Natural Language Processing and Machine Learning Approaches to Detect Bangla Hate Speech on Social Media

In this modern era, with the popularity of the Unicode system, many people use Bangla text to express their opinions, comments, emotions on Facebook, Twitter, and YouTube, etc. However, this social connectivity always does not create positive impacts. Trolling, cyberbullying, online harassment, sexual and political abuse, and cybercrime often create a huge negative physiological effect on victims. In most cases, women, children, and celebrities are the victims which may often lead them to depression and suicidal activity. Therefore, at the early stage, cyberbullying and hate speech detection can reduce these cyber crimes.A few researches have been carried out in the Bangla language to detect cyberbullying and hate speech due to annotated corpora, lexical and morphological analyzers. This research aims to design and develop an effective model to detect Bangla cyberbullying and hate speech text with the combination of Natural Language Processing (NLP) and Machine Learning Classifier algorithms. In this study, we have collected 6244 Bangla text data from various online social platforms consisting of 6 categories of hate speeches. Then, we labeled the data into two main classes-bullied and neutral that were used to train our classifiers after preprocessing and features extracted by the Bangla NLP tool and TF-IDF(Term Frequency –Inverse Document Frequency). Our comparative study outlines that, with an accuracy rate of 89% and considering parameters Recall, Precision, Specificity, F-Score, and ROC, the Random Forest Classifier achieves the best results compared to others.