부산대학교 도서관

In this research, we tackled the critical challenge of detecting fake news in a bilingual context, focusing on English and Urdu. This issue is particularly important in the digital age, where misinformation can impact society and politics. To address this, we developed “BiL-FaND,” an ensemble-based system integrating multiple models designed to analyze distinct aspects of news content. The system employed Multilingual BERT for textual analysis, LSTM models for categorical and numerical data processing, and a caption-generating model using ResNet-101 and GRU for multimedia content analysis. This diverse methodological approach was key in handling the complexity and nuances of multilingual fake news. Our findings were significant, demonstrating the effectiveness of our approach. The textual analysis layer achieved 87% accuracy, indicating strong linguistic analysis performance. The LSTM models for categorical and numerical data showed excellent accuracies of 97% and 94%, respectively, highlighting their capability in pattern recognition and data interpretation. The multimedia layer further augmented our system’s accuracy, as evidenced by BLEU scores of 0.82 for English and 0.75 for Urdu captions, with an accuracy of 92%, calculated based on cosine similarity. These results concluded that our multi-layered, ensemble approach was highly effective in the bilingual fake news detection domain and achieved an overall accuracy of 92.07%.