Bangla Hate Speech Detection System Using Transformer-Based NLP and Deep Learning Techniques
Omar Faruqe, Mubassir Jahan, Md. Ahasan Atick Faisal, Md. Shahidul Islam, Riasat Khan
Abstract
Hate speech is a form of discriminatory communication disrupts community standards and breaches the line of self-limitation, causing harm to others and occasionally leading to cyberbullying. Hate speech spreads hatred toward a person or a particular group based on various characteristics, e.g., race, religion, gender, and so on is referred to as bias. The offensive speech detection system is the frontier where researchers are battling to provide secure internet using natural language processing and machine learning approaches. In this research, we strive to create an automatic Bangla hate speech detection system using natural language processing (NLP) and deep learning approaches. We utilized our custom dataset and some labeled data from an open-access repository in this work. 4,978 data from both sources were merged and implemented in our proposed model. Different data preprocessing techniques, tokenization, stemming, and removal of stopwords have been applied. Four deep learning and NLP-based classifiers have been applied to detect Bangla hate speech. Google API has been employed to convert text from Bangla to English. The emojis were removed from the datasets and the data were translated into Bangla. The GRU and Attention techniques performed best with 98.87% and 98% accuracies, respectively.