ABSTRACT:
Hate speech on online news platforms has emerged as a critical societal challenge, influencing public discourse and impacting industries. However, existing detection methods often fail to capture its contextual nature and are grounded in weak theoretical foundations. To address this gap, we integrate agenda-setting theory with design science paradigms to develop a deep learning model for detecting hate speech on online news platforms. Leveraging bidirectional encoder representations from transformers (BERT), our model analyzes the interplay between news headlines, texts, and user comments, capturing both media issue salience and emotional agenda-setting effects. Empirical validation demonstrates that our model outperforms traditional baselines in hate speech detection and topic classification tasks. This contextualized approach enhances prediction accuracy, explainability, and domain adaptability, contributing to improved performance and broader applicability. Our study makes theoretical and methodological contributions to Information Systems research and offers practical insights for implementing ethical, real-time hate speech detection strategies.
Key words and phrases: Hate-speech detection, design science, agenda-setting theory, deep learning, sociotechnical perspective, BERT, online platforms