A Graph Convolutional Network Feature Learning Framework for Interpretable Geometry Problem Solving

Abstract

Geometry problem solving is a long-standing problem in artificial intelligence. The task requires generating explainable solving sequences based on text and diagram descriptions. Existing approaches have made great progress in geometry formal language extraction and interpretable solving. However, they neglect the graph structure information in formal language. This leads to poor prediction effect of the theorem, and too long reasoning time for problem solving and affects the accuracy of problem solving. In this paper, we construct the formal language graph and use a graph convolutional network to encode structure information of formal language. We propose an improved diagram parser for better diagram relation set extraction. The experimental results show that our method achieves better performance in interpretable geometry problem solving.