An Attention Based BiLSTM Approach for Detecting Abusive Comments in Bangla Discussion Threads

Abstract

The usage of social networking platforms has accelerated rapidly with technological advancement in the last few decades. Facebook, Twitter, and YouTube are extensively utilized platforms for social communication, interaction, marketing, and information sharing. So abusive text and comments are widespread on those platforms, negatively impacting the expected user experience. Besides, Bangla is the seventh most widely spoken language worldwide, and a significant number of users use Bangla on those platforms. In this study, I have concentrated on detecting abusive Bangla comments in discussion threads from social platforms, i.e., Facebook and YouTube. After extensive pre-processing, tokenization, sequencing and padding Bidirectional Long Short-Term Memory (BiLSTM) with attention layer is trained utilizing the dataset consisting of a total of 12,022 comments which are labeled into two categories, abusive and nonabusive. In this study, I have estimated the accuracy, precision, recall, and f1-score to assess our models. Our proposed model performed notably with an accuracy of 94%.