SIGAN: Self-inhibited Graph Attention Network for Text Classification

Text classification is a pivotal task within the field of natural language processing (NLP) that has many practical applications. However, there are several challenges of existing methods of Graph Convolutional Networks, such as difficulties in capturing long-range dependencies and overfitting issues. To address these challenges, we have proposed a novel model called Self-Inhibited Graph Attention Network (abbreviated as SIGAN), which builds upon the foundation of TextGCN but introduces two key contributions: Graph Attention Networks (GATs) and Self-Inhibition Structure (SIS). GATs enable adaptive weight allocation to nodes, facilitating a better understanding of complex relationships within graph data, including long-distance dependencies; SIS replaces the Dropout operation used in TextGCN and preserves the overfitting prevention properties of Dropout while avoiding the loss of crucial features by considering the importance levels of individual neurons. Our model is evaluate on four large-scale datasets (R52, MR, R8, Ohsumed), and we compare our methods with several baselines such as TextGCN and TWPGCN. Our results show that SIGAN achieves better improvements in accuracy, with gains of 0.93%, 1.27%, 0.88%, and 0.39%, respectively. Experimental results show the effectiveness of the model enhancements and the structure of self-inhibition.