Graph-Encoder and Multi-decoders Solution Framework with Multi-attention

Hui Cai, Tiancheng Zhang, Xianghui Sun, Minghe Yu, Ge Yu

Abstract

The goal of math word problem (MWP) is to design an automatic solution model to answer the mathematical questions given in the text. In recent years, many neural network models have achieved good results. However, most models do not consider the unique language features of mathematical problems in feature extraction, and ignore the structural information of mathematical problems and equations in the design of encoder and decoder. In order to solve these problems, we propose a graph-encoder and multi-decoders solution framework with multi-attention. On the one hand, we introduce the multi-attention mechanism in the encoding part to capture multiple features, and construct graph to express quantitative information. On the other hand, in the decoding part, we combine tree-structured and sequence-structured decoders. The experimental results on the Math23K dataset show that our model outperforms the existing state-of-the-art methods.
BibTeX
@inproceedings{cai2021graph,
  author = {Hui Cai and Tiancheng Zhang and Xianghui Sun and Minghe Yu and Ge Yu},
  title = {{Graph-Encoder and Multi-decoders Solution Framework with Multi-attention}},
  booktitle = {WISA},
  year = {2021},
  doi = {10.1007/978-3-030-87571-8\_58},
}

Loading related content...

Figure