Search

CN-121860025-B - Educational knowledge graph construction method and system based on artificial intelligence

CN121860025BCN 121860025 BCN121860025 BCN 121860025BCN-121860025-B

Abstract

The application relates to the technical field of educational data processing and discloses an educational knowledge graph construction method and system based on artificial intelligence. The system firstly utilizes a convolutional neural network to extract multi-layer convolutional response distribution of a normalized teaching image, calculates semantic coverage attenuation indexes reflecting the change of information coverage along with the scale, maps the indexes into node bearing potential, calculates the quantity quota of visual anchor points and concept entities aiming at the image according to the index, positions a visual area according to the quantity quota and extracts related text concepts to construct a weighted image substructure, and finally carries out relation reasoning through a map completion model to shape a final map. The application can adaptively adjust the composition scale according to the knowledge density of the image content, solves the problem of density unbalance of the multi-mode map structure, and is suitable for intelligent education scenes.

Inventors

  • Lv Saidong

Assignees

  • 云南师范大学

Dates

Publication Date
20260508
Application Date
20260319

Claims (7)

  1. 1. An artificial intelligence based educational knowledge graph construction system, comprising: The multi-scale characterization analysis unit is configured to extract a multi-layer convolution response chart of the normalized teaching image by using a convolution neural network, count response energy distribution of the multi-layer convolution response chart, and calculate a semantic coverage attenuation index reflecting the change of image information along with the theoretical receptive field scale; The structural parameter configuration unit is configured to map the semantic coverage attenuation index into a node bearing potential value, and calculate the visual anchor point quantity quota and the concept entity quantity quota aiming at the normalized teaching image based on the node bearing potential value; The multi-mode graphic primitive assembling unit is configured to locate a visual attention area in the normalized teaching image according to the visual anchor point quantity quota to construct a weighted image substructure, and extract concept items from an associated text according to the concept entity quantity quota to establish cross-mode connection; The intelligent link optimization unit is configured to input the built initial triplet into the convolution spectrum complement model to perform relation reasoning and generate an education knowledge spectrum data structure containing multi-mode association data, so that the storage topology density of the multi-mode spectrum data is optimized and convergence of relation reasoning calculation is accelerated; extracting a multilayer convolution response map of the normalized teaching image by using a convolution neural network, including: Carrying out graying treatment, equal-proportion scaling and edge filling treatment on an original document page to generate a page image with uniform resolution, cutting a teaching image area through layout analysis on the page image, and carrying out mean value removal and variance normalization on the teaching image area to obtain the normalized teaching image; The normalized teaching image is input into a deep convolutional neural network, the deep convolutional neural network comprises a plurality of convolutional levels which are connected in sequence, and each convolutional level is configured with a convolutional kernel size and stride parameters; Extracting output characteristic tensors of the depth convolution neural network at preset levels as the multi-layer convolution response map, wherein the convolution response map of each preset level corresponds to a theoretical receptive field scale on the normalized teaching image; The response energy distribution of the multi-layer convolution response graph is counted, and a semantic coverage attenuation index reflecting the change of image information along with the theoretical receptive field scale is calculated, wherein the method comprises the following steps: Calculating the square sum of all channel response values of each preset level aiming at the multi-layer convolution response graph of each preset level to obtain local response energy, dividing the local response energy by the total energy of the full graph of the level to generate normalized response energy distribution, and further calculating the negative logarithm of the square sum of the response energy distribution to obtain a second-order Rayleigh entropy value; Acquiring the logarithmic value of the theoretical receptive field scale corresponding to each preset level, and establishing a corresponding sequence of the second-order Rayleigh entropy value relative to the logarithmic value of the theoretical receptive field scale; And performing linear regression fitting on the corresponding sequence by using a least square method, calculating the slope value of a regression line, and taking the opposite number of the slope value as the semantic coverage attenuation index.
  2. 2. The artificial intelligence based educational knowledge graph constructing system according to claim 1, wherein mapping the semantic coverage decay indicator to a node bearing potential value comprises: Weighting the semantic coverage attenuation index by using a preset scale calibration coefficient, and superposing a preset offset calibration coefficient to obtain a linear transformation result; and inputting the linear transformation result into an S-shaped logic function for nonlinear mapping, calculating to obtain a real value with a value range from zero to one, and taking the real value as the node bearing potential value.
  3. 3. The artificial intelligence based educational knowledge graph construction system according to claim 2, wherein calculating the visual anchor point quantity quota and the concept entity quantity quota for the normalized teaching image based on the node bearing potential value comprises: Obtaining a preset minimum visual anchor point number constant and a preset maximum visual anchor point number constant, calculating the difference value of the minimum visual anchor point number constant and the maximum visual anchor point number constant, multiplying the difference value by the node bearing potential value, executing upward rounding operation on the obtained product result, and adding the operation result and the minimum visual anchor point number constant to obtain the visual anchor point number quota; Obtaining a preset minimum concept entity number constant and a preset maximum concept entity number constant, calculating the difference value of the minimum concept entity number constant and the maximum concept entity number constant, multiplying the difference value by the node bearing potential value, executing upward rounding operation on the obtained product result, and adding the operation result and the minimum concept entity number constant to obtain the concept entity number quota.
  4. 4. The artificial intelligence based educational knowledge graph constructing system according to claim 3, wherein locating a visual attention area in the normalized teaching image according to the visual anchor point quantity quota to construct a weighted image substructure comprises: Selecting target level features in the multi-layer convolution response graph, performing single-pixel convolution operation and nonlinear activation operation on the target level features to generate a visual attention thermodynamic diagram, arranging all pixel positions in the visual attention thermodynamic diagram in descending order according to response scores, selecting the pixel positions with the number equal to the visual anchor point number quota as key points, and mapping key point coordinates back to the original space of the normalized teaching image to determine center coordinates of the visual attention area; Cutting out a local image patch from the normalized teaching image by taking the center coordinates as a reference, embedding the local image patch into a node embedding network to extract node feature vectors, and embedding the spliced input relationship of each pair of node feature vectors into the network to extract relationship feature vectors; And calculating an inner product of the relation feature vector and a preset projection vector, performing normalized exponential function operation on an inner product result to obtain a basic weight, and multiplying the basic weight by the node bearing potential value to obtain a final weight of a connecting edge in the weighted image substructure.
  5. 5. The artificial intelligence based educational knowledge graph construction system of claim 4, wherein extracting concept items from associated text based on the concept entity quantity quota to establish a cross-modal connection comprises: Obtaining a legend text bound with the standardized teaching image and a preset number of text blocks closest to the standardized teaching image in a document layout, and splicing the legend text and an optical character recognition result of the text blocks to generate the associated text; Segmenting the associated text by using a preset separator set, enumerating all continuous substrings in the segmented fragments, reserving substrings containing Chinese characters, letters or numbers to form a concept candidate set, and inputting each candidate word in the concept candidate set into a text convolutional neural network to extract a concept feature vector; Calculating Euclidean norms of the concept feature vectors, arranging candidate words in the concept candidate set in descending order according to the numerical values of the Euclidean norms, selecting the candidate words with the quantity equal to the quantity quota of the concept entities as the concept items, and establishing cross-modal connecting edges pointing to the concept items from the normalized teaching images.
  6. 6. The artificial intelligence based educational knowledge graph constructing system according to claim 5, wherein inputting the initial triplet into the convolution graph completion model for relational reasoning to generate an educational knowledge graph data structure containing multi-modal associated data, thereby optimizing the storage topology density of the multi-modal graph data and accelerating the convergence of relational reasoning calculation, comprising: Collecting a connecting edge in the weighted image substructure and the cross-modal connecting edge to construct an initial triplet set, and performing dimension remodeling and splicing on a head entity feature vector and a relationship type feature vector in each triplet to form a two-dimensional feature map; Performing two-dimensional convolution operation on the two-dimensional feature map by using a map convolution layer, performing random inactivation operation on a linear rectification activation function and the feature map on a convolution result, and mapping back to a feature space through a linear projection layer by flattening to obtain a projection feature vector; And calculating the inner products of the projection feature vectors and all candidate tail entity feature vectors in the atlas, obtaining the prediction values of the candidate triples through S-type activation function operation, selecting the complement edges with the highest prediction values and preset quantity to write the complement edges into the initial triplet set, and converting the expanded set into an attribute map data structure for persistence storage.
  7. 7. An artificial intelligence based educational knowledge graph construction method applied to the artificial intelligence based educational knowledge graph construction system of any one of claims 1-6, characterized by comprising: extracting a multilayer convolution response chart of the normalized teaching image by using a convolution neural network, counting response energy distribution of the multilayer convolution response chart, and calculating a semantic coverage attenuation index reflecting the change of image information along with the theoretical receptive field scale; Mapping the semantic coverage attenuation index into a node bearing potential value, and calculating a visual anchor point quantity quota and a concept entity quantity quota aiming at the normalized teaching image based on the node bearing potential value; positioning a visual attention area in the normalized teaching image according to the visual anchor point quantity quota to construct a weighted image sub-structure, and extracting concept items from the associated text according to the concept entity quantity quota to establish cross-modal connection; and (3) inputting the built initial triplet into a convolution spectrum complement model for relational reasoning to generate an educational knowledge spectrum data structure containing multi-mode associated data, thereby optimizing the storage topology density of the multi-mode spectrum data and accelerating the convergence of relational reasoning calculation.

Description

Educational knowledge graph construction method and system based on artificial intelligence Technical Field The invention relates to the technical field of educational data processing, in particular to an educational knowledge graph construction method and system based on artificial intelligence. Background With the development of online education and intelligent teaching, educational knowledge patterns play an increasingly important role as a core infrastructure for connecting teaching resources and planning learning paths. The traditional knowledge graph construction mainly relies on analysis of text data such as teaching materials, teaching plans and the like, and entity and relation are extracted through natural language processing technology. However, in an actual teaching scenario, a great deal of key knowledge does not exist only in text form, but is carried in visual images such as function diagrams, circuit diagrams, geometric schematic diagrams, flowcharts and the like. Therefore, multi-modal knowledge maps fusing vision and text are constructed as current research hotspots. In the existing multi-modal education knowledge graph construction process, a deep learning model is generally adopted to process teaching images. The mainstream scheme often regards an image as an independent node, or uses a general object detection algorithm to identify a plurality of objects in a preset category from the image, and links the objects to a knowledge graph. For example, in dealing with a problem that includes a circuit diagram, the system may attempt to identify the resistance, power supply, etc. elements in the diagram and establish their connection to the associated concepts. However, existing processing methods have significant limitations in coping with the complexity of educational scenes. The main problem is that existing algorithms typically use fixed or static strategies to decide how much information to extract from an image. For example, the system may set the top N objects with highest confidence in the fixed extraction of each image, or simply mount the image as an integral attribute under the text node. This static strategy ignores the large difference in content-bearing density of educational images. In an actual textbook layout, a chapter overview may cover more than ten key concepts and complex logical relationships, while a local thematic map may correspond to only a single knowledge point. If a unified strategy is adopted for extraction, serious structural imbalance is caused, namely, for a high-density overview chart, a fixed extraction quantity can cause a large amount of key information to be omitted, so that relationship links in the chart are incomplete, and for a low-density partial chart, irrelevant noise can be introduced to forcibly extract excessive information, so that chart structural redundancy is caused. The uncontrolled density of the map structure ultimately affects the accuracy and the calculation convergence efficiency of upper applications (such as test question recommendation and path planning). Therefore, how to enable a computer system to sense the richness of knowledge contained in an image and adaptively allocate computing resources and structural space for constructing a map according to the richness of knowledge is a technical problem to be solved in the current educational knowledge map construction. Disclosure of Invention The invention provides an educational knowledge graph construction method and system based on artificial intelligence, which solve the technical problems in the background technology. In a first aspect, an artificial intelligence based educational knowledge graph construction system includes: The multi-scale characterization analysis unit is configured to extract a multi-layer convolution response chart of the normalized teaching image by using a convolution neural network, count response energy distribution of the multi-layer convolution response chart, and calculate a semantic coverage attenuation index reflecting the change of image information along with the theoretical receptive field scale; The structural parameter configuration unit is configured to map the semantic coverage attenuation index into a node bearing potential value and calculate a visual anchor point quantity quota and a concept entity quantity quota aiming at the teaching image based on the node bearing potential value; The multi-mode graphic primitive assembling unit is configured to locate a visual attention area in the teaching image according to the visual anchor point quantity quota to form a weighted image sub-structure, and extract concept items from the associated text according to the concept entity quantity quota to establish cross-mode connection; The intelligent link optimization unit is configured to input the built initial triplet into the convolution spectrum complement model to perform relation reasoning and generate an education knowledge spectrum data struct