A Systematic Review of Deep Learning Approaches for Information Disorder Detection on Social Media: The Context of Misinformation, Disinformation, and Mal-information

Abstract

Social media platforms have become powerful tools for communication, information exchange, networking, and community building. However, their widespread use has facilitated the proliferation of misinformation, disinformation, and mal-information, posing significant challenges to societal harmony and individual well-being. Attacks such as social pressure, bullying, extortion, manipulation, abuse, prejudice, and violence can deepen divisions and conflicts, demonize minorities, and undermine the integrity of public opinion and democratic elections. This study presents a systematic review with quantitative comparative analysis of machine learning and deep learning approaches for information disorder detection, with a unique emphasis on low-resource multilingual contexts, particularly Amharic and Afaan Oromo, while addressing critical challenges related to multimodality, data scarcity, and contextual linguistic complexity. Using the PRISMA framework, 963 studies were screened, resulting in the selection of 99 high-quality articles for in-depth analysis. This study provides a dataset-level synthesis, comparing benchmark and low-resource datasets across multiple dimensions such as data modalities (text, image, memes, audio) and learning models. The findings indicate an increasing adoption of transformer-based and deep learning models, which demonstrated strong performance, particularly on large benchmark and multimodal datasets. Importantly, the review identifies critical research gaps, including limited multimodal fusion techniques, insufficient multi-class classification approaches, and a lack of localized context-aware models for low-resource languages. Unlike global surveys, this study provides a localized Ethiopian-context analysis that addresses linguistic diversity, resource constraints, and real-world deployment challenges. The study contributes an integrated perspective encompassing multilingual, multimodal, and multiclass dimensions and outlines future directions for developing scalable, context-aware, and inclusive detection models.