A model agnostic explainable approach for detecting Cyber bullying in Bangla language using transformer based models

Abstract

Almost every facet of social communication has changed as a result of the exponential growth of social media platforms usage. Meanwhile, evidence is accumulating that the rising usage of social networks in the digital realm has given rise to an unsettling problem that has resurrected in new contexts: cyberbullying. The majority of current cyberbullying detection research focuses on English texts. On the other hand, while being spoken by 230 million people globally and being rich in diversity, the Bengali language is under-resourced for natural language processing (NLP). Recently, there has been an alarming surge in the number of incidences of gender-based discrimination or sexual harassment expressed on social media sites. In this study, we presented the cyberbullying detection under different categories in low-resourced Bangla language using transformer based models. We created our own dataset on gender discrimination and appended it to another open-source Bangla dataset with 4 classes. In our proposed approach, we used five different models to train our augmented dataset, followed by an ensembling technique on those five models.Then we make the models explainable using model agnostic approaches. Finally, we compared the individual prediction accuracies with the ensembled prediction accuracies. While training the dataset, we followed the stratified k-fold cross validation technique. Our evaluations yield up to an Accuracy of 75% in cyberbullying detection on emsembling.