Spam detection using sentence-BERT with attention-enhanced GRU and SVM
Leila Boussaad · Applied Computing and Informatics · 2025
Purpose This paper introduces a cutting-edge framework for spam detection that integrates Sentence-BERT embeddings, an attention-augmented Gated Recurrent Unit (GRU) and a Support Vector Machine (SVM) for refined classification. Design/methodology/approach Our approach involves generating richly contextualized embeddings via Sentence-BERT. These embeddings are processed by a GRU network enhanced with a self-attention mechanism to effectively capture intricate long-range dependencies and salient features. The final classification is performed by an SVM, leveraging its robust capacity for binary decision-making between spam and non-spam categories. Additionally, data augmentation is employed using the T5 model to generate paraphrased spam instances, enhancing dataset diversity and model robustness. The evaluation framework includes precision, recall, and F-measure metrics for comprehensive validation. Findings The proposed model demonstrates exceptional performance, achieving a notable accuracy of 99.88%, with a 95% confidence interval of [99.70%, 99.99%], highlighting its strong effectiveness in spam detection. Originality/value The integration of Sentence-BERT, attention-augmented GRU, and SVM presents a novel combination for spam detection, effectively capturing intricate dependencies and enhancing classification accuracy. The use of T5-based data augmentation further fortifies the model's robustness, offering valuable insights into advanced spam detection methodologies.