Weakly Supervised Formula Learner for Solving Mathematical Problems

Yuxuan Wu, Hideki Nakayama

Keywords: mathematical reasoning, weakly supervised learning

Abstract: Mathematical reasoning task is a subset of the natural language question answering task. Several approaches have been proposed in existing work to solve mathematical reasoning problems. Among them, the two-phase solution to first predict formulas from questions and then calculate answers from formulas has achieved desirable performance. However, this design results in the reliance on annotated formulas as the intermediate labels for training. In this work, we put forward a brand-new idea to enable the models to explore the formulas by themselves to eliminate the reliance on formula annotations. To realize this, we proposed Weakly Supervised Formula Leaner, a learning framework that can autonomously search for the optimal formulas through the training process and continuously update itself. Our experiment is conducted on a typical mathematical dataset MathQA. The result shows that our models learning with weak supervision outperform the baseline methods.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

One-sentence Summary: This work proposed a learning framework for exploring formulas for solving mathematical problems with weak supervision.

Reviewed Version (pdf): https://openreview.net/references/pdf?id=tlAIJVdq-