Detecting Arabic Misinformation Using an Attention Mechanism-Based Model

Abstract

The proliferation of fake news or misinformation, commonly referred to as fake news, has a significant effect on a global scale, as it is aimed at influencing public opinion as well as crowd decision-making. It is therefore crucial to verify the truthfulness of news before it is released to the public. Today, most studies on early detection of Arabic misinformation rely on machine learning methods and transformer-based models. Therefore, in the current study, we used deep learning techniques to propose a model for detecting Arabic misinformation by leveraging the contextual features of news article content. The proposed model was built based on BiLSTM and the attention mechanism. To extract features from Arabic text, we utilized a pre-trained AraBERT model, which generates contextual embeddings from text, then are fed to the BiLSTM layer as input features. Moreover, we investigated the effectiveness of the attention mechanism in improving the overall performance of the model by configuring model architecture to exclude the attention mechanism and comparing the results. Two datasets were utilized to train and evaluate the proposed model, namely, the AraNews and ArCovid19-Rumors datasets. Experimental results showed that the proposed model outperformed other existing models, achieving an accuracy of 0.96 on the ArCovid19-Rumors dataset and 0.90 on the AraNews dataset. This remarkable performance was due to the capability of the attention mechanism to enhance the overall performance, allowing the model to capture the relationship between textual features.