Multimodal hate content detection using deep learning

Abstract

Over the past two decades, social media platforms have revolutionized global communication by enabling billions of users to share and consume content in real-time across geographic and cultural boundaries. With the rise of video-first platforms, such as TikTok, Instagram, and YouTube, communication has increasingly become multimodal, combining text, images, and audio into complex and expressive formats. While this evolution enriches user interaction, it also complicates content moderation, particularly in detecting subtle and context-dependent forms of hate speech. Implicit hate speech, unlike its explicit counterpart, often relies on coded language, cultural references, sarcasm, or multimodal cues, making it significantly harder to detect using conventional, unimodal systems. In this thesis, we address this critical and underexplored problem by introducing a novel task of hate speech detection in videos. To facilitate research in this domain, we present a new dataset curated specifically for this task, consisting of approximately 2,000 annotated videos. Each video is enriched with aligned modality-specific inputs, including textual transcripts, extracted audio features, and visual frames, along with auxiliary features such as sentiment scores, emotion cues, and image captions.