ABSTRACT
Online platforms, while fostering global communication and interaction, have inadvertently become breeding grounds for antisocial conduct such as cyberbullying, trolling, and hate speech. The imperative to detect such behaviours particularly in the context of cyberbullying—has intensified, necessitating the deployment of automated systems due to the limitations of manual moderation in handling vast volumes of social media content. Cyberbullying encompasses hostile and derogatory language aimed at harming individuals through digital interactions. This study proposes an advanced detection methodology leveraging the Robustly Optimized Bidirectional Encoder Representations from Transformers (RoBERTa), augmented with Global Vectors for Word Representation (GloVe) and dimensionality reduction via Principal Component Analysis (PCA). The performance of the proposed framework is rigorously benchmarked against a range of machine learning, deep learning, and transformer-based techniques. Statistical evaluations, including k-fold cross-validation, demonstrate that the proposed model significantly outperforms existing methods, attaining an accuracy and recall of 0.98, and a precision and F1-score of 0.97. These findings substantiate the model’s efficacy in detecting cyberbullying with superior reliability and consistency.
Index Terms: Cyberbullying detection, global vectors for word representation (GLOVE), natural language processing (NLP) for social media analysis, principle component analysis (PCA), robustly optimized bidirectional encoder representations from transformers approach (RoBERTa), transformer-based learning.