Cyberbullying Detection using BERT and Sentiment-Aware Deep Learning

Abstract

Cyberbullying is a pervasive threat in digital communication, often leading to severe psychological consequences such as depression, anxiety, and suicidal tendencies. Traditional detection systems primarily rely on lexical and syntactic features and often fail to detect implicit forms of abuse like sarcasm, emotional manipulation, and passive aggression. This paper introduces a novel emotion-aware cyberbullying detection system that enhances performance by integrating emotional intelligence into deep learning models. The proposed system leverages a hybrid architecture combining BERT-based contextual embedding’s, an Emotion Detection Model (EDM), and sentiment analysis tools like AFINN and the NRC Emotion Lexicon. A curated and validated dataset labelled with emotional attributes (anger, fear, guilt, etc.) was developed using Twitter and Wikipedia comments. The system effectively classifies cyberbullying content with an F1-score of 0.97, outperforming baseline and state-of-the-art models. This research provides three key contributions: (1) a comprehensive emotion-annotated dataset for cyberbullying detection, (2) empirical evidence that emotional features significantly enhance detection accuracy, and (3) a scalable and platform-independent detection framework suitable for real-time deployment. The integration of affective computing marks a significant advancement in detecting nuanced and implicit cyberbullying behavior online.