AI-Driven Fake News Detection: Applying Transformers /Large Language Models

Abstract

The proliferation of fake news across digital platforms has emerged as a major concern, particularly when misinformation is expressed in low-resource dialects such as Algerian Arabic. The Algerian dialect poses specific challenges for automatic detection due to its rich linguistic variability, lack of standardized orthography, and limited availability of natural language processing (NLP) tools tailored to it. This thesis explores the problem of fake news detection in the Algerian dialect, not only in social media content but also in other digital sources such as online articles and news websites. We investigate the effectiveness of transformer-based models and large language models (LLMs) for this task, conducting extensive experiments using a variety of pre-trained models including AraGPT2, DziriBERT, CaMelBERT, LLaMa3, Qwen and others. The experimental findings confirm that transformer and LLM-based approaches provide promising performance for fake news detection in under-resourced languages, with AraGPT2 achieving the highest accuracy in our evaluation benchmarks. Additionally, we propose a hybrid strategy that combines transformer-based embeddings with LLM inference to better capture the contextual and linguistic nuances of the dialect.