Multimodal hate speech detection: a novel deep learning framework for multilingual text and images

Abstract

The rapid proliferation of social media platforms has facilitated the expression of opinions but also enabled the spread of hate speech. Detecting multimodal hate speech in low-resource multilingual contexts poses significant challenges. This study presents a deep learning framework that integrates bidirectional long short-term memory (BiLSTM) and EfficientNetB1 to classify hate speech in Urdu-English tweets, leveraging both text and image modalities. We introduce multimodal multilingual hate speech (MMHS11K), a manually annotated dataset comprising 11,000 multimodal tweets. Using an early fusion strategy, text and image features were combined for classification. Experimental results demonstrate that the BiLSTM+EfficientNetB1 model outperforms unimodal and baseline multimodal approaches, achieving an F1-score of 81.2% for Urdu tweets and 75.5% for English tweets. This research addresses critical gaps in multilingual and multimodal hate speech detection, offering a foundation for future advancements.