SENTIMENT AND CONTENT-BASED CYBERBULLYING DETECTION USING MACHINE LEARNING

Abstract

Social media's rapid growth has exacerbated cyberbullying, necessitating the implementation of automated detection techniques. Tokenization, noise reduction, and normalization are among the techniques employed to process content on the internet. In order to identify abusive language patterns, it is possible to locate emotional indicators and sentiment polarity. TF-IDF vectors, semantic embeddings, and N-grams are among the manifestations of contextual meaning. Predictive classifiers that are instructed include KNN, Random Forest, and Deep Neural Networks. A strategy that employs ensemble learning enhances accuracy and reduces the incidence of false positives. The paradigm effectively differentiates between content that is non-bullying, harassment, and bullying. In an effort to evaluate performance, we evaluate the confusion matrix, F1-score, precision, recall, and accuracy. The results indicate that the detection process is more effective when multiple models are employed than when only one is employed. In general, the algorithm is effective in generalizing to a variety of forms of cyberbullying. Abuseful sentences that alter predictions are identified by interpretability techniques. In general, the strategy simplifies the identification of cyberbullying in a variety of online scenarios.