A Feature Engineering Approach for Detecting Cyberbullying in Bangla Text Using Machine Learning

Cyberbullying has become an alarming issue with the prodigious rise of different social media platforms. The consequences of cyberbullying are uncountable and thus it should be detected and prevented immediately. Previous research has demonstrated great accuracy in detecting textual cyberbullying data. But the detection of the severity level of cyberbullying specially in Bangla text is yet to be performed. To mitigate this gap, this research is conducted to assess the severity level of Bangla cyberbullying text using feature engineering approach. We performed feature engineering to introduce new features such as PMI-SO (Pointwise Mutual Information - Semantic Orientation) and domain specific features to perform the detection. Two datasets were created in order to detect the severity level and to generate the features. The datasets contained 9721 and 33736 Bangla texts respectively. The Severity Dataset was annotated as low, medium, high and non-bullying. Final performance analysis showed that XG Boost classifier was the most accurate classifier with 87 % accuracy. The performance of SVM was followed by Logistic Regression, SVM and Naive Bayes classifiers. These classifiers obtained 85%, 81 % and 77% accuracy respectively.