Phatthachdau at SemEval-2026 Task 9: A Multi-Stage Augment-Judge-Train Pipeline for Multilingual Online Polarization Detection
Year: 2026
Type: conference-paper
Abstract: For SemEval-2026 Task 9 (Detecting Online Polarization), we address two contrasting challenges: the extreme label imbalance in lowresource languages (e.g., Hausa, with only 11% polarized instances) and the sophistication of implicit hostility in high-resource languages (English).We propose the Augment-Judge-Train (AJT) pipeline to overcome these barriers.By leveraging Gemini 2.0 Flash for taxonomydriven synthetic data generation and Gemini 2.5 Pro as an LLM-as-a-Judge layer for quality control, we effectively expand the minority class while preserving cultural nuances.Our final system utilizes a weighted soft-voting ensemble, combining specialized Encoders (for local linguistic patterns) with LLM-LoRA (for deep contextual reasoning).Evaluated on the official test set, our system achieved 1st Place in Hausa (0.8336 Macro-F1) and ranked in the top 10 for English (0.8092 Macro-F1), demonstrating the efficacy of culture-aware synthetic data in enhancing social NLP.
Author: Phan Phat
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
SDG: Reduced inequalities
Open Access status: hybrid
OpenAlex ID: W7166887357