Article

BERT-Based Emotion and Sarcasm-Aware Classification of Harmful Online Content for Cyber Law Enforcement

---

Abstract

The rise of cyberbullying on social media has created urgent demands for automated systems that can detect harmful digital content in both explicit and subtle forms. This study presents an exploratory analysis of a multimodal dataset consisting of 5,793 social media posts annotated with four key dimensions: harmfulness level (Harmless, Partially-Harmful, Harmful), emotion, sentiment, and sarcasm. Our goal was to understand the distributional, emotional, and linguistic features of harmful content and how they interact in context, with implications for cyber law enforcement and content moderation. The analysis reveals a significant class imbalance in harmfulness distribution, with 63.18% of posts labeled as Harmless, 33.51% as Partially-Harmful, and only 3.28% as Harmful. Emotionally negative expressions were dominant, particularly Disgust (1,565 posts) and Sadness (1,321 posts). Sarcasm was present in 1,023 posts, accounting for 17.66% of the dataset, indicating that indirect or veiled forms of hostility are widespread. Representative samples further demonstrate the ambiguity of harmful intent in posts that blend humor, sarcasm, and emotional undertones. These findings emphasize the need for context-aware classification models that incorporate affective signals and pragmatic cues to accurately identify both overt and covert cyberbullying. Such models are essential not only for enhancing the performance of content moderation systems but also for strengthening digital evidence collection and enforcement under cybercrime regulations. The study concludes with recommendations for integrating emotion and sarcasm detection into transformer-based architectures to improve interpretability and legal relevance in automated harmful content assessment.

Keywords: Cyberbullying Detection, Harmful Content, Sarcasm, Emotion Analysis, Cyber Law Enforcement

How to Cite:

Chantanasut, S., (2025) “BERT-Based Emotion and Sarcasm-Aware Classification of Harmful Online Content for Cyber Law Enforcement”, Journal of Cyber Law 1(4), 300-313. doi: https://doi.org/10.63913/jcl.v1i4.73