Detecting Offensive Language on Arabic Social Media Using Deep Learning

Offensive content on social media such as verbal attacks, demeaning comments or hate speech has many negative effects on its users. The automatic detection of offensive language on Arabic social media is an important step towards the regulation of such content for Arabic speaking users of social media. This paper presents the results of evaluating the performance of four different neural network architectures for this task: Convolutional Neural Network (CNN), Bidirectional Long Short-Term Memory (Bi-LSTM), Bi-LSTM with attention mechanism, and a combined CNN-LSTM architecture. These networks are trained and tested on a labeled dataset of Arabic YouTube comments. We run this dataset through a series of pre-processing steps and use Arabic word embeddings to represent the comments. We also apply Bayesian optimization techniques to tune the hyperparameters of the neural network models. We train and test each network using 5-fold cross validation. The CNN-LSTM achieves the highest recall (83.46%), followed by the CNN (82.24%), the Bi-LSTM with attention (81.51%) and the Bi-LSTM (80.97%).