Year: 2026
Type: conference-paper
Abstract: As online discourse becomes increasingly polarized, the need for automated systems to detect and categorize divisive content grows.In this paper we describe our participation in SemEval-2026 Task 9, addressing Subtask 2: Polarization Type Classification (Naseem et al., 2026a).We use a fine-tuned TwHIN-BERT model, a multilingual encoder-based transformer pre-trained on social media data, to classify text across five polarization categories.Significant label imbalance and label distribution across languages are central challenges to the task.Our system achieves an average macro-score of 0.60 across all nine languages (see 3).Our results suggest that performance in such a multi-label polarization detection task is more strongly driven by local distributional properties than by resource size.
Authors: Ilinca Vandici, Ådne Jøssing, Lukas Viestädt
Institution: University of Tübingen
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7166874347