Enhancing Multilingual Fake News Detection Through LLM-Based Data Augmentation

The rapid growth of online news consumption has intensified the spread of misinformation, underscoring the critical need for effective fake news detection methods. Despite significant advancements in this area, the scarcity and inadequacy of high-quality labeled datasets necessary for training effective detection models remains a major challenge. In this paper, we introduce a novel approach to address this issue by leveraging large language models (LLMs) for data augmentation. Specifically, we employ Llama 3 to generate multiple synthetic news samples per original article, enriching existing fake news datasets to enhance fake news detection. We explore various augmentation strategies like different augmentation rates, random or similarity-based subsampling, and selectively augmenting data from specific classes to optimize the augmented datasets to train better classifiers. We evaluate the efficacy of our approach using BERT-based classifiers on two multilingual datasets. Our findings reveal notable improvements particularly when augmenting only the fake class with rate 1.