Document Classification with Hierarchical Graph Neural Networks

Adrien Guille, Hugo Attali

Keywords: Document Classification, Graph Neural Networks

Abstract: Various neural architectures have been explored for document classification, such as convolutional and recurrent networks or as of late, Transformers. In parallel, graph neural networks have vastly improved over the recent years. In this paper, we present preliminary results obtain with a novel, parameter-efficient, graph neural network that operates on documents encoded as directed graphs. These graphs describe document-wise word co-occurrence as well as the composition of sentences and thus make the neural network learn word, sentence and document representations in a hierarchical manner. Experiments conducted on various datasets show that it outperforms recent CNNs, RNNs and GNNs designed for document classification.