BullySpotter: Online Bullying Detection In Bengali Text Using Transform Learning

Prevalence of virtual entertainment has expanded quickly and presently it is very simple to cooperate with various people across web-based entertainment. Subsequently, online bullying towards individuals across web-based entertainment has additionally expanded. As online bullying can cause critical mental trouble, it ought to be prevented. There's been a lot of work done by experts to identify online bullying in Bengali language but they have many limitations. In this paper, we developed a detection system utilizing Deep Learning and Transform Learning algorithms that can detect bully word from Bengali text, which we called BullySpotter. We created a dataset with over 10,000 Bengali texts from the YouTube comment section. Bengali texts are first completely preprocessed, then manually classified into multiple classes (6) preparing them for transformer-based neural network (i.e., multilingual Bidirectional Encoder Representations (cased), bangla Bidirectional Encoder Representations (base)). At long last, we figure thoroughness and adequacy scores to gauge the nature of clarifications with respect to dependability. The banglaBERT model achieved the best accuracy of 89 percent in comparison to deep learning (BLSTM and CNN with word embeddings) and transform-based models (i.e., banglaBERT, mBERT).