Chinese Math Word Problems Generation Network

Abstract

Aiming at the feature vector bottleneck problem and the high time cost of the training process in the automatic generation of Chinese math word problems under the end-to-end architecture, we proposed an automatic generation method of Chinese math word problems based on the pre-training model combined with the integration of encoder and decoder. We used a deep neural network to model the mathematical equation sequence and Chinese keyword information, and used the stepped attention matrix to generate word problems. For training and testing on the Ape210K data set, compared with the end-to-end method, the Rouge-1 and Rouge-L evaluation indicators in our method was increased by 14.1% and 12.5%, as well as the training time cost was reduced by nearly 50%.