Two-Stage Classification of Offensive Meme Content and Analysis

Memes are generally understood as a blend of visual and textual elements and have become an effective medium of communication on social media. However, this ease of communication has also led to challenges in content moderation, particularly with the spread of hate speech, aggressive content, and cyberbullying. Traditional moderation tools often rely on models that focus solely on text or image data, which is inadequate for the multimodal nature of memes. This research proposes a multimodal meme classifier that leverages Stacked LSTM for textual analysis and VGG-16 for visual features, integrating both modalities to enhance classification accuracy. By integrating lexical and visual information, the model captures the full context of memes, enabling it to effectively distinguish between offensive and non-offensive content. The dual-pathway approach offers a significant improvement over single-pathway methods, demonstrating promising classification performance on a curated meme dataset. This solution shows great potential for real-time integration in content moderation systems, enabling efficient detection of offensive memes. Its multimodal framework is very promising on the growing demand for robust systems capable of moderating multimodal content in harmful online activities.