Method, system and device for constructing physical discipline knowledge graph based on heterogeneous graph neural network and medium

Abstract

The application discloses a physics discipline knowledge graph construction method and system based on a heterogeneous graph neural network, and the method comprises the following steps: collecting structured and unstructured data from teaching resources of the physics discipline, and constructing a knowledge graph ontology comprising at least one of heterogeneous nodes, such as concept nodes, example question nodes and experiment nodes; using a hybrid graph neural network HGNN architecture integrating at least two models of MPNN, GAT, GIN, Transformer and GraphSAGE to perform representation learning on the knowledge graph; dynamically expanding and enhancing the associated edges between knowledge point nodes according to the wrong question data of students, and applying the trained knowledge graph to at least one of the following teaching scenarios: personalized learning path planning, experiment vulnerability detection or dynamic test question generation. Through the advanced graph neural network component, the application effectively captures and models the rigorous formula derivation logic chain in the physics discipline, ensures the correctness of the reasoning path between knowledge points, and solves the problem of insufficient logical dependence expression in the traditional method.

Description

Technical Field

[0001] This invention relates to the field of intelligent education, and in particular to a method and system for constructing subject knowledge graphs based on heterogeneous graph neural networks. Background Technology

[0002] Knowledge graph technology, as a core infrastructure in the field of artificial intelligence, has been widely used in intelligent education to represent subject concepts, attributes, and interrelationships. Currently, the construction of knowledge graphs for junior and senior high school physics mainly relies on traditional graph structure models and representation learning techniques. However, in dealing with the rigorous logic, multi-layered nature, and practicality unique to physics, the following significant problems exist: 1. Insufficient logical dependency modeling: Traditional knowledge graphs struggle to handle the rigorous logical chains of formula derivation in physics (such as the reasoning path from Newton's laws to kinematic formulas), and lack effective expression of isomorphic derivation relationships, resulting in low accuracy of knowledge reasoning.

[0003] 2. Inefficient hierarchical structure compression: The multi-level system of "macro-micro" in physics (such as mechanics-thermodynamics-quantum physics) is difficult to represent efficiently in traditional Euclidean space. The semantic distance between nodes deviates greatly from the actual hierarchical relationship, which affects the efficiency of knowledge retrieval.

[0004] 3. Weak cross-modal association: Heterogeneous data such as experiments, concepts, and examples lack dynamic association mechanisms, making it impossible to distinguish the importance of different types of nodes (such as the association weight between electromagnetic induction experiments and background knowledge), resulting in the knowledge graph failing to reflect the key points in teaching practice.

[0005] 4. Deficiencies in mathematical expression modeling: Long-range dependencies of symbols in physical formulas (such as the semantic consistency of integral symbols in different chapters) are difficult to capture. Traditional models cannot effectively handle the cross-chapter semantic associations of mathematical symbols, which limits the ability to transfer knowledge. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a method and system for constructing subject knowledge graphs based on heterogeneous graph neural networks.

[0007] To achieve the above objectives, the present invention employs the following method: A method for constructing a subject knowledge graph based on heterogeneous graph neural networks includes the following steps: collecting structured and unstructured data, and constructing a knowledge graph ontology containing at least one heterogeneous node among concept nodes, example nodes, and experiment nodes; The knowledge graph is represented using a hybrid graph neural network (HGNN) architecture. Based on students' incorrect answer data, the connections between knowledge point nodes are dynamically expanded and strengthened. The weight of the newly added connections is calculated using the following formula: e new =α∙LeakyReLU( W∙[h 错题 || h 关联知识点 ] ) Among them, e new The weights of the newly added edges are defined by α, which is an adjustable weight coefficient, LeakyReLU is the activation function, W is the feature transformation weight matrix, and h is the weight of the newly added edges. 错题 h represents the feature vector of the knowledge points corresponding to the incorrect questions. 关联知识点 This is the feature vector of the correct knowledge points associated with the incorrect questions, where [·||·] represents the vector concatenation operation; The trained knowledge graph can be applied to at least one of the following teaching scenarios: personalized learning path planning, experimental vulnerability detection, or dynamic test question generation.

[0008] As a preferred embodiment of the present invention, the specific method for performing representation learning on the knowledge graph using the Hybrid Graph Neural Network (HGNN) architecture includes: The concept nodes and the logical chains of formula derivation between them are encoded using the graph isomorphic network (GIN) module to maintain the isomorphism of the formula derivation. The Graph Attention Network (GAT) module is used to dynamically weight the connections between experimental nodes and concept nodes, and learnable attention weights λ are assigned to the experimental nodes. gat .

[0009] As a preferred embodiment of the present invention, the specific method for learning the representation of the knowledge graph using the hybrid graph neural network (HGNN) architecture further includes: projecting the hierarchical structure of the knowledge graph using a Poincaré disk model in hyperbolic space to compress the semantic distance between nodes with parent-child relationships.

[0010] As a preferred embodiment of the present invention, the method further includes training the HGNN model using a multi-task learning framework and dynamically adjusting the weights of edges in the knowledge graph based on the feedback of the loss function, wherein the update formula is: A ij ^(t+1) = A ij ^(t) + η∙(∂L / ∂A ij ) Among them, A ij Let represent the edge weight between node i and node j, t be the training iteration number, η be the learning rate, and L be the loss function.

[0011] As a preferred embodiment of the present invention, the personalized learning path planning in the knowledge graph application step specifically involves: searching and planning the optimal learning path from the starting knowledge point to the target knowledge point in the knowledge graph based on the Dijkstra algorithm.

[0012] As a preferred embodiment of the present invention, the experimental vulnerability detection in the graph application step specifically involves: when the attention weight α between nodes is calculated by the graph attention network GAT... ij When the value falls below a preset threshold, an experimental design vulnerability warning is triggered.

[0013] As a preferred embodiment of the present invention, the dynamic question generation in the graph application step specifically involves: sampling and associating knowledge points through random walks on the graph, and generating comprehensive application question stems by combining them with a generative adversarial network (GAN).

[0014] This invention also discloses a physics discipline knowledge graph construction system based on heterogeneous graph neural networks, comprising: The data acquisition and processing module is used to collect and process structured and unstructured data from various teaching resources in order to construct a knowledge graph ontology. The graph neural network computing module is used to load and run the hybrid graph neural network (HGNN) model, and perform representation learning, dynamic relation expansion and edge weight update of the knowledge graph. The knowledge graph storage module is used to store and update physics discipline knowledge graphs containing nodes, edges, and their weights. The teaching application interface module is used to receive teaching requests and provide personalized learning paths, experimental vulnerability detection reports, or dynamically generated test questions.

[0015] The present invention also discloses an electronic device, including a processor and a memory, the memory storing a computer program, wherein the processor executes the computer program to implement the steps of the method according to any one of claims 1-7. The present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the methods described above.

[0016] In summary, this invention patent includes at least one of the following beneficial technical effects: 1. Improved accuracy of logical reasoning: This invention effectively captures and models the rigorous formula derivation logic chain in physics through advanced graph neural network components, ensuring the correctness of the reasoning path between knowledge points and solving the problem of insufficient logical dependency expression in traditional methods.

[0017] 2. Improved the representation efficiency of knowledge hierarchy structure: By projecting and embedding the inherent tree-like hierarchical system of physics into hyperbolic geometric space, the semantic distance between widely existing parent-child hierarchical nodes is significantly compressed, realizing a more efficient and natural representation of the "macro-micro" multi-scale knowledge system, greatly improving the efficiency of knowledge retrieval and tracing.

[0018] 3. Enhanced dynamic association capabilities of multimodal knowledge elements: The introduction of a dynamic relationship expansion and weight update mechanism based on learning feedback can automatically distinguish and strengthen important associations between heterogeneous nodes such as concepts, formulas, experiments, and examples, enabling the constructed knowledge graph to dynamically adapt to teaching practice and accurately reflect the learning focus and difficulties.

[0019] 4. Expanding the depth and breadth of intelligent teaching applications: The dynamic, accurate, and teaching-adept knowledge graph constructed by this invention provides a solid data foundation for advanced teaching applications such as personalized learning path planning, real-time detection of experimental loopholes, and intelligent generation of dynamic test questions, significantly improving the intelligence level and teaching assistance effectiveness of the teaching system. Attached Figure Description

[0020] Figure 1 This is a flowchart of a method for constructing a physics discipline knowledge graph based on a heterogeneous graph neural network, according to Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of a physics discipline knowledge graph construction system framework based on heterogeneous graph neural networks in Embodiment 2 of the present invention.

[0021] Figure 3 This is a structural block diagram of an electronic device according to Embodiment 3 of the present invention. Detailed Implementation

[0022] The following is in conjunction with the appendix Figure 1 -Appendix Figure 3 The present invention will be described in further detail below. Example

[0023] like Figure 1 As shown in the figure, this embodiment discloses a method for constructing a subject knowledge graph based on heterogeneous graph neural networks, including the following steps: Step S01: Collect structured and unstructured data from physics teaching resources and construct a knowledge graph ontology containing at least one heterogeneous node among concept nodes, example nodes, and experiment nodes. First, structured and unstructured data are collected from diverse teaching resources. Structured data includes the names, attributes, hierarchical relationships, and logical chains of formula derivations from textbooks, supplementary teaching materials, and exercise banks; unstructured data includes experimental operation videos, audio explanations, and illustrated descriptions.

[0024] Subsequently, natural language processing (NLP) and computer vision (CV) technologies were used to parse and extract these data, construct heterogeneous nodes of types such as "concept", "formula", "experiment" and "example", and construct edges with relationships such as "belongs to", "derives", "related to", and "verified" to form the initial ontology of the physics knowledge graph.

[0025] Taking the construction of a knowledge graph of Newton's Second Law as an example: The rigorous formula derivation chain of "Newton's Second Law - F=ma - acceleration calculation" is extracted from high school physics textbooks. Based on this, "Newton's Second Law," "Force (F)," "Mass (m)," and "Acceleration (a)" are defined as "concept" nodes, and logical relationships of "derivation" are established between them. Simultaneously, the "Example of Slider Force Analysis" is defined as an "Example" node, and an "Application" relationship is established between it and the "Newton's Second Law" concept node. For unstructured data, the "Slider Experiment" video is analyzed, and visual features such as "force sensor readings" and "slider displacement" are extracted from keyframes using a 3D-CNN neural network model. These features are assigned as attributes to newly created "Experiment" nodes, and these nodes are connected to the "Newton's Second Law" concept node through a "Verified" relationship.

[0026] Step S02: Use the Hybrid Graph Neural Network (HGNN) architecture to perform representation learning on the knowledge graph.

[0027] In this step, we first construct a hybrid graph neural network (HGNN) architecture that integrates at least two models from MPNN, GAT, GIN, Transformer, and GraphSAGE. In this implementation, the model includes at least: Graph Isomorphic Network (GIN) component: used to process subgraphs with strict logical isomorphism (such as formula derivation chains), ensuring their strong graph structure discrimination ability.

[0028] Graph Attention Network (GAT) component: Used to handle interactions between heterogeneous nodes, dynamically distinguishing the importance of relationships between different types of nodes by calculating attention weights.

[0029] Hyperbolic Space Projection Component: Using the Poincaré disk model, the tree-like hierarchical structure of the knowledge graph is projected onto hyperbolic space to achieve a more efficient and more natural hierarchical representation.

[0030] Taking the construction of a knowledge graph of Newton's second law as an example: all concept nodes in the tree-like hierarchical structure of "Classical Mechanics - Newton's Law - F=ma" are mapped onto a Poincaré disk with curvature k=-0.7 using a hyperbolic projection component. In this space, the semantic distance between the parent node "Newton's Law" and the child node "F=ma" is significantly shortened. Simultaneously, the GAT component is enabled for the connection between "experiment" nodes and "concept" nodes in the model, and a higher initial attention weight λ is set for the experimental nodes. gat =0.6, indicating the importance of experimental data in reasoning.

[0031] Step S03: Based on students' incorrect answer data, dynamically expand and strengthen the connections between knowledge point nodes, specifically as follows: Data on students' incorrect answers during the learning process is collected and used as feedback signals to dynamically enhance and optimize the relationships within the knowledge graph. For frequently missed questions, the corresponding feature vector h of the incorrect knowledge points is extracted. 错题 The correct knowledge point feature vector h that should be associated with 关联知识点 The weight e of the newly added associated edge is calculated using the dynamic relation expansion formula. new : e new =α∙LeakyReLU( W∙[h 错题 || h 关联知识点 ] ) Where α is the weight coefficient, W is the learnable weight matrix, LeakyReLU is the activation function (λ=0.01), and [·||·] represents the vector concatenation operation. This weight is added to the edge of the corresponding node, or the weight of an existing edge is enhanced.

[0032] Taking the construction of a knowledge graph of Newton's second law as an example: During the training process, a large amount of data from students' error logs is input, specifically those containing "confusion about the applicable conditions of the F=ma formula (such as non-inertial reference frames)". The system extracts the feature vector of the "F=ma" node as h. 错题 Extract the feature vectors of the "inertial reference frame" nodes as h 关联知识点 A higher e can be obtained by calculating using the above formula. new The value is then added. Subsequently, a "Note" relationship edge is added between the "F=ma" node and the "Inertial Reference Frame" node, or the weight of an existing edge is increased to enhance this value, so that the graph can accurately reflect this learning pain point.

[0033] Step S04: Train the HGNN model using a multi-task learning framework, and dynamically adjust the weights of edges in the knowledge graph based on the feedback from the loss function. Specifically: The HGNN model is trained using a multi-task learning framework, with the main task being the prediction of knowledge point mastery. The prediction result is given by the following formula: ỿ = Softmax( MLP( [h v ||T 时序 ] ) ) Among them, h v T is the node feature vector of the current knowledge point. 时序 For example, MLP is the temporal feature vector in the learning behavior (such as the time, order, interval, etc. of learning the knowledge point), Softmax is the normalized exponential function.

[0034] During training, the weights of edges in the knowledge graph are dynamically adjusted based on the backpropagation signal of the loss function L, and the update formula is as follows: Aij ^(t+1) = A ij ^(t)+η∙(∂L / ∂A ij ) This mechanism makes the knowledge graph no longer static, but capable of continuously evolving based on student learning feedback. When the system detects that a student's mastery of a certain knowledge point has improved, it automatically reduces the weight of related incorrect edges while strengthening the weight of correct reasoning path edges.

[0035] Step S05: Apply the trained knowledge graph to at least one of the following teaching scenarios: personalized learning path planning, experimental vulnerability detection, or dynamic test question generation. In this step, a well-trained and continuously updated knowledge graph is deployed into a practical teaching system to provide intelligent applications.

[0036] Personalized learning path planning: Based on the Dijkstra graph search algorithm, the optimal path is calculated between the student's current node and the target node.

[0037] For example, to help students who have not mastered the concepts, an optimal learning path was planned starting from the "inertial concept" node, passing through the "derivation of F=ma", "slider experiment" and "variant example" nodes. Actual tests showed that this improved learning efficiency by 22%.

[0038] Experimental vulnerability detection: Real-time monitoring of students' access to experimental nodes and reasoning processes; when GAT calculates the attention weight α of the associated edges... ij When the value falls below a threshold (e.g., 0.1), an alert is triggered.

[0039] For example, when a student performs a force analysis but ignores the concept of "net external force," the system detects that the attention weight of the "net external force" node and the current node is too low (α). ij If the value is less than 0.1, an early warning will be triggered, prompting students to supplement their learning of the knowledge point of "composition of forces".

[0040] Dynamic question generation: By performing random walks on the knowledge graph, a set of closely related knowledge points are sampled and input into a pre-trained generative adversarial network (GAN) to automatically generate comprehensive application questions, options, and answers.

[0041] For example, starting from the node of "Newton's Third Law", the system samples and moves to related nodes such as "action and reaction forces" and "conservation of momentum" to generate a comprehensive application problem about "sloping plane collision problem", effectively covering high-frequency test points. Example

[0042] This embodiment discloses a subject knowledge graph construction system based on heterogeneous graph neural networks, including the following modules: The data acquisition and processing module 21 is used to collect and process structured and unstructured data from various teaching resources in order to construct a knowledge graph ontology; The graph neural network computing module 22 is used to load and run the hybrid graph neural network (HGNN) model, and perform representation learning, dynamic relation expansion and edge weight update of the knowledge graph. The knowledge graph storage module 23 is used to store and update a physics discipline knowledge graph containing nodes, edges and their weights; The teaching application interface module 24 is used to receive teaching requests and provide personalized learning paths, experimental vulnerability detection reports, or dynamically generated test questions. Example

[0043] Reference Figure 3 This embodiment discloses an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the subject knowledge graph construction method based on heterogeneous graph neural network in the above embodiment. Figure 3 The electronic device shown is merely an example and does not impose any limitation on the functionality and scope of use of the embodiments of the present invention.

[0044] Electronic devices can take the form of general-purpose computing devices, such as server devices. Components of an electronic device may include, but are not limited to, at least one processor and at least one memory. Example

[0045] This embodiment also provides a computer-readable storage medium storing a computer program thereon. When the program is executed by a processor, it implements the steps in the subject knowledge graph construction method based on heterogeneous graph neural networks as described in the above embodiments. The readable storage medium may specifically include, but is not limited to, a hard disk, random access memory, read-only memory, erasable programmable read-only memory, optical storage device, magnetic storage device, or any suitable combination thereof.

[0046] In a possible implementation, the present invention can also be implemented as a program product comprising program code that, when run on a terminal device, causes the terminal device to perform the steps of the method for constructing a physics knowledge graph based on heterogeneous graph neural networks as described above. The program code for executing the present invention can be written in any combination of one or more programming languages, and can be executed entirely on the user device, partially on the user device, as a standalone software package, partially on the user device and partially on a remote device, or entirely on a remote device.

[0047] The above are all preferred embodiments of the present invention, and are not intended to limit the scope of protection of the present invention. The scope of protection of the present invention is defined by the appended claims. Those skilled in the art will understand the implications without departing from the scope of the invention. Based on the principles and essence of this invention, various changes or modifications can be made to these embodiments, but all such changes and modifications fall within the protection scope of this invention.

Claims

1. A method for constructing a subject knowledge graph based on heterogeneous graph neural networks, characterized in that, Includes the following steps: Collect structured and unstructured data to construct a knowledge graph ontology containing at least one type of heterogeneous nodes, including concept nodes, example nodes, and experiment nodes; The knowledge graph is represented using a hybrid graph neural network (HGNN) architecture. Based on students' incorrect answer data, the connections between knowledge point nodes are dynamically expanded and strengthened. The weight of the newly added connections is calculated using the following formula: e new =α∙LeakyReLU( W∙[h 错题 ||h 关联知识点 ]) Among them, e new The weights of the newly added edges are defined by α, which is an adjustable weight coefficient, LeakyReLU is the activation function, W is the feature transformation weight matrix, and h is the weight of the newly added edges. 错题 h represents the feature vector of the knowledge points corresponding to the incorrect questions. 关联知识点 This is the feature vector of the correct knowledge points associated with the incorrect questions, where [·||·] represents the vector concatenation operation; The trained knowledge graph can be applied to at least one of the following teaching scenarios: personalized learning path planning, experimental vulnerability detection, or dynamic test question generation.

2. The method according to claim 1, characterized in that, The specific method for learning representations of the knowledge graph using the Hybrid Graph Neural Network (HGNN) architecture includes: The concept nodes and the logical chains of formula derivation between them are encoded using the graph isomorphic network (GIN) module to maintain the isomorphism of the formula derivation. The Graph Attention Network (GAT) module is used to dynamically weight the connections between experimental nodes and concept nodes, and learnable attention weights λ are assigned to the experimental nodes. gat .

3. The method according to claim 1 or 2, characterized in that: The specific method for learning representations of the knowledge graph using the Hybrid Graph Neural Network (HGNN) architecture further includes: projecting the hierarchical structure of the knowledge graph using a Poincaré disk model in hyperbolic space to compress the semantic distance between nodes with parent-child relationships.

4. The method according to claim 1, characterized in that, The method also includes training the HGNN model using a multi-task learning framework and dynamically adjusting the weights of edges in the knowledge graph based on the feedback from the loss function. The update formula is as follows: A ij ^(t+1) = A ij ^(t) + η∙(∂L / ∂A ij ) Among them, A ij Let represent the edge weight between node i and node j, t be the training iteration number, η be the learning rate, and L be the loss function.

5. The method according to claim 1, characterized in that, The personalized learning path planning in the knowledge graph application steps specifically involves: searching and planning the optimal learning path from the starting knowledge point to the target knowledge point in the knowledge graph based on the Dijkstra algorithm.

6. The method according to claim 1, characterized in that, The experimental vulnerability detection in the graph application steps specifically involves: When the attention weights α between nodes are calculated by the graph attention network GAT ij When the value falls below a preset threshold, an experimental design vulnerability warning is triggered.

7. The method according to claim 1, characterized in that, The dynamic question generation in the atlas application steps specifically involves: Knowledge points are sampled and associated by random walks on the graph, and combined with generative adversarial networks (GANs) to generate comprehensive application problem stems.

8. A system for implementing the method as described in any one of claims 1 to 7, characterized in that, include: The data acquisition and processing module is used to collect and process structured and unstructured data from various teaching resources in order to construct a knowledge graph ontology. The graph neural network computing module is used to load and run the hybrid graph neural network (HGNN) model, and perform representation learning, dynamic relation expansion and edge weight update of the knowledge graph. The knowledge graph storage module is used to store and update physics discipline knowledge graphs containing nodes, edges, and their weights. The teaching application interface module is used to receive teaching requests and provide personalized learning paths, experimental vulnerability detection reports, or dynamically generated test questions.

9. An electronic device, characterized in that, It includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The device contains a computer program that, when executed by a processor, implements the steps of the method according to any one of claims 1-7.

Images