Cross-Platform Hate Speech Detection Using an Attention-Enhanced BiLSTM Model

Hate speech is rapidly spreading across digital platforms, appearing in diverse forms driven by regional, cultural, and linguistic differences. This growing trend presents serious challenges to social harmony and online safety. Existing hate speech detection models often fall short because they rely on limited and homogeneous datasets, making them less effective in real-world, culturally diverse settings. Handling large-scale, diverse datasets adds notable complexity to capturing contextual nuances, as different populations and cultures demonstrate unique language patterns and expressions. This study addresses the necessity for a more universal solution by proposing a deep learning model trained on an extensive and diverse dataset comprising 842,000 samples collected from various digital platforms. The approach combines a Bidirectional Long Short-Term Memory (BiLSTM) model with a self-attention mechanism to capture contextual depth. Various data embedding techniques were used to assess their impact, along with data resampling and standard Natural Language Processing (NLP) pre-processing steps. The proposed model achieved 93% accuracy with an F1-score of 0.92, outperforming several baseline and state-of-the-art models. This work provides a comprehensive and scalable framework for the detection of hate speech across various online platforms.