Construction and Mathematical Application of Document-Level Relationship Extraction Model Combining R-GCN and Text Features

With the development of smart education, building a knowledge graph to integrate educational resources has become particularly important. However, existing sentence level relationship extraction methods are difficult to address the complex relationship extraction of cross sentence and long-distance dependencies in educational texts. Therefore, a document-level relationship extraction model based on educational text features is proposed, which combines the graph convolutional neural network with text features to optimize the recognition ability of cross-sentence relationships. Unlike conventional methods, this study first constructs a document-level relationship extraction dataset for mathematics. Five types of mathematical knowledge point relationships are defined to construct a mathematical knowledge graph. This model uses graph convolutional neural networks to capture complex semantic relationships between entities in sentences, and extracts contextual information through a pre-trained language model BERT, demonstrating excellent performance in mathematical knowledge point relationship modeling. The model achieved an accuracy of 96.34%. In the extraction of cross-sentence relationships, the F1 value reached 92.05%. Regarding the relationship between mathematical knowledge points, the accuracy and recall in five types were all 0.87 to 0.92, and the accuracy and recall in complex nested relationships were 0.95 and 0.93, respectively. In low resource scenarios, the model still maintained a precision of 0.81 and an F1 value of 0.83 at a data scale of 10%, demonstrating excellent robustness. This research has improved the efficiency and accuracy of extracting educational text relationships, and expanded the application of document-level relationship extraction technology in the field of education, which has practical value for constructing mathematical knowledge graphs.