Hostility Detection in UK Politics: A Dataset on Online Abuse Targeting MPs

AbstractSocial media platforms, particularly X, enable direct interaction between politicians and constituents but also expose politicians to hostile responses targetting both their governmental role and personal identity. This online hostility can undermine public trust and potentially incite offline violence. While general hostility detection models exist, they lack the specificity needed for political contexts and country-specific issues. We address this gap by creating a dataset of 3,320 English tweets directed at UK Members of Parliament (MPs) over two years, annotated for hostility and targeted identity characteristics (race, gender, religion). Through linguistic and topical analyses, we examine the unique features of UK political discourse and evaluate pre-trained language models and large language models on binary hostility detection and multi-class targeted identity type classification tasks. Our work provides essential data and insights for studying politics-related hostility in the UK.