Abstract
This paper presents a comprehensive study on the effectiveness of a novel hybrid model of LSTM and CNN in cyberbullying detection in online social media text. The effectiveness of the proposed model is tested against other traditional models of machine learning, namely SVM, Random Forest, and Decision Trees, in terms of accuracy, precision, recall, F-Score, and AUC-ROC curves. The proposed model is a hybrid of Long Short-Term Memory and Convolutional Neural Network, which combines the contextual intelligence of Long Short-Term Memory and the proficiency of Convolutional Neural Network in feature extraction. The results of the experiment show that the proposed model of LSTM and CNN is significantly superior to other models of machine learning in all evaluation parameters. Moreover, the results of the ROC curve also affirm that the proposed model is significantly more sensitive and specific in distinguishing cyberbullying and non-cyberbullying text. This paper also presents a novel tool that can help social media platforms in mitigating cyberbullying and harassment in a most efficient and effective way. Moreover, it also presents a novel area of research that can help in exploring the ethical and legal implications of deploying such efficient and sophisticated tools in cyberbullying and harassment.