Arabic Fake News Detection in Social Media Context Using Word Embeddings and Pre-trained Transformers

The quick spread of fake news in different languages on social platforms has become a global scourge threatening societal security and the government. Fake news is usually written to deceive readers and convince them that this false information is correct; therefore, stopping the spread of this false information becomes a priority of governments and societies. Building fake news detection models for the Arabic language comes with its own set of challenges and limitations. Some of the main limitations include 1) lack of annotated data, 2) dialectal variations where each dialect can vary significantly in terms of vocabulary, grammar, and syntax, 3) morphological complexity with complex word formations and root-and-pattern morphology, 4) semantic ambiguity that make models fail to accurately discern the intent and context of a given piece of information, 5) cultural context and 6) diacrasy. The objective of this paper is twofold: first, we design a large corpus of annotated fake new data for the Arabic language from multiple sources. The corpus is collected from multiple sources to include different dialects and cultures. Second, we build fake detection by building machine learning models as model head over the fine-tuned large language models. These large language models were trained on Arabic language, such as ARBERT, AraBERT, CAMeLBERT, and the popular word embedding technique AraVec. The results showed that the text representations produced by the CAMeLBERT transformer are the most accurate because all models have outstanding evaluation results. We found that using the built deep learning classifiers with the transformer is generally better than classical machine learning classifiers. Finally, we could not find a stable conclusion concerning which model works well with each text representation method because each evaluation measure has a different favored model.