Cyberbullying Detection Based on Semantic-Enhanced Marginalized Denoising Auto-Encoder

Abstract

This paper introduces a novel approach, Semantic-enhanced Marginalized Stacked Denoising Autoencoder (smSDA), for cyberbullying detection on social media platforms. The escalating prevalence of cyberbullying necessitates effective detection methods to create a safer online environment, especially for children and young adults. Our proposed method addresses the challenge of developing robust numerical representations for text messages, crucial for accurate cyberbullying detection. By extending the capabilities of deep learning models with semantic enhancements, specifically semantic dropout noise and sparsity constraints, smSDA captures underlying feature structures related to bullying content. Comprehensive experiments conducted on publicly available datasets from Twitter and MySpace demonstrate the superiority of smSDA over baseline methods, indicating its potential as an effective tool for combating cyberbullying.