Graph-Guided Multi-view Text Classification: Advanced Solutions for Fast Inference

Recently, traditional large-scale pre-trained models such as BERT perform well in text classification tasks. However, these models have a large number of parameters and high memory requirements, making it difficult to implement in some real-time scenarios or limited resources. Therefore, researchers attempt to use lightweight Graph Neural Networks(GNN) with excellent feature expression as an alternative solution. However, current GNN-based methods solely focus on the structure information of texts, but ignore the sequence information and long-distance dependency relationships between nodes. To solve the above problem, we propose a lightweight network structure G2TX based on multi view feature fusion, which can achieve a balance between model performance and parameters. First, to address the challenge of unordered nodes in graph structure, we introduce a Multi Sequence Fusion Module (MSF) to enhance node sequence information. It integrates features from multiple views through diverse strategies for both word-level and text-level fusion. Secondly, to expand the receptive field of nodes, we propose a Remote Feature Extraction Module (RFE) to bridge the difficult interaction gap between word nodes and remote nodes. Finally, we use KL divergence to integrate the features of both MSF and RFE. The experimental results demonstrate that our model achieves state-of-the-art performance under smaller parameter settings and fast inference conditions.