Researchers have developed a novel two-stage model for detecting harmful content in Arabic, addressing challenges posed by dialectal variations and semantic overlap. The system first utilizes GLiNER to identify harm-related entities within user prompts. These extracted entities are then appended to the original text and processed by a fine-tuned MARBERT classifier for final classification, achieving a macro-F1 score of 0.886 on the ArGuard 2026 development set. The approach demonstrates the effectiveness of integrating Named Entity Recognition (NER) with Arabic language models for fine-grained harmful content detection, though distinguishing between closely related harm categories remains an area for improvement. AI
IMPACT This research advances methods for detecting harmful content in Arabic, potentially improving online safety and moderation tools.
RANK_REASON This is a research paper detailing a new model for a specific NLP task. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
