DETECTION OF CYBERBULLYING TEXT USING HYBRID NEURAL NETWORK ARCHITECTURE

Abstract

The rise of digital technology in the modern era and the proliferation of online social media platforms and different online forums have led to unparalleled degrees of communication and sharing of information. Amidst the benefits, it has also led to a pervasive issue in the digital world known as cyberbullying, leading to significant challenges to the well-being of individuals and a threat to societal harmony. This project research work involves a thorough review of the recent advancements in deep learning techniques by various researchers to automate the process of cyberbullying detection. This project work investigates various deep learning techniques like Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM), transformers, Graph Convolution Networks (GCN), “Hybrid” based deep models and also a few Machine Learning and AI-based techniques to encounter cyber hate text. The survey is conducted on various private and publicly available datasets to gain insights into these diverse Deep Learning techniques, highlighting their performance, strengths, and limitations. My research experiment reveals that LSTM and Bi-LSTM deep models achieved exceptional performance and BERT, m-BERT and modified BERT models achieved good F-1 scores in detecting toxic content across multiple languages. The hybrid-based models and the introduction of the GCN model are also showing promising results in this domain. Based on a SWOT analysis approach, this study looks at the phenomenon of cyber hate texts spreading on social sites in depth which will provide valuable insights for researchers, practitioners, and policymakers that will guide in combating cyberbullying detection and the selection and choice of appropriate models. This comprehensive in-depth analysis will provide valuable insights for researchers, practitioners, and policymakers that will guide in combating cyberbullying detection and the selection and choice of appropriate models.