Taien at SemEval-2026 Task 9: Multilingual Polarization Detection Using Transformer-based Models

AbstractThis submission describes a multilingual polarization detection system for SemEval-2026 Task 9. The system leverages parallel fine-tuning of XLM-RoBERTa and mDeBERTa-v3 transformer models with a probability-level ensemble to improve prediction reliability. We employ language-independent preprocessing, subword tokenization, and a standardized classification head for all 22 languages to ensure a consistent modeling framework across the multilingual setting. Experimental results demonstrate strong performance on both high-resource and low-resource languages, highlighting the effectiveness of the ensemble approach in stabilizing predictions and improving multilingual polarization detection.