Ensemble Gemma Models Achieve Top-2 in Multilingual Polarization Detection

A team of researchers has created a system for SemEval-2026 Task 9: Multilingual Polarization Detection, which involves binary classification in 22 languages. Their methodology involves fine-tuning individual Gemma 3 models (with 12B and 27B parameters) for each language through Low-Rank Adaptation (LoRA), enhanced by synthetic data generated from GPT-4o-mini using three methods: direct generation, paraphrasing, and creating contrastive pairs. They implemented a multi-stage quality filtering process that included deduplication based on embeddings. By adjusting thresholds for each language on the development set, they achieved F1 improvements of 2–4% without retraining. The weighted ensembles of predictions from the 12B and 27B models, along with language-specific strategy selection, resulted in a mean macro-F1 score of 0.811, placing them 2nd overall, with top rankings in 3 languages and among the top 3 in 8 languages.