LANS: A Layout-Aware Neural Solver for Plane Geometry Problem

Ming-Liang Zhang,Zhong-Zhi Li,Fei Yin,Cheng-Lin Liu

DOI: https://doi.org/10.48550/arXiv.2311.16476

2023-11-25

Computer Vision and Pattern Recognition

Abstract:Geometry problem solving (GPS) is a challenging mathematical reasoning task requiring multi-modal understanding, fusion and reasoning. Existing neural solvers take GPS as a vision-language task but be short in the representation of geometry diagrams which carry rich and complex layout information. In this paper, we propose a layout-aware neural solver named LANS, integrated with two new modules: multimodal layout-aware pre-trained language model (MLA-PLM) and layout-aware fusion attention (LA-FA). MLA-PLM adopts structural and semantic pre-training (SSP) to implement global relationship modeling, and point matching pre-training (PMP) to achieve alignment between visual points and textual points. LA-FA employs a layout-aware attention mask to realize point-guided cross-modal fusion for further boosting layout awareness of LANS. Extensive experiments on datasets Geometry3K and PGPS9K validate the effectiveness of the layout-aware modules and superior problem solving performance of our LANS solver, over existing symbolic solvers and neural solvers. The code will make public available soon.