Tralaleros at SemEval-2026 Task 9: Multilingual Polarization Detection with Transformer-based Models
Year: 2026
Type: conference-paper
Abstract: Online polarization, the sharp division of public opinion into opposing and often hostile groups, has become a multilingual phenomenon, but most existing corpora and detectors are restricted to a single language or political event.SemEval 2026 Task 9 addresses this gap by releasing a 22-language polarization corpus drawn from political discourse on social media.This paper describes our submission to Subtask 1, the binary polarization-detection track.We systematically investigate four training strategies for transformer-based classifiers, namely data rebalancing through undersampling and back-translation, hyperparameter optimisation with focal loss, model scaling from mBERT to larger multilingual encoders, and ensemble aggregation across architectures.A perlanguage analysis shows that no single model dominates and that the gain from ensembling is small once individual models are well tuned.Our final weighted-average ensemble reaches 0.7936 Macro F1 on the official leaderboard, with top-3 placements on Hindi and Telugu.All four fine-tuned models are released on Hug-gingFace.
Authors: Adrian Dahl, Bado Völckers, Adam Mierzwa
Institutions: Christian-Albrechts-Universität zu Kiel, Universität Hamburg
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7166866634