Classification of Arabic Comments to Detect Cyberbullying from Social Media Using Convolutional Neural Network and Meta-Learning

As a result of the proliferation of social media, cyberbullying has become widespread in the Arab community on social media and cyberbullying has become a concern targeting individuals and may cause some serious side effects. The problems of Natural Language Processing (NLP) for the Arabic language and its complexity make it difficult to classify texts accurately. In recent years, deep learning models have emerged as a viable option for solving some of the NLP problems. In this study, we constructed a hybrid approach of Convolutional Neural Network (CNN) and Meta-learning for classifying cyberbullying Arabic comments. A set of electronic text data in Libyan dialect and Modern Standard Arabic was collected from several Libyan social media platforms such as Facebook, YouTube, and other online platforms to identify instances of cyberbullying on social media. Pre-processing is a vital part of the data preparation process for detecting cyberbullying, where a CNN model was trained on the data. Finally, the model was evaluated for accuracy, recall, precision, and F1 scores. Thus, the results showed that CNN outperforms better when combined with Meta and gave higher results than CNN only. We obtained the best classification accuracy of 98, 91, and 84% for three datasets. The accuracy of CNN alone was 71, 69, and 52% respectively for the three experiments. These results confirm the success of the model and the improvement in CNN performance with Meta and that it gives better results than CNN. These results confirm the potential of neural networks in developing and succeeding in cyberbullying detection systems.