Leveraging Arabic BERT for High-Accuracy Fake News Detection

Abstract

Media platforms have become essential for staying informed about events and activities around the globe. While there has been research on identifying news in English, detecting it in Arabic has been relatively overlooked. The unique linguistic characteristics and diverse slang expressions in Arabic have contributed to a scarcity of studies in this area. This research examines the effectiveness of deep learning algorithms in identifying fake news, specifically in the Arabic language. In this study, Global Vectors for Word Representation (GloVe) were used to capture the semantic relationships between words in order to improve the performance of the models. Furthermore, the utilization of Long Short-Term Memory (LSTM) and Bidirectional LSTM (BiLSTM) algorithms has shown potential in the field of neural networks for text classification purposes. The study also delves into incorporating a trained Arabic BERT (Bidirectional Encoded Representations from Transformers) model, which is widely recognized for its outstanding performance, in various natural language processing tasks. The current research utilizes the Arabic Fake News Dataset (AFND). It is a large, fully annotated dataset of Arabic fake news. The results demonstrated that, among all the investigated algorithms, Arabic BERT achieved accuracy with a score of 0.98 on the dataset. According to the results obtained, LSTM and BiLSTM achieved scores of 0.94 and 0.93, respectively, implementing GloVe word embeddings. This research showcases the effectiveness of the Arabic BERT model, alongside the LSTM and BiLSTM models, in detecting information. It highlights the contribution of BERT in enhancing accuracy when dealing with the identification and mitigation of challenges related to identifying news in Arabic-language contexts.