Text Classification with Attention Gated Graph Neural Network

Zhaoyang Deng, Chenxiang Sun, Guoqiang Zhong, Yuxu Mao

Abstract: Text classification is a fundamental and important task in natural language processing. There have been many graph-based neural networks for this task with the capacity of learning complicated relational information between word nodes. However, existing approaches are potentially insufficient in capturing semantic relationships between the words. In this paper, to address the above issue, we propose a novel graph-based model where every document is represented as a text graph. Specifically, we devise an attention gated graph neural network (AGGNN) to propagate and update the semantic information of each word node from their 1-hop neighbors. Keyword nodes with discriminative semantic information are extracted via our proposed attention-based text pooling layer (TextPool), which also aggregates the document embedding. In this case, text classification is transformed into a graph classification task. Extensive experiments on four benchmark datasets demonstrate that the proposed model outperforms other previous text classification approaches.