This paper presents the systems developed for SemEval-2026 Task 9, which targets the detection and categorization of multilingual, multicultural, and multi-event online polarization across 22 languages. To address the challenges posed by linguistic diversity and short, heterogeneous texts, we evaluate several Transformer-based architectures for multilingual polarization detection. Our approach models the task as a multi-label classification problem and incorporates mean pooling for sentence representation, focal loss to mitigate severe label imbalance, and label-wise attention mechanisms to capture polarization-specific linguistic cues. Experimental results show that combining robust multilingual encoders with label-aware modelling substantially improves the detection of polarized content across diverse communities and events