Harmful content classification in social media using gated recurrent units and bidirectional encoder representations from transformer
V. Sujatha, Yarramreddy Tejaswi, V. B. Pravalika, Pasupuleti Pavani, Ch. Sravani · 2025
Harmful content on the internet poses a significant challenge to maintaining a safe and inclusive online environment. To remove this issue, we propose a novel approach for harmful content classification, combining the power of recurrent neural networks and pre-trained transformers. Specifically, we utilize gated recurrent units (GRUs) for text preprocessing and bidirectional encoder representations from transformers (BERT) for the final classification task. In the preprocessing stage, GRUs are employed to capture sequential dependencies in textual data, allowing for the effective extraction of context-based features from the input text. This preprocessing step is crucial for understanding the nuanced structure and context of potentially harmful content, such as hate speech, cyberbullying, or offensive language. Subsequently, the pre-processed data is fed into a fine-tuned BERT model for classification. BERT is a state-of-the-art transformer model that excels in understanding the semantics of text, making it particularly well-suited for the nuanced and context-dependent nature of harmful content detection. By leveraging BERT’s pre-trained contextual embeddings, we can efficiently classify text into different categories of harm, including hate speech, harassment, or misinformation. Our proposed approach is capable of handling multi-class classification, making it versatile for a huge range of harmful content identification tasks. Furthermore, the combination of GRU-based preprocessing and BERT- based classification offers a powerful and adaptable solution for detecting harmful content, contributing to the ongoing efforts to create safer online spaces, and promoting responsible content moderation.