Transformer-Based Manipulative Span Identification with SHAP Explainability in Multilingual Telegram Data

The proliferation of manipulative content on social media poses significant challenges to information integrity, particularly in sensitive geopolitical contexts. During the Rus-sia-Ukraine war, such narratives have the potential to distort public perception and shape international opinion. To address this issue, a shared task UNLP 2025 was formulated to detect such manipulative spans in Ukrainian and Russian 9,500 Telegram posts. This study leverages the dataset provided by the task organizers and explores a range of Machine Learning (ML), Deep Learning (DL) and Transformer-based approaches to investigate the detection of manipulative spans in text. SHapley Additive ex-Planations (SHAP) are further employed to interpret token-level contributions towards manipulativeness, providing transparency into model predictions. Our experimental results reveal that the RemBERT transformer model achieves…