Multi-Level Contextual Semantic Modeling for Implicit Hate Speech Recognition

Abstract

Social media platforms are key spaces for sharing opinions, experiences, and daily events. However, the rapid rise in hate speech, especially implicit hate, has become a significant concern. Implicit hate is subtle, context-dependent, and more difficult to detect than explicit hate. Despite the growing body of work on Hate Speech Detection (HSD), most existing models are developed for resource-rich languages like English. Unlike resource-rich languages, Hindi lacks sufficient annotated data and dedicated research, making it a low-resource language. To address this limitation, this work presents a deep learning–driven approach for the classification of implicit and explicit hate speech in Hindi. The approach uses FastText embeddings to capture semantic and subword-level information, which suits noisy social media text. A Bidirectional Long Short-Term Memory (BiLSTM) network models contextual dependencies, while an attention mechanism assigns higher importance to discriminative words and sentences, enabling effective identification of implicit hate. Extensive experiments on the Hindi Hate Speech Dataset (HHSD) demonstrate the model’s effectiveness. Compared with the widely used multilingual BERT (mBERT) baseline, the proposed approach improves classification performance, with a 2.51% increase in accuracy and a 17.84% improvement in weighted F1-score. Additionally, the model requires significantly fewer parameters than transformer-based models, making it more parameter-efficient.