NEURAL APPROACHES TO HATE SPEECH AND CYBERBULLYING DETECTION

Abstract

The widespread use of social media has resulted in a surge in hate speech and cyberbullying, therefore robust detection methods are critical. The dynamic character of the language and the ambiguity included in categorization problems are too much for conventional approaches to handle. Investigating if neural networks and uncertainty estimations together can increase detection accuracy is the aim of this study. Using Bayesian deep learning and probabilistic models is one method to facilitate scenario management. The system is more resistant to unwanted inputs and configuration modifications when this tactic is used. A dataset created from real social media comments is used in the evaluation. The findings imply that both memorization and accuracy have improved. This approach can be modified to take into account recently developed hate speech formats. The automatic moderation mechanism performs better as a result. Real-time deployment and multilingual support will be features of future versions of the application.