MILE: Memory-Interactive Learning Engine for Neuro-Symbolic Solutions to Mathematical Problems

Mathematical problem solving is a task that examines the capacity of machine learning systems to perform quantitative and logical reasoning. Existing work employed formulas as intermediate labels in this task to implement a neuro-symbolic approach and achieved remarkable performance. However, we are questioning the limitations of existing methods from two perspectives: the expressive capacity of formulas and the learning capacity of existing models. In this paper, we proposed the Memory-Interactive Learning Engine (MILE), a new framework for the neuro-symbolic solution to mathematical problems. The main contributions of this work include a new formula-representing technique and a new decoding method. In our experiment on the Math23K dataset, MILE outperformed existing methods on not only question-answering accuracy but also robustness and generalization capacity (the software is available at https://github.com/evan-ak/mile ).