# PSK at SemEval-2026 Task 9: Multilingual Polarization Detection Using Ensemble Gemma Models with Synthetic Data Augmentation Source: [PSK at SemEval-2026 Task 9: Multilingual Polarization Detection Using Ensemble Gemma Models with Synthetic Data Augmentation]() Collection index: [llms.txt]() Data updated: 2026-05-08T05:00:14.135Z Authors: Srikar Kashyap Pulipaka Published: 2026-05-06T17:29:14.000Z ## Abstract 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. ## Recorded benchmark results A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage. - [Subtask 1 on POLAR amh (test)](): Subtask 1; POLAR amh (test); Score: 0.8. - [Subtask 1 on POLAR hin (test)](): Subtask 1; POLAR hin (test); Score: 0.824. - [Subtask 1 on POLAR swa (test)](): Subtask 1; POLAR swa (test); Score: 0.811. - [Subtask 1 on POLAR arb (test)](): Subtask 1; POLAR arb (test); Score: 84.8. - [Subtask 1 on POLAR ben (test)](): Subtask 1; POLAR ben (test); Score: 83.7. - [Subtask 1 on POLAR deu (test)](): Subtask 1; POLAR deu (test); Score: 72.8. - [Subtask 1 on POLAR eng (test)](): Subtask 1; POLAR eng (test); Score: 81.8. - [Subtask 1 on POLAR fas (test)](): Subtask 1; POLAR fas (test); Score: 82.8. - [Subtask 1 on POLAR hau (test)](): Subtask 1; POLAR hau (test); Score: 80. - [Subtask 1 on POLAR ita (test)](): Subtask 1; POLAR ita (test); Score: 56.3. - [Subtask 1 on POLAR khm (test)](): Subtask 1; POLAR khm (test); Score: 74.3. - [Subtask 1 on POLAR mya (test)](): Subtask 1; POLAR mya (test); Score: 87.7. - [Subtask 1 on POLAR nep (test)](): Subtask 1; POLAR nep (test); Score: 90.8. - [Subtask 1 on POLAR ori (test)](): Subtask 1; POLAR ori (test); Score: 81.1. - [Subtask 1 on POLAR pan (test)](): Subtask 1; POLAR pan (test); Score: 81.2. - [Subtask 1 on POLAR pol (test)](): Subtask 1; POLAR pol (test); Score: 83.5. - [Subtask 1 on POLAR rus (test)](): Subtask 1; POLAR rus (test); Score: 80.7. - [Subtask 1 on POLAR spa (test)](): Subtask 1; POLAR spa (test); Score: 77.9. - [Subtask 1 on POLAR tel (test)](): Subtask 1; POLAR tel (test); Score: 89.3. - [Subtask 1 on POLAR tur (test)](): Subtask 1; POLAR tur (test); Score: 0.809. - [Subtask 1 on POLAR urd (test)](): Subtask 1; POLAR urd (test); Score: 0.803. - [Subtask 1 on POLAR zho (test)](): Subtask 1; POLAR zho (test); Score: 91.9. - [Subtask 1 on POLAR Average (test)](): Subtask 1; POLAR Average (test); Score: 81.1. - [Polarization Detection on SemEval Task 9 2026 (test)](): Polarization detection; SemEval Task 9 2026 (test); Macro-F1: 81.2. ## Original sources and implementation [Original paper]() [PDF]()