Spatio-Temporal Fusion Knowledge Tracking Method Based on Large Model Annotation Enhancement

Abstract

The present invention discloses a spatio-temporal fusion knowledge tracking method based on large model annotation enhancement. This method first collects and organizes the historical interaction data of students in the intelligent teaching system to determine the questions, knowledge points and their corresponding relationships. Then, various relationships between knowledge points are annotated by the large model to construct an initial graph and annotate the units and the relationships between units. Next, the annotation results are detected and optimized to correct errors and improve the annotation quality. Then, the knowledge point embeddings are processed by a temporal learning module based on the attention mechanism and combined with the difficulty to obtain the question embeddings, and the temporal embedding codes are obtained. After that, feature extraction, integration and contrast learning are carried out by the fusion learning module. Finally, the node embeddings are dynamically updated based on the unit relationship graph, multiple information is fused, and the neural network is used to combine the forgetting factor to predict and calculate the loss function. This method improves the accuracy and interpretability of knowledge tracking and helps with personalized teaching.

Description

Technical Field

[0001] The present invention relates to the field of educational data mining, and particularly to a spatio-temporal fusion knowledge tracking method based on large model annotation enhancement. Background Art

[0002] In the current era of rapid development of digital education, accurately grasping students' knowledge mastery and realizing personalized teaching have become key demands in the education field. Knowledge tracking technology has emerged as the times require. Its core goal is to dynamically and accurately depict students' understanding and mastery of various knowledge points based on historical interaction data during the learning process, providing strong support for the adjustment of teaching strategies and the push of personalized learning resources, thereby improving the overall teaching effect and learning efficiency.

[0003] However, traditional knowledge tracking technologies have exposed many limitations in practical applications. On the one hand, traditional methods highly rely on educational experts to manually annotate the relationships between knowledge points. This process not only consumes a large amount of human, material, and time costs, but also has strong subjectivity, and there may be significant differences in the annotation results among different experts. On the other hand, when facing a large-scale knowledge point system, the efficiency of manual annotation is extremely low, making it difficult to meet the rapidly developing educational needs. In addition, although some automatic annotation methods based on statistics or simple graph theory have reduced the human cost to a certain extent, the generated knowledge point relationship edges are often sparse and cannot comprehensively and accurately reflect the complex logical connections between knowledge points, seriously affecting the analysis and prediction accuracy of the knowledge tracking model for students' learning situations.

[0004] In recent years, with the rise of large model technology, its powerful language understanding and knowledge processing capabilities have brought new ideas to knowledge tracking. However, knowledge tracking technologies based on large models also face challenges. When large models annotate the relationships between knowledge points, there are often annotation errors and hallucination phenomena, that is, generating some knowledge point relationships that do not conform to the actual educational logic, resulting in doubts about the reliability of the annotation results. At the same time, due to the limitations of the large model's own structure and training methods, it is difficult to handle cross-level and long-distance knowledge point associations, and it is difficult to effectively capture the complex hierarchical information and spatial structure in the knowledge system, making the knowledge tracking model insufficient in reflecting students' overall mastery and application abilities of knowledge and unable to provide accurate and effective guidance for teaching practice. Summary of the Invention

[0005] In order to overcome the disadvantages and deficiencies of the existing technology, the present invention provides a spatio-temporal fusion knowledge tracking method based on large model annotation enhancement.

[0006] The technical solution adopted by the present invention is a spatio-temporal fusion knowledge tracking method based on large model annotation enhancement, and this method includes the following steps:

[0007] Step S1, Data Preparation: Collect the historical interaction data of students in the intelligent teaching system, organize the historical interaction data into a learning historical trajectory containing question-and-answer interactions, and determine the question set, knowledge point set, and the corresponding relationship between questions and knowledge points;

[0008] Step S2, Large Model Annotation: Use the large model to traverse the knowledge points in the dataset through preset prompts, annotate the precursor, successor, parallel, and other relationships between knowledge points, and construct an initial knowledge point relationship graph; at the same time, use the large model to annotate the units to which the knowledge points belong and the relationships between units to optimize the long-distance knowledge point spatial association and reflect hierarchical information;

[0009] Step S3, Annotation Optimization: Detect the large model annotation results. For problems such as mutual precursors, unilateral precursors, the formation of a cycle in the precursor results, and the non-connectivity of the graph, use new prompts for secondary confirmation or preset solutions for correction to improve the annotation quality;

[0010] Step S4, Sequential Learning: Adopt a temporal learning module based on the attention mechanism to process the knowledge point embeddings involved in the questions, calculate the question embeddings in combination with the question difficulty, and obtain the temporal embedding encoding through the attention encoding layer;

[0011] Step S5, Fusion Learning: Through the knowledge point static graph embedding learning module, use the graph attention encoder to extract the features of the knowledge point relationship graph, and integrate it with the original sequence model to obtain the joint static embedding; through the dual-view embedding contrast learning module, construct the collaborative contrast of the sequence view and the graph view to distinguish the effective knowledge point associations and noise relationships;

[0012] Step S6, Dynamic Learning and Prediction: Based on the student personalized information dynamic learning module of the unit relationship graph, dynamically update the node embeddings according to the student interaction history, and fuse the student dynamic information of the temporal and graph structure models; at the prediction layer, use the neural network to learn the parameters of the modified 2PL model, combine the forgetting factor for the final prediction, and calculate the loss function.

[0013] Furthermore, in step S2, when annotating the relationships between knowledge points, for the precursor knowledge points, the large model judges the knowledge points that need to be mastered before learning a certain knowledge point based on the knowledge point name and educational domain knowledge; for the successor knowledge points, it judges the knowledge points that need to be mastered after learning this knowledge point; the parallel knowledge points are annotated and the learning levels are within the preset range; the remaining knowledge points are annotated as those that the large model cannot clearly classify into the first three relationships, and the knowledge point relationship annotation results are represented as directed edge relationships , where, respectively represent different knowledge points, and construct an initial knowledge point relationship graph is the knowledge point set, It is a set of edge relationships between knowledge points, and the edge relationships are determined based on the marked precursor, successor, and parallel relationships.

[0014] Furthermore, in step S3, for the case where the annotation results are precursors / successors to each other, assume there are knowledge points and . If the large model's preliminary annotation is 's precursor and is 's precursor, use a new prompt word , with the format "Please clearly judge the accurate precursor-successor relationship between knowledge points and ", and let the large model re-judge; for the case of single-sided precursor / successor, if the annotated knowledge point is the precursor of the knowledge point , but is not marked as 's successor, the new prompt word is "Please confirm whether there is a precursor / successor relationship between knowledge points and , and if so, please clearly mark it"; for the case where the precursor / successor results form a loop, the large model disconnects the loop structure at a preset position, and the new prompt word is "Currently, the precursor relationships of the following knowledge point list {A, B, C,...} are connected end to end to form a loop. Please select a preset node among them as the learning starting point and disconnect the loop structure"; for the non-connected case, the large model judges to reconnect the graph structure, and the new prompt word is "Currently, the knowledge point sets P: {A, B, C,...} and Q: {D, E, F,...} are non-connected. Try to find associated nodes in the two sets to form precursor / successor or parallel relationships between the nodes."

[0015] Furthermore, in step S4, perform embedding processing on the knowledge points involved in the question, and calculate the question embedding in combination with the question difficulty. The specific process is as follows:

[0016] Weighted aggregation of knowledge point embeddings. For each given question, there is a corresponding set of knowledge point embeddings , where is the number of knowledge points involved in this question, and weights are assigned to each knowledge point embedding;

[0017] Calculate the attention coefficient through the attention mechanism. The calculation formula is:

[0018] ;

[0019] Among them, is the query vector, is the key information currently being focused on, is the key vector, which is used to match with the query vector and measure the relevance between different knowledge points, is the traversed key vector. Through the exponential function and normalization operation, the relevance is converted into an attention coefficient. The larger the coefficient, the more important the knowledge point is in the overall representation;

[0020] According to the calculated attention coefficient, a weighted sum is performed on the knowledge point embeddings to obtain the weighted knowledge point embeddings :

[0021] ;

[0022] Among them, is the attention coefficient, is the value vector;

[0023] Residual enhancement of multi-knowledge point embeddings. For the weighted knowledge point embeddings perform residual connection and activation function processing. After passing through the activation function, the final multi-knowledge point embeddings are obtained , and the calculation formula is:

[0024] ;

[0025] Among them, is the learnable weight matrix, which is continuously adjusted during the training process to optimize the weight of the residual connection, The activation function introduces non-linearity;

[0026] Quantification and conversion of question difficulty. The question difficulty is calculated by the ratio of the number of times the question is answered wrong to the total number of times it is answered , and the formula is as follows:

[0027] ;

[0028] Perform a linear transformation to obtain the difficulty factor , and the specific transformation formula is:

[0029] ;

[0030] Among them, and [[ID=6*]]are the learnable weight matrices, enabling the difficulty factor to be fused with the multi-knowledge point embeddings;<*

[0031] Generation of question embeddings. Add the final multi-knowledge point embeddings to the difficulty factor to obtain the embedding of this question :

[0032] ;

[0033] The generated question embeddings are used for the subsequent attention encoding layer to obtain a sequential embedding encoding that can reflect the question features and difficulty information.

[0034] Furthermore, in step S5, a graph attention network (GAT) encoder with edge type awareness is used to extract features from the knowledge point relationship graph. Let the feature of node in the knowledge point graph be , the feature of node in the knowledge point graph be , the edge type related parameter matrix be , the relationship type embedding be , the leaky rectified linear unit be , the attention coefficient is calculated as where is the set of neighbor nodes of node . Through double-layer GAT encoding, a static embedding matrix of knowledge points is obtained. When integrated with the original sequence model, let the embedding result of the sequence model after the attention encoding layer be , the GAT embedding of the set of knowledge points corresponding to the current question be .

[0035] Furthermore, in step S5, positive and negative sample pairs are defined: The positive sample is the same knowledge point in different views. The pair is ; The negative sample is the embedding combination of knowledge points belonging to different units under the corresponding view. Let the knowledge point belong to unit , the knowledge point belong to unit , and , the negative sample pair is or . The contrastive loss function is designed as

[0036] ;

[0037] is the cosine similarity function, is the temperature coefficient hyperparameter, is the number of samples.

[0038] Furthermore, in step S5, a heterogeneous graph is a static node of knowledge points, is a unit dynamic node, is a knowledge point - unit relationship edge, is a unit - unit relationship edge. The node update process introduces real - time interaction features. Let the formula for updating the node embedding at time be the activation function, vector concatenation operation, be the weight matrix, be the joint static embedding at time be the interaction feature at time For the set of knowledge points involved in the current question calculate the associated unit state , be the set of units to which belongs, be the attention coefficient, be the current time - series hidden state, be the weight matrix, and be the state vectors of the associated units.

[0039] Furthermore, in step S6, the parameters in the IRT model are learned using a neural network, and the 2PL model is modified for final prediction. The discrimination of the question and the mastery level of the student are modeled through two fully - connected layers respectively, represent different multi - layer perceptrons, be the fused student dynamic information embedding, be the joint static embedding. The final prediction is obtained according to the IRT model, and the formula is be the question difficulty parameter. The loss function of the model is the cross - entropy loss function, be the true label, be the predicted label, be the weight coefficient of the contrast loss, be the contrast loss function.

[0040] Furthermore, in step S1, for the collected historical interaction data of students, if there is student data with the number of interactions less than the preset value , the data is discarded. When conducting experimental verification, the length of the input interaction sequence is unified to the preset value, , for sequences with a length greater than 200, they are split into multiple sequences with a length of 200; for sequences with a length less than 200, zero vectors are used to pad from the beginning of the sequence to ensure the same length. When dividing the dataset, the data in the dataset is used as the test set, and the other data is used as the training set and the validation set at the ratios of and respectively.

[0041] Furthermore, in the step of large model annotation, different versions of the large model are used for annotation experiments. When analyzing the impact of the relabeling strategy on the annotation quality of the large model, the healthy relationships, conflict relationships, weak relationships, the number of loops, and the number of non-connected components in the preliminary labeling results are counted, as well as the proportion of healthy relationships , the proportion of conflict relationships , and the proportion of weak relationships . The calculation formulas are , where are the numbers of healthy relationships, conflict relationships, and weak relationships respectively, is the total number of labeled relationships; the number of loops is the count of loops formed by the relationship annotation results between knowledge points, and the non-connected components is the count of non-connected subgraphs in the results.

[0042] Beneficial effects:

[0043] The present invention proposes a spatio-temporal fusion knowledge tracking method based on large model annotation enhancement. This method uses a large model to complete the annotation of knowledge point relationships and the units to which they belong, and through a detection and re-annotation strategy, effectively reduces annotation errors, improves annotation quality, overcomes the problems of traditional annotation relying on a large amount of manpower and many errors in automatic annotation, and provides a reliable basis for subsequent model learning. In terms of model construction, the design advantage of integrating spatio-temporal information is prominent. The sequence learning module fully explores the temporal information of students' historical learning, combines knowledge point embedding and question difficulty to obtain question embedding, reflecting the dynamic changes in the students' learning process. The knowledge point static graph embedding learning module and the dual-view contrast learning module use the graph relationships annotated by the large model to generate static embeddings containing spatial structure information, and optimize the embedding distribution through contrast learning to enhance the model's perception ability of knowledge point structure features. The student personalized information dynamic learning module based on the unit relationship graph, aiming at the problem of sparse interaction between students and knowledge points, dynamically updates the node embedding according to the students' interaction history, effectively integrates hierarchical spatial information and temporal information, and more accurately reflects the students' personalized knowledge states. In practical applications, this method is experimented on multiple real education datasets, and the results show that it can significantly improve the performance of the model in the knowledge tracking task, and the AUC index is significantly better than most baseline models. At the same time, the large model annotation enhances the interpretability of the model, clearly presenting the organizational relationship of knowledge points and the learning paths of students, providing valuable insights for educators, helping to realize personalized teaching resource recommendation, and ultimately improving teaching quality and students' learning effects. Brief Description of the Drawings

[0044] Figure 1 It is a general step flow chart of the present invention. Detailed Embodiments

[0045] It should be noted that, without conflict, the embodiments in this application and the features in the embodiments can be combined with each other. The following further describes this application in detail with reference to the drawings and specific embodiments.

[0046] As Figure 1 shown, the spatio-temporal fusion knowledge tracking method based on large model annotation enhancement includes the following steps:

[0047] Step S1, data preparation: Collect the historical interaction data of students in the intelligent teaching system, organize it into a learning history trajectory containing question and answer interactions of questions, determine the question set, knowledge point set, and the corresponding relationship between questions and knowledge points; [[ID=

[0048] Specifically, data preparation is the cornerstone of the entire knowledge tracing method. It collects the historical interaction data of students in the intelligent teaching system. These data cover the Q&A interactions between students and questions, and are organized into learning history trajectories. Each interaction record contains the questions answered by the students and information on whether the answers are correct or not. Determining the question set, knowledge point set, and their corresponding relationships helps to accurately locate the students' mastery of different knowledge points. For example, in the study of mathematics courses, the question set includes various types of questions such as algebra and geometry, and the knowledge point set covers functions, equations, graphic properties, etc. Through the corresponding relationship, it is possible to know the specific knowledge points examined by a certain function question. This step provides the original material for subsequent analysis. Just like the cornerstone of building a high-rise building, only when the data preparation is sufficient and accurate, the subsequent analysis and model training are more reliable. Its significance lies in providing a solid data foundation for knowledge tracing and is applicable to various online education platforms and intelligent teaching systems.

[0049] Step S2, large model annotation: Using the large model, traverse the knowledge points in the dataset through preset prompt words, annotate the precursor, successor, parallel, and other relationships between knowledge points, and construct an initial graph of knowledge point relationships; at the same time, use the large model to annotate the units to which the knowledge points belong and the relationships between units, optimize the long-distance spatial association of knowledge points, and reflect hierarchical information;

[0050] Specifically, large model annotation traverses the knowledge points in the dataset through the powerful knowledge understanding ability of the large model and preset prompt words. Annotating the precursor, successor, parallel, and other relationships between knowledge points can clearly show the sequence and degree of association of knowledge. For example, in the field of physics, mechanical knowledge is usually the precursor of electromagnetic knowledge. The annotation of this relationship helps to understand the construction logic of the knowledge system. At the same time, annotating the units to which the knowledge points belong and the relationships between units optimizes the long-distance spatial association of knowledge points. Taking the history discipline as an example, historical events in different periods are divided into corresponding units, and the precursor-successor relationship between units reflects the context of historical development, which helps students learn systematically. This step provides rich structural information for knowledge tracing. Its significance lies in providing more comprehensive knowledge association information for the model and is applicable to the sorting and teaching assistance of various disciplinary knowledge, enabling the model to better simulate the learning path of students.

[0051] Step S3, annotation optimization: Detect the results of large model annotation. For problems such as the annotation results being precursors / successors to each other, unilateral precursors / successors, the formation of loops in the precursor / successor results, and the graph being non-connected, use new prompt words for secondary confirmation or preset solutions for correction to improve the annotation quality;

[0052] Specifically, annotation optimization is a quality assurance process for the annotation results of large models. There may be problems in large model annotation such as annotation results being predecessors / successors of each other, unilateral predecessors / successors, the formation of loops by predecessor / successor results, and non-connected graphs. To address these issues, new prompts are used for secondary confirmation or preset solutions are used for correction. For example, when it is found that the annotations of two knowledge points are predecessors and successors of each other, a new prompt is used to let the large model re-judge to ensure a reasonable knowledge order. If there is a non-connected graph situation, a unit marking scheme is adopted to solve it, making the graph structure complete and ensuring smooth information transmission. This step is of great significance. It can effectively improve the annotation quality, avoid interference of incorrect information on model learning, ensure that the model is trained based on accurate knowledge relationships, and is applicable to any knowledge tracking scenario that relies on the annotation results of large models, laying a foundation for subsequent accurate analysis and prediction.

[0053] Step S4, Sequence Learning: Use a time-series learning module based on the attention mechanism to process the embeddings of the knowledge points involved in the question, calculate the question embedding in combination with the question difficulty, and obtain the time-series embedding encoding through the attention encoding layer;

[0054] Specifically, sequence learning uses a time-series learning module based on the attention mechanism to process the embeddings of the knowledge points involved in the question. It considers the importance of each knowledge point in the question and assigns weights to different knowledge points through attention calculation to obtain a more representative multi-knowledge-point embedding. Calculate the question embedding in combination with the question difficulty. For example, use the question wrong-answer rate as a difficulty indicator and incorporate it into the question embedding through linear transformation, enabling the model to perceive the differences in question difficulty. Through the attention encoding layer, transform the question embedding sequence into a time-series embedding encoding to capture the time-series characteristics of the student's learning process. In an actual scenario, if a student continuously answers questions of a certain difficulty level incorrectly, this encoding can reflect the student's learning difficulties at this knowledge point or this difficulty level. This step is crucial for understanding the dynamic changes in students' learning and provides key time-series information for subsequent prediction of students' answering situations, and is widely used in educational scenarios for analyzing students' learning processes.

[0055] Step S5, Fusion Learning: Through the knowledge point static graph embedding learning module, use the graph attention encoder to extract features from the knowledge point relationship graph and integrate it with the original sequence model to obtain the joint static embedding; through the dual-view embedding contrast learning module, construct the collaborative contrast of the sequence view and the graph view to distinguish the effective knowledge point associations from the noise relationships;

[0056] Specifically, fusion learning achieves spatio-temporal information fusion through a knowledge point static graph embedding learning module and a dual-view embedding contrast learning module. The knowledge point static graph embedding learning module uses a graph attention encoder to extract features from the knowledge point relationship graph, which can capture the spatial structure information between knowledge points, such as which knowledge points are logically closer. It is integrated with the original sequence model to obtain a joint static embedding, combining sequence information and spatial structure information. The dual-view embedding contrast learning module constructs a collaborative contrast between the sequence view and the graph view, taking the embeddings of the same knowledge point under different views as positive samples and the embedding combinations of different unit knowledge points as negative samples. Through optimization using the contrast loss function, the model can better distinguish effective knowledge point associations from noise relationships. In practical applications, it helps the model to more accurately understand the relationships between knowledge points, improve the accuracy of knowledge tracking, and is applicable to educational scenarios that require in-depth analysis of knowledge point relationships.

[0057] Step S6, Dynamic learning and prediction: Based on the student personalized information dynamic learning module of the unit relationship graph, the node embeddings are dynamically updated according to the student interaction history, integrating the student dynamic information of the time series and graph structure models; at the prediction layer, a neural network is used to learn the parameters of the modified 2PL model, and the final prediction is made by combining the forgetting factor, and the loss function is calculated.

[0058] Specifically, the dynamic learning and prediction is based on the student personalized information dynamic learning module of the unit relationship graph, and the node embeddings are dynamically updated according to the student interaction history. Considering the sparsity of the student's interaction with knowledge points, this module focuses on the knowledge points that the student has interacted with, combines the unit relationship graph, and makes the node embeddings change with the learning process, better reflecting the student's personalized information. For example, if a student performs well on a certain unit knowledge point, the module can strengthen the relevant node embeddings. At the prediction layer, a neural network is used to learn the parameters of the modified 2PL model, and the forgetting factor is combined, such as the student may forget knowledge points if not reviewed for a long time, for the final prediction. By calculating the cross-entropy loss and contrast loss, the model is continuously optimized. This step directly serves to predict the student's answering situation and provides targeted suggestions for teaching, playing a key role in educational scenarios such as personalized learning recommendations and intelligent tutoring.

[0059] Preferably, in the large model annotation step, when annotating the relationships between knowledge points, for the predecessor knowledge points, the large model judges the knowledge points that need to be mastered before learning a certain knowledge point based on the knowledge point name and educational domain knowledge; for the successor knowledge points, it judges the knowledge points that need to be mastered after learning this knowledge point; the parallel knowledge points are annotated and the learning levels are within a preset range; the remaining knowledge points are annotated as those that the large model cannot clearly classify into the first three relationships. The knowledge point relationship annotation result is represented as a directed edge relationship , where, respectively represent different knowledge points. Construct the initial knowledge point relationship graph is the knowledge point set, is a set of edge relationships between knowledge points, and the edge relationships are determined based on the marked predecessor, successor, and parallel relationships.

[0060] Specifically, in the spatio-temporal fusion knowledge tracking method enhanced by large model annotation, the annotation of the relationships between knowledge points in the large model annotation step plays a crucial role.

[0061] For predecessor knowledge points, the large model, relying on its understanding of the knowledge point names and the educational domain knowledge contained therein, determines the prerequisite knowledge points that must be mastered before learning a certain knowledge point. For example, in mathematics learning, the large model will judge that before learning the knowledge point of the derivative of a function, knowledge points such as the basic concepts of functions and limits need to be mastered in advance, which is like marking the necessary paths when planning the learning path for students.

[0062] For successor knowledge points, the large model then judges the knowledge content that should be further mastered after learning this knowledge point. For example, after learning the linear equation with one variable, knowledge points such as the linear equation with two variables and inequalities are usually the subsequent learning directions, enabling students to clearly understand the knowledge progression route.

[0063] Parallel knowledge points are preset knowledge points with similar learning levels. Taking the physics subject as an example, in the mechanics part, the two knowledge points of friction and elasticity, their prerequisite knowledge all involves the basic concept of force, and the difficulty and learning levels are quite similar. Marking them as parallel knowledge points helps students conduct analogical learning and deepen their understanding of knowledge.

[0064] For knowledge points that the large model cannot clearly classify into the first three relationships, they are marked as other knowledge points, which can avoid forced association and ensure the accuracy of annotation.

[0065] The annotation results of these knowledge point relationships are presented in the form of directed edge relationships to construct an initial graph of knowledge point relationships . As a set of knowledge points, it contains all knowledge points in the learning process; is a set of edge relationships between knowledge points, and its edge relationships are determined based on the marked predecessor, successor, parallel, and other relationships. This initial graph provides a basic knowledge structure framework for the subsequent knowledge tracking model, just like the Figure 1 ground, enabling the model to clearly "see" the associations between knowledge points, thereby more accurately analyzing students' knowledge mastery situation, providing strong support for personalized learning recommendations and teaching strategy adjustments, and being widely applied to the knowledge tracking scenarios of various online education platforms and intelligent teaching systems.

[0066] Preferably, in the annotation optimization step, for the situation where the annotation results are predecessors / successors of each other, assume that there are knowledge points and , if the large model is initially marked yes The forerunner and yes The predecessor of , the new prompt word is , its format is "Please clearly judge the knowledge point and The accurate predecessor and successor relationship between them" allows the large model to re-judge; in the case of unilateral predecessor / successor, if the knowledge point is marked It's a knowledge point The predecessor of , but no corresponding mark yes The successor of Please confirm the knowledge points and Is there a predecessor / successor relationship between them? If so, please clearly mark it. If the predecessor / successor results form a loop, the large model will break the loop structure at the preset position. The new prompt word For example, "The predecessor relationship between the following knowledge points {A, B, C, ...} is connected end to end to form a ring. Please select the preset node as the starting point for learning and break the ring structure." For disconnected situations, the large model determines to reconnect the graph structure and the new prompt word "The current knowledge point set P: {A, B, C, ...} and Q: {D, E, F, ...} are not connected. Try to find related nodes in the two sets so that the nodes form a predecessor / successor or parallel relationship."

[0067] Preferably, in the sequence learning step, the knowledge points involved in the question are embedded, and the question embedding is calculated in combination with the difficulty of the question. The specific process is as follows:

[0068] Knowledge point embedding weighted aggregation, for each given question, there is a corresponding knowledge point embedding set ,in, is the number of knowledge points involved in the question. In order to obtain more representative multi-knowledge point embeddings from this set, it is necessary to assign weights to each knowledge point embedding.

[0069] Calculate the attention coefficient through the attention mechanism , and its calculation formula is:

[0070] ;

[0071] in, is the query vector, Used to indicate the current focus of information. is the key vector, which is used to match the query vector to measure the relevance between different knowledge points. It is the key vector for traversal. Through the exponential function and normalization operation, the correlation is converted into attention coefficients. The larger the coefficient, the more important the knowledge point is in the overall representation.

[0072] Then, according to the calculated attention coefficients, a weighted sum of the knowledge point embeddings is performed to obtain the weighted knowledge point embeddings :

[0073] ;

[0074] Among them, is the attention coefficient, is the value vector;

[0075] Residual enhancement of multi-knowledge point embeddings. To enhance the ability of feature representation and the training effect of the model, residual connection and activation function processing are performed on the weighted knowledge point embeddings . After passing through the activation function, the final multi-knowledge point embeddings are obtained, and its calculation formula is:

[0076] ;

[0077] Among them, is the learnable weight matrix, which is continuously adjusted during the training process to optimize the weight of the residual connection. The activation function introduces non-linear factors, which helps the model learn more complex feature representations.

[0078] Quantification and conversion of question difficulty. To consider the difficulty factor of the question in the question embedding, first, the question difficulty is calculated by the ratio of the number of times the question is answered wrong to the total number of times it is answered . The formula is as follows:

[0079] ;

[0080] To better integrate the quantified difficulty into the question embedding, a linear transformation is performed on it to obtain the difficulty factor . The specific transformation formula is:

[0081] ;

[0082] Among them, and are learnable weight matrices, which are continuously adjusted during the training process to achieve an effective linear transformation of the difficulty value, so that the difficulty factor can be reasonably integrated with the multi-knowledge point embeddings.

[0083] Generation of the question embedding, embedding the final multi-knowledge points and the difficulty factor are added together to obtain the embedding of this question :

[0084] ;

[0085] Through the above steps, the embedding processing of the knowledge points involved in the question and the fusion of the question difficulty are completed. The generated question embedding will be used in the subsequent attention encoding layer to obtain a temporal embedding encoding that can reflect the question characteristics and difficulty information.

[0086] Preferably, in the knowledge point static graph embedding learning module, a graph attention network (GAT) encoder with edge type awareness is used to extract features from the knowledge point relationship graph. Let the feature of node in the knowledge point graph be and the feature of node in the knowledge point graph be The leaky rectified linear unit is The edge type related parameter matrix is The relationship type embedding is The attention coefficient The calculation formula is is the set of neighbor nodes of node The knowledge point static embedding matrix is obtained through double-layer GAT encoding. When integrated with the original sequence model, let the embedding result of the sequence model after passing through the attention encoding layer be The GAT embedding of the set of knowledge points corresponding to the current question is The joint static embedding[[ID=*]] is obtained through residual attention operation aggregation.

[0087] Preferably, in the dual-view embedding contrastive learning module, in step S5, positive and negative sample pairs are defined: The positive sample is the same knowledge point in different views. The pair is ; The negative sample is the embedding combination of knowledge points belonging to different units under the corresponding view. Let the knowledge point belong to unit and the knowledge point belong to unit and The negative sample pair is or The contrastive loss function is designed as

[0088] ;

[0089] is the cosine similarity function, is the temperature coefficient hyperparameter, is the number of samples.

[0090] Preferably, in the student personalized information dynamic learning module based on the unit relationship graph, an heterogeneous graph is defined is the static knowledge point node, is the dynamic unit node, is the knowledge point-unit relationship edge, is the unit-unit relationship edge. The real-time interaction feature is introduced in the node update process. Let the node embedding update formula at time be the activation function, vector concatenation operation, be the weight matrix, be the joint static embedding at time be the interaction feature at time For the set of knowledge points involved in the current question, its associated unit state is calculated as follows: be the set of units to which belongs, be the attention coefficient, be the matrix transpose operation, be the weight matrix, and are the state vectors of the associated units.

[0091] Preferably, in the prediction layer, the parameters in the IRT model are learned using a neural network, and the 2PL model is modified for the final prediction. The discrimination degree of the question and the mastery level of the student are modeled through two fully connected layers respectively, represents different multi-layer perceptrons, is the embedded student dynamic information after fusion, is the joint static embedding. The final prediction is obtained according to the IRT model, and the formula is is the question difficulty parameter. The loss function of the model is the cross-entropy loss function, is the true label, is the predicted label, is the weight coefficient of the contrast loss, is the contrast loss function.

[0092] Preferably, in the data preparation step, for the collected historical interaction data of students, if there is student data with the number of interactions less than the preset value , discard it. When conducting experimental verification, unify the input interaction sequence length to the preset value ( ). For sequences with a length greater than 200, split them into multiple sequences with a length of 200; for sequences with a length less than 200, pad them with zero vectors from the beginning position of the sequence to ensure the same length. When dividing the dataset, use the data in the dataset as the test set, and use the other data as the training set and validation set at the ratios of and respectively.

[0093] Preferably, in the large model annotation step, use different versions of large models for annotation experiments, such as the Spark model of iFlytek (call it using the official commercial Spark-lite version api) and the open-source DeepSeek large model (use the DeepSeek-V3 model in knowledge point relationship annotation and detection re-annotation, and use the DeepSeek-R1 model in knowledge point belonging unit and inter-unit relationship graph annotation). When analyzing the impact of the re-annotation strategy on the large model annotation quality, count the healthy relationships, conflict relationships, weak relationships, number of loops, non-connected components, etc. in the preliminary marking results. The proportion of healthy relationships , the proportion of conflict relationships , and the proportion of weak relationships are calculated by the formulas , are the numbers of healthy relationships, conflict relationships, and weak relationships respectively, is the total number of marked relationships; the number of loops is the count of loops formed by the knowledge point relationship annotation results, and the non-connected component is the count of non-connected subgraphs in the results.

[0094] In the field of knowledge tracing, there are many deficiencies in the existing technologies, while the spatio-temporal fusion knowledge tracing method based on large model annotation has obvious advantages.

[0095] Traditional knowledge tracing relies on education experts to annotate knowledge point relationships, which is costly and inefficient. As the number of knowledge points increases, the annotation difficulty increases exponentially. Some methods using statistical or graph theory for annotation, although reducing the labor cost, have sparse edge relationships and poor quality, and it is difficult to accurately reflect the true connections between knowledge points. Although the emerging large model annotation in recent years brings new opportunities, it faces problems such as annotation errors, hallucinations, and difficulties in cross-level interactions.

[0096] The method of the present application effectively solves the above-mentioned problems. On the one hand, through an innovative detection and re-labeling strategy, the errors and hallucinations in the large model labeling are greatly reduced. Using preset prompt words and multiple rounds of verification, the labeling results are strictly screened and corrected to ensure the accuracy of the knowledge point relationship and the labeling of the unit to which they belong, and to generate high-quality knowledge point and unit relationship diagrams, thereby overcoming the defect of unstable quality of large model labeling. On the other hand, the method introduces a spatiotemporal fusion mechanism, a static graph embedding method and a dual-view comparative learning module to fully explore the large model's understanding of the spatial structure of knowledge points, optimize the embedding quality, and break the cross-level interaction barriers. At the same time, the dynamic learning module based on the graph annotation results fully considers the students' dynamic learning process, and carries out dynamic information dissemination based on the learning order and hierarchical information between knowledge points, accurately reflecting the students' true knowledge status. This method significantly improves the model performance and interpretability, provides educators with more accurate analysis of students' knowledge mastery, helps recommend personalized teaching resources, and strongly promotes the development of educational data mining and knowledge tracking technology.

[0097] While embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A spatiotemporal fusion knowledge tracking method based on large model annotation enhancement, characterized by: The method includes: Step S1, data preparation: Collect historical interaction data of students in the intelligent teaching system, organize the historical interaction data into learning history tracks including question-answering interactions, and determine the set of questions, the set of knowledge points, and the corresponding relationship between questions and knowledge points; Step S2, large model annotation: Using the large model, traverse the knowledge points in the dataset using preset prompt words, annotate the predecessor, successor, parallel, and other relationships between knowledge points, and construct an initial graph of knowledge point relationships. At the same time, use the large model to annotate the units to which the knowledge points belong and the relationships between units, optimize the spatial association of long-distance knowledge points, and reflect hierarchical information. Step S3, annotation optimization: The large model annotation results are checked. For problems such as annotated results being predecessors of each other, unilateral predecessors, predecessor results forming loops, and graph disconnection, new prompt words are used for secondary confirmation or preset solutions are used for correction to improve the annotation quality. Step S4, sequence learning: Using a temporal learning module based on an attention mechanism, the knowledge points involved in the question are processed, the question embedding is calculated based on the difficulty of the question, and the temporal embedding code is obtained through the attention encoding layer; Step S5, fusion learning: Through the knowledge point static graph embedding learning module, the graph attention encoder is used to extract features from the knowledge point relationship graph, and integrated with the original sequence model to obtain a joint static embedding; through the dual-view embedding comparison learning module, a collaborative comparison of the sequence view and the graph view is constructed to distinguish between valid knowledge point associations and noise relationships; Step S6, dynamic learning and prediction: The module for dynamic learning of personalized student information based on the unit relationship graph dynamically updates node embeddings according to the student interaction history, integrating the student dynamic information of the time series and graph structure models; in the prediction layer, the modified 2PL model parameters are learned using a neural network, the final prediction is made in combination with the forgetting factor, and the loss function is calculated.

2. The spatiotemporal fusion knowledge tracking method based on large model annotation enhancement according to claim 1 is characterized in that: In step S2, when marking the relationship between knowledge points, for the predecessor knowledge points, the big model judges the knowledge points that need to be mastered before learning a certain knowledge point based on the name of the knowledge point and the knowledge in the education field; for the subsequent knowledge points, it judges the knowledge points that need to be mastered after learning the knowledge point; the parallel knowledge point marking and learning level are within the preset range; the remaining knowledge points are marked by the big model. The knowledge points that cannot be clearly classified into the first three relationships are represented as directed edge relationships. ,in, Represent different knowledge points respectively and construct the initial graph of knowledge point relationships is a collection of knowledge points, It is a set of edge relationships between knowledge points. The edge relationships are determined based on the annotated predecessor, successor, and parallel relationships.

3. The spatiotemporal fusion knowledge tracking method based on large model annotation enhancement according to claim 1 is characterized in that: In step S3, if the labeling results are predecessors / successors of each other, there is a knowledge point and , if the large model is initially marked yes The forerunner and yes The predecessor of , the new prompt word is , the format is "Please clearly judge the knowledge point and The accurate predecessor and successor relationship between them" allows the large model to re-judge; in the case of unilateral predecessor / successor, if the knowledge point is marked It's a knowledge point The predecessor of , but no corresponding mark yes The successor of Please confirm the knowledge points and Is there a predecessor / successor relationship between them? If so, please clearly mark it. If the predecessor / successor results form a loop, the large model will break the loop structure at the preset position, and the new prompt word For example, "The predecessor relationship between the following knowledge points {A, B, C, ...} is connected end to end to form a ring. Please select the preset node as the starting point for learning and break the ring structure." For disconnected situations, the large model determines to reconnect the graph structure and the new prompt word "The current knowledge point set P: {A, B, C, ...} and Q: {D, E, F, ...} are not connected. Try to find related nodes in the two sets so that the nodes form a predecessor / successor or parallel relationship." 4. The spatiotemporal fusion knowledge tracking method based on large model annotation enhancement according to claim 1 is characterized in that: In step S4, the knowledge points involved in the question are embedded and the question embedding is calculated based on the question difficulty. The specific process is as follows: Knowledge point embedding weighted aggregation, for each given question, there is a corresponding knowledge point embedding set ,in, Assign a weight to each knowledge point embedding based on the number of knowledge points involved in the question; Calculate the attention coefficient through the attention mechanism , the calculation formula is: ； in, is the query vector, This is the key information that we are currently focusing on. Is the key vector, used to match the query vector, used to measure the relevance between different knowledge points, through the exponential function and normalization operation, is the traversed key vector; the correlation is converted into an attention coefficient. The larger the coefficient, the more important the knowledge point is in the overall representation. According to the calculated attention coefficient, the knowledge point embedding is weighted and summed to obtain the weighted knowledge point embedding : ； in, is the attention coefficient, is a value vector; Residual enhancement of multi-knowledge point embedding, weighted knowledge point embedding Perform residual connection and activation function processing, after After the activation function, the final multi-knowledge point embedding is obtained , the calculation formula is: ； in, is a learnable weight matrix that is continuously adjusted during training to optimize the weights of the residual connections. The activation function introduces nonlinear factors; Quantification and conversion of problem difficulty, problem difficulty The number of incorrect answers to this question The total number of answers The ratio is calculated as follows: ； Perform linear transformation to obtain difficulty factor , the specific conversion formula is: ； in, and It is a learnable weight matrix that enables the difficulty factor to be integrated with the embedding of multiple knowledge points; The generation of topic embedding embeds the final multiple knowledge points and difficulty factor Add together to get the embedding of the question : ； Generated title embedding Used in the subsequent attention encoding layer to obtain temporal embedding code that can reflect the characteristics and difficulty of the question.

5. The spatiotemporal fusion knowledge tracking method based on large model annotation enhancement according to claim 1 is characterized in that: In step S5, edge type-aware GAT encoder is used to extract features from the knowledge point relationship graph. The characteristics of , nodes in the knowledge point graph The characteristics of , the edge type related parameter matrix is , the relationship type is embedded as , the leaky rectified linear unit is , attention coefficient The calculation formula is For nodes The neighbor node set is obtained by double-layer GAT encoding to obtain the static embedding matrix of the knowledge point , when integrated with the original sequence model, the embedding result of the sequence model after the attention encoding layer is , the current question corresponds to the knowledge point set The GAT is embedded as , aggregated by residual attention operation to obtain joint static embedding .

6. The spatiotemporal fusion knowledge tracking method based on large model annotation enhancement according to claim 1 is characterized in that: In step S5, a positive and negative sample pair is defined: a positive sample is a pair of the same knowledge point in different visual senses. ; Negative samples are the embedding combinations of knowledge points belonging to different units under the corresponding views. Let knowledge points Belong to the unit , knowledge points Belong to the unit ,and , the negative sample pair is or , contrastive loss function Designed for ； in, is the cosine similarity function is the temperature coefficient hyperparameter, is the sample size.

7. The spatiotemporal fusion knowledge tracking method based on large model annotation enhancement according to claim 1 is characterized in that: In step S5, a heterogeneous graph is defined. It is a static node of knowledge point. is a unit dynamic node, is the knowledge point-unit relationship edge, It is a unit-unit relationship edge. The node update process introduces real-time interaction features. The update formula of node embedding at a certain moment is: is the activation function, Vector concatenation operations, is the weight matrix, for Joint static embedding of moments, for The interactive features of the moment, the knowledge point set involved in the current question , associated unit status The calculation is as follows: , for The unit collection, is the attention coefficient, is the current temporal hidden state, is the matrix transpose operation, is the weight matrix, and is the state vector of the associated unit.

8. The spatiotemporal fusion knowledge tracking method based on large model annotation enhancement according to claim 1 is characterized in that: In step S6, the neural network is used to learn the parameters of the IRT model, and the 2PL model is modified for the final prediction. The discrimination of the questions is respectively calculated through two fully connected layers. and students' mastery level Modeling, Represents different multilayer perceptrons, To embed the integrated student dynamic information, To combine static embedding, the final prediction is obtained according to the IRT model , the formula is is the difficulty parameter of the question, the loss function of the model is the cross-picking loss function, is the true label, To predict the label, is the weight coefficient of contrast loss, is the contrast loss function.

9. The spatiotemporal fusion knowledge tracking method based on large model annotation enhancement according to claim 1 is characterized in that: In step S1, if the number of student interactions in the collected historical interaction data is less than the preset value, The student data of Unified to the default value, For sequences with a length greater than 200, split them into multiple sequences with a length of 200; for sequences with a length less than 200, fill them with zero vectors from the beginning of the sequence to ensure the same length. When dividing the data set, The data of and The ratio of is used as training set and validation set.

10. The spatiotemporal fusion knowledge tracking method based on large model annotation enhancement according to claim 1 is characterized in that: In the large model annotation step, different versions of the large model are used for annotation experiments. When analyzing the impact of the re-annotation strategy on the annotation quality of the large model, the healthy relationships, conflicting relationships, weak relationships, number of rings, non-connected components, and the proportion of healthy relationships in the preliminary annotation results are counted. , the proportion of conflict relations , weak tie ratio The calculation formulas are , are the number of healthy relationships, conflict relationships, and weak relationships, is the total number of labeled relationships; the number of rings Count of the rings formed by the annotation results of the relationship between knowledge points, non-connected components is the count of non-connected subgraphs in the result.

Images