Automatic Cyberbullying Detection

Abstract

Cyberbullying is defined in many ways but there is vast consensus in the current literature which have cruel, intentional, and repetitive behavior between peers, through social media networks that are online social applications. The incredible fame of social media platform like Facebook, Twitter has been providing more freedom for online users to upload their content on social media sites without any constraints. Many researchers proposed many solutions for the detection of cyberbullying in English, Arabic, Dutch, Hindi, and Marathi. However, little work is done in the detection of cyberbullying in the Roman Urdu language. Many techniques are used in the detection of cyberbullying, mostly Machine Learning and Natural Language Processing. Our proposed method implemented for cyberbullying detection. For this purpose we crawl data from twitter normalized it with some preprocessing techniques. Six different ML classifiers that is logistic regression, random forest, SVM, decision tree, Naïve Bayes and KNN are trained on our benchmark balanced binary class (bullied-1, non-bullied-0) dataset of 3786. Feature are extracted by TF-IDF feature extraction technique and n-gram (unigram and bigram) method is also applied. The benchmark dataset is train and six different ML classifiers are applied and performance is evaluated by accuracy, precision, recall, F1 score and confusion matrix are plotted separately for uni-gram and bi-gram. Also cross validation with K-fold 10 is applied against each ML classifiers. It is noticed that if the N-gram range increases, the performance of supervised ML classifier change. Naïve Bayes performs well among all other classifiers having accuracy 76% and recall score is 78% with unigram. On the other hand SVM have 74.3% accuracy with uni gram and increase in recall score when bi-gram is apply. Recall score with Bi-gram for SVM with k-fold is 90%. It is noticed that SVM recall score is greater than baseline studies with our proposed methodology.