A Math Word Problem Solver Based on Deep Adaptive Graph Convolutional Networks

Among the tasks of natural language processing, the solution of math word problems is an important research topic. In this paper, we propose a new math word problem solver (DGT). Our solver uses the Seq2Seq model, which consists of two parts: an encoder with joint BiLSTM and DAGCN, and a tree-based decoder. We design four relationship graphs about quantity nodes and word nodes as the input of DAGCN in the model, namely, the relationship graph of adjacent word nodes, the relationship graph of quantity nodes and surrounding word nodes, the relationship graph of quantity node sizes, and the relationship graph of the same surrounding word nodes, which can better help the model understand the relationship between quantities and thus improve the accuracy of math word problem solving.