A HYBRID MULTI-LEVEL DEEP LEARNING FRAMEWORK FOR HATE AND OFFENSIVE SPEECH DETECTION ON SOCIAL MEDIA

Abstract

The rapid growth of social media has increased the volume of user-generated text, including harmful content such as hate and offensive speech. This study proposes a hybrid multi-level approach that integrates Term Frequency-Inverted Document Frequency (TF–IDF) to extract feature, Chi-Square to select features, k-means clustering for grouping, Synthetic Minority Oversampling Technique (SMOTE) for balancing the dataset, and an Artificial Neural Network (ANN) as the final classifier. The framework is designed to enhance the accuracy and reliability of distinguishing between hate speech and offensive language. Experiments on a public Twitter dataset show that the proposed approach delivers an F1-score of 0.914 in the best configuration, outperforming Support Vector Machine (SVM) and Random Forest baselines.