Cyberbullying detection on multi-modal data using pre-trained deep learning architectures

Cyberbullying is a big challenging task in the social media era. The forms of bullying are increasing with the increase of digital technologies. In the past, most of the bullying happened through text messages. Now bullies take advantage of technology, they try bullying others in different forms such as images, videos, and emojis. In this paper, we proposed an approach to identify cyberbullying on both text and image data combinations. We used RoBERTa and Xception deep learning architectures to generate word embeddings from the text data and the image respectively. LightGBM classifier is used to classify bullying and non-bullying tweets. The experiments conducted on 2100 samples of combined data of text and image. The proposed approach efficiently classifies bullying data with F1-score of 80% and outperforms as compared to existing approaches.