Study on the Effectiveness of Multilingual Models in Fake News Classification

This study aims to delve into the application of BERT-based foundational models and currently popular large-scale models in the field of fake news detection. By comprehensively evaluating the classification performance of these two approaches across different datasets, we aim to provide an useful reference for both research and practical applications in fake news detection. The results indicate that on the smaller-scale Dataset 1, six popular large models, including GPT-2, QWEN, and RENIE, perform better, with the best average F1 score reaching 95.4% and ROC achieving 95%. On the larger-scale Dataset 2, BERT-based models and their seven variants show superior classification performance, with the best average F1 score reaching 90.4% and ROC achieving 78.5%. Although large-scale models may outperform in certain scenarios, BERT-based foundational models offer advantages in terms of resource efficiency and are more effective in training and deployment. Therefore, our study provides researchers with valuable practical reference when choosing models.