We describe our system for SemEval-2026 Task9 (POLAR), Subtask 1 - binary polarizationdetection. Our approach investigates polariza-tion detection through monolingual and cross-lingual experimental settings. We first utilizea RoBERTa-based architecture enhanced withfeature fusion, combining contextual sentencerepresentations with handcrafted sentiment andintensity cues. As for multilingual joint train-ing, we explore it within the Indo-Europeanfamily to test whether cross-lingual transfer canelevate performance in data-scarce scenarios.Our final fine-tuned model achieves averageF1-score of 0.763 on the test set, compared to0.491 for a random baseline. We also reportablations for augmentation, feature fusion, andclass weighting to quantify each component’scontribution.