Character-Level Convolutional Neural Networks for Cyberbullying Detection: A Robust Approach to Handling Noisy Social Media Text

Abstract

With the increasing prevalence of cyberbullying on social media, there is a pressing need for effective detection methods that can handle the noisy, unstructured nature of online text. Traditional machine learning models often struggle with the informal language, misspellings, and emoticons commonly used in cyberbullying messages. In this paper, we propose a novel approach for detecting cyberbullying using Character-level Convolutional Neural Networks (Char-CNNs). Unlike word-based models, the Char-CNN model operates at the character level, allowing it to effectively handle spelling errors, intentional distortions, and the use of emojis. We evaluate the performance of Char-CNN on a publicly available social media dataset and compare it with a traditional Word-CNN model. Our results show that Char-CNN outperforms the word-based approach across key performance metrics, including accuracy, precision, recall, and F-measure. The model's ability to generalize well in the presence of noisy data makes it a promising tool for real-time cyberbullying detection. Furthermore, we discuss the limitations of the current model and future directions for enhancing its performance, particularly in detecting more subtle forms of cyberbullying.