From Memes to Harassment Automated Detection of Cyberbullying for Cyberlaw Enforcement

This study investigates the automated detection of cyberbullying in multimodal online content, focusing on the intersection of technical classification and legal relevance. Using a dataset of 5,793 entries consisting of both image and text components, we analyze several key attributes including bullying classification, sentiment polarity, sarcasm, emotional tone, harmfulness score, and target type. The results reveal that 55% of the content is labeled as bullying when both image and text are considered together, compared to only 27% when using image-only classification. Negative sentiment dominates (2,657 entries), with sarcasm present in 2,154 entries, highlighting the prevalence of veiled or implicit abuse. Emotional annotations show that disgust (913) and sadness (580) are among the most common emotional tones associated with harmful content. Furthermore, the majority of abusive posts (2,405) are targeted at individuals, underscoring the need for stronger personal protection within cyberlaw frameworks. These findings support the development of contextaware and legally informed detection systems capable of addressing the nuanced and often implicit nature of online harassment.