A Multimodal Approach to Bangla Cyberbullying Meme Detection in Social Media

Cyberbullying, also known as cyber harassment, has been a very common but severe issue among social media users, especially young people. Meme, a popular feature of social media websites, has been a medium for cyber harassment. Therefore, early detection of bullying memes can prevent social media users from facing harassment or trolling. A significant amount of previous research has been done on bully meme detection using textual or visual features. However, bully memes have become very difficult to detect using only textual or visual features. In our proposed research work, we have developed a multimodal model to detect bully memes using both textual and visual features. Our proposed approach leverages the power of Vision Transformer for extracting visual features and pre-trained transformer models for extracting textual features. The proposed approach is evaluated on the MUTE dataset of memes labeled for cyberbullying content, demonstrating significant improvements over unimodal methods. We have found a state-of-the-art result achieving an F1 score of 0.76 on the MUTE dataset.