Hate Speech Classification Using Mixed Language Feature Modification and Machine Learning Approach

Hate speech is often considered vague or contradictory in its definitions, generally referring to various kinds of negative expressions by users, including insults and harsh criticism. To clarify, it is suggested to focus on references to violence in a statement and discriminatory content. Things like negative stereotypes, expressions of violence, dehumanization, hurting or killing, and insults that use offensive language, slurs, denigrating with metaphors, and wordplays. This study investigates the effectiveness of different machine learning and deep learning models for hate speech classification, evaluating the impact of stopword removal preprocessing and feature expansion techniques (FastText and Word2Vec). Our findings indicate that not performing stopword removal can slightly improve accuracy compared to preprocessing with stopword removal. Furthermore, feature expansion had minimal impact on classification accuracy. Among the models tested, BERT-Large achieved the highest accuracy of 85.16%, significantly outperforming traditional machine learning models (Linear Support Vector Classifier (SVC), SVC, Random Forest) and other deep learning models (Long Short-Term Memory (LSTM), Convolutional Neural Network (CNN)). These results suggest that transformer-based models like Bidirectional Encoder Representation of Transformers (BERT) are particularly effective for hate speech classification, while feature expansion techniques may not be crucial for enhancing model performance.