Enhancing Cyberbullying Detection Through Sarcasm Recognition and Advanced Text Preprocessing Using RoBERTa

Abstract

Cyberbullying is an increasingly common problem in digital realms where offenders often use sarcastic language to disguise insult messages and mock the automated moderation system. This study looks at how sarcasm detection models can improve the detection of cyberbullying, which can identify abusive texts identified as sarcastic. The study suggests creating a thorough pre-processing of data, addressing the limitations reported in the literature review, such as text cleanliness and the use of advanced models. Techniques such as replacing abbreviations, removing emojis and emoticons, and filtering multilingual and nonsensical texts are applied. In addition, RoBERTa, a Transformer-based model, is used to detect sarcasm and cyberbullying due to its greater contextual understanding. The results aim to improve the detection of aggressive content hidden behind sarcasm, reduce false negatives, and strengthen content moderation systems. This approach helps to create a safer digital environment and reduces the psychological impact on victims of online harassment. Keywords-Cyberbullying, Deep Learning, RoBERTa, Sarcasm detection.