Self-Competitive Learning for Solving Math Word Problem

Anonymous

Abstract: Math word problem (MWP) aims to automatically solve mathematical questions given in texts. Most previous MWP models tend to fit the sole ground-truth solution provided by the dataset, without considering the diverse but equivalent solution expressions. To mitigate this issue, we propose a self-competitive learning framework (called SCL), which attempts to get different predictions and improve the generalization ability of the model by cooperatively learning a source network and a pruned competitor network. The competitor network is created by pruning a source network, which perturbs the source network’s structure and is conducive to generate diverse solutions. The source network and the competitor network learn collaboratively and teach each other throughout the training process. Extensive experiments on two large-scale benchmarks demonstrate that our model substantially outperforms the strong baseline methods. In particular, our method improves the best performance (accuracy) by 8.4% (78.4% $\rightarrow$ 86.8%) for Math23k and 6.2% (70.5% $\rightarrow$ 76.7%) for Ape210K.