Research on Machine Learning Techniques for Unmasking Online Negative Speech

Abstract

The proliferation of online platforms has facilitated global communication but has also led to a concerning rise in hate speech across digital spaces. This research addresses the critical challenge of detecting hate speech in Banglish, a blend of Bangla and English, which presents unique linguistic and cultural complexities. We propose a comprehensive approach leveraging machine learning techniques to identify and mitigate hate speech in this bilingual context. Our methodology involves data collection, prepossessing, feature extraction, and the evaluation of multiple machine learning models, including SVM, KNN, Naive Bayes, Decision Trees, Logistic Regression, Random Forest, AdaBoost, Bagging, Extra Trees, Gradient Boosting, XGBoost, and QDA. The dataset, meticulously curated to capture Banglish nuances, was balanced using Random Over Sampler to address class imbalance. Among the evaluated models, Naive Bayes emerged as the top performer, achieving an accuracy of 78.33%, precision of 81.95%, recall of 74.40%, F1 score of 77.88%, specificity of 82.51%, and an AUC-ROC score of 86.18%. The study highlights the effectiveness of traditional machine learning models in handling high-dimensional sparse data and provides a foundation for developing robust moderation tools to foster safer digital environments for Banglish-speaking communities. The research not only bridges a linguistic gap in hate speech detection but also contributes to broader efforts in combating online hate speech across diverse linguistic contexts.