Scholarlink

<p>The phenomenon of disinformative tweet propagation on social networks poses a significant challenge for researchers in the field of natural language processing (NLP) due to its real-world impact on various aspects of society, including health, economics, and politics. This paper addresses the automated identification of disinformative tweets within the context of social networks, specifically focusing on Arabic tweets published during the COVID-19 pandemic. A hybrid deep learning model combining convolutional neural networks (CNN) and bidirectional long short-term memory (BiLSTM) was tested on the ArAIEval dataset. Arabic BERT variants, including AraBERT, MARBERT, and CaMelBERT, were fine-tuned and used as embedding layers. To address the issue of data imbalance, three data augmentation techniques were implemented: (1) synonym replacement, (2) random deletion, and (3) contextualized embedding-based replacement. The hybrid CNN–BiLSTM architecture, including the AraBERT embedding layer, achieved a micro-F1 score of 91.00% with the utilization of a BERT-based data augmentation strategy on the minority class.</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

Fine-Tuning Arabic BERT Models with Hybrid CNN–BiLSTM Architecture for Disinformative Tweet Detection

  • 期刊论文

  • 发表日期：

    2026年1月13日

查看全文
  • Areej Jaber,
  • Tasneem Duridi,
  • Eman Daraghmi,
  • Paloma Martínez

摘要

The phenomenon of disinformative tweet propagation on social networks poses a significant challenge for researchers in the field of natural language processing (NLP) due to its real-world impact on various aspects of society, including health, economics, and politics. This paper addresses the automated identification of disinformative tweets within the context of social networks, specifically focusing on Arabic tweets published during the COVID-19 pandemic. A hybrid deep learning model combining convolutional neural networks (CNN) and bidirectional long short-term memory (BiLSTM) was tested on the ArAIEval dataset. Arabic BERT variants, including AraBERT, MARBERT, and CaMelBERT, were fine-tuned and used as embedding layers. To address the issue of data imbalance, three data augmentation techniques were implemented: (1) synonym replacement, (2) random deletion, and (3) contextualized embedding-based replacement. The hybrid CNN–BiLSTM architecture, including the AraBERT embedding layer, achieved a micro-F1 score of 91.00% with the utilization of a BERT-based data augmentation strategy on the minority class.