READABLE SOLVING METHOD FOR GEOMETRY PROBLEMS BASED ON RETINANET

This paper proposes a readable solving method for geometry problems, in order to solve the problem of inaccurate recognition of geometry diagrams and the difficulty of text and diagram fusion. The framework of RetinaNet was improved by using DenseNet-121, and auxiliary tasks were introduced to improve the accuracy of geometry feature extraction. The geometry relations in diagram were extracted by using Inter-GPS and described by formal language, so that the recognition and readable expression of geometry diagrams could be achieved. Thenodes, diagram and text in the geometry relation set were encoded by using graph convolutional neural network(GCN), ResNet101 and gate recurrent unit(GRU) respectively. The readable solving equations were generated through an equation generator. The experiment was conducted on dataset Geometry3K and GeoQA, and the average accuracy of diagram recognition reached 83.83% and the average accuracy of readable solutions reached 59.6%, which verified the feasibility and effectiveness of the proposed method.