Knowledge tracking method and system based on knowledge hierarchical structure modeling

Abstract

The invention belongs to the technical field of education theory and computer science, and particularly discloses a knowledge tracking method and system based on knowledge hierarchical structure modeling, and the method comprises the steps: obtaining a knowledge hierarchical tree and a knowledge association graph; obtaining an embedded representation of each node; aggregating cross-hierarchy information of each node based on the embedding representation of each node in combination with the knowledge hierarchy tree to obtain an embedding vector corresponding to the knowledge hierarchy tree; aggregating node information in the same hierarchy based on the embedded representation of each node in combination with the knowledge association graph to obtain an embedded vector corresponding to the knowledge association graph; graph reconstruction and cross-graph hierarchical comparison learning are carried out on the knowledge hierarchical tree and the embedded vector of the knowledge association graph; and based on the historical answering condition of the learner, combining the updated knowledge hierarchy tree or knowledge association graph embedding vector, and predicting the answering performance of the learner in the next time step. According to the invention, the multi-level structure of the knowledge concept is fully utilized, and the reliability of knowledge tracking is improved.

Description

Technical Field

[0001] This application belongs to the cross - field of educational theory and computer science technology. More specifically, it relates to a knowledge tracing method and system based on knowledge hierarchy structure modeling. Background Art

[0002] Knowledge Tracing (KT) is a core task in intelligent tutoring systems, aiming to dynamically infer learners' knowledge mastery status based on their learning interaction behaviors. With the large - scale accumulation of educational data and the development of deep learning technology, KT methods have evolved from traditional statistical models to complex deep models. According to different ways of knowledge concept modeling, existing KT methods are mainly divided into three categories: independent - association - modeling knowledge tracing, implicit - association - modeling knowledge tracing, and explicit - association - modeling knowledge tracing.

[0003] Early independent - association - modeling methods, such as Bayesian Knowledge Tracing, regarded each knowledge concept as an independent variable and used the Bayesian probability framework to update learners' mastery status of specific knowledge. These methods have simple structures and are easy to interpret, but they ignore the correlations between knowledge concepts, making it difficult to handle complex knowledge networks and having limited prediction accuracy.

[0004] With the rapid development of deep learning technology, researchers have introduced more complex models to capture the complex temporal patterns in learners' learning behaviors and the hidden associations between knowledge concepts. For example, representative models of implicit - association - modeling knowledge tracing, such as Deep Knowledge Tracing (DKT), Attentive Knowledge Tracing (AKT), etc. Among them, DKT uses a Recurrent Neural Network (RNN) to model learners' learning sequences and can capture personalized learning trajectories. However, DKT relies on an implicit network structure to learn the associations between knowledge concepts and lacks explicit modeling of knowledge relationships, resulting in poor interpretability of the model. AKT is based on a multi - layer attention mechanism to model the mastery and forgetting effects of knowledge concepts to improve KT performance. Although these implicit - association - modeling methods have improved the prediction performance of KT models through deep learning technology, they rely on implicit neural network structures and lack explicit representation of knowledge relationships, limiting the application effects of the models in teaching scenarios with complex knowledge hierarchy structures.

[0005] To address the deficiencies of implicit association modeling methods in representing knowledge relationships, researchers have introduced explicit association modeling knowledge tracing methods, which use graph structures to explicitly construct the association relationships between knowledge concepts. These methods are mainly based on Graph Neural Networks (GNNs), and their representative models include Graph-based Knowledge Tracing (GKT). GKT constructs learners, questions, and knowledge concepts as multi-type nodes and uses graph convolutional networks to capture multi-dimensional relationship features, explicitly modeling the relationships between knowledge concepts. Although GKT can more accurately reflect the structure of knowledge concepts and the cognitive state of learners, it still has limitations in dealing with knowledge hierarchical structures and multi-level association relationships. Existing explicit association modeling knowledge tracing methods mainly focus on modeling knowledge relationships at the same level or in a flattened manner, failing to fully utilize the multi-level structure of knowledge concepts, which limits their application effects in complex cognitive architectures. Summary of the Invention

[0006] Aiming at the deficiencies of the prior art, the purpose of this application is to provide a knowledge tracing method and system based on knowledge hierarchical structure modeling, aiming to solve the problem that existing knowledge tracing fails to fully utilize the multi-level structure of knowledge concepts and has limited application effects.

[0007] To achieve the above purpose, in the first aspect, this application provides a knowledge tracing method based on knowledge hierarchical structure modeling, including: Obtain a knowledge hierarchy tree and a knowledge association graph; the knowledge hierarchy tree is used to represent the hierarchical relationship of knowledge concepts, and the knowledge association graph is used to represent the intra-layer relationship of knowledge concepts. The knowledge hierarchy tree and the knowledge association graph include the same multiple nodes, and one node represents one knowledge concept; Obtain the embedding representations of each node; the embedding representations are used to identify the level to which the node belongs and reflect the similarity between different levels. Based on the embedding representations of each node and in combination with the knowledge hierarchy tree, aggregate the cross-level information of each node to obtain the first aggregated embedding representation of each node, so as to obtain the embedding vector corresponding to the knowledge hierarchy tree. Based on the embedding representations of each node and in combination with the knowledge association graph, aggregate the node information within the same level to obtain the second aggregated embedding representation of each node, so as to obtain the embedding vector corresponding to the knowledge association graph; Perform graph reconstruction and cross-graph hierarchical contrast learning on the embedding vector of the knowledge hierarchy tree and the embedding vector of the knowledge association graph to increase the similarity of the aggregated embedding representations of the same node in the knowledge hierarchy tree and the knowledge association graph, and obtain the updated embedding vector of the knowledge hierarchy tree and the embedding vector of the knowledge association graph; Based on the learner's historical answering situation, in combination with the updated embedding vector of the knowledge hierarchy tree or the embedding vector of the knowledge association graph, predict the learner's answering performance at the next time step.

[0008] In a possible implementation, obtaining a knowledge hierarchy tree and a knowledge association graph includes: Obtaining a knowledge hierarchy tree constructed according to the hierarchical relationship of knowledge concepts; each node in the knowledge hierarchy tree represents a knowledge concept, and each edge represents the association relationship between knowledge concepts in adjacent layers; Constructing a knowledge association graph based on the historical answer sequence of the learner, where each node in the knowledge association graph represents a knowledge concept, and each edge represents the association relationship between knowledge concepts in the same layer.

[0009] In a possible implementation, obtaining the embedding representation of each node includes: Obtaining the embedding representation of a node based on the initial embedding representation randomly initialized for the node and the hierarchical embedding representation of the node; the hierarchical embedding representation of the node is obtained by constructing the Fourier feature representation of knowledge concepts in each layer.

[0010] In a possible implementation, aggregating the cross-hierarchical information of each node based on the embedding representation of each node and the knowledge hierarchy tree includes: Determining a set of paths from all leaf nodes to the root node in combination with the knowledge hierarchy tree; Determining the embedding matrix of each path according to the embedding representation of each node in each path; Using a multi-head attention mechanism to extract the features of each path embedding matrix to obtain the feature matrix of each path; For a node that appears in multiple paths, based on the embedding representation of the node and the feature matrices of each path in which the node appears, determining the attention weight of the node in each path in which it appears; Based on the embedding representation of each node, the feature matrices of each path, and the attention weight of each node in each path, determining the first aggregated embedding representation of each node to obtain the embedding vector of the knowledge hierarchy tree.

[0011] In a possible implementation, aggregating the node information within the same layer based on the embedding representation of each node and the knowledge association graph includes: Determining the set of neighbor nodes of each node in combination with the knowledge association graph; the neighbor node is a node connected to the node; Concatenating the embedding representations of each node and its respective neighbor nodes after mapping, and then obtaining the attention weights of each node and its respective neighbor nodes after mapping through an embedding matrix; Performing weighted summation on the embedding representations of the respective neighbor nodes corresponding to each node and the corresponding attention weights, and combining the embedding representation of each node to obtain the second aggregated embedding representation of each node to obtain the embedding vector of the knowledge association graph.

[0012] In a possible implementation, the graph reconstruction and cross-graph hierarchical contrastive learning includes: Determining a hierarchical structure loss corresponding to the knowledge hierarchical tree embedding vector based on a first positive example sample and a first negative example sample, and performing graph reconstruction on the knowledge hierarchical tree, so that the parent-child relationship in the knowledge hierarchical tree is similar to the relationship in the corresponding embedding space of its embedding vector; the first positive example sample includes a plurality of first node pairs, each first node pair consists of two directly connected nodes in the knowledge hierarchical tree, and the first negative example sample includes a plurality of second node pairs, each second node pair consists of two non-directly connected nodes in the knowledge hierarchical tree; Determining an intra-layer structure loss corresponding to the knowledge association graph embedding vector based on a second positive example sample and a second negative example sample, and performing graph reconstruction on the knowledge association graph, so that the neighbor relationship in the knowledge association graph is similar to the relationship in the corresponding embedding space of its embedding vector; the second positive example sample includes a plurality of third node pairs, each third node pair consists of two directly connected nodes in the knowledge association graph, and the second negative example sample includes a plurality of fourth node pairs, each fourth node pair consists of two non-directly connected nodes in the knowledge association graph; Performing cross-graph hierarchical contrastive learning on the graph-reconstructed knowledge hierarchical tree and the graph-reconstructed knowledge association graph, increasing the similarity of the aggregated embedding representations of the same node after reconstruction in the knowledge hierarchical tree and the knowledge association graph, and obtaining updated knowledge hierarchical tree embedding vectors and knowledge association graph embedding vectors.

[0013] In a possible implementation, predicting the answering performance of the learner at the next time step includes: Using a long short-term memory network to obtain the hidden state of the learner based on the learner's historical answering situation; the hidden state is used to reflect the learner's mastery of knowledge concepts; Based on the hidden state of the learner and the updated knowledge hierarchical tree embedding vector or the updated knowledge association graph embedding vector, comparing the similarity between the knowledge concepts mastered by the learner and the knowledge concepts of each layer, and determining the weights of the knowledge concepts at each level; Based on the embedding representation of each layer of knowledge concepts in the updated knowledge hierarchical tree embedding vector or the updated knowledge association graph embedding vector, predicting the answering performance of the learner at the next time step, and obtaining the prediction results of each layer of knowledge concepts; Combining the weights of the knowledge concepts of each layer to perform weighted summation on the corresponding prediction results to obtain the final prediction result.

[0014] In a second aspect, the present application provides a knowledge tracking system based on knowledge hierarchical structure modeling, including: Graph construction unit, used to obtain a knowledge hierarchy tree and a knowledge association graph; the knowledge hierarchy tree is used to represent the hierarchical relationship of knowledge concepts, and the knowledge association graph is used to represent the intra-layer relationship of knowledge concepts. The knowledge hierarchy tree and the knowledge association graph include the same multiple nodes, and one node represents one knowledge concept; Graph aggregation unit, used to obtain the embedding representations of each node; the embedding representations are used to identify the level to which the node belongs and reflect the similarity between different levels; based on the embedding representations of each node and combined with the knowledge hierarchy tree, the cross-level information of each node is aggregated to obtain the first aggregated embedding representation of each node, so as to obtain the embedding vector corresponding to the knowledge hierarchy tree; based on the embedding representations of each node and combined with the knowledge association graph, the node information within the same level is aggregated to obtain the second aggregated embedding representation of each node, so as to obtain the embedding vector corresponding to the knowledge association graph; Graph reconstruction unit, used to perform graph reconstruction and cross-graph hierarchical contrast learning on the embedding vector of the knowledge hierarchy tree and the embedding vector of the knowledge association graph, increase the similarity of the aggregated embedding representations of the same node in the knowledge hierarchy tree and the knowledge association graph, and obtain the updated embedding vectors of the knowledge hierarchy tree and the knowledge association graph; Knowledge prediction unit, used to predict the answering performance of the learner at the next time step based on the learner's historical answering situation, combined with the updated embedding vector of the knowledge hierarchy tree or the embedding vector of the knowledge association graph.

[0015] In a possible implementation manner, the graph reconstruction unit is used to determine the hierarchical structure loss corresponding to the embedding vector of the knowledge hierarchy tree based on the first positive example samples and the first negative example samples, and perform graph reconstruction on the knowledge hierarchy tree, so that the parent-child relationship in the knowledge hierarchy tree is similar to the relationship in the corresponding embedding space of its embedding vector; the first positive example samples include multiple first node pairs, and each first node pair is composed of two directly connected nodes in the knowledge hierarchy tree. The first negative example samples include multiple second node pairs, and each second node pair is composed of two nodes that are not directly connected in the knowledge hierarchy tree; determine the intra-layer structure loss corresponding to the embedding vector of the knowledge association graph based on the second positive example samples and the second negative example samples, and perform graph reconstruction on the knowledge association graph, so that the neighbor relationship in the knowledge association graph is similar to the relationship in the corresponding embedding space of its embedding vector; the second positive example samples include multiple third node pairs, and each third node pair is composed of two directly connected nodes in the knowledge association graph. The second negative example samples include multiple fourth node pairs, and each fourth node pair is composed of two nodes that are not directly connected in the knowledge association graph; and perform cross-graph hierarchical contrast learning on the graph-reconstructed knowledge hierarchy tree and the graph-reconstructed knowledge association graph, increase the similarity of the reconstructed aggregated embedding representations of the same node in the knowledge hierarchy tree and the knowledge association graph, and obtain the updated embedding vectors of the knowledge hierarchy tree and the knowledge association graph.

[0016] In a third aspect, the present application provides an electronic device, including: at least one memory for storing a program; at least one processor for executing the program stored in the memory, and when the program stored in the memory is executed, the processor is configured to execute the method described in the first aspect or any possible implementation manner of the first aspect.

[0017] In a fourth aspect, the present application provides a computer-readable storage medium storing a computer program, and when the computer program runs on a processor, it causes the processor to execute the method described in the first aspect or any possible implementation manner of the first aspect.

[0018] In a fifth aspect, the present application provides a computer program product, and when the computer program product runs on a processor, it causes the processor to execute the method described in the first aspect or any possible implementation manner of the first aspect.

[0019] Generally speaking, compared with the prior art by the above technical solutions conceived in the present application, the following beneficial effects are achieved: The present application provides a knowledge tracking method and system based on knowledge hierarchy structure modeling. By explicitly modeling the hierarchy structure of knowledge, a knowledge hierarchy tree and a knowledge association graph are created, which clarifies the organization form of knowledge concepts and provides a basis for the structured representation of knowledge concepts. By aggregating the information between and within levels in the knowledge structure respectively, the semantics of the embedded representation are clarified from the vertical subordination relationship and the horizontal association relationship, so that the structural information can be fully expressed. In the prediction stage, the traditional mixed knowledge state is decoupled, the knowledge states of each knowledge level are separated, and the performance of the learner is comprehensively predicted by combining the knowledge concepts of the corresponding level, breaking the limitation of the previous single-level evaluation. The knowledge tracking method based on the knowledge structure provided by the present application has better performance, can predict the answering situation of students more accurately and comprehensively, and can provide data support for teachers' precise teaching. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is a flowchart of the knowledge tracking method based on knowledge hierarchy structure modeling provided by an embodiment of the present application; Figure 2 is a flow architecture diagram of the knowledge tracking based on knowledge hierarchy structure modeling provided by an embodiment of the present application; Figure 3 is a schematic diagram of the prediction result of using only single-level knowledge concepts or questions for each question provided by an embodiment of the present application; Figure 4 is an architecture diagram of the knowledge tracking system based on knowledge hierarchy structure modeling provided by an embodiment of the present application; Figure 5It is a schematic structural diagram of an electronic device provided by an embodiment of the present application. Detailed implementation manners

[0021] In order to make the objectives, technical solutions and advantages of the present application more clear and understandable, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0022] In the embodiments of the present application, words such as "exemplary" or "for example" are used to indicate examples, illustrations or descriptions. Any embodiment or design solution described as "exemplary" or "for example" in the embodiments of the present application should not be construed as being more preferred or having more advantages than other embodiments or design solutions. Specifically, the use of words such as "exemplary" or "for example" is intended to present relevant concepts in a specific manner.

[0023] In the description of the embodiments of the present application, unless otherwise specified, the meaning of "a plurality of" refers to two or more. For example, a plurality of node pairs refers to two or more node pairs.

[0024] The embodiments of the present application will be described below with reference to the accompanying drawings in the embodiments of the present application.

[0025] To make up for the above deficiencies, the present application explicitly models the hierarchical structure of knowledge concepts, fully explores the hierarchical relationships of knowledge concepts, comprehensively considers the influence of knowledge concepts at each level on the prediction task, and significantly improves the accuracy of predicting the answering performance of learners.

[0026] The following gives a brief introduction to the knowledge hierarchy tree and the knowledge association graph: Knowledge tracing, as an important technology in intelligent education systems, aims to dynamically evaluate and predict learners' knowledge mastery to support personalized teaching and resource recommendation. Let represent a set of learners, and is a set of questions. The knowledge tracing task can be described as predicting the performance of a learner at time step based on the learner's historical interaction information sequence , usually predicting whether the learner can answer the question at that time step correctly. Among them, is an interaction pair composed of the question answered by the learner and the answering situation, and is generally constructed as , where is the question answered at time step , and is the learner's response to the question. If the answer is correct, then . The goal of knowledge tracing is to estimate the probability of answering the question correctly at time step and the problem in the case of, the probability that the learner answers correctly, that is .

[0027] The knowledge concept is represented as , is a set of knowledge concepts. In the real situation, the knowledge concepts have a hierarchical division, and the hierarchical structure of the knowledge concepts can be described as a tree (i.e., the knowledge hierarchy tree), where represents the set of nodes. For any , in there is exactly one knowledge concept corresponding to it. represents the set of edges. For any , it represents the edge between node and node . is the set of levels. For any , it represents the level number of the node, and a mapping is defined, represents the level number of node , and the level number of the root node is 1.

[0028] In order to capture the horizontal complex correlation relationships within the knowledge concept layer, a horizontal relationship graph, called the knowledge association graph, can be introduced. For the knowledge concepts at the th layer, the knowledge association graph formed by them is represented as , where represents the set of nodes at the th layer, represents the set of edges between the nodes at the th layer. The edges are constructed based on the context dependence of the knowledge concepts in the learner's answer sequence, similar to the semantic associations revealed by the lexical associations in the natural language context. The context of the knowledge concepts in the learner's answer sequence also reflects their internal association relationships. By counting the co-occurrence or non-occurrence of the knowledge concepts in the answer sequence, edges are established between the knowledge concepts to accurately depict their mutual correlations. The knowledge association graph is composed of multiple subgraphs, and each subgraph corresponds to the association relationship of the knowledge concepts at one level. The overall knowledge association graph is defined as , where represents the subgraph at the th layer.

[0029] In summary, the hierarchical relationship of knowledge concepts is modeled through a knowledge hierarchy tree, forming a top-down hierarchical knowledge organizational structure. At the same time, a knowledge association graph is introduced to depict the association relationship of knowledge concepts at the same level. On this basis, a hierarchical embedding representation is constructed separately to identify the level to which a knowledge concept belongs and reflect the distance relationship between levels. In addition, this application optimizes the quality of the embedding representation through a cross-graph hierarchical contrast learning strategy. Finally, by introducing a weight assignment model, the impact of knowledge concepts at different levels on the prediction result is dynamically evaluated, and the weights of each level are integrated to comprehensively evaluate the learner's response. Experiments on real datasets have significantly improved the accuracy of predicting the learner's answering performance.

[0030] As Figure 1 shown, an embodiment of this application provides a flowchart of a knowledge tracking method based on knowledge hierarchy structure modeling, including: Step S101, obtaining a knowledge hierarchy tree and a knowledge association graph; the knowledge hierarchy tree is used to represent the hierarchical relationship of knowledge concepts, and the knowledge association graph is used to represent the intra-level relationship of knowledge concepts. The knowledge hierarchy tree and the knowledge association graph include the same multiple nodes, and one node represents one knowledge concept.

[0031] Specifically, a knowledge hierarchy tree can be constructed according to the hierarchical relationship of knowledge concepts; each node in the knowledge hierarchy tree represents one knowledge concept, and each edge represents the association relationship of knowledge concepts between adjacent levels.

[0032] Among them, the hierarchical relationship of the above knowledge concepts can be calibrated by expert prior knowledge or directly adopt the existing general knowledge hierarchical relationship. For example: for the questions related to the continuity of functions, the corresponding knowledge concepts include: function continuity, and the knowledge concepts at the next lower level include: discontinuous points; the knowledge concepts at the next lower level include: the first type of discontinuous points and the second type of discontinuous points.

[0033] It can be seen that a question may simultaneously examine knowledge points at different levels. For example, a question examining the first type of discontinuous points requires the learner not only to master the concepts related to the first type of discontinuous points, but also to have a certain understanding of the two knowledge concepts of discontinuous points and function continuity at the upper level. Because when the learner answers this question, first, the learner needs to call the relevant knowledge of function continuity to judge that this question examines discontinuous points, rather than continuous functions. Second, the learner needs to have a certain grasp of the concept of discontinuous points to judge that this question examines the first type of discontinuous points, rather than the second type of discontinuous points, and then use the relevant knowledge of the first type of discontinuous points to answer. This level refers to different layers of the knowledge graph.

[0034] Exemplarily, a knowledge association graph can be constructed based on the historical answer sequence of the learner. Each node in the knowledge association graph represents a knowledge concept, and each edge represents the association relationship between knowledge concepts within the same layer.

[0035] Under normal circumstances, the learner answers questions in the intelligent tutoring system.

[0036] Step S102, obtain the embedding representation of each node; the embedding representation is used to identify the level to which the node belongs and reflect the similarity between different levels.

[0037] Exemplarily, the embedding representation of a node can be obtained based on the initial embedding representation randomly initialized for the node and the hierarchical embedding representation of the node; the above-mentioned hierarchical embedding representation of the node is obtained by constructing the Fourier feature representation of each layer of knowledge concepts.

[0038] In the work of this application, the initial feature matrix of the knowledge concept is determined by the embedding corresponding to the node and the hierarchical embedding of the node as determined.

[0039] (1) Among them, , is a learnable parameter and is a randomly initialized embedding matrix. represents the hierarchical identification embedding corresponding to the node. In this application, it is fused so that the new node embedding has hierarchical information. The hierarchical identification is an important attribute of nodes in the knowledge hierarchy tree and the knowledge interaction graph. From a semantic perspective, the semantic similarity between knowledge concepts in adjacent levels is usually higher than that between knowledge concepts in levels that are far apart. The hierarchical embedding needs to be able to reflect the similarity degree between different levels to a certain extent. According to Bochner's theorem, any translation-invariant kernel function can be represented as the Fourier transform of a non-negative measure. In this application, the hierarchical identification embedding is constructed by constructing the Fourier feature representation. For the hierarchical identification embedding of the th layer: (2) Among them, is a learnable parameter, is the dimension of the hierarchical identification embedding. Specifically, the hierarchical embedding is made to be able to approximately represent the translation-invariant kernel between levels by introducing the learnable frequency parameter , so as to meet the conditions of Bochner's theorem. This generation method ensures the continuous change and translation invariance of the hierarchical identification in the vector space, enables the embedding vectors of adjacent levels to have a smooth transition, and thus better captures and reflects the hierarchical relationship. The similarity between hierarchical identification embeddings can be defined as: (3) If all the final values are the same or have comparable effects, then the similarity between hierarchical identity embeddings and the cosine function of a single frequency is directly proportional. If the final value distributions are different, then the hierarchical effect is a weighted average of different frequencies on the cosine function, and can still be regarded as a smooth proportional approximation, so it can effectively fit the embeddings of hierarchical identities and accurately reflect the distance relationship between hierarchies.

[0040] It can be understood that the hierarchical embedding of the node at the i-th level is I i , as shown in formula (2). The hierarchical embedding of the node is composed of the hierarchical embeddings I1, I2,... I i of the nodes at each level, etc.

[0041] Step S103: Aggregate the cross-level information of each node based on the embedding representation of each node and the knowledge hierarchy tree to obtain the first aggregated embedding representation of each node, so as to obtain the embedding vector corresponding to the knowledge hierarchy tree.

[0042] Exemplarily, aggregating the cross-level information of each node based on the embedding representation of each node and the knowledge hierarchy tree includes: Determine the set of paths from all leaf nodes to the root node in combination with the knowledge hierarchy tree; Determine the embedding matrix of each path according to the embedding representation of each node in each path; Use the multi-head attention mechanism to extract the features of the embedding matrix of each path to obtain the feature matrix of each path; For nodes that appear in multiple paths, based on the embedding representation of the node and the feature matrices of each path where the node appears, determine the attention weight of the node in each path where it appears; Based on the embedding representation of each node, the feature matrices of each path, and the attention weight of each node in each path, determine the first aggregated embedding representation of each node to obtain the embedding vector of the knowledge hierarchy tree.

[0043] Figure 2 Shows the main architecture of the knowledge tracking process. (a) represents the embedding training process through graph reconstruction and cross-graph hierarchical contrast learning after constructing the knowledge hierarchy tree and the knowledge association graph. (d) represents the structure of the weight assignment network. (b) represents the details of the encoding layer of the cross-level aggregation network. (c) represents the details of the encoding layer of the intra-level aggregation network. (e) represents the process of calculating the assignment prediction unit.

[0044] See Figure 2As shown, the cross-level aggregation network performs cross-level aggregation on the node information of the knowledge hierarchy tree. To effectively represent the information of nodes in the hierarchical structure, this application proposes a hierarchical encoding method based on path sequences. This application regards the sequence formed by the leaf nodes to the root node of the knowledge hierarchy tree as a hierarchical path, and uses the self-attention mechanism to construct the sequence features of path nodes to encode the hierarchical associations between nodes.

[0045] It should be noted that the cross-level aggregation network is composed of multiple overlapping encoding layers. Each encoding layer performs aggregation encoding on the cross-level node information on the path through attention weights, and the information of multiple encoding layers is fused to obtain the final representation.

[0046] For any leaf node , there is one or more paths from to its root node. For any path where the leaf node is and and satisfies (( ), and for any , there is always ; where respectively represent the different nodes passed from the leaf node to the root node in path P, represents the leaf node, represents the corresponding root node, represents the vector corresponding to the leaf node is the embedding representation of node as . The embedding matrix of path can be expressed as: (4) represents the edge between node P i and node P i+1 in path P of the knowledge hierarchy tree, represents the set of edges in the knowledge hierarchy tree.

[0047] Among them, between different levels, this application constructs the sequence features of each node through the multi-head attention mechanism using path data: (5) (6) where , represents the learnable parameter matrix, represents the concatenation operation, respectively represent the learnable weight matrices. Next, perform residual connection and output the path feature matrix: (7) Among them, , represents a learnable parameter matrix. Since there are many paths leading to the root node, the same node may appear at the same position in multiple path sequences. In this application, the attention weights for each path are calculated for the nodes that appear multiple times: (8) Among them, , represents a learnable parameter vector, represents the set of all paths containing node , represents a path containing node . represents the output path feature of node in path . In this application, the attention weights are used to update the features of node : (9) Among them, , represents the final embedding of node . Since there is more than one encoding layer, the final feature representation is obtained by averaging after iterative output by multiple encoding layers (a total of L encoding layers): (10) The set The representations of all nodes can be compactly represented as , where each row of the matrix corresponds to the final embedding vector of node .

[0048] Step S104, based on the embedding representations of each node, combine the knowledge association graph to aggregate the node information within the same level, and obtain the second embedding representation of each node after aggregation, so as to obtain the embedding vector corresponding to the knowledge association graph.

[0049] Exemplarily, aggregating the node information within the same level based on the embedding representations of each node and combining the knowledge association graph includes: Determine the set of neighbor nodes of each node in combination with the knowledge association graph; the neighbor nodes are the nodes connected to this node; Concatenate the embedding representations of each node and its respective neighbor nodes after mapping, and then obtain the attention weights of each node and its respective neighbor nodes after mapping through the embedding matrix; Perform weighted summation on the embedding representations of the respective neighbor nodes corresponding to each node and the corresponding attention weights, and combine the embedding representations of each node to obtain the second embedding representation of each node after aggregation, so as to obtain the embedding vector of the knowledge association graph.

[0050] It should be noted that the intra-level aggregation network is composed of multiple overlapping encoding layers. Each encoding layer aggregates and encodes the node information at the same level through attention weights, and the information of multiple encoding layers is fused to obtain the final representation.

[0051] See Figure 2 As shown, in the knowledge hierarchy structure, in addition to the inter-layer relationships in different vertical directions, the association relationships between knowledge concepts within the same level are equally important. The intra-level aggregation network realizes the mining of intra-level association relationships in the knowledge hierarchy structure through a knowledge association graph. For any node , the neighbor node set is , which contains all neighbor nodes of node . Calculate its attention coefficient with neighbor nodes, and calculate the attention coefficient of node itself: (11) Among them, , represents a learnable parameter matrix for mapping the embedding representation of the node, , refers to the embedding matrix for mapping to obtain the attention coefficient of neighbor nodes, represents the concatenation operation, represents the union; is the embedding representation of node , u is the neighbor node of node v . Then activate and normalize it: (12) Use the normalized attention weights to weight-sum the features of neighbor nodes, and concatenate the outputs of all attention heads to form the final node features: (13) Among them, , is the number of multi-head attentions, represents the multi-head attention mechanism. Similarly, since there is more than one encoding layer in the cross-level aggregation network, the final feature representation is obtained by averaging after iterative output by multiple encoding layers (a total of L encoding layers): (14) The set The representations of all nodes can be compactly represented as , where each row of the matrix corresponds to the final embedding vector of node .

[0052] Step S105: Perform graph reconstruction and cross-graph hierarchical contrast learning on the embedding vectors of the knowledge hierarchy tree and the knowledge association graph, increase the similarity of the aggregated embedding representations of the same node in the knowledge hierarchy tree and the knowledge association graph, and obtain the updated embedding vectors of the knowledge hierarchy tree and the knowledge association graph.

[0053] Optionally, perform graph reconstruction on the knowledge hierarchy tree, and construct positive and negative sample pairs in combination with the structure of the knowledge hierarchy tree. The positive sample pairs are composed of two connected nodes in the knowledge hierarchy tree, and the negative sample pairs are composed of two nodes that are not directly connected. Perform graph reconstruction on the knowledge association graph, and construct positive and negative sample pairs in combination with the structure of the knowledge association graph. The positive sample pairs are composed of two connected nodes in the knowledge association graph, and the negative sample pairs are composed of two nodes that are not directly connected.

[0054] Calculate the scores of each positive and negative sample pair through dot product, calculate the loss functions of the knowledge hierarchy tree and the knowledge association graph based on the scores, and then optimize their embedding representations.

[0055] Use cross-graph hierarchical contrast learning to optimize the knowledge hierarchy tree and the knowledge association graph simultaneously. Calculate the similarity of the node embedding representations of the same knowledge and the node embedding representations of the same layer but different knowledge in the knowledge hierarchy tree and the knowledge association graph respectively. By constructing hierarchical contrast learning constraints, enhance the consistency of the same knowledge concept in different graphs.

[0056] See Figure 2 As shown, in order to fully exploit the structural information contained in the knowledge hierarchy tree and the knowledge association graph, the present application designs a hierarchical knowledge contrast learning strategy. First, process the knowledge hierarchy tree and the knowledge association graph respectively through graph reconstruction methods to enhance their hierarchical structure characteristics. Subsequently, in order to overcome the limitations under a single graph representation and enable the embeddings of the same knowledge concept to be aligned with each other in different graphs, the present application introduces a cross-graph hierarchical contrast learning mechanism, aiming to integrate the respective information of the knowledge hierarchy graph and the knowledge association graph, thereby improving the unity of the embeddings.

[0057] In the knowledge hierarchy tree, by introducing a hierarchical structure loss, enhance the hierarchical characteristics in the knowledge concept, so that the parent-child relationship in the tree maintains similarity in the embedding space: (15) Among them, is the Sigmoid function, represents the positive sample score in the knowledge hierarchy tree, represents the negative sample score in the knowledge hierarchy tree, represents the negative sample set in the knowledge hierarchy tree; among them, two connected nodes together form a pair of positive samples, such as node vAnd nodes u are interconnected, then nodes v and nodes u constitute a pair of positive example samples; two nodes that are not directly connected together constitute a pair of negative example samples. For example, nodes v and nodes k are not directly connected, then nodes v and nodes k constitute a pair of negative example samples. The dot product is used to calculate the score of the sample pair: (16) Among them, is the temperature coefficient, represents the dot product operation. Similarly, a similar enhancement strategy is used in the knowledge relationship graph: (17) Among them, is the Sigmoid function, represents the score of the positive example samples in the knowledge association graph, represents the score of the negative example samples in the knowledge association graph, represents the set of negative examples in the knowledge association graph.

[0058] To overcome the limitations under a single graph representation, a cross-graph hierarchical contrastive learning strategy is adopted to fuse different information embedded in the two graphs, promoting the embeddings of the same knowledge concept in the two graphs to be more similar, so as to enhance the consistency of the knowledge concept embeddings in different graphs.

[0059] (18) Among them, γ is the scaling factor, which is a hyperparameter. is the total number of nodes, is the indicator function, and respectively represent the hierarchical embedding identifiers of nodes i and j. If and are the same, the value is 1, otherwise it is 0; the purpose is to enable contrastive learning between knowledge concepts at the same level and isolate knowledge concepts at different levels.

[0060] Finally, the overall optimization objective can be expressed as the weighted sum of the following loss functions: (19) Step S106, based on the learner's historical answer situation, combined with the updated knowledge hierarchical tree embedding vector or knowledge association graph embedding vector, predict the learner's answering performance at the next time step.

[0061] Exemplarily, predicting the learner's answering performance at the next time step includes: Calculate the knowledge state representation (i.e., the hidden state) of the learner through a Long Short-Term Memory (LSTM) network, where the knowledge state is determined according to the learner's historical answering data; Calculate the weights of knowledge concepts at each level, where the weights are determined by the knowledge state representation and the embedding representations of knowledge concepts at each level; Calculate the prediction results of knowledge concepts at each level, where the prediction results are determined by the knowledge state and the embedding representations of knowledge concepts at each level; Perform a weighted sum of the prediction results at each level as the final prediction result, where the prediction result is determined by the weights at each level and the predicted values at each level; Construct a loss function based on the comparison between the predicted values at each level and the learner's true answering results to optimize the parameters of the entire prediction part, thereby achieving accurate prediction of the learner's answering performance.

[0062] See Figure 2 As shown, the embedding representations of knowledge concepts at different levels are obtained through previous work. However, the influence of knowledge concepts at each level on the learner's performance varies at different times. How to assign weights to knowledge concepts at different levels is an important task. For this reason, a weight assignment model is designed to dynamically assign the influence weights of knowledge concepts at different levels.

[0063] First, the hidden state of the learner needs to be obtained: (20) Among them, represents the latent state, and are parameters of represents the interaction context at time step t, and its splitting method is: (21) Among them, represents the splitting function, represents the total number of levels of knowledge concepts, and respectively represent the latent states of the th layer after splitting.

[0064] Subsequently, the formula for calculating the influence weight of knowledge concepts at each level is: (22) Among them, represents the weight matrix, which can be denoted as the abbreviation MLP, represents the level of node , represents on the hierarchical path The weight of the knowledge concept at the th level corresponding to the node at the t th time step.

[0065] The training of the weight assignment network is a multi-class classification task. The optimization work includes comparing the prediction results of knowledge concepts at different levels with their ground truths. Initially, the calculation formula for the prediction result of the knowledge concept at each level is: (23) The weight is regarded as the probability distribution of the multi-class classification task. Its optimization process is as follows: (24) Among them, represents the weight of the knowledge concept at the th level corresponding to the node in the hierarchical path , represents the loss of the weight classification task, represents the predicted value and the closest level to the actual value.

[0066] The assignment of weights must meet specific conditions to ensure that the prediction results at different levels are combined according to their respective weights, so that the results at each level closely correspond to the actual results: (25) Among them, represents the loss of predicting the performance of the learner, represents the set of learners, and s is one of the learners.

[0067] Finally, the loss function of the entire weight assignment network is (26) It can be understood that the final predicted output is: .

[0068] In a specific embodiment, the embodiment of the present application uses the answering information of 80% of the users in each dataset for training to predict the learning performance of the remaining learners. Table 1 shows the experimental results of this method and other baseline models. Compared with other mainstream baseline methods, this method has improvements on each dataset. Among them, the improvement range of AUC (Area Under Curve) is from 0.31% to 0.41%, the improvement range of accuracy (Accuracy, ACC) is from 0.18% to 0.24%, and the reduction range of Root Mean Squared Error (RMSE) is from 0.11% to 0.48%. This proves the effectiveness of the knowledge-level modeling in this paper and its superiority in actual experiments. Compared with other datasets, the advantage of this method on the Junyi dataset is not significant, which may be because the number of questions in this dataset is small and the discrimination between questions is low, thus affecting the performance of this method to a certain extent. Generally speaking, the method of the present application has achieved the best results on multiple evaluation indicators and shows excellent performance. Even in the Assistment dataset with relatively scarce knowledge-level structure, it also shows excellent performance, which indicates that it can also play a role in sparse knowledge-level structure data.

[0069] Table 1

[0070] To verify the effectiveness of hierarchical aggregation, the present application conducts independent knowledge tracking experiments using knowledge at different levels respectively. In the prediction stage, an evaluator is used to evaluate the contribution of each level to the learner's answering results, and the final answering results of the learner are comprehensively predicted by combining the contributions of each level. And comprehensive prediction is carried out by combining the contributions of each level.

[0071] Specifically, the present application uses only the knowledge concepts or questions at a single level to predict each question to test the utility of integrating the contributions of each level. The experimental results are as Figure 3 shown, and the following conclusions can be drawn: 1. There are differences in the effects of single levels: When only using a single level for prediction, the performance on each dataset has a certain degree of decline. This indicates that relying solely on the information of a certain level is not sufficient to accurately predict the final answering results of learners.

[0072] 2. Differences in the characteristics of different datasets: Although the overall trend is the same, there are differences in the AUC levels of each dataset at different levels. For example, the improvements of Eedi and Junyi data at deeper levels are more significant, which may be related to the complexity of their knowledge concept distributions, richer knowledge-level structures, or more refined annotation methods.

[0073] Generally speaking, after integrating the contributions of each level, optimal performance has been achieved on each dataset. This result verifies the effectiveness of using hierarchical evaluation of contributions in this application, indicating that the hierarchical knowledge structure helps to better support complex reasoning and problem-solving, which is in line with the view in the cognitive architecture theory that knowledge information at different levels should work together to improve the overall efficiency of learning and reasoning. Compared with the single-level "flat" driven prediction, this application has achieved ideal results in both form and performance by comprehensively evaluating the contributions of knowledge concepts at each level and dynamically integrating the influences of each level.

[0074] Figure 4 This is the architecture diagram of the knowledge tracking system based on knowledge hierarchical structure modeling provided by the embodiments of this application, as Figure 4 shown, including: A graph construction unit 410, configured to obtain a knowledge hierarchical tree and a knowledge association graph; the knowledge hierarchical tree is used to represent the hierarchical relationship of knowledge concepts, and the knowledge association graph is used to represent the intra-level relationship of knowledge concepts. The knowledge hierarchical tree and the knowledge association graph include the same multiple nodes, and one node represents one knowledge concept; A graph aggregation unit 420, configured to obtain the embedding representations of each node; the embedding representations are used to identify the level to which the node belongs and reflect the similarity between different levels; based on the embedding representations of each node and in combination with the knowledge hierarchical tree, aggregate the cross-level information of each node to obtain the first aggregated embedding representation of each node, so as to obtain the embedding vector corresponding to the knowledge hierarchical tree; based on the embedding representations of each node and in combination with the knowledge association graph, aggregate the node information within the same level to obtain the second aggregated embedding representation of each node, so as to obtain the embedding vector corresponding to the knowledge association graph; A graph reconstruction unit 430, configured to perform graph reconstruction and cross-graph hierarchical contrast learning on the embedding vector of the knowledge hierarchical tree and the embedding vector of the knowledge association graph, increase the similarity of the aggregated embedding representations of the same node in the knowledge hierarchical tree and the knowledge association graph, and obtain the updated knowledge hierarchical tree embedding vector and knowledge association graph embedding vector; A knowledge prediction unit 440, configured to predict the answering performance of the learner at the next time step based on the learner's historical answering situation and in combination with the updated knowledge hierarchical tree embedding vector or knowledge association graph embedding vector.

[0075] Optionally, the above-mentioned graph reconstruction unit 430 is used to determine the hierarchical structure loss corresponding to the knowledge hierarchical tree embedding vector based on the first positive example samples and the first negative example samples, and reconstruct the knowledge hierarchical tree, so that the parent-child relationship in the knowledge hierarchical tree is similar to the relationship in the corresponding embedding space of its embedding vector; the first positive example samples include multiple first node pairs, each first node pair consists of two directly connected nodes in the knowledge hierarchical tree, and the first negative example samples include multiple second node pairs, each second node pair consists of two nodes that are not directly connected in the knowledge hierarchical tree; determine the intra-layer structure loss corresponding to the knowledge association graph embedding vector based on the second positive example samples and the second negative example samples, and reconstruct the knowledge association graph, so that the neighbor relationship in the knowledge association graph is similar to the relationship in the corresponding embedding space of its embedding vector; the second positive example samples include multiple third node pairs, each third node pair consists of two directly connected nodes in the knowledge association graph, and the second negative example samples include multiple fourth node pairs, each fourth node pair consists of two nodes that are not directly connected in the knowledge association graph; and perform cross-graph hierarchical contrast learning on the reconstructed knowledge hierarchical tree and the reconstructed knowledge association graph to increase the similarity of the aggregated embedding representations of the same node after reconstruction in the knowledge hierarchical tree and the knowledge association graph, and obtain the updated knowledge hierarchical tree embedding vector and knowledge association graph embedding vector.

[0076] It should be understood that the above system is used to execute the method in the above embodiment. For the corresponding program units in the system, their implementation principles and technical effects are similar to those described in the above method. The working process of this system can refer to the corresponding process in the above method, which will not be elaborated here.

[0077] Based on the method in the above embodiment, an embodiment of the present application provides an electronic device, as Figure 5 shown, the electronic device may include: a processor 510, a communication interface 520, a memory 530, and a communication bus 540. Among them, the processor 510, the communication interface 520, and the memory 530 communicate with each other through the communication bus 540. The processor 510 can call the logical instructions in the memory 530 to execute the method in the above embodiment.

[0078] In addition, when the above logical instructions in the memory 530 are implemented in the form of a software functional unit and sold or used as an independent product, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application.

[0079] Based on the method in the above embodiments, an embodiment of the present application provides a computer-readable storage medium storing a computer program, which, when running on a processor, causes the processor to execute the method in the above embodiments.

[0080] Based on the method in the above embodiments, an embodiment of the present application provides a computer program product, which, when running on a processor, causes the processor to execute the method in the above embodiments.

[0081] It can be understood that the processor in the embodiments of the present application may be a central processing unit (CPU), or may be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. The general-purpose processor may be a microprocessor or any conventional processor.

[0082] The method steps in the embodiments of the present application may be implemented in a hardware manner or by a processor executing software instructions. The software instructions may be composed of corresponding software modules, and the software modules may be stored in a random access memory (RAM), flash memory, read-only memory (ROM), programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), registers, hard disk, removable hard disk, CD-ROM, or any other form of storage medium well known in the art. An exemplary storage medium is coupled to the processor so that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium may also be a component of the processor. The processor and the storage medium may be located in an ASIC.

[0083] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted through the computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center in a wired manner (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wirelessly (such as infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or a data center that includes one or more integrated available media. The available medium can be a magnetic medium (such as a floppy disk, a hard disk, a magnetic tape), an optical medium (such as a DVD), or a semiconductor medium (such as a solid state disk (SSD)), etc.

[0084] It can be understood that the various numerical numbers involved in the embodiments of the present application are only for the convenience of description and are not used to limit the scope of the embodiments of the present application.

[0085] Those skilled in the art can easily understand that the above are only the preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A knowledge tracking method based on knowledge hierarchical structure modeling, characterized in that: include: Obtain knowledge hierarchy tree and knowledge association graph; The knowledge hierarchy tree is used to represent the hierarchical relationship of knowledge concepts, and the knowledge association graph is used to represent the intra-layer relationship of knowledge concepts. The knowledge hierarchy tree and the knowledge association graph include the same multiple nodes, and one node represents one knowledge concept. Obtain an embedding representation of each node; the embedding representation is used to identify the level to which the node belongs and reflect the similarity between different levels; based on the embedding representation of each node combined with the knowledge hierarchy tree, the cross-level information of each node is aggregated to obtain a first embedding representation of each node after aggregation, so as to obtain an embedding vector corresponding to the knowledge hierarchy tree; based on the embedding representation of each node combined with the knowledge association graph, the node information within the same level is aggregated to obtain a second embedding representation of each node after aggregation, so as to obtain an embedding vector corresponding to the knowledge association graph; Perform graph reconstruction and cross-graph hierarchical comparative learning on the embedding vectors of the knowledge hierarchy tree and the knowledge association graph, increase the similarity of the embedding representations of the same node after aggregation in the knowledge hierarchy tree and the knowledge association graph, and obtain updated embedding vectors of the knowledge hierarchy tree and the knowledge association graph; Based on the learner's historical answer situation, combined with the updated knowledge hierarchy tree embedding vector or knowledge association graph embedding vector, the learner's answering performance in the next time step is predicted.

2. The method according to claim 1, characterized in that Get the knowledge hierarchy tree and knowledge association graph, including: Obtaining a knowledge hierarchy tree constructed according to the hierarchical relationship of knowledge concepts; each node in the knowledge hierarchy tree represents a knowledge concept, and each edge represents the association relationship between knowledge concepts in adjacent layers; A knowledge association graph is constructed based on the learner's historical answer sequence, in which each node represents a knowledge concept, and each edge represents the association relationship between knowledge concepts in the same layer.

3. The method according to claim 1, characterized in that Get the embedding representation of each node, including: The embedded representation of the node is obtained based on the randomly initialized initial embedded representation of the node and the hierarchical embedded representation of the node; the hierarchical embedded representation of the node is obtained by constructing the Fourier feature representation of each layer of knowledge concepts.

4. The method according to claim 1, characterized in that Based on the embedding representation of each node and the knowledge hierarchy tree, the cross-level information of each node is aggregated, including: Combine the knowledge hierarchy tree to determine the set of paths from all leaf nodes to the root node; Determine the embedding matrix of each path according to the embedding representation of each node in each path; Use the multi-head attention mechanism to extract the features of each path embedding matrix and obtain the feature matrix of each path; For a node that appears in multiple paths, the attention weight of the node in each path in which it appears is determined based on the embedding representation of the node and the feature matrix of each path in which the node appears; Based on the embedding representation of each node, the feature matrix of each path, and the attention weight of each node in each path, the first embedding representation of each node after aggregation is determined to obtain the embedding vector of the knowledge hierarchy tree.

5. The method according to claim 1, characterized in that Based on the embedding representation of each node and the knowledge association graph, the node information in the same level is aggregated, including: Determine the neighbor node set of each node in combination with the knowledge association graph; the neighbor node is a node connected to the node; The embedded representations of each node and its neighboring nodes are concatenated after mapping, and then the attention weights of each node and its neighboring nodes are obtained after mapping through the embedding matrix; The embedding representations of each neighboring node corresponding to each node and the corresponding attention weights are weighted summed, and the second embedding representation of each node after aggregation is obtained by combining the embedding representation of each node to obtain the embedding vector of the knowledge association graph.

6. The method according to claim 1, characterized in that The graph reconstruction and cross-graph hierarchical comparative learning include: Determine the hierarchical structure loss corresponding to the embedding vector of the knowledge hierarchy tree based on the first positive sample and the first negative sample, and reconstruct the graph of the knowledge hierarchy tree so that the parent-child relationship in the knowledge hierarchy tree and the relationship in the embedding space corresponding to its embedding vector maintain similarity; the first positive sample includes a plurality of first node pairs, each of which is composed of two directly connected nodes in the knowledge hierarchy tree, and the first negative sample includes a plurality of second node pairs, each of which is composed of two nodes that are not directly connected in the knowledge hierarchy tree; Determine the intra-layer structural loss corresponding to the embedding vector of the knowledge association graph based on the second positive sample and the second negative sample, and reconstruct the knowledge association graph so that the neighbor relationship in the knowledge association graph and the relationship in the embedding space corresponding to its embedding vector maintain similarity; the second positive sample includes a plurality of third node pairs, each of which is composed of two directly connected nodes in the knowledge association graph, and the second negative sample includes a plurality of fourth node pairs, each of which is composed of two nodes that are not directly connected in the knowledge association graph; The knowledge hierarchy tree after graph reconstruction and the knowledge association graph after graph reconstruction are subjected to cross-graph hierarchical comparative learning to increase the similarity of the reconstructed aggregate embedding representation of the same node in the knowledge hierarchy tree and the knowledge association graph, and obtain the updated knowledge hierarchy tree embedding vector and knowledge association graph embedding vector.

7. The method according to claim 1, characterized in that Predict the learner's performance in answering questions at the next time step, including: Using a long short-term memory network to obtain the learner's hidden state based on the learner's historical answer situation; the hidden state is used to reflect the learner's mastery of knowledge concepts; Based on the learner's hidden state and the updated knowledge hierarchy tree embedding vector or the updated knowledge association graph embedding vector, the similarity between the knowledge concepts mastered by the learner and the knowledge concepts at each level is compared to determine the weights of the knowledge concepts at each level; Based on the embedded representation of each layer of knowledge concepts in the updated knowledge hierarchy tree embedding vector or the updated knowledge association graph embedding vector, the learner's answering performance at the next time step is predicted to obtain the prediction results of each layer of knowledge concepts; The weight of each layer of knowledge concepts is combined to weight the corresponding prediction results and obtain the final prediction result.

8. A knowledge tracking system based on knowledge hierarchical structure modeling, characterized in that: include: A graph construction unit, used to obtain a knowledge hierarchy tree and a knowledge association graph; The knowledge hierarchy tree is used to represent the hierarchical relationship of knowledge concepts, and the knowledge association graph is used to represent the intra-layer relationship of knowledge concepts. The knowledge hierarchy tree and the knowledge association graph include the same multiple nodes, and one node represents one knowledge concept. A graph aggregation unit is used to obtain an embedded representation of each node; the embedded representation is used to identify the level to which the node belongs and reflect the similarity between different levels; based on the embedded representation of each node, the cross-level information of each node is aggregated in combination with the knowledge hierarchy tree to obtain a first embedded representation of each node after aggregation, so as to obtain an embedded vector corresponding to the knowledge hierarchy tree; based on the embedded representation of each node, the node information in the same level is aggregated in combination with the knowledge association graph to obtain a second embedded representation of each node after aggregation, so as to obtain an embedded vector corresponding to the knowledge association graph; A graph reconstruction unit is used to reconstruct the embedding vectors of the knowledge hierarchy tree and the knowledge association graph and perform cross-graph hierarchical comparative learning, increase the similarity of the embedding representations of the same node after aggregation in the knowledge hierarchy tree and the knowledge association graph, and obtain updated embedding vectors of the knowledge hierarchy tree and the knowledge association graph; The knowledge prediction unit is used to predict the learner's answering performance in the next time step based on the learner's historical answering situation combined with the updated knowledge hierarchy tree embedding vector or knowledge association graph embedding vector.

9. The system according to claim 8, characterized in that The graph reconstruction unit is used to determine the hierarchical structure loss corresponding to the embedding vector of the knowledge hierarchy tree based on the first positive sample and the first negative sample, and to reconstruct the graph of the knowledge hierarchy tree so that the parent-child relationship in the knowledge hierarchy tree and the relationship in the embedding space corresponding to its embedding vector maintain similarity; the first positive sample includes a plurality of first node pairs, each first node pair is composed of two directly connected nodes in the knowledge hierarchy tree, and the first negative sample includes a plurality of second node pairs, each second node pair is composed of two nodes that are not directly connected in the knowledge hierarchy tree; based on the second positive sample and the second negative sample, determine the intra-layer structure loss corresponding to the embedding vector of the knowledge association graph, and reconstruct the graph of the knowledge association graph so that the neighbor relationship in the knowledge association graph and the relationship in the embedding space corresponding to its embedding vector maintain similarity; The second positive example samples include multiple third node pairs, each third node pair consists of two nodes directly connected in the knowledge association graph, and the second negative example samples include multiple fourth node pairs, each fourth node pair consists of two nodes that are not directly connected in the knowledge association graph; and the knowledge hierarchy tree after graph reconstruction and the knowledge association graph after graph reconstruction are subjected to cross-graph hierarchical comparative learning to increase the similarity of the reconstructed aggregate embedding representation of the same node in the knowledge hierarchy tree and the knowledge association graph, and obtain updated knowledge hierarchy tree embedding vector and knowledge association graph embedding vector.

10. An electronic device, characterized in that: include: at least one memory for storing a computer program; At least one processor is used to execute the program stored in the memory. When the program stored in the memory is executed, the processor is used to execute the method according to any one of claims 1 to 7.

Images