UniGeo: Unifying Geometry Logical Reasoning via Reformulating Mathematical Expression

• DOI: 10.18653/v1/2022.emnlp-main.218
• Publication Date: 2023-08-04T20:21:02Z
• Authors: Jiaqi Chen; Tong Li; Jinghui Qin; Pan Lu; Liang Lin; Chongyu Chen; Xiaodan Liang
• Keywords: Expression (computer science); Benchmark (surveying); Sequence (biology); Qualitative reasoning

Geometry problem solving is a well-recognized testbed for evaluating the high-level multi-modal reasoning capability of deep models. In most existing works, two main geometry problems: calculation and proving, are usually treated as specific tasks, hindering model to unify its on multiple math tasks. However, in essence, these tasks have similar representations overlapped knowledge which can improve understanding ability both Therefore, we construct large-scale Unified benchmark, UniGeo, contains 4,998 problems 9,543 proving problems. Each annotated with multi-step proof reasons mathematical expressions. The be easily reformulated sequence that shares same formats program Naturally, also present unified multi-task Geometric Transformer framework, Geoformer, tackle simultaneously form generation, finally shows improved by unifying formulation. Furthermore, propose Mathematical Expression Pretraining (MEP) method aims predict expressions solution, thus improving Geoformer model. Experiments UniGeo demonstrate our proposed obtains state-of-the-art performance outperforming task-specific NGS over 5.6% 3.2% accuracies problems, respectively.