PolDeck at SemEval-2026 Task 9: Multilingual Online Polarization Detection via Hybrid Model Ensembling and Data Augmentation

AbstractIn this paper, we address SemEval 2026 Task 9: Multilingual Online Polarization Detection. We present our hybrid ensemble framework, integrating few-shot prompting with Qwen3-30B, a native multilingual XLM-R encoder, and a translation-augmented DeBERTa encoder. To mitigate label imbalance, we implement a multi-stage augmentation pipeline leveraging LLMs for synthetic paraphrasing and cross-lingual translation. Our system ranked in the Top 10 on the English and German leaderboards, proving that integrating both high-capacity monolingual models and flexible multilingual models in a holistic system is a viable method for detecting online polarization. Our code is available on GitHub.