GeoExpert: Geometry Problem Solving Based on RoBERTa and Graph Attention Networks

Geometry problem solving always attracts much attention of research and engineering.Solving geometry problems requires the capacity of parsing multi-modal information and utilizing theorem knowledge for comprehensive reasoning.The lack of high-quality datasets and interpretable symbolic reasoning is a great challenge in automated geometry problem solving.Based on Inter-GPS, we reannotate all wrong annotations in Geometry3K dataset to form a new dataset Geometry3K+ with a higher quality.We further perform a data augmentation method to expand the dataset scale to 7, 200.We propose GeoExpert, an automated geometry problem solving framework, which employs RoBERTa and Graph Attention Networks to recommend theorems.The experimental results demonstrate the effectiveness of Geometry3K+ and GeoExpert, and the accuracy is improved to 92.9%.