Enhancing Geometry Problem Solving with Attention Mechanism and Super-Resolution

As deep learning techniques advance, they have shown great potential in solving complex mathematical problems. However, most research has focused on solving mathematical word problems. In this study, we aim to leverage deep learning to automatically solve geometry problems, which is a significant challenge in the fields of mathematical education and automated reasoning. Although various studies have attempted to apply different algorithms to automatically solve geometry problems, the impact of diagram quality on model performance has often been overlooked. In this work, we effectively address the issue of low-resolution geometric diagrams by introducing image super-resolution technology, thereby enhancing model performance and generalization capability. Additionally, we propose a novel dual parallel encoder structure incorporating an attention mechanism, improving the model’s understanding of problem contexts. Our model achieves an accuracy of 66.2% on the GeoQA test set and $\mathbf{5 2. 7 \%}$ on the GeoQA+ test set. To the best of our knowledge, this is the highest accuracy reported for these datasets. These results provide new insights and tools for research and applications in this field.