Improving Equation Set Problems with Label Augmentation

Anonymous

Abstract: Math word problems solving has received considerable attention from many NLP researchers. Inspired by the encoder-decoder structure, they created a series of neural network models to solve arithmetic word problems and equation set problems. However, these encoder-decoder models used the ground truth as the only generation target, resulting in shallow heuristics to generate expressions. In this paper, we propose a simple and effective label augmentation method for equation set problems. Specifically, we transform the ground truth into several equivalent labels by normalization rules, and these new labels will be used as additional generation targets for model training. Experimental results on the English dataset DRAW1K and Chinese dataset HMWP show that the label augmentation method has at most 4.5% improvement over the state-of-the-art (SoTA) models.