Deep Learning-Based Cyberbullying Detection in Kurdish Language

Abstract

Abstract Cyberbullying is a significant concern in this digital age due to its harmful effects on individuals and society. Sadly, social media platforms have only exacerbated the problem, making it imperative to find effective ways to identify and prevent offensive content. While previous research has extensively focused on English and explored machine learning techniques to tackle this issue. To address this gap, this paper introduces a new hybrid deep learning model called Gray Wolf Algorithm-convolutional neural network (GWA-CNN), explicitly designed to detect cyberbullying in the Kurdish language on Twitter. The proposed model combines the CNN framework with an optimised GWA version to improve CNN’s parameters and reduce training time. We evaluated GWA-CNN thoroughly utilizing the first-ever manually annotated Kurdish dataset of 30k tweets that have been meticulously curated and divided into three categories, namely sexism, racism and neutral expressions, and compared its performance to those of state-of-the-art algorithms such as Naïve Bayes, K-Nearest Neighbors, Recurrent Neural Networks, Gated Recurrent Units and attention-based transformer. The experimental results demonstrate that GWA-CNN exhibited superior performance in all scenarios, outperforming other approaches in detecting cyberbullying on Twitter.