ABARUAH at SemEval-2026 Task 9: Multilingual Polarization Detection across Seven Indic Languages using Qwen3

AbstractOnline polarization creates division within the society. As such, it is important to detect and remove polarized messages from social media. This study presents fine-tuned Qwen3-8B Large Language Model (LLM) based models to identify online polarization, its specific categories, and its manifestation types. This study used Quantized Low-Rank Adaptation (QLoRA) to fine-tune the model in seven Indic languages: Bengali, Hindi, Nepali, Oriya, Punjabi, Telugu, and Urdu. The experimental results demonstrate the efficacy of this approach, achieving macro F1-scores of 0.82, 0.78, 0.90, 0.76, 0.78, 0.87, and 0.79, respectively, for polarization detection. The proposed model surpassed the established baseline systems in several of the subtasks, suggesting that parameter-efficient fine-tuning is a viable and powerful strategy for addressing linguistic diversity in low-resource and high-variability Indic language datasets. To leverage cross-lingual transfer, a unified model was developed by fine-tuning on a concatenated dataset of seven Indic languages. This approach proved superior to standalone language-specific models, yielding substantial improvements in F1-score (most notably a 28.76 point gain in Subtask 2 for Punjabi language). This provides strong evidence for the benefits of cross-lingual knowledge transfer in low-resource settings.