In this paper, we present our system for SemEval-2026 Task 9, which focuses on the fine-grained identification of polarization manifestations in multilingual social media content. Our approach combines transformer-based encoders (RoBERTa-base for English and Afro-XLM-R-small for Hausa) within a One-vs-Rest (OvR) framework, complemented by controlled oversampling, Adaptive Focal Loss, and label-wise threshold optimization. To mitigate severe class imbalance and label sparsity, we adopt language-specific optimization strategies supported by pairwise χ2 independence analysis. Our system achieves macro-F1 scores of 0.464 in English and 0.192 in Hausa on the official test sets, ranking 5th in Hausa and 14th in English on the official leaderboard.