sequence by the learner through the knowledge tracking model;

[0061] S3: based on the trained knowledge demand model, selecting a preset nearest related knowledge concept based on the weight of a single learned knowledge concept, outputting the recommendation probability value of the nearest related knowledge concept, and outputting at least one recommended learning target corresponding to the related knowledge concept to the learner in descending order of the recommendation probability value;

[0062] S4: predicting the mastery of the learner to multiple recommended learning targets based on the trained knowledge demand model, outputting the predicted answer correct probability, arranging the corresponding recommended learning targets in descending order of the predicted correct probability, and generating a recommended learning path.

[0063] It should be noted that the recommended learning target can be a corresponding exercise or other related reading material. In order to facilitate quantitative training of the model, the exercise is generally taken as the recommended learning target, and one recommended learning target can include 1-5 knowledge concepts, which is not limited by the present application.

[0064] Specifically, in step S1, the learner's answer sequence is obtained as follows:

[0065] S i ={(e1,r1),(e2,r2),(e3,r3),…,(e t ,r t ),};

[0066] wherein e t ​for the tth question in the question sequence, r t for the student S i the answer to question e t ;

[0067] if the answer is correct, r t is 1, otherwise 0;

[0068] It should be noted that the knowledge of the cell part in the biological discipline is taken as an example.

[0069] For example, question 1: the process of photosynthesis in cells; question 2: the law of inheritance of cells; question 3: the basic composition of cells. The matrix corresponding to the knowledge concept and the question can be represented as:

[0070] Topic 1 Topic 2 Topic 3 Cells 1 1 1 Metabolism of cells 1 0 0 Genetics of cells 0 1 0

[0071] The matrix composed of 0 and 1 is the Q matrix corresponding to the question and the knowledge concept. As can be seen from the matrix, one practice question may be associated with multiple knowledge points, such as question 1 in the first column of the above table. Question 1 covers two knowledge concept points, including cells and cell metabolism.

[0072] In step S1, each question and the corresponding knowledge concept are output as a Q matrix. For the element q ij in the Q matrix, if it is 1, the question e i in the ith row includes the corresponding knowledge concept k j ; if it is 0, the question e i in the ith row does not include the corresponding knowledge concept k j .

[0073] In step S2, the question e i includes at least one knowledge concept k j .

[0074] As an example, the representation matrix Q corresponding to the question-knowledge concept is as follows:

[0075]

[0076] where e t represents the serial number of the practice question t, k t represents the serial number of the knowledge concept t, 0 indicates that there is no association between the relevant practice question and the corresponding knowledge concept, and 1 indicates that there is an association between the relevant practice and the corresponding knowledge concept, i.e. the practice question covers this knowledge concept.

[0077] It should be noted that one practice question may be associated with multiple knowledge concepts.

[0078] Specifically, in one embodiment, the knowledge mastery state of the learner is predicted by an LSTM model (Long Short-Term Memory), including three aspects:

[0079] In a first aspect, after obtaining the answer sequence and the corresponding Q matrix in step S2, the student embedding and the knowledge concept embedding are generated, including:

[0080] The student embedding includes learner information and question text information;

[0081] The knowledge concept embedding includes the question text information and the corresponding knowledge concept information;

[0082] The knowledge concept embedding s k，t and the student embedding c k，t are concatenated and operated, and the formula is applied:

[0083]

[0084] Further, for the student embedding, the knowledge state of the learner in the student embedding is affected by each exercise question. If the learner's answer to the question is correct, it means that the learner has mastered the knowledge point, otherwise it means that the learner has not mastered the knowledge point.

[0085] For the knowledge embedding, the learning of part of the knowledge concept in the knowledge embedding will affect the learning of other knowledge concepts. For example, if the knowledge concept of addition is mastered, the knowledge concept of subtraction is relatively not a completely new and unfamiliar concept. Subtraction is essentially the addition of a positive number and a negative number. Accordingly, if it is concluded from the results of doing exercises that the learner has mastered the knowledge concept of addition, the weight of the exercises covering the concept of subtraction can be correspondingly reduced.

[0086] Further, the learner's answer result y k，t to the question is obtained, and the embedding a k，t fusing the student's answer is generated. k，t a k，t is input into the knowledge tracking model:

[0087]

[0088] When y k，t is 1, the learner's answer to the question is correct, and when it is 0, the learner's answer to the question is incorrect.

[0089] Specifically, by embedding the students' personal information and / or answer information together with the text information of the questions, the knowledge tracking model learns the potential information between the students' answer and the question text; and the knowledge concept embedding mainly combines the text information of the question with the knowledge concept information involved in the question through semantic analysis to let the model learn. Through the two embeddings, the training difficulty of the knowledge tracking model is reduced, and the connection between the learners and the questions, the knowledge concepts and the questions is emphasized respectively, which is conducive to the convergence process of the knowledge tracking model training and conducive to subsequent prediction based on the knowledge tracking model;

[0090] In a second aspect, the knowledge state of the learner in the process of continuously training the questions is updated by the LSTM neural network, including:

[0091] For each exercise question, each question e t is converted into a one-hot encoding k t ∈{0,1}, and the dimension of the one-hot encoding is the total number K of all knowledge concepts.

[0092] Based on the embedding matrix , the one-hot encoding k t is converted into a low-dimensional vector v The continuous value of v t is W k T k t , thereby avoiding the sparsity of the learner's answer data.

[0093] It should be noted that the sparsity represents the frequency and density of the learner's questions. Some question data is long, and some question data is short, which easily causes waste of data storage space and increases the difficulty of calculation, leading to difficulty in quickly converging the model training.

[0094] Based on the inner product of v t and the knowledge memory vector in the memory module M i , the influence weight of the i-th concept in the knowledge concept k t of the exercise e t is calculated by softmax calculation.

[0095] Among them, the knowledge memory vector refers to the mastery state of a specific knowledge concept, that is, the predicted probability of answering the question related to the knowledge concept correctly.

[0096]

[0097] Where 1≤i≤K.

[0098] Specifically, through the two embeddings mentioned above, the knowledge acquired by learners during practice can be further refined, and the influence weight of the knowledge concepts acquired in each practice can be used to track the learner's mastery of different knowledge concepts.

[0099] Specifically, if the influence weight is large, it indicates that the student's current grasp of this knowledge concept is poor, and the predicted accuracy rate of the student's answers to related knowledge concept questions is low; if the influence weight is small, it indicates that the student has basically mastered this knowledge concept, and the predicted accuracy rate of the student's answers to related knowledge concept questions is high.

[0100] It should be noted that during the learning process, the learner's current knowledge state or mastery of knowledge concepts is related to their previous knowledge state and is highly correlated with the benefits gained from learning new knowledge concepts. In other words, the learner's current knowledge state depends on their knowledge base, learning process, and the learning benefits gained after learning according to the recommended learning objectives.

[0101] Therefore, in the third aspect, the knowledge tracking module is used to predict the learner's knowledge mastery status and calculate the loss function;

[0102] The final output is:

[0103] r T+1 =σ(W2·y T+1 +b2);

[0104] in

[0105] y T+1 This represents the overall probability predicted at step T+1;

[0106] This represents a join operation.

[0107] Furthermore, in step S2, a learner practice question e is updated through an LSTM network. t Post-knowledge state Application formula:

[0108]

[0109] It should be noted that the state of knowledge here... In essence, it is a collection of the mastery status of multiple knowledge concepts;

[0110] The current knowledge state h of student k output by LSTM k，t and the embedding of knowledge concepts k，t The concatenated vectors are used for prediction. The concatenated vectors are then fed into a fully connected layer to obtain a summary vector s.k，t , and then the vector is sent into a sigmoid activation layer to obtain the probability that the learner k answers the question e t+1 The probability of correct result is calculated by the formula:

[0111]

[0112] p k，t+1 =σ(W s s k，t +b s );

[0113] where h is the dimension of the hidden state vector, g k，t , f k，t , o k，t , r k，t , h k，t are vector parameters, W i , W f , W fc , W s , W o , W c and U i , U f , U o , U c are weight matrices, b i , b f , b o , b c , b fc , b s are bias vectors, and represents the multiplication operation, and σ(.) and tanh(.) represent sigmoid and tanh activation functions, respectively;

[0114] p k，t+1 is the probability that the learner k answers the question e t+1 correctly, i.e., the probability determined by the calculation, which determines the mastery of the learner on different knowledge concepts;

[0115] The knowledge state is essentially the combination of the mastery of the learner on multiple knowledge concepts, reflecting the overall knowledge mastery of the learner or the set of correct answer probabilities for multiple questions e t ;

[0116] Further, before step S3, the loss function is calculated:

[0117]

[0118] where r t is the true answer of the learner to the question, and The predicted answer result of the learner to the question predicted by the knowledge tracking model;

[0119] The model parameters are further optimized by the loss function, and if the loss function value is greater than the threshold value, it indicates that the prediction effect of the model output is not good, and if the loss function value is less than the threshold value, it indicates that the prediction effect of the model output is good;

[0120] Further, in step S3, the nearest learning target is recommended according to the knowledge state of the learner through the knowledge demand model; the association between the knowledge concepts is preset, and the application is not limited thereto; the real-time knowledge concept mastery degree of a single learner is obtained through the knowledge tracking model, and the nearest associated knowledge concept is recommended to the learner according to the predicted learning of the learner to the knowledge concept:

[0121] For example, learning concepts A, B and C, A and B are the nearest knowledge concepts, when it is judged that the learning of the learner to the knowledge concept A is not ideal, it is predicted that the learning of the learner to the knowledge concept B is also not good, and therefore, compared with the unassociated knowledge concept C, the learning resource or exercise related to the knowledge concept B is preferentially recommended to the learner, thereby facilitating the learning of the learner to consolidate the knowledge concept;

[0122] The output of the knowledge demand model is a one-dimensional vector, and each item represents the recommendation probability of each knowledge concept, that is, the recommendation probability of the learning target or exercise related to the knowledge concept; the knowledge demand model KDM is a fully connected neural network, and in the last layer of the knowledge demand model, a softmax is introduced as an activation function to output the recommendation probability value of each knowledge concept, and the value range is 0-1, and the sum of all probability values is 1.

[0123] Specifically, a knowledge concept and the exercise related to the knowledge concept are selected for recommendation by a weighted random function;

[0124] Optionally, the application also includes a reward mechanism in the training process of the knowledge demand model, so as to avoid repeatedly recommending the same learning resource or exercise of a knowledge concept, and the reward mechanism comprises:

[0125] The difference between the mastery degrees of the learner to the knowledge concept after learning the adjacent two recommended learning targets is obtained, and the formula is applied:

[0126]

[0127] Wherein, d k，t is the mastery state of the knowledge concept k at time t, and m k，t is the number of recommended learning targets including the knowledge concept k.

[0128] By calculating g_t, if the model repeatedly recommends the same learning resource or exercise corresponding to a knowledge concept, the calculated value is Otherwise, the calculated value is d k，t -d k，t-1 Thus, a reward and punishment mechanism for the model is established for further training of the model.

[0129] Further, in step S5, after the learner learns according to the recommended learning target and learning resource, the new knowledge mastery state of the learner is modeled and predicted according to the above steps S1-S4, the predicted correct answer probability of the learner to the related knowledge concept exercise, that is, the predicted mastery probability of the related knowledge concept, is predicted, and the learning target with the highest predicted mastery probability corresponding to multiple recommended learning targets is taken as the optimal learning target; thus, the adaptation of the nearest learning target of the learner is realized.

[0130] The present application also provides an electronic device, which can include a processor, a communications interface, a memory 830 and a communications bus, wherein the processor, the communications interface, the memory and the communications bus complete mutual communication through the communications bus. The processor can invoke the logic instructions in the memory to execute the above method.

[0131] In addition, the logic instructions in the memory described above can be implemented in the form of a software functional unit and sold or used as an independent product, and can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0132] On the other hand, the present application also provides a computer program product, which includes a computer program stored on a non-transitory computer readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the above method.

[0133] In yet another aspect, the present application also provides a non-transitory computer readable storage medium having stored thereon a computer program which, when executed by a processor, implements the method described above.

[0134] The device embodiments described above are merely illustrative, wherein the units described as separate components can or can not be physically separate, and the components displayed as units can or can not be physical units, i.e., can be located in one place, or can be distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the embodiment scheme according to actual needs. Those skilled in the art can understand and implement without creative labor.

[0135] Through the description of the above embodiments, those skilled in the art can clearly understand that the embodiments can be realized by means of software plus necessary universal hardware platforms, and of course can also be realized by hardware. Based on such understanding, the above technical solutions, essentially or in other words, the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, etc., and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0136] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, rather than limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for some technical features; and these modifications or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A learning objective recommendation method based on a two-stream knowledge embedding network, characterized in that, include: S1 obtains the learner's answer sequence and the knowledge concept corresponding to each question. The learner's answer sequence is as follows: ={( ),( ), ( ), …, ( )};in, For the t-th question in the answer sequence, For students For the question The answer; if the answer is correct, then If the value is 1, then it is 0; output each question and its corresponding knowledge concept as a Q-matrix. For each element in the Q-matrix... If it is 1, then the question in the i-th row... Including relevant knowledge concepts If it is 0, then the question in the i-th row... Excluding the corresponding knowledge concepts Among them, the title Includes at least one knowledge concept; S2 inputs the answer sequence and the corresponding Q matrix into the trained knowledge tracking model, obtains the weight of a single knowledge concept in the knowledge concepts of the answer sequence, outputs the learner's mastery status of each learned knowledge concept, and predicts the learner's correct answer probability for the next answer sequence through the knowledge tracking model; S3 is based on a trained knowledge demand model. It selects the nearest neighbor related knowledge concepts based on the weight of a single learned knowledge concept, outputs the recommendation probability value of the nearest neighbor related knowledge concept, and outputs at least one recommended learning target corresponding to the related knowledge concept to the learner in descending order of recommendation probability value. S4 predicts the learner's mastery of multiple recommended learning objectives based on the trained knowledge demand model, outputs the predicted correct answer probability, arranges the corresponding recommended learning objectives in descending order of the predicted correct answer probability, and generates a recommended learning path. In step S2, after obtaining the answer sequence and the corresponding Q matrix, student embeddings and knowledge concept embeddings are generated, including: The student embedding includes learner information and question text information; The knowledge concept embedding includes the question text information and the corresponding knowledge concept information; Embed the knowledge concepts and the student embedded Serial operation, applying the formula: = ； To further obtain learners' answers to the questions. Generate an embedding that integrates student answers ,Will Input into the knowledge tracing model: = ； when When the value is 1, the learner's answer to the question is correct; when the value is 0, the learner's answer to the question is incorrect.

2. The learning objective recommendation method based on a two-stream knowledge embedding network according to claim 1, characterized in that, In step S2, the questions in the answer sequence are obtained. knowledge concepts The weight of the i-th knowledge concept in the knowledge concepts of the answer sequence includes: Each question The corresponding knowledge concepts are transformed into one-hot encoding. For all knowledge concepts in the one-hot encoding ∈{0,1}, the dimension is the total number of knowledge concepts, K. Based on embedding matrix The one-hot encoding Convert to low-dimensional vector ∈ Its continuous values ​​are: = ; based on The inner product of the knowledge memory vector Perform softmax calculation to obtain the weights. : =Softmax( )= ; Where 1≤i≤K.

3. The learning objective recommendation method based on a two-stream knowledge embedding network according to claim 2, characterized in that, In step S2, a learner's practice answer sequence is updated using an LSTM network. Post-knowledge state Apply the formula: ； Output learner k's current knowledge state , knowledge state and the embedding of knowledge concepts Inputting the sigmoid function, we obtain the answers given by learner k. The probability of a correct result can be calculated using the formula: = ； = ( )； Where h is the dimension of the hidden state vector. , All are vector parameters. , , , as well as , , Both are weight matrices. , , , , For bias vectors, This represents the product operation. (.) and tanh(.) represent the sigmoid and tanh activation functions, respectively; learner k Able to answer questions correctly The probability of answering the question correctly.

4. The learning objective recommendation method based on a two-stream knowledge embedding network according to claim 3, characterized in that, Before step S3, the loss function is calculated: ； in, This represents the learner's true answer to the question. This represents the learner's predicted answer to the question, as predicted by the knowledge tracking model.

5. The learning objective recommendation method based on a two-stream knowledge embedding network according to claim 3, characterized in that, To obtain the difference in learner's grasp of knowledge concepts after studying two adjacent recommended learning objectives, the formula is: ； in, Let be the probability of correctly answering questions related to knowledge concept k at time t. The number of recommendations for the learning objective, which includes knowledge concept k.

6. A learning objective recommendation system based on a two-stream knowledge embedding network, characterized in that, include: The data preprocessing module is used to obtain the learner's answer sequence and the knowledge concept corresponding to each question. The learner's answer sequence is as follows: ={( ),( ), ( ), …, ( )};in, For the t-th question in the answer sequence, For students For the question The answer; if the answer is correct, then If the value is 1, then it is 0; output each question and its corresponding knowledge concept as a Q-matrix. For each element in the Q-matrix... If it is 1, then the question in the i-th row... Including relevant knowledge concepts If it is 0, then the question in the i-th row... Excluding the corresponding knowledge concepts Among them, the title Includes at least one knowledge concept; The answer prediction module is used to input the answer sequence and the corresponding Q matrix into the trained knowledge tracking model, obtain the weight of a single knowledge concept in the knowledge concepts of the answer sequence, output the learner's mastery status of each learned knowledge concept, and predict the learner's correct answer probability for the next answer sequence through the knowledge tracking model. The learning objective recommendation module is used to select the nearest neighboring knowledge concept based on the weight of a single learned knowledge concept, output the recommendation probability value of the nearest neighboring knowledge concept, and output at least one recommended learning objective corresponding to the related knowledge concept to the learner in descending order of recommendation probability value. The learning path generation module is used to predict the learner's mastery of multiple recommended learning objectives based on a trained knowledge demand model, output the predicted correct answer probability, arrange the corresponding recommended learning objectives in descending order of the predicted correct answer probability, and generate a recommended learning path. The answer prediction module is further configured to generate student embeddings and knowledge concept embeddings after obtaining the answer sequence and the corresponding Q matrix, including: The student embedding includes learner information and question text information; The knowledge concept embedding includes the question text information and the corresponding knowledge concept information; Embed the knowledge concepts and the student embedded Serial operation, applying the formula: = ； To further obtain learners' answers to the questions. Generate an embedding that integrates student answers ,Will Input into the knowledge tracing model: = ； when When the value is 1, the learner's answer to the question is correct; when the value is 0, the learner's answer to the question is incorrect.

7. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the learning objective recommendation method as described in any one of claims 1 to 5.

Images