Verifying Urdu news authenticity using deep learning with concatenated BERT and GloVe embedding

TL;DR

This paper presents a deep learning model using BERT and GloVe embeddings to detect fake news in Urdu, achieving high accuracy.

Contribution

A novel Urdu fake news detection system using concatenated BERT and GloVe embeddings with XLM-RoBERTa.

Findings

XLM-RoBERTa with concatenated GloVe achieved an F1 score of 0.956 and accuracy of 0.962.

The model outperformed other deep learning and machine learning approaches for Urdu fake news detection.

Abstract

Fake news has a significant effect on reader perceptions and is therefore a serious problem. It is difficult to differentiate between fake and real news as the number of News platforms on social media are growing daily. This work aims to create a complete fake news detection mechanism for Pakistani news by using many fact-checked APIs. A total of 14,178 real and fake news items from fifteen different regions and sections of reputable and authoritative Urdu newspapers and news broadcasting websites in Pakistan are included in the dataset. We evaluate the dataset via three deep learning models RoBERTa, XLM-RoBERTa and mBERT are embedded with concatenated BERT and GloVe. For the training of the selected models, we first use an extensive multilingual dataset. The results of the proposed models are as follows: evaluated via performance metrics such as the F1 score, accuracy, precision and recall. XLM-RoBERTa with concatenated GloVe embedding outperforms with an F1 score of 0.956, accuracy of 0.962, precision of 0.932 and recall of 0.940. A comparative analysis of the latest machine learning and deep learning models is also done. The latest Urdu benchmark dataset shows that the XLM-RoBERTa model with Concatenated BERT and GloVe embedding outperforms these models for Urdu fake news detection.

Figures