Detection of Arabic Offensive Content Using Graph Neural Networks

Abstract

The detection of offensive content on social media has become a critical task in the field of Natural Language Processing (NLP), particularly for underrepresented languages such as Arabic. This thesis investigates the application of Graph Neural Networks (GNNs) for Arabic offensive content classification, leveraging the rich contextual representations offered by pre-trained language models. To address data limitations and improve model performance, we constructed a concatenated dataset by merging two publicly available Arabic offensive language corpora from political and social discussions. Each comment was converted into a graph structure, where nodes correspond to tokens and edges represent contextual relations. Contextualized embeddings were generated using AraBERT, a transformer-based model specifically trained for the Arabic language. We developed and evaluated three GNN architectures—Graph Convolutional Networks (GCN), Graph Attention Networks (GAT), and GraphSAGE—for the classification task. These models were assessed against traditional baselines including Long Short-Term Memory (LSTM) networks and Support Vector Machines (SVM). Experimental results demonstrate that GNN-based models, particularly GraphSAGE and GAT, outperform conventional approaches, owing to their ability to incorporate structural dependencies and semantic information through message passing mechanisms. This research highlights the effectiveness of graph-based deep learning in Arabic text classification and provides a foundation for further exploration in offensive language detection, especially within morphologically rich and low-resource languages.