A Swish-Activated Deep Neural Bigru Framework With Contextual Information For Fine-Grained Multilingual Hate Speech Detection

Abstract

The growing prevalence of hate speech in multilingual and code-mixed social media environments necessitates robust fine-grained detection models capable of capturing semantic, contextual, and language-specific variations. In this study, a Swish Activated Deep Neural Bi-directional Gated Recurrent Unit with Contextual Information (SADNB-CI) is developed for fine-grained hate speech detection across Tamil, English, and Tanglish social media text. The proposed framework addresses the limitations of existing hate speech detection approaches by integrating Global Dense Vector Representation-based embedding, Swish-activated convolution, max pooling, CNN-based feature representation, BiGRU-based sequential modelling, and contextual information learning for multi-label classification. Separate embedding representations are generated for Tamil, English, and Tanglish inputs, while convolutional filters extract discriminative local features and the BiGRU layer captures forward and backward contextual dependencies. Contextual hate speech categories, including community-based, religion-based, gender-based, and political hate speech, are incorporated to enhance the distinction between hate speech and non-hate speech and to support fine-grained classification. The SADNB-CI model is evaluated using the AI Tamil Hate Speech Detector dataset with precision, recall, accuracy, and false positive rate as performance metrics. Experimental findings demonstrate that the proposed model outperforms k-means+textGCN and t-HateNet across Tamil, English, and Tanglish samples by achieving higher precision, recall, and accuracy while reducing the false positive rate. The results indicate that the integration of Swish-activated CNN feature extraction with BiGRU-based contextual learning improves multilingual and code-mixed hate speech detection, thereby establishing SADNB-CI as an effective approach for fine-grained social media hate speech classification.