.jpeg)
In August 2025, Iknoor Singh, Carolina Scarton, and Kalina Bontcheva from the University of Sheffield published a paper titled “GateNLP at SemEval-2025 Task 10: Hierarchical Three-Step Prompting for Multilingual Narrative Classification.” The article presents a new approach for automatically classifying news articles into structured narratives and sub-narratives, a task critical for fact-checkers, policy makers, and researchers monitoring online misinformation.
The authors describe their Hierarchical Three-Step Prompting (H3Prompt) method, which uses a three-step framework: first, categorizing articles into broad domains such as the Ukraine-Russia war or climate change; second, identifying the most relevant main narratives; and third, assigning fine-grained sub-narratives. Their approach relies on fine-tuning a LLaMA 3.2 language model with both annotated training data and synthetically generated news articles, which enhances performance across multiple languages including English, Portuguese, Russian, Bulgarian, and Hindi.
The paper also summarises experimental results, showing that H3Prompt outperforms baseline machine learning and zero-shot models, achieving first place on the English test set and high rankings in several other languages. The authors additionally highlight the benefits of ensemble methods and synthetic data in boosting classification accuracy and handling complex multilingual content.
Link to Zenodo: https://zenodo.org/records/17723009