Multi-expert collaborative knowledge tracking method and system for cognitive strategy perception

Abstract

The invention provides a multi-expert collaborative knowledge tracking method and system for cognitive strategy perception, and relates to the technical field of knowledge tracking, and the method comprises the steps: building a memory consolidation expert module through combining a knowledge obtaining process with a forgetting gate and absorptivity parameters, building an attention regulation expert module through combining attention distribution of knowledge points with a time attention mechanism, and building a memory consolidation expert module; creating a structure adaptation expert module based on a correlation evolution process between knowledge points; generating a global memory matrix by combining an attention mechanism based on the long-term memory of the historical cognitive state of the student and the current interaction embedded information, and determining a corresponding expert weight; performing weighted fusion on knowledge state parameters generated by the memory consolidation expert module, the attention regulation and control expert module and the structure adaptation expert module on the basis of expert weights, splicing the obtained fused knowledge state parameters and feature embedding vectors of test questions to be predicted, and inputting the spliced information to a preset feedforward neural network, and obtaining the answer accuracy corresponding to the test question to be predicted.

Description

Technical Field

[0001] This invention relates to the field of knowledge tracing technology, and in particular to a multi-expert collaborative knowledge tracing method and system based on cognitive strategy perception. Background Technology

[0002] With the increasing demand for personalized education, how to monitor and assess the dynamic evolution of learners' knowledge status in real time has become a key issue that urgently needs to be addressed in the field of educational technology. Against this backdrop, knowledge tracing methods have emerged. These methods construct dynamic models reflecting learners' cognitive development processes by analyzing their historical interaction data. Based on inferences about learners' potential mastery of specific concepts, knowledge tracing can help intelligent tutoring systems accurately predict their subsequent performance and support the implementation of effective personalized interventions. As a cognitive modeling method, the core objective of knowledge tracing technology is to track the dynamic changes in learners' potential cognitive states over time. Research in this field is developing towards a deeper integration of human cognitive elements, resulting in various models with clear cognitive mechanisms. For example, in terms of memory mechanisms, Deep Knowledge Tracing (DKT) uses recurrent neural networks to implicitly simulate temporal patterns in learning behavior, while Dynamic Key-Value Memory Networks (DKVMN) introduce external memory modules to achieve explicit storage and updating of conceptual knowledge. In terms of attention mechanisms, Attentional Knowledge Tracing (AKT) and Self-Attentional Knowledge Tracing (SAKT) dynamically capture changes in learners' focus through attention weights.

[0003] Despite significant advancements in existing knowledge tracing technologies, most methods still encode learning dynamics into a single latent representation. This holistic design not only struggles to explicitly decouple different cognitive processes but also ignores individual differences in learners' regulatory mechanisms, such as attention control, information integration, and conceptual reconstruction. The ability to model these mechanisms separately remains insufficient, limiting the accuracy and predictive validity of their representations of cognitive states. Summary of the Invention

[0004] In view of this, the present invention proposes a multi-expert collaborative knowledge tracing method and system for cognitive strategy perception.

[0005] The technical solution of this invention is implemented as follows: The first aspect of this invention provides a multi-expert collaborative knowledge tracing method for cognitive strategy perception, comprising:

[0006] A memory consolidation expert module is created by combining the knowledge acquisition process with the forgetting gate and the absorption rate parameter; an attention regulation expert module is created by combining the attention distribution of knowledge points with the time attention mechanism; and a structural adaptation expert module is created based on the evolutionary process of the association between knowledge points. The absorption rate parameter is related to students' learning ability and the discrimination of test questions.

[0007] Based on the long-term memory of students’ historical cognitive state and the embedded information of current interaction, a global memory matrix is ​​generated by combining the attention mechanism, and corresponding expert weights are assigned to the memory consolidation expert module, the attention regulation expert module and the structural adaptation expert module through the global memory matrix.

[0008] Based on the expert weights, the first knowledge state parameter generated by the memory consolidation expert module, the second knowledge state parameter generated by the attention regulation expert module, and the third knowledge state parameter generated by the structure adaptation expert module are weighted and fused to obtain the fused knowledge state parameter.

[0009] The fused knowledge state parameters are concatenated with the feature embedding vector of the question to be tested, and then input into a preset feedforward neural network to obtain the answer accuracy corresponding to the question to be tested.

[0010] Based on the above technical solutions, preferably, the step of creating a memory consolidation expert module by combining the knowledge acquisition process with forgetting gates and absorption rate parameters includes:

[0011] The knowledge acquisition process is divided into a perception stage, a comprehension stage, and a retention stage. The amount of new knowledge acquired by learners in the perception and comprehension stages is fitted by an absorption rate parameter. The decay of knowledge over time is simulated by a forgetting gate mechanism to obtain the decayed knowledge amount. Combined with the learner's initial knowledge amount, a memory consolidation expert module is created.

[0012] Based on the above technical solutions, preferably, the attention distribution of knowledge points combined with the time attention mechanism to create an attention regulation expert module includes:

[0013] We use structural attention mechanisms to weight the current knowledge set, simulating the initial attention distribution of learners on different knowledge areas;

[0014] Based on the initial attention distribution, a time-based attention mechanism is introduced to generate a dynamic attention distribution and create an attention regulation expert module.

[0015] Based on the above technical solutions, preferably, the module for creating structural adaptation experts based on the correlation and evolution process between knowledge points includes:

[0016] Based on global memory, an adaptive adjacency matrix is ​​dynamically constructed to model the positive and negative dependencies between knowledge concepts;

[0017] Multi-order graph convolution is used to transfer conceptual information, aggregate local and global knowledge association information, and model high-order dependencies between knowledge points.

[0018] By integrating the positive dependencies, the negative dependencies, and the higher-order dependencies through a multi-head attention mechanism, a structural adaptation expert module is created.

[0019] Based on the above technical solutions, preferably, the step of assigning corresponding expert weights to the memory consolidation expert module, the attention regulation expert module, and the structural adaptation expert module through the global memory matrix includes:

[0020] The contextual memories of the memory consolidation expert module, the attention regulation expert module, and the structural adaptation expert module are obtained based on the global memory matrix;

[0021] The inherent pattern features within each expert model are compared with the context memory to calculate the similarity, thereby determining the expert weight corresponding to each expert model.

[0022] Based on the above technical solutions, preferably, the step of calculating the similarity between the inherent pattern features within each expert model and the context memory to determine the expert weight corresponding to each expert model includes:

[0023] An Einstein summation operation is performed on the intrinsic pattern features within each expert model, and the similarity between the intrinsic pattern features and the context memory is calculated using a cosine similarity function. Based on the similarity, the expert weights corresponding to each expert model are determined.

[0024] Based on the above technical solution, preferably, the weighted fusion of the first knowledge state parameter generated by the memory consolidation expert module, the second knowledge state parameter generated by the attention regulation expert module, and the third knowledge state parameter generated by the structure adaptation expert module, based on the expert weights, to obtain the fused knowledge state parameter, includes:

[0025] Based on the relevance weights of the test questions and knowledge points, the knowledge state related to the current answer is extracted and input into the memory consolidation expert module to generate the first knowledge state parameter;

[0026] The first knowledge state parameter is input into the attention regulation expert module to generate the second knowledge state parameter;

[0027] The global memory representation parameters are input into the structural adaptation expert module to generate the third knowledge state parameters.

[0028] Furthermore, a second aspect of the present invention provides a multi-expert collaborative knowledge tracing system for cognitive strategy perception, comprising: an expert creation module, a weight allocation module, a parameter fusion module, and a result prediction module; wherein,

[0029] The expert creation module is configured to create a memory consolidation expert module by combining the knowledge acquisition process with the forgetting gate and the absorption rate parameter; to create an attention regulation expert module by combining the attention distribution of knowledge points with the time attention mechanism; and to create a structural adaptation expert module based on the correlation and evolution process between knowledge points. The absorption rate parameter is related to students' learning ability and test question discrimination.

[0030] The weight allocation module is configured to generate a global memory matrix based on the long-term memory of the student's historical cognitive state and the current interaction embedded information, combined with the attention mechanism, and to assign corresponding expert weights to the memory consolidation expert module, the attention regulation expert module and the structural adaptation expert module through the global memory matrix.

[0031] The parameter fusion module is configured to perform weighted fusion of the first knowledge state parameter generated by the memory consolidation expert module, the second knowledge state parameter generated by the attention regulation expert module, and the third knowledge state parameter generated by the structure adaptation expert module based on the expert weights, to obtain the fused knowledge state parameter.

[0032] The result prediction module is configured to concatenate the fused knowledge state parameter with the feature embedding vector of the question to be tested, and input it into a preset feedforward neural network to obtain the answer accuracy corresponding to the question to be tested.

[0033] More preferably, a third aspect of the present invention provides an electronic device, including a processor and a memory; the memory stores a computer program, wherein the computer program, when executed by the processor, implements the multi-expert collaborative knowledge tracing method for cognitive strategy perception described in the first aspect.

[0034] More preferably, a fourth aspect of the present invention provides a computer storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the multi-expert collaborative knowledge tracing method for cognitive strategy perception described in the first aspect.

[0035] The cognitive strategy perception-based multi-expert collaborative knowledge tracing method and system of the present invention has the following advantages over the prior art:

[0036] 1. By simulating learners' core cognitive mechanisms in knowledge perception, information integration, and structural construction, this model comprehensively represents the dynamic changes at different levels during the learning process. In terms of model architecture, a globally memory-driven expert gating scheduling mechanism is introduced. This dynamically allocates the contribution weights of each cognitive expert based on individual student differences and task characteristics, achieving knowledge state modeling that aligns with individual cognitive strategies. By integrating the output states of three types of cognitive experts, personalized knowledge tracking results are generated, thereby improving the prediction accuracy of student response behavior and potential cognitive states. While maintaining model interpretability, this model supports adaptive modeling of learner characteristics and cognitive strategy configuration, providing a more scientific and accurate basis for learning analysis and personalized intervention in intelligent teaching systems.

[0037] 2. By dividing the learning process into three stages—perception, understanding, and retention—it accurately matches learners' cognitive patterns. Combined with dynamic forgetting control and active absorption regulation, it dynamically adjusts the intensity of knowledge absorption and forgetting patterns based on student ability differences and test question discrimination, achieving personalized modeling for students at different cognitive levels. It utilizes a structural attention mechanism to weight the set of knowledge concepts, characterizing the distribution of students' attention across different concepts. Introducing a temporal attention mechanism generates a dynamic attention distribution, reflecting the spatiotemporal characteristics of learners' cognitive resource allocation, thus highlighting the impact of key learning segments on the current cognitive state. Using a multi-head attention mechanism, it captures the adaptive process of learners' knowledge state along the time dimension, integrating temporal and spatial evolutionary characteristics to ensure the reliability of knowledge state assessment. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 A flowchart illustrating a multi-expert collaborative knowledge tracing method for cognitive strategy perception provided in an embodiment of the present invention;

[0040] Figure 2 A schematic diagram illustrating the principle of the memory consolidation expert module provided in an embodiment of the present invention;

[0041] Figure 3 A schematic diagram illustrating the principle of the attention regulation expert module provided in an embodiment of the present invention;

[0042] Figure 4 A schematic diagram illustrating the principle of the structural adaptation expert module provided in an embodiment of the present invention;

[0043] Figure 5A schematic diagram illustrating the principle of the fusion expert module provided in an embodiment of the present invention;

[0044] Figure 6 This is a schematic diagram of the structure of a multi-expert collaborative knowledge tracing system for cognitive strategy perception provided in an embodiment of the present invention;

[0045] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0046] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0047] In some embodiments, such as Figure 1 As shown, Figure 1 This is a flowchart illustrating a multi-expert collaborative knowledge tracing method based on cognitive strategy perception, provided in an embodiment of the present invention. The multi-expert collaborative knowledge tracing method based on cognitive strategy perception provided by the present invention includes:

[0048] S110 creates a memory consolidation expert module by combining the knowledge acquisition process with the forgetting gate and the absorption rate parameter, creates an attention regulation expert module by combining the attention distribution of knowledge points with the time attention mechanism, and creates a structural adaptation expert module based on the evolutionary process of the association between knowledge points; the absorption rate parameter is related to students' learning ability and the test question discrimination.

[0049] In some embodiments, a memory consolidation expert module is created by combining the knowledge acquisition process with forgetting gates and absorption rate parameters, including:

[0050] The knowledge acquisition process is divided into perception, understanding and retention stages. The amount of new knowledge acquired by learners in the perception and understanding stages is fitted by the absorption rate parameter. The decay of knowledge over time is simulated by the forgetting gate mechanism to obtain the decayed knowledge amount. Combined with the learner's initial knowledge amount, a memory consolidation expert module is created.

[0051] In this embodiment, please refer to Figure 2 , Figure 2 This is a schematic diagram illustrating the principle of the memory consolidation expert module provided in this embodiment of the invention. The knowledge acquisition process is divided into three stages: perception, understanding, and retention, focusing on simulating the knowledge transformation and forgetting patterns experienced by students during the learning process. Based on the relevance weights between test questions and knowledge points, the knowledge state relevant to the current answer is extracted as perceptual input.

[0052] ;

[0053] in, This represents the learner's prior knowledge state regarding all concepts. It refers to the knowledge state related to the previous answer after activation. It is a question-concept matrix, and This represents the choice of concept representation at time t. Next, we will discuss the activated knowledge state. and the current interaction vector Combined, we obtain the knowledge concept perception vector. ,in, and These represent the weights and biases, respectively.

[0054] Secondly, an absorption rate parameter is introduced to dynamically adjust the intensity of knowledge absorption based on differences in student abilities and the discriminatory power of test questions, thereby achieving personalized modeling for students with different cognitive levels; learning rate The calculation formula is as follows, where, and These represent the discrimination and difficulty of the test questions, respectively. and These represent the weights and biases, respectively.

[0055] ;

[0056] Through absorption rate and knowledge concept perception vector Multiplication determines the amount of knowledge a learner has actually acquired. :

[0057] .

[0058] Finally, by simulating the decay of knowledge over time using the forgetting gate mechanism, the portion of historical knowledge retained after forgetting is calculated. By adding the consolidated historical state to the newly acquired knowledge to form a staged consolidated knowledge representation, the current knowledge state is updated. ,in, and These represent the weights and biases, respectively.

[0059] ;

[0060] ;

[0061] Finally, the knowledge states at each time point are concatenated to obtain the total knowledge state for the time interval 0-t, which is the final output of the memory consolidation expert. .

[0062] In some embodiments, an attention regulation expert module is created by combining the attention distribution of knowledge points with a time-based attention mechanism, including:

[0063] We use structural attention mechanisms to weight the current knowledge set, simulating the initial attention distribution of learners on different knowledge areas;

[0064] Based on the initial attention distribution, a time-based attention mechanism is introduced to generate a dynamic attention distribution and create an attention regulation expert module.

[0065] This module employs a dual spatiotemporal weighted attention mechanism to dynamically characterize the learner's cognitive focus.

[0066] In this embodiment, please refer to Figure 3 , Figure 3 This is a schematic diagram illustrating the principle of the attention regulation expert module provided in this embodiment of the invention. Considering that learners will focus on specific knowledge concepts based on their current learning needs, rather than comprehensively covering all concepts, a structured attention mechanism is used to weight the current set of concepts, characterizing the distribution of learners' attention across different concepts.

[0067]

[0068] in, Represents the weight parameters. The hidden vector dimension for each attention head.

[0069] A temporal attention mechanism is introduced, integrating the results of structural attention and temporal attention to generate a dynamic attention distribution that reflects the spatiotemporal characteristics of students' cognitive resource allocation. Historical learning behaviors are retrospectively reviewed and focused on, with weights assigned to past states based on time intervals and question similarity, thereby highlighting the impact of key learning segments on the current cognitive state.

[0070]

[0071] in, It is a state of knowledge acquired through time-based attention regulation. The hidden vector dimension for each attention head, It is a time-enhanced weight, which depends on t and The distance between time steps It is the question similarity weight, composed of t and The similarity of questions within a time step is calculated. (Function) and The specific calculation formula is as follows:

[0072]

[0073] ;

[0074] To differentiate attention at different time intervals, a Gaussian function is used to weight time steps closer to the target. This Gaussian function follows a normal distribution trend, assigning higher weights to positions closer to the target. Its local enhancement properties make it particularly suitable for emphasizing proximity values. Furthermore, learner attention modulation is closely related to the similarity between test items. The similarity attention of test items answered at corresponding time points is calculated, with higher weights assigned to items with stronger correlation. Here, an attention mechanism is used to model the similarity relationship between test items.

[0075] In some embodiments, a structural adaptation expert module is created based on the evolutionary process of relationships between knowledge points, including:

[0076] Based on global memory, an adaptive adjacency matrix is ​​dynamically constructed to model the positive and negative dependencies between knowledge concepts;

[0077] Multi-order graph convolution is used to transfer conceptual information, aggregate local and global knowledge association information, and model high-order dependencies between knowledge points.

[0078] By integrating positive dependencies, negative dependencies, and higher-order dependencies through a multi-head attention mechanism, a structural adaptation expert module is created.

[0079] This module introduces a global memory representation to generate an adaptive adjacency matrix for the concept graph, capturing the dynamic evolution of the knowledge structure.

[0080] In this embodiment, please refer to Figure 4 , Figure 4 This is a schematic diagram illustrating the principle of the structural adaptation expert module provided in this embodiment of the invention. Considering that during the learning process, learners do not simply fill their brains with new knowledge, but rather actively connect it with existing knowledge, sometimes even reconstructing their original understanding to achieve the integration of the knowledge system, this continuous process of adjusting and reorganizing the internal knowledge structure is called structural adaptation. As the mastery of individual concepts increases, the learner's knowledge structure dynamically reorganizes to accommodate new understandings. However, these structural changes are latent and cannot be directly observed. Therefore, an adaptive adjacency matrix is ​​dynamically constructed based on global memory, multi-order graph convolution is used for concept information transfer, and a multi-head attention mechanism is used to integrate structural features and temporal representations.

[0081] First, an adaptive weighted adjacency matrix is ​​constructed by calculating the semantic relevance between concepts. In order to dynamically reveal the mutual influence between concepts, the adjacency matrix is ​​adaptively generated by converting the global memory representation into the corresponding source node and target node embedding.

[0082] By representing global memory Mapping to different vector spaces to represent source nodes and target node .in, , and , These represent the weights and biases, respectively.

[0083] ;

[0084] The spatial adaptation matrix between the source and target nodes is calculated by multiplying the two nodes. and Since the positive influence of a predecessor concept on a subsequent concept is usually different from the negative influence of a subsequent concept on a predecessor, calculating these two relationships separately allows for explicit modeling of positive and negative dependencies between knowledge concepts.

[0085] ;

[0086] Secondly, multi-order graph convolution operations are employed to aggregate local and global knowledge association information, modeling high-order dependencies between knowledge points. Inspired by diffusing convolution, we use a multi-order convolution method to simulate the multi-step diffusion process. The K-hop neighbors of the diffusion propagation are obtained through a power series. express:

[0087] ;

[0088] Finally, by combining residual connections and normalization operations, the knowledge structure is dynamically reconstructed during the learning process, resulting in a knowledge state representation after structural evolution. The learner's knowledge state continuously evolves at the knowledge structure level and is also constantly updated over time. To capture this temporal evolution characteristic, an attention mechanism is employed, which performs excellently in long-term sequence modeling. Finally, the structural model, temporal model, and residual normalization layer are concatenated to construct a network architecture that integrates spatiotemporal dimensions. Its mathematical expression is:

[0089]

[0090] in, It is the final output of the structural experts.

[0091] Multi-head attention mechanisms can be used to capture adaptive processes along the time dimension, integrating temporal and spatial evolutionary features.

[0092] S120: Based on the long-term memory of the student's historical cognitive state and the current interactive embedded information, a global memory matrix is ​​generated in combination with the attention mechanism, and corresponding expert weights are assigned to the memory consolidation expert module, the attention regulation expert module and the structural adaptation expert module through the global memory matrix.

[0093] In some embodiments, a global memory matrix is ​​used to assign corresponding expert weights to the memory consolidation expert module, the attention regulation expert module, and the structural adaptation expert module, including:

[0094] Contextual memory of the memory consolidation expert module, attention regulation expert module, and structural adaptation expert module is obtained based on the global memory matrix;

[0095] The similarity between the inherent pattern features within each expert model and the contextual memory is calculated to determine the expert weight corresponding to each expert model.

[0096] In some embodiments, the inherent pattern features within each expert model are compared with the contextual memory to calculate the similarity, thereby determining the expert weights corresponding to each expert model, including:

[0097] Perform an Einstein summation operation on the intrinsic pattern features within each expert model, and use the cosine similarity function to calculate the similarity between the intrinsic pattern features and the contextual memory. Based on the similarity, determine the expert weight corresponding to each expert model.

[0098] In this embodiment, please refer to Figure 5 , Figure 5 This is a schematic diagram of the principle of the fusion expert module provided in the embodiment of the present invention; the expert gating scheduling mechanism is used to dynamically allocate the contribution weight of each cognitive expert to adapt to the individual differences of different students, and finally use the contribution weight to weighted sum the knowledge states output by each expert to realize the probability prediction of the student's answer result.

[0099] First, a globally trainable memory matrix is ​​constructed to store the model's long-term memory of students' historical cognitive states. Combined with the information embedded in the current interaction, the global memory matrix is ​​generated through an attention mechanism, enabling the model to dynamically retrieve relevant experiences based on the current learning task.

[0100] ;

[0101] in, Represents contextual memory, It is the learner's embedded student interaction from time 0-T. It is a learnable empty matrix that maps the input to the memory dimension through an attention mechanism.

[0102] Simultaneously, features of each expert are extracted, and the inter-expert dependencies are enhanced through a self-attention mechanism, reinforcing the inherent pattern features within each expert. Then, similarity with contextual memory is calculated to generate a score for each expert, and normalization is applied to ensure that the sum of all scores is 1. This process reflects the degree of matching between different experts and the current task.

[0103] ;

[0104] ;

[0105] in, This indicates the hidden output state before the e-th expert is activated; inactive variables contain more historical information. Denotes the dimension of the hidden layer, and set E contains three experts. , Representative issue The corresponding concept list, after performing the Einstein summation operation, is combined with the test questions to obtain the enhanced knowledge status of each expert. Subsequently, a self-attention mechanism is computed to enhance the differentiated capabilities of each expert. The cosine similarity function `sim()` is used to calculate the similarity with contextual memory, thereby obtaining a weighted score for each expert. .

[0106] S130, the first knowledge state parameter generated by the memory consolidation expert module, the second knowledge state parameter generated by the attention regulation expert module, and the third knowledge state parameter generated by the structure adaptation expert module are weighted and fused based on expert weights to obtain the fused knowledge state parameter.

[0107] In some embodiments, the first knowledge state parameter generated by the memory consolidation expert module, the second knowledge state parameter generated by the attention regulation expert module, and the third knowledge state parameter generated by the structural adaptation expert module are weighted and fused based on expert weights to obtain a fused knowledge state parameter, including:

[0108] Based on the relevance weights of the test questions and knowledge points, the knowledge state related to the current answer is extracted and input into the memory consolidation expert module to generate the first knowledge state parameter;

[0109] The first knowledge state parameter is input into the attention regulation expert module to generate the second knowledge state parameter;

[0110] The global memory representation parameters are input into the structural adaptation expert module to generate the third knowledge state parameters.

[0111] S140: The knowledge state parameter is concatenated with the feature embedding vector of the question to be tested, and then input into the preset feedforward neural network to obtain the answer accuracy corresponding to the question to be tested.

[0112] In this embodiment, the calculation process of the feature embedding vector of the test item to be tested includes: individual feature embedding and learner interaction embedding. For individual feature embedding, firstly, given a test item set... Each question is embedded in a matrix. Represented as ,in, The total number of questions. This is the dimension of question embedding. Similarly, for concept sets... We introduce the corresponding embedding matrix. Characterize it as in, Represents the number of concepts. Learner response sequence. Each This represents the learner's response at time step t. Each response is embedded into a dimension of... vector To establish the connection between test questions and related concepts, we constructed a test question-concept matrix. This is the Q-matrix. The Q-matrix can be obtained by domain experts based on the syllabus and test question annotations. Each element has a value of 1 indicating that the test question relates to that knowledge point, and a value of 0 indicating that it does not. By statistically analyzing the concept tags corresponding to the test questions, a matrix with the following dimensions is ultimately formed. A sparse matrix is ​​used to characterize the relationship structure between questions and concepts.

[0113] For learner interaction embeddings representing the historical records of responses over a period of time, the questions answered by the learner at each time step t are captured. And the corresponding answers To obtain a learner interaction vector containing this information, the questions and answers at each time step are concatenated, and a linear transformation is applied:

[0114] ;

[0115] in, It is the learner's response interaction embedding representation at time t, and the weight matrix. , It is a bias term. , These represent the embeddings of the questions and responses, respectively.

[0116] The output state vectors of the memory consolidation expert module, attention regulation expert module, and structural adaptation expert module are fused and then weighted and summed according to the expert weights generated by gating scheduling to obtain the final knowledge state representation at the current time step. The final fused knowledge state parameter is concatenated with the feature embedding vector of the current test question and input into a feedforward neural network. The network contains activation functions and linear layers to capture non-linear mapping relationships and outputs the probability that the student answers the question correctly. The formula is expressed as follows, where... Output representing each expert:

[0117] .

[0118] In some embodiments, please refer to Figure 6 , Figure 6 This is a schematic diagram of the structure of a multi-expert collaborative knowledge tracing system with cognitive strategy perception provided in an embodiment of the present invention. The present invention provides a multi-expert collaborative knowledge tracing system 600 with cognitive strategy perception, comprising: an expert creation module 610, a weight allocation module 620, a parameter fusion module 630, and a result prediction module 640; wherein,

[0119] The expert creation module 610 is configured to create a memory consolidation expert module by combining the knowledge acquisition process with the forgetting gate and the absorption rate parameter; an attention regulation expert module by combining the attention distribution of knowledge points with the time attention mechanism; and a structure adaptation expert module based on the evolutionary process of the relationship between knowledge points. The absorption rate parameter is related to students' learning ability and the discrimination of test questions.

[0120] The weight allocation module 620 is configured to generate a global memory matrix based on the long-term memory of the student's historical cognitive state and the current interaction embedded information, combined with the attention mechanism, and to assign corresponding expert weights to the memory consolidation expert module, the attention regulation expert module and the structural adaptation expert module through the global memory matrix.

[0121] The parameter fusion module 630 is configured to perform weighted fusion of the first knowledge state parameter generated by the memory consolidation expert module, the second knowledge state parameter generated by the attention regulation expert module, and the third knowledge state parameter generated by the structure adaptation expert module based on expert weights, to obtain the fused knowledge state parameter.

[0122] The result prediction module 640 is configured to concatenate the fused knowledge state parameters with the feature embedding vector of the question to be tested, and input them into a preset feedforward neural network to obtain the answer accuracy corresponding to the question to be tested.

[0123] In some embodiments, the expert creation module 610 is specifically configured as follows:

[0124] The knowledge acquisition process is divided into perception, understanding and retention stages. The amount of new knowledge acquired by learners in the perception and understanding stages is fitted by the absorption rate parameter. The decay of knowledge over time is simulated by the forgetting gate mechanism to obtain the decayed knowledge amount. Combined with the learner's initial knowledge amount, a memory consolidation expert module is created.

[0125] In some embodiments, the expert creation module 610 is specifically configured as follows:

[0126] We use structural attention mechanisms to weight the current knowledge set, simulating the initial attention distribution of learners on different knowledge areas;

[0127] Based on the initial attention distribution, a time-based attention mechanism is introduced to generate a dynamic attention distribution and create an attention regulation expert module.

[0128] In some embodiments, the expert creation module 610 is specifically configured as follows:

[0129] Based on global memory, an adaptive adjacency matrix is ​​dynamically constructed to model the positive and negative dependencies between knowledge concepts;

[0130] Multi-order graph convolution is used to transfer conceptual information, aggregate local and global knowledge association information, and model high-order dependencies between knowledge points.

[0131] By integrating positive dependencies, negative dependencies, and higher-order dependencies through a multi-head attention mechanism, a structural adaptation expert module is created.

[0132] In some embodiments, the weight allocation module 620 is specifically configured as follows:

[0133] Contextual memory of the memory consolidation expert module, attention regulation expert module, and structural adaptation expert module is obtained based on the global memory matrix;

[0134] The similarity between the inherent pattern features within each expert model and the contextual memory is calculated to determine the expert weight corresponding to each expert model.

[0135] In some embodiments, the weight allocation module 620 is specifically configured as follows:

[0136] Perform an Einstein summation operation on the intrinsic pattern features within each expert model, and use the cosine similarity function to calculate the similarity between the intrinsic pattern features and the contextual memory. Based on the similarity, determine the expert weight corresponding to each expert model.

[0137] In some embodiments, the parameter fusion module 630 is specifically configured as follows:

[0138] Based on the relevance weights of the test questions and knowledge points, the knowledge state related to the current answer is extracted and input into the memory consolidation expert module to generate the first knowledge state parameter;

[0139] The first knowledge state parameter is input into the attention regulation expert module to generate the second knowledge state parameter;

[0140] The global memory representation parameters are input into the structural adaptation expert module to generate the third knowledge state parameters.

[0141] It should be noted that the cognitive strategy-aware multi-expert collaborative knowledge tracing system provided in this application embodiment and the cognitive strategy-aware multi-expert collaborative knowledge tracing method provided in this application embodiment are based on the same application concept. Therefore, the specific implementation of this embodiment can refer to the implementation of the aforementioned cognitive strategy-aware multi-expert collaborative knowledge tracing method, and the repeated parts will not be described again.

[0142] In some embodiments, please refer to Figure 7 , Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device 700 provided in this application includes a processor 710 and a memory 720; the memory 720 stores a computer program, wherein the computer program, when executed by the processor, implements the aforementioned multi-expert collaborative knowledge tracing method for cognitive strategy perception.

[0143] Specifically, processor 710 may include, for example, a general-purpose microprocessor, an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. Processor 710 may also include onboard memory for caching purposes. Processor 710 may be a single processing unit or multiple processing units for performing different actions of the method flow according to embodiments of this application.

[0144] The memory 720 can be any medium capable of containing, storing, transmitting, propagating, or transmitting instructions. For example, the memory 720 can be, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, apparatus, or propagation medium. Specific examples of the memory 720 include: magnetic storage devices such as magnetic tape or hard disk drives (HDDs); optical storage devices such as optical discs (CD-ROMs); and also random access memory (RAM) or flash memory; and / or wired / wireless communication links.

[0145] This application also provides a computer-readable medium storing a computer program thereon, which, when executed by a processor, implements the aforementioned multi-expert collaborative knowledge tracing method based on cognitive strategy perception. This computer-readable medium may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into that device / apparatus / system. The aforementioned computer-readable medium carries one or more programs, which, when executed, implement the method according to the embodiments of this application.

[0146] According to embodiments of this application, a computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media can also be any computer-readable medium other than computer-readable storage media, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wired, optical fiber, radio frequency signals, etc., or any suitable combination thereof.

[0147] Those skilled in the art will understand that the features described in the various embodiments and / or claims of this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, the features described in the various embodiments and / or claims of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application. All such combinations and / or combinations fall within the scope of this application. Therefore, the scope of this application should not be limited to the above embodiments, but should be defined not only by the appended claims, but also by their equivalents. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the protection scope of this invention.

Claims

1. A method for cognitive strategy-aware multi-expert collaborative knowledge tracing, characterized in that, The method comprises the steps of: respectively combining a knowledge acquisition process with a forgetting gate and an absorption rate parameter to create a memory consolidation expert module, combining an attention distribution of knowledge points with a time attention mechanism to create an attention regulation expert module, and creating a structure adaptation expert module based on an association evolution process between knowledge points; The absorption rate parameter is related to the learning ability of the student and the discrimination degree of the test question. Based on the long-term memory of the historical cognitive state of the student and the embedded information of the current interaction, a global memory matrix is generated by combining the attention mechanism, and the corresponding expert weights of the memory consolidation expert module, the attention regulation expert module and the structure adaptation expert module are allocated through the global memory matrix. Based on the expert weights, the first knowledge state parameter generated by the memory consolidation expert module, the second knowledge state parameter generated by the attention regulation expert module and the third knowledge state parameter generated by the structure adaptation expert module are weighted and fused to obtain a fused knowledge state parameter. The fused knowledge state parameter and the feature embedding vector of the test question to be predicted are spliced and input into a preset feedforward neural network to obtain the correct answer rate corresponding to the test question to be predicted.

2. The cognitive strategy-aware multi-expert collaborative knowledge tracing method of claim 1, wherein, The method comprises the steps of: The knowledge acquisition process is divided into a perception stage, an understanding stage and a retention stage, the absorption rate parameter is used to fit the amount of new knowledge mastered by the learner in the perception stage and the understanding stage, the forgetting gate mechanism is used to simulate the characteristics of knowledge decay over time, and the decay knowledge amount is obtained, and the memory consolidation expert module is created by combining the initial knowledge amount of the learner.

3. The cognitive strategy-aware multi-expert collaborative knowledge tracing method of claim 1, wherein, The method comprises the steps of: The structural attention mechanism is used to weight the current knowledge set to simulate the initial attention distribution of the learner on different knowledge. The time attention mechanism is introduced based on the initial attention distribution to generate a dynamic attention distribution and create an attention regulation expert module.

4. The cognitive strategy-aware multi-expert collaborative knowledge tracing method of claim 1, wherein, The method comprises the steps of: An adaptive adjacency matrix is dynamically constructed based on the global memory to model the positive and negative dependency relationships between knowledge concepts. Multi-order graph convolution is used to aggregate local and global knowledge association information to model high-order dependency relationships between knowledge points. The positive, negative and high-order dependency relationships are integrated through a multi-head attention mechanism to create a structure adaptation expert module.

5. The cognitive strategy-aware multi-expert collaborative knowledge tracing method of claim 1, wherein, The method comprises the steps of: The context memory of the memory consolidation expert module, the attention regulation expert module and the structure adaptation expert module is obtained based on the global memory matrix. The similarity between the inherent mode features of each expert model and the context memory is calculated to determine the expert weight corresponding to each expert model.

6. The cognitive strategy-aware multi-expert collaborative knowledge tracing method of claim 5, wherein, The method comprises the steps of: An Einstein summation operation is performed on inherent mode features inside each expert model, and a cosine similarity function is used to calculate the similarity of the inherent mode features and the context memory, and based on the similarity, an expert weight corresponding to each expert model is determined.

7. The cognitive strategy-aware multi-expert collaborative knowledge tracing method of claim 1, wherein, The first knowledge state parameter generated by the memory consolidation expert module, the second knowledge state parameter generated by the attention regulation expert module, and the third knowledge state parameter generated by the structure adaptation expert module are weighted and fused based on the expert weights to obtain a fused knowledge state parameter, which includes: According to the correlation weight of the test questions and the knowledge points, the knowledge state related to the current answer is extracted and input to the memory consolidation expert module to generate a first knowledge state parameter; The first knowledge state parameter is input to the attention regulation expert module to generate a second knowledge state parameter; The global memory representation parameter is input to the structure adaptation expert module to generate a third knowledge state parameter.

8. A cognitive strategy-aware multi-expert collaborative knowledge tracing system, characterized in that, It includes: An expert creation module, a weight distribution module, a parameter fusion module, and a result prediction module; wherein, The expert creation module is configured to create a memory consolidation expert module by combining the knowledge acquisition process with the forgetting gate and the absorption rate parameter, create an attention regulation expert module by combining the attention distribution of knowledge points with the time attention mechanism, and create a structure adaptation expert module based on the association evolution process between knowledge points; the absorption rate parameter is related to the learning ability of students and the discrimination of test questions; The weight distribution module is configured to generate a global memory matrix based on the long-term memory of the student's historical cognitive state and the current interactive embedded information, and to distribute corresponding expert weights to the memory consolidation expert module, the attention regulation expert module, and the structure adaptation expert module through the global memory matrix; The parameter fusion module is configured to weight and fuse the first knowledge state parameter generated by the memory consolidation expert module, the second knowledge state parameter generated by the attention regulation expert module, and the third knowledge state parameter generated by the structure adaptation expert module based on the expert weights to obtain a fused knowledge state parameter. The result prediction module is configured to concatenate the fused knowledge state parameter with the feature embedding vector of the test question to be predicted and input it to a preset feedforward neural network to obtain the correct answer rate of the test question to be predicted.

9. An electronic device comprising a processor and a memory; said memory having stored a computer program, wherein, The computer program, when executed by the processor, implements the cognitive strategy-aware multi-expert collaborative knowledge tracing method of any one of claims 1-7.

10. A computer storage medium, characterized in that, A computer program is stored thereon, wherein the computer program is executed by a processor to implement the cognitive strategy-aware multi-expert collaborative knowledge tracing method of any one of claims 1-7.

Images