PolaFusion at SemEval-2026 Task 9: Ensemble Transformers with Targeted Augmentation for Multilingual Polarization Detection

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

Summary

The PolaFusion system was developed for SemEval-2026 Task 9, focusing on detecting polarization in social media posts across 22 languages. It addresses severe class imbalance in three subtasks: binary polarization detection, type classification, and rhetorical manifestation identification. PolaFusion employs a hierarchical gating architecture, using a binary gatekeeper model to direct content to specialist classifiers trained on polarized data. A mega-ensemble combines fivefold mDeBERTa-v3-base and three-fold XLM-RoBERTa-large models, aggregating probabilities via soft-voting. Additionally, a Macro-F1-aware augmentation strategy leverages Qwen3-235B to generate synthetic minority-class examples for scarce and poorly learned language-label pairs. Inverse-frequency class weighting within BCEWithLogitsLoss further aids training. The system achieved Macro-F1 scores of 0.800, 0.576, and 0.502 on Subtasks 1, 2, and 3, respectively, surpassing the POLAR baseline by +0.040, +0.089, and +0.082 average Macro-F1. Its code is publicly available.

Key takeaway

For Machine Learning Engineers building multilingual classification systems with severe class imbalance, PolaFusion's approach offers a robust blueprint. You should consider implementing hierarchical gating, large Transformer ensembles, and targeted, Macro-F1-aware data augmentation using models like Qwen3-235B. This strategy, combined with inverse-frequency weighting, can significantly improve performance on rare labels and across diverse languages, as demonstrated by its superior Macro-F1 scores.

Key insights

PolaFusion effectively detects multilingual social media polarization by combining hierarchical gating, ensemble Transformers, and targeted data augmentation to counter class imbalance.

Principles

• Hierarchical gating improves specialized classification.
• Ensemble models enhance robustness.
• Targeted augmentation counters class imbalance.

Method

PolaFusion uses a binary gatekeeper to route content to specialist classifiers, then aggregates soft-vote probabilities from an 8-model Transformer ensemble. Qwen3-235B generates minority-class examples for scarce language-label pairs, with inverse-frequency weighting.

In practice

• Apply hierarchical models for complex tasks.
• Use Qwen3-235B for targeted data augmentation.
• Implement inverse-frequency weighting for imbalanced data.

Topics

• Polarization Detection
• Multilingual NLP
• Class Imbalance
• Ensemble Learning
• Data Augmentation
• Transformers

Code references

• Abdullah4152/PolaFuse

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

Related on AIssential

• Exploring Data Augmentation and Resampling Strategies for Transformer-Based Models to Address Class Imbalance in AI Scoring of Scientific Explanations in NGSS Classroom — Targeted data augmentation significantly improves transformer-based model performance in automated scoring of imbalanced educational text data.
• Multi-Stage Training for Abusive Comment Detection in Indic Languages — A multi-stage pipeline combining language preprocessing and model ensembles effectively detects abusive Indic language comments while minimizing false positives.
• CAMO: A Class-Aware Minority-Optimized Ensemble for Robust Language Model Evaluation on Imbalanced Data — CAMO is a hierarchical ensemble method that dynamically boosts minority class predictions using confidence and uncertainty.
• SURGELLM: Rethinking Multi-Task Evaluation through Task-Aware Feature Gating with Class-Balanced Normalization — SURGELLM enhances multi-task NLP performance by integrating task-aware feature gating and class-balanced normalization within a unified transformer framework.
• A machine learning study highlighting the challenges of fidgety movement recognition using vision and inertial sensors — Automated Fidgety Movement recognition in infants faces significant generalization challenges across unseen subjects.
• Ensembles of Ensembles of Ensembles: A Guide to Stacking — Ensembling diverse models across multiple layers enhances predictive performance and robustness for tabular and time series data.

Open in AIssential →

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