A Heterogeneous Directed Graph Attention Network for inductive text classification using multilevel semantic embeddings

Abstract

In the current study, a novel network model is proposed for text classification based on Graph Attention Networks (GATs) and sentence-transformer embeddings. Most existing methods with a pretraining model as an input layer still treat words as the minimum processing unit. However, word embedding is not an efficient and appropriate solution when dealing with long texts containing many professional words. This study aims to design a model capable of handling text classification tasks at multilevel semantic segmentation. The main contribution of this study is that a novel GAT variant is designed using global nodes and Squeeze-and-Excitation Networks (SENet) to capture semantic information. Moreover, a novel unidirectional attention mechanism is introduced for our model to avoid the message passing of irrelevant noisy information within global nodes. The numerical results show that according to the characteristics of datasets, specific semantic information combinations can effectively improve the accuracy and performance of text classification. Without fine-tuning of the pretrained encoder, the new state-of-the-art performance is achieved on three benchmark datasets. In addition, a comprehensive analysis of the graph attention mechanism in the model for specific cases suggests that the unidirectional attention mechanism and the use of global nodes are key contributing factors to multilevel semantic fusion.