Text Classification Model Based on Graph Attention Networks and Adversarial Training

Abstract

Text information on the internet often has a strong sense of immediacy, constantly reflecting societal dynamics and evolving events. This is especially crucial in the field of news text, where the classification and analysis of these immediate and varied text data become essential. Existing text classification models frequently struggle to effectively represent the semantic and local feature information of texts, limiting their effectiveness. The primary challenge lies in improving the representation of both semantic and local feature information in text classification models, which is critical for capturing the nuanced meanings in rapidly evolving news texts. This paper proposes a deep learning-driven framework designed to enhance the effectiveness of text classification models. The method incorporates noise perturbation during training for adversarial training, thereby enhancing the model’s generalization ability on original samples and increasing its robustness. A graph attention network is employed to extract the contextual semantic information of vocabulary from sequential texts. This information is then combined with extracted sentence feature information to enrich the feature representation of the sequence. An attention mechanism is also introduced to extract more critical feature information from the text, thereby deepening the understanding of textual semantic information. Experimental results demonstrate that this method successfully integrates the boundary and semantic information of vocabulary into the classification task. The approach comprehensively and deeply mines the semantic features of the text, leading to improved classification performance.