PSK at SemEval-2026 Task 9: Multilingual Polarization Detection Using Ensemble Gemma Models with Synthetic Data Augmentation
📰 ArXiv cs.AI
Learn how to detect multilingual polarization using ensemble Gemma models with synthetic data augmentation for improved performance
Action Steps
- Fine-tune separate Gemma models for each language using Low-Rank Adaptation (LoRA)
- Generate synthetic data using a large language model (LLM) such as GPT-4
- Apply three synthetic data strategies: direct generation, paraphrasing, and contrastive pair creation
- Ensemble the fine-tuned Gemma models to improve overall performance
- Evaluate the system on a multilingual polarization detection task
Who Needs to Know This
NLP engineers and researchers can benefit from this approach to improve the accuracy of polarization detection in multiple languages, while data scientists can apply the synthetic data augmentation strategies to other classification tasks
Key Insight
💡 Synthetic data augmentation using LLMs can improve the performance of multilingual polarization detection models
Share This
Boost polarization detection accuracy with ensemble Gemma models & synthetic data augmentation! #NLP #Multilingual
Key Takeaways
Learn how to detect multilingual polarization using ensemble Gemma models with synthetic data augmentation for improved performance
Full Article
Title: PSK at SemEval-2026 Task 9: Multilingual Polarization Detection Using Ensemble Gemma Models with Synthetic Data Augmentation
Abstract:
arXiv:2605.05159v1 Announce Type: cross 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-m
Abstract:
arXiv:2605.05159v1 Announce Type: cross 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-m
DeepCamp AI