Multi-domain Urdu fake news detection using pre-trained ensemble model.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Computational Linguistics

Background:

• Fake News (FN) dissemination impacts societal and political landscapes.
• Automated mechanisms are needed to counter online propaganda and biased news.
• Detecting authenticity in regional languages like Urdu is challenging due to limited data and research.

Purpose of the Study:

• To address the language bias in AI research by focusing on Urdu Fake News Detection (FND).
• To propose and evaluate a novel stacked ensemble learning methodology for Urdu FND.

Main Methods:

• Individual fine-tuning of pre-trained language models (PLMs): ELECTRA, mBERT, and XLM-RoBERTa.
• Application of stacked ensemble learning combining these fine-tuned PLMs.
• Hyperparameter optimization for enhanced model performance.

Main Results:

• The proposed stacked ensemble model achieved superior prediction performance compared to individual PLMs.
• Key performance metrics include an Accuracy of 0.914, a Matthews Correlation Co-efficient (MCC) of 0.898, and an F1-score of 0.904.
• The results validate the efficacy of the ensemble approach for Urdu FND.

Conclusions:

• The stacked ensemble learning of fine-tuned PLMs offers a robust solution for Urdu FND.
• This methodology effectively overcomes limitations of individual transformer models.
• The study contributes to reducing language bias in AI and improving fake news detection in under-resourced languages.