Real-Time Cyberbullying Detection in User-Generated Content Using NLP Techniques and SVM

Abstract

Cyberbullying has become a popular issue in online spaces, seriously affecting people’s mental well-being and overall safety on the digital front. As social media and other platforms keep growing, harmful behaviours are popping up more often, and old-school manual checks can’t keep up with the volume or the need for instant responses. In this work, we introduce a framework for detecting cyberbullying that’s both explainable and emotion-aware. It combines machine learning and natural language processing to recognise toxic text. The setup starts with solid text prep: normalising words, removing common stop words, and lemmatising to get everything into a standard form. Then it converts the text into features using TF-IDF (term frequency-inverse document frequency). But it doesn’t stop at basic word counts; the model also pulls in deeper context, such as overall sentiment, hints of sarcasm, mentions of gender, and clear signs of personal attacks, to make predictions more reliable. We train a supervised classifier to separate bullying posts from harmless ones, and it outputs a probability score indicating how damaging a message might be. To address the black-box problem with most classifiers, our system provides straightforward explanations in plain language, breaking down the words and context that led to each decision. On top of that, there’s a part that detects emotions like anger, sadness, or fear, which often show up in this kind of nasty exchange. We built the whole thing into a user-friendly web app with Streamlit, so it handles live user input or crunches through large batches of data at once. When we tested it, the framework held its own in terms of accuracy while remaining transparent and easy to use in practice. That makes it a good fit for tasks like moderating social media, teaching digital safety in schools, or monitoring online communities.