Projects |

This project focuses on multilingual, hierarchical text classification for narrative analysis, designed to categorize news articles into high-level narratives and sub-narratives across multiple languages. Using RoBERTa for English and XML-RoBERTa for Portuguese, the system applies multi-label, multi-class classification to detect and analyze propaganda, misinformation, and media framing. By implementing data balancing techniques such as oversampling rare classes and threshold optimization, the model significantly improves multi-label detection accuracy. The pipeline integrates rule-based filtering, transformer models, and fine-tuned classification strategies, ensuring robust and scalable AI-driven analysis. This work achieved 10th place worldwide in a competitive Samevel Challenge , with F1 scores of 0.70 for Climate Change sub-narratives, 0.69 for high-level classification, and 0.64 for Ukraine-Russia War sub-narratives.