Researchers have developed NepOOC-M, the first publicly available benchmark for detecting out-of-context (OOC) misinformation in Nepali and English. The dataset includes 1,090 image-caption pairs annotated across five typologies of misinformation. Evaluations showed that text-only models, specifically a multilingual-BERT (mBERT) variant, performed comparably to multimodal architectures, achieving a Macro-F1 score of 94.65%. The study suggests that expanding the dataset size is more impactful for progress than increasing architectural complexity. AI
IMPACT Provides a new resource for developing and evaluating AI models for misinformation detection in underrepresented languages.
RANK_REASON Publication of a new academic benchmark dataset and comparative analysis of models. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
