REAL‑TIME CYBERBULLYING DETECTION WITH NLP AND DEEP LEARNING MODELS

Abstract

Cyberbullying poses a significant challenge to online safety and mental health, creating barriers for healthy digital communication and reducing the well-being of social media users. In this study, we present a web-based cyberbullying detection system utilizing machine learning techniques, specifically leveraging TF-IDF feature extraction combined with multiple classification algorithms. The objective is to create a robust and efficient solution capable of real-time cyberbullying detection in social media posts, comments, and image-based content. The proposed system capitalizes on the powerful capabilities of Linear Support Vector Machine (SVM), a state-of-the-art text classification algorithm, known for its accuracy and generalization in high-dimensional feature spaces. By training the model on a comprehensive dataset of 768 preprocessed samples from social media platforms, labeled across multiple categories including harassment, hate speech, threats, and offensive language, we optimized the classifier to accurately identify and categorize cyberbullying content. The system achieved exceptional performance with 91.84% accuracy, 92.00% precision, 93.88% recall, and 96.80% AUC-ROC score. Additionally, to ensure the practicality and accessibility of the solution, the trained model was integrated into a Streamlit web application with OCR capabilities for meme analysis. This application provides a user-friendly interface for social media moderators, educators, parents, and platform administrators, enabling real-time cyberbullying analysis with interactive visualizations, offensive word highlighting, severity classification, and comprehensive metrics commonly accessed through web browsers.