CoPol at SemEval-2026 Task 9: Modeling Polarization Type Co-occurrence with Label Correlation Networks
Year: 2026
Type: conference-paper
Abstract: We present POLAR-LDA for SemEval-2026 Task 9 Subtask 2: multi-label classification of five polarization types across 22 languages.Two underexplored challenges define this task: absolute positive scarcity certain languagelabel pairs contain zero or one positive training instance, below the recoverable threshold for any supervised classifier and culturally variable label co-occurrence patterns that independent classifiers cannot capture.POLAR-LDA extends mDeBERTa-v3-base with a Label Correlation Network applying Graph Attention over language-specific co-occurrence matrices, Asymmetric Loss for positive-scarcity regimes, and language-family grouped ensemble training.The system achieves 0.567 macro F1 on the official evaluation (range: 0.784 Hindi to 0.256 Italian; per-label: 0.685 political to 0.444 other) ; per-label data volume dominates linguistic family as a predictor of performance (σ intra-family =0.119 vs. σ inter-family =0.086).Diagnostic analysis identifies data voids as the hard floor of multilingual polarization classification and shows that per-label training volume dominates linguistic family as a predictor of cross-lingual performance variance.
Author: Pushkar Arora
Institution: Delhi Skill and Entrepreneurship University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7166805481