CYBER BULLYING DETECTION IN SOCIAL NETWORKS

Abstract

Social networking and communication have been accelerated by information and communication technologies, yet cyberbullying presents serious difficulties. The current laborious and ineffective methods for reporting and stopping cyberbullying are user-dependent. For the automatic detection of cyberbullying, traditional machine learning and transfer learning techniques were investigated. The study made use of a structured annotation procedure and a large dataset. The Conventional Machine Learning method used textual, sentiment and emotional, static and contextual word embeddings, term lists, psycholinguistics, and toxicity characteristics. The use of toxicity features for the detection of cyberbullying was first presented in this study. Word Convolutional Neural Network (Word CNN) contextual embeddings performed similarly; embeddings were selected based on their higher Fmeasure. When given separately, toxicological features, embeddings, and textual features provide new standards. In terms of managing high-dimensionality features and training time, this performed better than Linear SVC. In contrast to the basis models, Transfer Learning achieved a faster training computation by fine-tuning using Word CNN. Additionally, Flask Web was used to detect cyberbullying, with a 97.06% accuracy rate. For privacy reasons, the name of the particular dataset was not mentioned.