PSK at SemEval-2026 Task 9: Multilingual Polarization Detection Using Ensemble Gemma Models with Synthetic Data Augmentation

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, medium

Summary

The PSK system for SemEval-2026 Task 9 achieved a 2nd place ranking out of 60 teams in multilingual polarization detection, a binary classification task across 22 languages. The approach involved fine-tuning individual Gemma 3 models (12B and 27B parameters) for each language using Low-Rank Adaptation (LoRA). A crucial component was synthetic data augmentation, generated by GPT-4o-mini through direct generation, paraphrasing, and contrastive pair creation, followed by a multi-stage quality filtering pipeline with embedding-based deduplication. The system further optimized performance by applying per-language threshold tuning on the development set, yielding 2 to 4% F1 improvements, and utilizing weighted ensembles of the 12B and 27B model predictions. This comprehensive strategy resulted in a mean macro-F1 of 0.811, with 1st place finishes in 2 languages and top-3 in 8 languages. The study also highlighted that other architectures like XLM-RoBERTa and Qwen3, despite strong development set performance, experienced significant 30 to 50% F1 drops on the test set, underscoring the importance of generalization.

Key takeaway

For Machine Learning Engineers developing multilingual NLP systems, this work demonstrates that combining LoRA-tuned Gemma models with GPT-4o-mini generated synthetic data and per-language threshold tuning is highly effective. You should prioritize robust generalization over peak development set performance, as alternative models showed significant F1 drops on test data. Consider implementing these strategies to achieve competitive results in cross-lingual classification tasks.

Key insights

Ensembling LoRA-tuned Gemma models with LLM-generated synthetic data and per-language thresholding significantly improves multilingual polarization detection generalization.

Principles

• Per-language tuning enhances F1 by 2-4% without retraining.
• Synthetic data augmentation improves model robustness.
• Generalization is critical; dev set performance can mislead.

Method

Fine-tune Gemma 3 (12B, 27B) per language via LoRA. Augment with GPT-4o-mini synthetic data (direct, paraphrase, contrastive) with quality filtering. Ensemble models and apply per-language threshold tuning.

In practice

• Use LoRA for efficient multilingual model adaptation.
• Generate synthetic data with LLMs like GPT-4o-mini.
• Implement per-language threshold tuning for F1 gains.

Topics

• Multilingual NLP
• Polarization Detection
• Gemma Models
• LoRA Fine-tuning
• Synthetic Data Augmentation
• Ensemble Learning

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• mdok-style at SemEval-2026 Task 9: Finetuning LLMs for Multilingual Polarization Detection — Finetuning mid-sized LLMs with QLoRA and augmented data improves multilingual online polarization detection.
• Efficient Financial Language Understanding via Distillation with Synthetic Data — Distillation with synthetic data, guided by clustering real examples, enables efficient financial NLP in low-resource settings.
• Negation-Aware Data Augmentation for Portuguese Natural Language Inference — Targeted negation-aware data augmentation improves BERT-based NLI model performance on Portuguese datasets by reducing negation bias.
• P3B3: A Multi-Turn Conversational Benchmark for Measuring European and Brazilian Portuguese Variety Bias in LLMs — Most LLMs exhibit a strong bias towards Brazilian Portuguese, highlighting the need for balanced multilingual representation.
• Data Synthesis and Parameter-Efficient Fine-Tuning for Low-Resource NMT: A Case Study on Q'eqchi' Mayan — Synthetic data effectively bootstraps NMT structural learning for low-resource languages, but requires authentic data for semantic depth.
• Global vs. Local Sentence Embeddings for Brazilian Portuguese: Revisiting Monolingual Models in the Age of Foundation Models — Optimal model choice for PT-BR sentence embeddings balances vocabulary coverage and adaptation flexibility.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.