Abstract
Deep Learning (DL) has been one of the preferred techniques for Natural Language Processing (NLP) applications. Due to its nature, a text can be better represented in a graph structure, when compared with the classical feature-based representations. Therefore, several researchers have explored the use of Graph Neural Networks (GNN) for text analysis. GNNs show excellent results in text classification tasks, given their property of capturing contextual and global information in a corpus. The Text Graph Convolutional Network (TGCN) showed the ability to outperform traditional NLP methods in benchmark classification tasks. However, this method has a very high memory cost for the text graph construction. By exploring the results of text representations, we propose a new method to generate a text graph, capable of influencing the result of the TGCN, leading to a reduced use of memory.