Advancing Online Safety: AI-Powered Hate Speech and Offensive Language Detection in Social Media

Hate speech encompasses verbal or written expressions that demean or attack individuals or communities due to their identity, including religion, race, gender, nationality, or other defining traits. Social media's emergence as the prevailing communication vector has coincidentally enabled faster distribution of damaging content. The escalation of hateful and abusive language online has transcended virtual spaces, frequently triggering tangible societal repercussions. The present work contributes to creating protective digital communities by implementing detection and intervention systems for toxic communication in fluid social media contexts. A key contribution of this research lies in advancing hate speech identification on social networks through systematic analysis of Twitter data, including tweets and retweets, to derive actionable insights for enhancing content moderation systems. A curated dataset of 21,010 instances—spanning hate speech, offensive language, and neutral speech—was assembled for evaluation. Multiple machine learning models, such as Logistic Regression, Random Forest, Support Vector Machines (SVM), Naïve Bayes, and XGBoost, were employed for classification. The Random Forest classifier outperformed others with 98% accuracy, whereas Logistic Regression, SVM, and XGBoost achieved moderate accuracy rates between 78% and 86%. Naïve Bayes trailed with 71% accuracy, highlighting varying efficacy across algorithmic approaches.