RoBERTa and AfroXLMR detect online polarization with 79% F1 in English and Swahili

A new research paper tackles the growing problem of online polarization across languages and cultures, submitted to SemEval-2026 Task 9. Author Aaron Bundi Anampiu introduces a system that uses transformer-based models to detect three levels of polarization: whether content is polarizing, what type (e.g., hate speech, disinformation), and how it manifests (e.g., dehumanization, lack of empathy). The method leverages RoBERTa-base for English and AfroXLMR-base for Swahili, employing class-weighted loss functions to combat severe label imbalance — a common challenge in polarization detection where certain categories are rare.

On the official test set, the system achieves F1 macro scores of 0.7901 for English and 0.7910 for Swahili on binary polarization detection (Subtask 1). For subtask 2 (polarization type classification), scores were 0.4615 (English) and 0.4808 (Swahili); for subtask 3 (manifestation identification), 0.4791 (English) and 0.5830 (Swahili). Error analysis shows the models particularly struggle with detecting dehumanization and lack of empathy — suggesting these nuanced forms require more sophisticated reasoning. The paper demonstrates that with targeted techniques like per-label threshold tuning, transformer models can effectively handle multi-label, imbalanced data across languages, a crucial step for building robust content moderation tools.