Solving Math Word Problems by Dual Feature Fusion Graph Neural Network

In recent years, graph neural networks have been widely used in solving math word problems. However, previous researches are still not deep enough to mine numerical values in text. Therefore, we propose a dual feature fusion graph neural network model. In terms of graph construction, we proposes two new numerical relationship graphs: one is to classify the numeric values in the text, and the other is to compare the numeric values in the text with the numeric values in the answer. The graph features are obtained by multi-head convolution and common graph convolution modules. BERT model is used to encode text at sentence level, and it is fused with graph features through attention mechanism to guide the model to pay attention to global information. Experiments on two well-known datasets show that both of our proposed algorithms have advantages over other state-of-the-art competitors.