Deep Neural Solver for Math Word Problems

Yan Wang,Xiaojiang Liu,Shuming Shi

DOI: https://doi.org/10.18653/v1/d17-1088

2017-01-01

Abstract:This paper presents a deep neural solver to automatically solve math word problems. In contrast to previous statistical learning approaches, we directly translate math word problems to equation templates using a recurrent neural network (RNN) model, without sophisticated feature engineering. We further design a hybrid model that combines the RNN model and a similarity-based retrieval model to achieve additional performance improvement. Experiments conducted on a large dataset show that the RNN model and the hybrid model significantly outperform state-of-the-art statistical learning methods for math word problem solving.