Exploring Hate Speech and Offensive Language in Social Media: Insights from Twitter Textual Content with Stacked BiLSTM

Detecting hate speech is a critical task in today’s digitally connected world, as it plays a significant role in maintaining societal harmony and individual well-being. Hate speech, when left unchecked, can foster hostility, escalate tension, and negatively affect individual well-being and public safety. It often creates to toxic online environments that contribute to psychological distress, such as anxiety, and isolation among targeted individuals or communities. As a result, timely and accurate identification of hate speech is crucial for several social, ethical, and legal reasons. To address this issue, a S-BiLSTM (Stacked Bidirectional Long Short-Term Memory) architecture is proposed to identify such speech and offensive sentences by integrating with Synthetic Minority Oversampling TEchnique (SMOTE). The SMOTE generates synthetic examples of the minority class instances not only enhances detection accuracy but also addresses both sequential dependencies of hate speech and the data imbalance typical in datasets taken from online social networks. The proposed model is evaluated on two benchmark datasets widely used for hate speech and offensive language detection. The model outperformed the existing works in identifying offensive sentences and accurately identifying hate speech with 94% and 95.47% respectively.