Paper Title
An Empirical Study on the Effectiveness and Limitations of Deep Learning Models for Cyber Bullying Detection on Social Media Text

Abstract
Cyber bullying has emerged as a serious challenge on social media platforms, where abusive, humiliating, and harmful interactions can spread rapidly and often anonymously. Automated cyber bullying detection has therefore become an important research problem in natural language processing and machine learning. This paper presents an empirical evaluation of multiple text-based cyber bullying detection approaches, including TF-IDF based machine learning models, deep learning models using word embeddings, and transformer-based architectures. The objective of this study is not only to compare classification performance using metrics such as accuracy, AUC, and F1-score, but also to examine the practical limitations of these models when exposed to implicit, context-dependent, and ambiguous bullying expressions. Experimental results show that although several models achieve strong benchmark performance, their ability to generalize to nuanced real-world bullying scenarios remains limited. The findings highlight the gap between dataset based evaluation and real-world applicability, and emphasize the need for future cyber bullying detection systems that incorporate contextual understanding, multimodal cues, and human-in-the loop moderation for more reliable and ethical deployment. Keywords – Cyber Bullying Detection, Natural Language Processing, Deep Learning, Transformer Models, Social Media Analysis