Text Classification Method Based on Gating Mechanism and Graph Convolutional Networks

Graph convolutional networks (GCN), in contrast to traditional neural networks, demonstrate advantages in handling non-Euclidean text classification problems by accommodating nodes and graphs with arbitrary topological structures. However, existing text classification models based on GCN often aggregate neighboring node information with equal or preset edge weights, neglecting effective consideration of local interactions between words. This limitation leads to insufficient expression of semantic information for nodes. Therefore, this paper proposes a text classification method based on the gating mechanism and graph convolutional networks. In this approach, termed graph convolutional networks based on the gating mechanism (GM-GCN), the gating mechanism is integrated with the two layers of GCN in TextGCN. GM-GCN utilizes a select matrix with gating functionality to control information transmission, allowing for the selective fusion of neighborhood information at multiple orders for nodes in the graph. This approach preserves prior orders' information, enabling a more effective focus on local interactions between words. Consequently, it enriches nodes' feature representation in capturing textual semantics. The experimental results demonstrate significant improvements in accuracy, precision, recall, GPU memory consumption, and runtime when compared with other classical and latest text classification models across four benchmark datasets: 20NG, R8, R52, and Ohsumed. For accuracy, this proposed method achieved notable enhancements of 0.34%, 0.27%, 0.32%, and 0.25%, respectively. Regarding precision, improvements were observed at 0.33%, 0.25%, 0.29%, and 0.24%, while recall showed enhancements of 0.28%, 0.31%, 0.26%, and 0.22%, respectively. In GPU memory consumption, the method demonstrated reductions of 137MB, 73MB, 86MB, and 88MB in space, and in terms of runtime, it exhibited decreases of 26s, 9s, 13s, and 16s, respectively. As a result of these findings, the proposed approach was found to be effective in improving text classification performance.