TOXIC COMMENT CLASSIFICATION USING DEEP LEARNING

Online abuse and harassment are escalating challenges in the cyber community, requiring robust and efficient content moderation strategies. This study investigates two approaches for identifying toxic comments: training individual classifiers for each facet of toxicity and treating the task as a multi-label classification problem. The facets of toxicity include behaviors like threats, insults, and hate speech, often exhibited simultaneously in a single comment. Using a Kaggle dataset and 10-fold cross-validation, machine learning models such as logistic regression, Naive Bayes, and decision trees are evaluated for performance. A novel preprocessing technique is proposed, transforming multi-label into multi-class classification, which significantly enhances model accuracy. Logistic regression emerges as the most effective model, outperforming others in both binary and multi-class classification tasks. These findings suggest the potential of this preprocessing strategy to boost the performance of more sophisticated neural network-based models, offering a scalable solution for real-world toxic content moderation systems.