Adaptive learning resource recommendation method based on dynamic knowledge state

Abstract

This invention discloses an adaptive learning resource recommendation method based on dynamic knowledge state, comprising the following steps: Step 1, assessing the objective difficulty of the questions by considering the difficulty of the questions themselves and the hierarchical difficulty of the knowledge concepts; Step 2, assessing the learner's subjective perception of difficulty based on the assessment results of the objective difficulty of the questions; Step 3, assessing the learner's dynamic knowledge state based on the assessment results of the learner's subjective perception of difficulty, and updating the learner's dynamic knowledge state by considering forgetting factors; Step 4, employing a traction mechanism to bridge the gap between the learner's current knowledge state and the difficulty of the next input question, matching questions of appropriate difficulty to the learner, and implementing adaptive learning resource recommendation. This invention dynamically adjusts the learner's knowledge state by considering both the objective difficulty of the questions and the learner's subjective perception of difficulty, ensuring that the recommended questions align with the learner's knowledge state, thus achieving personalized learning resource recommendation.

Description

Technical Field

[0001] This invention belongs to the technical field of smart education recommendation methods, specifically relating to an adaptive learning resource recommendation method based on dynamic knowledge states. Background Technology

[0002] Knowledge tracking, a key branch of educational technology, aims to monitor and understand learners' mastery of specific concepts. Effective knowledge tracking allows educators to personalize their teaching methods and improve learners' learning outcomes. Current knowledge tracking methods primarily focus on accurately predicting learners' performance, which to some extent ignores individual differences and dynamic changes in the learning process. However, these methods generally face a number of challenges. Irregular learner behavior and uneven time allocation lead to data sparsity, limiting situations where only a small portion of concepts are practiced or learned. This data sparsity hinders the accurate capture of learners' overall knowledge status, thus affecting the accuracy and application of knowledge tracking. Furthermore, existing methods often ignore the correlation between changes in learners' knowledge status and test difficulty. Learners' knowledge status may change during the learning process, and test difficulty has a significant impact on learners' performance, but current knowledge tracking methods lack comprehensive consideration of these factors, reducing the accuracy and comprehensiveness of knowledge tracking. Personalized instruction has been recognized as an effective way to improve learning outcomes. To achieve more targeted and personalized teaching, a novel knowledge tracking method is needed that can fully consider the correlation between learners' individual differences, dynamically changing knowledge status, and test difficulty, thereby more accurately understanding learners' needs and potential challenges. Summary of the Invention

[0003] The purpose of this invention is to provide an adaptive learning resource recommendation method for dynamic knowledge states, which solves the problem that existing methods reduce the accuracy and comprehensiveness of knowledge tracking.

[0004] The technical solution adopted in this invention is: an adaptive learning resource recommendation method based on dynamic knowledge state, comprising the following steps:

[0005] Step 1: Assess the objective difficulty of the test questions by considering both the inherent difficulty of the questions themselves and the difficulty of the hierarchical knowledge concepts.

[0006] Step 2: Assess learners' subjective perception of difficulty based on the assessment results of the objective difficulty of the test questions;

[0007] Step 3: Based on the assessment results of learners' subjective perception of difficulty, assess the learners' dynamic knowledge status and update the learners' dynamic knowledge status by taking into account forgetting factors.

[0008] Step 4: Employ a traction mechanism to bridge the gap between the learner's current knowledge level and the difficulty of the next input question, matching learners with questions of appropriate difficulty, and implementing adaptive learning resource recommendation.

[0009] The invention is further characterized in that,

[0010] Step 1 specifically includes the following steps:

[0011] Step 1.1: Use a difficulty level method to measure the difficulty of the question itself. as follows:

[0012]

[0013]

[0014] in, It is the learner answering the question The set, Representing learners First incorrect answer to the question ,constant It is the predefined difficulty level of the test questions themselves. It's a problem The difficulty;

[0015] Step 1.2: Use objective statistical methods to calculate the basic difficulty of each knowledge concept. as follows:

[0016]

[0017]

[0018] in, It is the learner's answer to the knowledge concept The set, Representing learners The first incorrect answer to the knowledge concept ,constant It is a predefined level of difficulty for knowledge concepts. It is a knowledge concept The difficulty;

[0019] Step 1.3: Using Bayesian network reasoning learners Whether or not the prerequisites are known Time for knowledge concepts The cognitive levels are as follows:

[0020]

[0021]

[0022] in, This indicates that learners have mastered the prerequisites. In the case of knowledge concepts cognitive level This indicates that learners have not mastered the prerequisites. In the case of knowledge concepts The level of cognition, if A higher value indicates that the learner has mastered the prerequisites even without prior knowledge. There is also a relatively high probability of mastering it. Indirectly explaining knowledge concepts right The degree of dependence is low; however, A lower value indicates that the learner has mastered the prerequisites. It is also unlikely that they will be able to master it. It can explain the concept of knowledge. High difficulty;

[0023] Step 1.4: Based on the basic difficulty obtained in Step 1.2 Compared with the learner's understanding of knowledge concepts obtained in step 1.3 cognitive level or Computational Hierarchical Knowledge Concept Difficulty as follows:

[0024]

[0025] Step 1.5: Adjust the difficulty of the questions obtained in Step 1.1. The difficulty of the hierarchical knowledge concepts obtained in step 1.4 By performing correlation, we can obtain the objective difficulty of the test questions:

[0026]

[0027]

[0028] in, and These represent the question and the question-answer pair, respectively. Indicates the learner at any given time Answer the questions The knowledge concept, Indicates the learner at any given time For the test questions The response vector, This indicates element-wise addition. and It is a scalar difficulty parameter.

[0029] Step 2 uses an attention mechanism to model learners' subjective perception of difficulty, establishing connections between the knowledge concepts in the test questions and each knowledge concept the learner has previously answered. The attention weight of the current knowledge concept to past knowledge concepts depends not only on the similarity between the corresponding queries and keys, but also on the relative time steps between the current knowledge concept and past knowledge concepts. Specifically, it includes the following steps:

[0030] Step 2.1: Calculate the relative time step between the current knowledge concept and past knowledge concepts. :

[0031]

[0032] in, It adjusts the distance between continuous-time indicators based on the correlation between past practice knowledge concepts and current knowledge concepts; and They represent the time steps respectively. The query and key for the problem It is an exponential function. It is a scaling factor. Indicates transpose;

[0033] Step 2.2: Based on the relative time step obtained in Step 2.1 Calculate the attention weight of the current knowledge concept to the past knowledge concept. :

[0034]

[0035] Step 2.3: Based on the attention weights obtained in Step 2.2 AND value By taking the inner product, we obtain the attention score:

[0036]

[0037] in, , and They represent the time steps respectively. The query, key, and value for the problem; It is a penalty factor, determined by the similarity between the query and the key. It is a learnable decay parameter. It is an activation function. It is an exponential function. It is a scaling factor. It is the attention score;

[0038] Step 2.4, at time The questions that take into account the objective difficulty of the test items obtained in step 1.5 and the attention scores in step 2.3 are used as the embedding vector for learners' subjective perception of difficulty. and :

[0039]

[0040]

[0041] in, and Represents the weight matrix. and These are query, key, and value. It is the attention score. This indicates element-wise addition.

[0042] Step 3 specifically includes the following steps:

[0043] Step 3.1: Embed the vector obtained in step 2.4 First, with the embedding matrix Multiplying them yields a continuous embedding vector. By taking With each keyway The attention weights are obtained by performing an inner product and passing it through an activation function. :

[0044]

[0045] in, , It is a weight matrix. It is a bias vector, a matrix Indicate the relationship between potential concepts;

[0046] Step 3.2: Based on the attention weights obtained in Step 3.1 Knowledge acquisition by computational learners :

[0047]

[0048] Among them, matrix Indicates the learner at each time step The knowledge state for each knowledge concept;

[0049] Step 3.3: Update the value matrix based on the correctness of the learner's answer, using the embedding matrix. Embedded tuples The knowledge gain obtained by learners after completing the test questions. When writing the learner's knowledge gain into the value matrix, the memory is erased before adding new information; the erase signal is: The matrix of values ​​from the previous timestamp The memory vector is modified as follows: After erasing, use the added vector. To update each memory slot, and at each timestamp Consider forget signals to update value memory :

[0050]

[0051]

[0052]

[0053]

[0054] in, , It is the time interval between two consecutively learned embeddings. It is a forgetting factor. It is element-wise multiplication. and It is a weight matrix. and This is the bias vector.

[0055] The guiding mechanism used in step 4 is specifically based on the learner's knowledge state sequence in the previous step. Adjust the difficulty of the next learning task to bridge the gap in learners' current knowledge level. Difficulty of the next input question The difference between them, calculate the traction value as follows:

[0056]

[0057]

[0058]

[0059] in, and For activation function, and It is a weight matrix. and It is a bias vector. yes The dynamic state of knowledge at any given moment. yes The question of time, It is a direct output of the difference between the learner's previous knowledge level and the current test question. The current test question includes the learner's subjective perception of difficulty, and a gate has been further designed. To select and retain Important features of the middle;

[0060] Using a time step of -1 knowledge state vector and time step are The difficulty of embedding The inner product is used to simulate the learner's practical process of applying learned knowledge to answer questions, and then based on the traction value... A feedforward response network is used to predict the probability of answering the question correctly at each time step. Match learners with questions of appropriate difficulty and implement adaptive learning resource recommendations:

[0061]

[0062] in, For activation function, This is the weight matrix. For bias vectors, yes Dynamic knowledge state at time -1 yes The question is about timing.

[0063] The beneficial effects of this invention are as follows: The adaptive learning resource recommendation method based on dynamic knowledge state of this invention considers the difficulty of the questions themselves and the hierarchical difficulty of knowledge concepts to model the objective test difficulty. Specifically, a Bayesian network is used to model the learner's cognitive level at the knowledge concept level; an attention mechanism is embedded to embed the learner's subjective difficulty perception, and the learner's personalized knowledge acquisition when answering objective test questions of different difficulties is evaluated based on the learner's subjective difficulty perception. This mechanism combines forgetting factors with the learner's knowledge state through a memory matrix, comprehensively updating the learner's constantly changing knowledge state during the learning process; to maintain consistency between knowledge state and test difficulty, an adaptive neural network is used to match questions of appropriate difficulty to the learner, achieving personalized learning resource recommendation. Attached Figure Description

[0064] Figure 1 This is a flowchart illustrating the adaptive learning resource recommendation method based on dynamic knowledge state of the present invention. Detailed Implementation

[0065] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0066] Example 1

[0067] This invention provides an adaptive learning resource recommendation method based on dynamic knowledge state, such as... Figure 1 As shown, specifically: First, an objective difficulty assessment of the test questions is conducted, taking into account both the difficulty of the questions themselves and the difficulty of the hierarchical knowledge concepts. Second, based on learners' subjective perception of difficulty, the assessment evaluates learners' personalized knowledge acquisition when answering questions of varying objective difficulty, combining forgetting factors with learners' knowledge status to comprehensively update learners' constantly changing knowledge status during the learning process. Finally, maintaining consistency between knowledge status and test question difficulty is crucial for knowledge tracing tasks. On the one hand, research on test question difficulty allows for the assessment of learners' knowledge status; on the other hand, knowledge status-based assessment utilizes adaptive neural networks to dynamically match questions of appropriate difficulty to learners.

[0068] This invention addresses the problem of data sparsity caused by irregular learner behavior and uneven time allocation in existing technologies. Furthermore, it resolves the impact of test difficulty on knowledge status, specifically:

[0069] (1) The difficulty of objective test questions is assessed by considering the difficulty of the questions themselves and the difficulty of hierarchical knowledge concepts. By modeling learners’ cognitive level in the knowledge concept hierarchy through Bayesian network modeling, the problem of data sparsity is solved.

[0070] (2) When modeling learners’ subjective difficulty perception, attention mechanism is used to evaluate learners’ personalized knowledge acquisition when answering different objective test questions based on subjective difficulty perception. This mechanism combines forgetting factors with learners’ knowledge status and comprehensively updates learners’ constantly changing knowledge status through memory matrix, thus solving the problem of test question difficulty on knowledge status.

[0071] (3) In order to maintain the consistency between knowledge status and test difficulty, an adaptive neural network is used to dynamically match questions of appropriate difficulty to learners, so as to achieve the purpose of personalized learning.

[0072] Through the above methods, the adaptive learning resource recommendation method based on dynamic knowledge state provided by this invention mainly adjusts the learner's knowledge state dynamically by considering the objective difficulty of the test questions and the learner's subjective perception of difficulty, so that the recommended test questions are consistent with the learner's knowledge state, thereby achieving personalized learning resource recommendation.

[0073] Example 2

[0074] This invention provides an adaptive learning resource recommendation method based on dynamic knowledge state, which is implemented according to the following steps:

[0075] Step 1: Assess the objective difficulty of the test questions by considering both the inherent difficulty of the questions themselves and the hierarchical difficulty of the knowledge concepts presented.

[0076] Assessing test difficulty is a quantitative or qualitative method. Research on test difficulty mainly focuses on two aspects: firstly, it measures the difficulty level of test questions in exams or tests, helping teachers and assessors understand the difficulty level of questions and thus better organize teaching and assessment activities. Secondly, it uses test difficulty to assess learners' knowledge status during the learning process, a process of great significance in the fields of education and psychology. Generally, there is a close relationship between test difficulty and learners' knowledge level. When learners face relatively difficult questions, those with higher knowledge levels are usually more likely to answer correctly, while those with lower knowledge levels find it more difficult. From the perspective of test design, the difficulty of the questions themselves has a significant impact on measuring students' knowledge levels. If the questions are too easy or too difficult, they cannot effectively differentiate learners with different knowledge levels. Furthermore, the difficulty of test questions is closely related to the knowledge concepts involved. Generally, the more difficult the question, the more knowledge concepts need to be mastered to answer correctly. It is worth noting that there are complex relationships between different knowledge concepts. For example, if a learner has mastered quadratic functions, they are likely to also master linear functions, but the reverse is not true. Therefore, in assessing the difficulty of test questions, this invention comprehensively considers both the inherent difficulty of the questions themselves and the difficulty of the hierarchical knowledge concepts. The assessment of test question difficulty will provide an important basis for subsequent resource recommendations, ensuring that the recommended resources are suitable for the learner's level. Specifically, it is implemented according to the following steps:

[0077] Step 1.1, Difficulty assessment of the question itself:

[0078] Specific indicators of question difficulty should be determined, including the complexity of the question stem, the difficulty of the answer, and the level of abstraction of the question. Consider using a difficulty level method to measure the difficulty of the question.

[0079]

[0080]

[0081] in, It is the learner answering the questions The set, Representing learners First incorrect answer to the question ,constant This refers to the predefined difficulty level of the questions themselves. For questions answered by fewer than five different learners in the dataset, this invention sets the difficulty level of these questions to a constant. .

[0082] Step 1.2, the hierarchical measurement of knowledge concept difficulty is divided into three stages: (1) defining the basic difficulty of each concept; (2) considering the correlation between knowledge concepts to ensure that learning the previous knowledge concept has certain prerequisites for learning the next knowledge concept; (3) gradually increasing the difficulty of knowledge concepts at each level:

[0083] Step 1.2.1: Use objective statistical methods to calculate the basic difficulty of each knowledge concept. :

[0084]

[0085]

[0086] in, It is the learner's answer to the knowledge concept The set, Representing learners The first incorrect answer to the knowledge concept ,constant This refers to a predefined level of difficulty for the knowledge concepts. For knowledge concepts in the dataset that are answered by fewer than three different students, this invention sets the difficulty level of these knowledge concepts to a constant. .

[0087] Step 1.2.2, using Bayesian network inference learners Whether or not the prerequisites are known Time for knowledge concepts The cognitive levels are as follows:

[0088]

[0089]

[0090] in, This indicates that learners have mastered the prerequisites. In the case of knowledge concepts cognitive level This indicates that learners have not mastered the prerequisites. In the case of knowledge concepts The level of cognition, if A higher value indicates that the learner has mastered the prerequisites even without prior knowledge. There is also a relatively high probability of mastering it. Indirectly explaining knowledge concepts right The degree of dependence is low; however, A lower value indicates that the learner has mastered the prerequisites. It is also unlikely that they will be able to master it. It can explain the concept of knowledge. High difficulty;

[0091] Step 1.2.3: Gradually increase the difficulty of the knowledge concepts at each level:

[0092]

[0093] Step 1.3 involves correlating the objective difficulty of the test questions by considering both the inherent difficulty of the questions themselves and the difficulty of the hierarchical knowledge concepts, thus obtaining the objective difficulty of the test questions:

[0094]

[0095]

[0096] in, and These represent the question and the question-answer pair, respectively. Indicates the learner at any given time Answer the questions The knowledge concept, Indicates the learner at any given time For the test questions The response vector, Indicate the problem The difficulty It is a hierarchical knowledge concept The difficulty This indicates element-wise addition. and It is a scalar difficulty parameter.

[0097] Step 2: Assess learners' subjective perception of difficulty based on the assessment results of the objective difficulty of the test items.

[0098] This paper uses an attention mechanism to model learners' subjective perception of difficulty, establishing connections between knowledge concepts in the test questions and each knowledge concept the learner has previously answered. Each knowledge concept and knowledge concept-answer pair has a key, query, and value embedding layer. Therefore, the attention weight of the current knowledge concept to past knowledge concepts depends not only on the similarity between the corresponding queries and keys, but also on the relative time steps between the current and past knowledge concepts.

[0099] First, calculate the relative time step between the current knowledge concept and past knowledge concepts. :

[0100]

[0101] in, It adjusts the distance between continuous time indicators based on the relevance between past practice knowledge concepts and current knowledge concepts. and They represent the time steps respectively. The query and key for the problem It is an exponential function. It is a scaling factor. This indicates transpose.

[0102] Secondly, based on the relative time step Calculate the attention weight of the current knowledge concept to the past knowledge concept. Then, the attention weights obtained will be... AND value By taking the inner product, we obtain the attention score:

[0103]

[0104]

[0105] in, , and They represent the time steps respectively. The query, key, and value for the problem. It is a penalty factor, determined by the similarity between the query and the key. It is a learnable decay parameter. It is an activation function. It is an exponential function. It is a scaling factor. It is the attention score.

[0106] Finally, at the moment The question scores, which take into account the objective difficulty of the test questions, and the attention scores are used as the embedding vectors for learners' subjective perception of difficulty. and :

[0107]

[0108]

[0109] in, and Represents the weight matrix. and These are query, key, and value. It is the attention score. This indicates element-wise addition.

[0110] Step 3: Based on the assessment results of learners' subjective perception of difficulty, evaluate the learners' dynamic knowledge status and update the learners' dynamic knowledge status by taking into account forgetting factors.

[0111] Exploring changes in learners' knowledge states during the learning process is a key task in adaptive resource recommendation. By considering question difficulty, this study uses dynamic key-value pairs to store learners' knowledge states in a matrix. A matrix represents the relationship between latent concepts. Indicates the learner at each time step The knowledge state for each knowledge concept.

[0112] Step 3.1, Embedding Vector First, with the embedding matrix Multiplying them yields a continuous embedding vector. By taking With each keyway The attention weights are obtained by performing an inner product and passing it through an activation function. :

[0113]

[0114] Among them, and , It is a weight matrix. It is the bias vector.

[0115] Step 3.2: Calculate the learner's knowledge state based on attention weights. :

[0116]

[0117] Step 3.3: As learners continue learning, their proficiency will decline over time, causing their knowledge status to be affected by both forgetting factors and knowledge acquisition. Therefore, it is necessary to continuously update and adjust the dynamic knowledge status model to reflect the learner's latest state. This can be achieved through real-time learner behavior data and test feedback information, ensuring the accuracy and timeliness of the knowledge status. Therefore, this invention uses a forgetting gate to simulate the forgetting phenomenon, combining forgetting factors with the learner's knowledge status to comprehensively update the learner's constantly changing knowledge status during the learning process. After the learner answers the test questions, the value matrix is ​​updated based on the correctness of the answer, and the embedding matrix B is used to embed the tuples. To obtain the knowledge gain that learners gain after completing this test question. When writing the learner's knowledge gain into the value matrix, the memory is erased before adding new information. The erase signal is: The matrix of values ​​from the previous timestamp The memory vector is modified as follows: After erasing, use the added vector. To update each memory slot, and at each timestamp Consider forget signals to update value memory :

[0118]

[0119]

[0120]

[0121]

[0122] in , It is the time interval between two consecutively learned embeddings. It is a forgetting factor. It is element-wise multiplication. and It is a weight matrix. and This is the bias vector.

[0123] Step 4: Employ a traction mechanism to bridge the gap between the learner's current knowledge level and the difficulty of the next input question, matching learners with questions of appropriate difficulty, and implementing adaptive learning resource recommendation.

[0124] Adaptive resource recommendation is implemented based on learners' dynamic knowledge status and test difficulty. This includes recommending educational resources that suit learners' current knowledge level and considering the degree of matching in terms of test difficulty. Recommended resources may include textbooks, exercises, multimedia teaching materials, etc., to better meet learners' personalized needs. The specific implementation follows these steps:

[0125] Based on the learner's knowledge state sequence in the previous step Adjust the difficulty of the next learning task to bridge the gap in learners' current knowledge level. Difficulty of the next input question The difference between them, calculate the traction value as follows:

[0126]

[0127]

[0128]

[0129] in, and For activation function, and It is a weight matrix. and It is a bias vector. yes The dynamic state of knowledge at any given moment. yes The question of time, It is a direct output of the difference between the learner's previous knowledge level and the current test question. The current test question includes the learner's subjective perception of difficulty, and a gate has been further designed. To select and retain The key feature of the time step is -1 knowledge state vector and time step are The difficulty of embedding The inner product is used to simulate the learner's practical process of applying learned knowledge to answer questions, and then based on the traction value... A feedforward response network is used to predict the probability of answering the question correctly at each time step. Match learners with questions of appropriate difficulty and implement adaptive learning resource recommendations:

[0130]

[0131] in, For activation function, This is the weight matrix. For bias vectors, yes Dynamic knowledge state at time -1 yes The question is about timing.

[0132] This invention presents an adaptive learning resource recommendation method based on dynamic knowledge state. By considering the influence of objective test difficulty and learner's subjective perception of difficulty on dynamic knowledge state, it explores a novel knowledge tracking paradigm. Specifically, this invention first models the objective test difficulty by considering the difficulty of the questions themselves and the hierarchical difficulty of knowledge concepts; then, it assesses the learner's personalized acquisition when answering objective test questions of different difficulties based on the learner's subjective perception of difficulty, and dynamically updates the learner's knowledge state by incorporating forgetting factors; finally, it uses a feedforward adaptive neural network to keep the knowledge state consistent with the difficulty of the test questions faced by the learner, providing a basis for learning resource recommendation and ensuring that the test questions appearing in the recommendation list better match the learner's target needs.

[0133] Example 3

[0134] This invention provides an adaptive learning resource recommendation method based on dynamic knowledge state, which divides the dataset into training, validation, and test sets in a 6:2:2 ratio, such as... Figure 1 As shown, the method includes the following steps:

[0135] Step 1: Data Collection and Preprocessing

[0136] The datasets used in this example can be obtained from publicly available smart education datasets. Four datasets—ASSISTment2009, ASSISTment2015, ASSISTment2017, and STATICS2011—were preprocessed. Taking ASSISTment2009 as an example, this dataset, provided by the ASSISTMENTS online tutoring platform, contains 325,637 response records from 4,151 learners on 110 knowledge concepts and 16,891 questions. To better estimate the difficulty distribution of each question and knowledge concept, questions with fewer than 5 responses and knowledge concepts with fewer than 3 responses were filtered out.

[0137] For computational efficiency, the sequence length seqlen was set to 200. For responses exceeding 200, the sequence was split into several unique subsequences based on the sequence length. All parameters were initialized randomly using a uniform distribution. The dimensions of the knowledge state and question embedding were both 256. The learning rate was 1e-5. The batch size was set to 32. The number of heads in the attention layer was 8. The Adam optimizer was used to train all models. Dropout layers were added to prevent overfitting. All experiments were performed on an NVIDIA GeForce RTX 3050 Laptop GPU.

[0138] Step 2: Objective Test Difficulty Assessment

[0139] Step 2.1: The difficulty of the question itself

[0140] The difficulty of a question is a metric used to assess how easy or difficult a question is relative to a group of test takers. This difficulty is typically measured by the average performance of the test takers. Specifically, the difficulty of questions is divided into 10 levels, calculated by determining the percentage of learners attempting the question for the first time out of all learners who attempt it. For questions that few learners attempt, a fixed difficulty level is assigned to ensure the difficulty remains consistent with other questions.

[0141]

[0142]

[0143] in, It is the learner answering the questions The set, Representing learners First incorrect answer to the question ,constant This refers to the predefined difficulty level of the questions themselves. For questions answered by fewer than five different learners in the dataset, this invention sets the difficulty level of these questions to a constant. .

[0144] Step 2.2: Hierarchical Difficulty of Knowledge Concepts

[0145] Because the different knowledge concepts involved in the problem have their own levels of difficulty—for example, quadratic functions are more difficult than linear functions, and composite functions are more challenging than quadratic functions—an objective statistical method is used to calculate the difficulty of the knowledge concepts, inspired by classical testing theory and previous research.

[0146]

[0147]

[0148] in, It is the learner's answer to the knowledge concept The set, Representing learners The first incorrect answer to the knowledge concept ,constant This refers to a predefined level of difficulty for the knowledge concepts. For knowledge concepts in the dataset that are answered by fewer than three different students, this invention sets the difficulty level of these knowledge concepts to a constant. .

[0149] Bayesian networks are used to effectively model learners' cognitive states at the knowledge concept level. For example, when a student wants to solve a problem about "composite functions," they first need to master the knowledge concepts of "quadratic functions" and "the parity of functions." In other words, the knowledge concepts of "quadratic functions" and "the parity of functions" are prerequisites for the knowledge concept of "composite functions." Whether or not the prerequisites are known Time for knowledge concepts Cognitive level is represented as:

[0150]

[0151]

[0152] in, This indicates that learners have mastered the prerequisites. In the case of knowledge concepts cognitive level This indicates that learners have not mastered the prerequisites. In the case of knowledge concepts The level of cognition, if A higher value indicates that the learner has mastered the prerequisites even without prior knowledge. There is also a relatively high probability of mastering it. Indirectly explaining knowledge concepts right The degree of dependence is low; however, A lower value indicates that the learner has mastered the prerequisites. It is also unlikely that they will be able to master it. It can explain the concept of knowledge. It is quite difficult.

[0153] The difficulty of the knowledge concepts is gradually increased at each level:

[0154]

[0155] in, It represents the basic difficulty level of each knowledge concept. Indicate learner Knowledge concepts given the prerequisites cognitive level Indicate learner Knowledge concepts without prior knowledge cognitive level

[0156] Step 2.3: Correlation Analysis

[0157] Set the sequence length seqlen to 200. For responses exceeding 200, divide the sequence into several unique subsequences based on the sequence length. Then, embed the difficulty of the question itself and the difficulty of the knowledge concept into the model's input. The calculation formula is as follows:

[0158]

[0159]

[0160] in, and These represent the question and the question-answer pair, respectively. Indicates the learner at any given time Answer the questions The knowledge concept, Indicates the learner at any given time For the test questions The response vector, Indicate the problem The difficulty It is a hierarchical knowledge concept The difficulty This indicates element-wise addition. and It is a scalar difficulty parameter.

[0161] Step 3: Learner's subjective perception of difficulty:

[0162] Attention mechanisms are used to model learners' subjective perception of difficulty, establishing connections between knowledge concepts in the test questions and each knowledge concept the learner has previously answered. Considering learner memory decay, when predicting a learner's answer to the current question, knowledge concepts answered long ago are less relevant than those answered more recently. Each knowledge concept and knowledge concept-answer pair has a key, query, and value embedding layer; therefore, the attention weight of the current knowledge concept to past knowledge concepts depends not only on the similarity between the corresponding queries and keys but also on their relative time steps.

[0163] Calculate the relative time step between the current knowledge concept and the past knowledge concepts. :

[0164]

[0165] in, It adjusts the distance between continuous time indicators based on the relevance between past practice knowledge concepts and current knowledge concepts. and They represent the time steps respectively. The query and key for the problem It is an exponential function. It is a scaling factor. This indicates transpose.

[0166] Based on relative time step Calculate the attention weight of the current knowledge concept to the past knowledge concept. Based on the weight of attention AND value By taking the inner product, we obtain the attention score:

[0167]

[0168]

[0169] in, , and They represent the time steps respectively. The query, key, and value for the problem. It is a penalty factor, determined by the similarity between the query and the key. It is a learnable decay parameter. It is an activation function. It is an exponential function. It is a scaling factor. It is the attention score.

[0170] At any moment The question scores, which take into account the objective difficulty of the test questions, and the attention scores are used as the embedding vectors for learners' subjective perception of difficulty. and :

[0171]

[0172]

[0173] in, and Represents the weight matrix. and These are query, key, and value. It is the attention score. This indicates element-wise addition.

[0174] Step 4: Assess the learner's dynamic knowledge status

[0175] Based on the assessment results of objective test difficulty and learners' subjective perception of difficulty, a dynamic knowledge state modeling method is adopted, using a matrix... A matrix represents the relationship between latent concepts. Indicates the learner at each time step For each knowledge concept, accurately capture the changes in the learner's knowledge state over time, assess the learner's personalized knowledge acquisition when answering questions and knowledge concepts of varying difficulty, and dynamically update the learner's knowledge state by considering forgetting factors. Specifically, implement this according to the following steps:

[0176] Step 4.1: Embedding Vector First, with the embedding matrix Multiplying them yields a continuous embedding vector. By taking With each keyway The attention weights are obtained by performing an inner product and passing it through an activation function. :

[0177]

[0178] in, , It is a bias vector, a matrix It indicates the relationship between potential concepts.

[0179] Step 4.2: Calculate the learner's knowledge state based on attention weights. :

[0180]

[0181] Step 4.3: Learners' proficiency also declines over time, causing the updating of their knowledge state to be influenced by both forgetting factors and knowledge acquisition. Therefore, this example uses a forgetting gate to simulate the forgetting phenomenon, combining forgetting factors with the learner's knowledge state to comprehensively update the learner's constantly changing knowledge state during the learning process. After the learner answers the questions, the value matrix is ​​updated based on the correctness of the learner's answers, and the embedding matrix B is used to embed the tuples. To obtain the knowledge gain that learners gain after completing this test question. When writing the learner's knowledge gain into the value matrix, the memory is erased before adding new information. The erase signal is: The matrix of values ​​from the previous timestamp The memory vector is modified as follows: After erasing, use the added vector. To update each memory slot, and at each timestamp Consider forget signals to update value memory :

[0182]

[0183]

[0184]

[0185]

[0186] in, It is the time interval between two consecutively learned embeddings. It is a forgetting factor. It is element-wise multiplication. and It is a weight matrix. and This is the bias vector.

[0187] Step 5: Adaptive Resource Recommendation

[0188] Based on the learner's knowledge state sequence in the previous step Adjust the difficulty of the next learning task to bridge the gap in learners' current knowledge level. Difficulty of the next input question The difference between them, calculate the traction value as follows:

[0189]

[0190]

[0191]

[0192] in, and For activation function, and It is a weight matrix. and It is a bias vector. yes The dynamic state of knowledge at any given moment. yes The question of time, It is a direct output of the difference between the learner's previous knowledge level and the current test question. The current test question includes the learner's subjective perception of difficulty, and a gate has been further designed. To select and retain Important features of the middle;

[0193] Using a time step of -1 knowledge state vector and time step are The difficulty of embedding The inner product is used to simulate the learner's practical process of applying learned knowledge to answer questions, and then based on the traction value... A feedforward response network is used to predict the probability of answering the question correctly at each time step. Match learners with questions of appropriate difficulty and implement adaptive learning resource recommendations:

[0194]

[0195] in, For activation function, This is the weight matrix. For bias vectors, yes Dynamic knowledge state at time -1 yes The question is about timing.

Claims

1. An adaptive learning resource recommendation method based on dynamic knowledge state, characterized in that, Includes the following steps: Step 1: Assess the objective difficulty of the test questions by considering both the inherent difficulty of the questions themselves and the difficulty of the hierarchical knowledge concepts. Step 2: Assess learners' subjective perception of difficulty based on the assessment results of the objective difficulty of the test questions; Step 3: Based on the assessment results of learners' subjective perception of difficulty, assess the learners' dynamic knowledge status and update the learners' dynamic knowledge status by taking into account forgetting factors. Step 4: Employ a traction mechanism to bridge the gap between the learner's current knowledge state and the difficulty of the next input question, matching the learner with questions of appropriate difficulty, and implementing adaptive learning resource recommendation; the traction mechanism specifically involves: based on the learner's knowledge state sequence in the previous step... Adjust the difficulty of the next learning task to bridge the gap in learners' current knowledge level. Difficulty of the next input question The difference between them, calculate the traction value as follows: in, and For activation function, and It is a weight matrix. and It is a bias vector. yes The dynamic state of knowledge at any given moment. yes The question of time, It is a direct output of the difference between the learner's previous knowledge level and the current test question. The current test question includes the learner's subjective perception of difficulty, and a gate has been further designed. To select and retain Important features of the middle; Using a time step of -1 knowledge state vector and time step are The difficulty of embedding The inner product is used to simulate the learner's practical process of applying learned knowledge to answer questions, and then based on the traction value... A feedforward response network is used to predict the probability of answering the question correctly at each time step. Match learners with questions of appropriate difficulty and implement adaptive learning resource recommendations: in, For activation function, This is the weight matrix. For bias vectors, yes Dynamic knowledge state at time -1 yes The question is about timing.

2. The adaptive learning resource recommendation method based on dynamic knowledge state as described in claim 1, characterized in that, Step 1 specifically includes the following steps: Step 1.1: Use a difficulty level method to measure the difficulty of the question itself. as follows: in, It is the learner answering the question The set, Representing learners First incorrect answer to the question ,constant It is the predefined difficulty level of the test questions themselves. It's a problem The difficulty; Step 1.2: Use objective statistical methods to calculate the basic difficulty of each knowledge concept. as follows: in, It is the learner's answer to the knowledge concept The set, Representing learners The first incorrect answer to the knowledge concept ,constant It is a predefined level of difficulty for knowledge concepts. It is a knowledge concept The difficulty; Step 1.3: Using Bayesian network reasoning learners Whether or not the prerequisites are known Time for knowledge concepts The cognitive levels are as follows: in, This indicates that learners have mastered the prerequisites. In the case of knowledge concepts cognitive level This indicates that learners have not mastered the prerequisites. In the case of knowledge concepts The level of cognition, if A higher value indicates that the learner has mastered the prerequisites even without prior knowledge. There is also a relatively high probability of mastering it. Indirectly explaining knowledge concepts right The degree of dependence is low; however, A lower value indicates that the learner has mastered the prerequisites. It is also unlikely that they will be able to master it. It can explain the concept of knowledge. High difficulty; Step 1.4: Based on the basic difficulty obtained in Step 1.2 Compared with the learner's understanding of knowledge concepts obtained in step 1.3 cognitive level or Computational Hierarchical Knowledge Concept Difficulty as follows: Step 1.5: Adjust the difficulty of the questions obtained in Step 1.

1. The difficulty of the hierarchical knowledge concepts obtained in step 1.4 By performing correlation, we can obtain the objective difficulty of the test questions: in, and These represent the question and the question-answer pair, respectively. Indicates the learner at any given time Answer the questions The knowledge concept, Indicates the learner at any given time For the test questions The response vector, This indicates element-wise addition. and It is a scalar difficulty parameter.

3. The adaptive learning resource recommendation method based on dynamic knowledge state as described in claim 2, characterized in that, Step 2 uses an attention mechanism to model learners' subjective perception of difficulty, establishing a connection between the knowledge concepts in the test questions and each knowledge concept the learner has previously answered. The attention weight of the current knowledge concept to past knowledge concepts depends not only on the similarity between the corresponding queries and keys, but also on the relative time steps between the current knowledge concept and past knowledge concepts. Specifically, it includes the following steps: Step 2.1: Calculate the relative time step between the current knowledge concept and past knowledge concepts. : in, It adjusts the distance between continuous-time indicators based on the correlation between past practice knowledge concepts and current knowledge concepts; and They represent the time steps respectively. The query and key for the problem It is an exponential function. It is a scaling factor. Indicates transpose; Step 2.2: Based on the relative time step obtained in Step 2.1 Calculate the attention weight of the current knowledge concept to the past knowledge concept. : Step 2.3: Based on the attention weights obtained in Step 2.2 AND value By taking the inner product, we obtain the attention score: in, , and They represent the time steps respectively. The query, key, and value for the problem; It is a penalty factor, determined by the similarity between the query and the key. It is a learnable decay parameter. It is an activation function. It is an exponential function. It is a scaling factor. It is the attention score; Step 2.4, at time The questions that take into account the objective difficulty of the test items obtained in step 1.5 and the attention scores in step 2.3 are used as the embedding vector for learners' subjective perception of difficulty. and : in, and Represents the weight matrix. and These are query, key, and value. It is the attention score. This indicates element-wise addition.

4. The adaptive learning resource recommendation method based on dynamic knowledge state as described in claim 3, characterized in that, Step 3 specifically includes the following steps: Step 3.1: Embed the vector obtained in step 2.4 First, with the embedding matrix Multiplying them yields a continuous embedding vector. By taking With each keyway The attention weights are obtained by performing an inner product and passing it through an activation function. : in, , It is a weight matrix. It is a bias vector, a matrix Indicate the relationship between potential concepts; Step 3.2: Based on the attention weights obtained in Step 3.1 Knowledge acquisition by computational learners : Among them, matrix Indicates the learner at each time step The knowledge state for each knowledge concept; Step 3.3: Update the value matrix based on the correctness of the learner's answer, using the embedding matrix. Embedded tuples The knowledge gain obtained by learners after completing the test questions. When writing the learner's knowledge gain into the value matrix, the memory is erased before adding new information; the erase signal is: The matrix of values ​​from the previous timestamp The memory vector is modified as follows: After erasing, use the added vector. To update each memory slot, and at each timestamp Consider forget signals to update value memory : in, , It is the time interval between two consecutively learned embeddings. It is a forgetting factor. It is element-wise multiplication. and It is a weight matrix. and This is the bias vector.

Images