Multilingual Fake News Detection using Transformer Models

The spread of fake news across linguistic boundaries poses a global threat, yet most automated detection systems are designed for a single language, leaving low resource languages highly vulnerable. This paper presents a multilingual fake news detection framework built on a transformer based cross lingual model. The system employs XLM RoBERTa, a pre trained multilingual transformer, fine tuned on a composite corpus of 48,000 news articles spanning English, Hindi, Bengali, and Tamil. Each article is labelled as real or fake, with careful annotation and balanced class distribution. The model processes raw text through a shared tokenizer and transformer encoder, generating a language agnostic contextual representation that is fed into a binary classification head. The proposed approach achieves an average accuracy of 92.4% across all four languages, outperforming monolingual BERT baselines by 6–11 percentage points on low resource languages. Cross lingual zero shot transfer experiments demonstrate that a model trained only on English data can detect fake news in Hindi with 78.2% accuracy, highlighting strong cross lingual generalisation. The framework is lightweight enough for inference on a single GPU and can be extended to new languages without retraining. These results indicate that multilingual transformers offer a scalable, effective solution for combating fake news in linguistically diverse environments.