We present our system for SemEval-2026 Task 9: Multilingual Polarization Detection, a binary classification task spanning 22 languages. Our approach fine-tunes separate Gemma 3 models (12B and 27B parameters) per language using Low-Rank Adaptation (LoRA), augmented with synthetic data generated by a large language model (LLM). We employ three synthetic data strategies (direct generation, paraphrasing, and contrastive pair creation) using GPT-4o-mini, with a multi-stage quality filtering pipeline including embedding-based deduplication. We find that per-language threshold tuning on the development set yields 2 to 4% F1 improvements without retraining. We also use weighted ensembles of 12B and 27B model predictions with per-language strategy selection. Our final system achieves a mean macro-F1 of 0.811 across all 22 languages, ranking 2nd overall of the participating teams, with 1st place finishes in 3 languages and top-3 in 8 languages. We also find that alternative architectures (XLM-RoBERTa, Qwen3) that showed strong development set performance suffered 30 to 50% F1 drops on the test set, highlighting the importance of generalization.
Key points
- Task: SemEval-2026 Task 9 — multilingual polarization detection, binary classification across 22 languages.
- Models: Per-language fine-tuned Gemma 3 (12B and 27B) using LoRA.
- Synthetic data: Three strategies via GPT-4o-mini — direct generation, paraphrasing, and contrastive pair creation — with multi-stage quality filtering including embedding-based deduplication.
- Threshold tuning: Per-language threshold optimization on the dev set yields 2–4% F1 gains without retraining.
- Ensembling: Weighted ensemble of 12B and 27B predictions with per-language strategy selection.
- Results: Mean macro-F1 of 0.811 across 22 languages; 2nd place overall, 1st place in 3 languages, top-3 in 8 languages.
- Generalization finding: XLM-RoBERTa and Qwen3 baselines with strong dev performance dropped 30–50% F1 on the test set.