HIVMID: A Multilingual and Multimodal Dataset for HIV/AIDS Misinformation Detection in English and isiZulu

Abstract

Digital health misinformation can affect health outcomes and informed decision-making, yet resources for African low-resource languages (LRLs) and HIV/AIDS misinformation remain scarce. Existing health misinformation benchmarks also focus largely on English and COVID-19-related content. To address this gap, we developed a multilingual, multimodal dataset covering English and isiZulu. The dataset was constructed through a six-stage methodology informed by the Datasheets for Datasets and Data Statements for Natural Language Processing frameworks and contains 1,147 records (673 factual and 474 misinformation) from 92 normalised sources. It covers five thematic categories, two languages, and both text-only and text-plus-image modalities. Six classical machine learning classifiers were evaluated using TF-IDF representations, stratified cross-validation, and an independent test set. The best-performing model, Random Forest with word-level TF-IDF, achieved 93.45% accuracy, a 93.49% weighted F1-score, and a Matthews Correlation Coefficient of 0.874, substantially outperforming a majority-class baseline while maintaining comparable performance across both language subsets. The paper introduces the Health Misinformation Dataset Construction Framework (HMD-CF), which generalises the dataset construction process into a transferable methodology for multilingual health misinformation resources. The proposed dataset provides one of the first benchmark resources for multilingual HIV/AIDS misinformation detection(MID) in an African LRL setting and supports future research on health MID, multimodal learning, and LRL technologies.