Scholarlink

<p>In digital communication, persistent spam poses significant risks to user privacy and security. Classifying short texts (e.g., SMS, tweets) is challenging due to their brevity, ambiguity, and lack of contextual depth. To address this, we propose a novel dual-attention-based hybrid deep learning framework that synergistically integrates Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory Networks (BiLSTM). The CNN branch employs location-based attention to highlight spatially significant features. In contrast, the BiLSTM branch uses co-attention to model sequential relationships, enabling comprehensive feature extraction from sparse, noisy short texts. The fused features are classified through an optimized, fully connected layer, achieving state-of-the-art accuracy of 99.74% (SMS) and 99.36% (Twitter) while maintaining computational efficiency (12ms latency, 8.5G FLOPs). Our model’s parallel architecture and attention mechanisms uniquely balance accuracy and deployability, outperforming transformers (e.g., BERT) with 10 × fewer computational resources. This work advances spam detection by addressing feature sparsity and real-time processing constraints, offering a robust solution for practical applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Dual attention-based hybrid deep learning framework for short text classification

  • 期刊论文

  • 发表日期：

    2025年8月12日

查看全文
  • Hussein Ala’a Alkaabi,
  • Russul Hazim Abed,
  • Fuqdan Al-Ibraheemi

摘要

In digital communication, persistent spam poses significant risks to user privacy and security. Classifying short texts (e.g., SMS, tweets) is challenging due to their brevity, ambiguity, and lack of contextual depth. To address this, we propose a novel dual-attention-based hybrid deep learning framework that synergistically integrates Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory Networks (BiLSTM). The CNN branch employs location-based attention to highlight spatially significant features. In contrast, the BiLSTM branch uses co-attention to model sequential relationships, enabling comprehensive feature extraction from sparse, noisy short texts. The fused features are classified through an optimized, fully connected layer, achieving state-of-the-art accuracy of 99.74% (SMS) and 99.36% (Twitter) while maintaining computational efficiency (12ms latency, 8.5G FLOPs). Our model’s parallel architecture and attention mechanisms uniquely balance accuracy and deployability, outperforming transformers (e.g., BERT) with 10 × fewer computational resources. This work advances spam detection by addressing feature sparsity and real-time processing constraints, offering a robust solution for practical applications.