Enhancing Arabic Hate Speech Detection: The Role of Word Embedding Techniques with Deep Learning Models

Hate speech is phenomena of concern that has taken a special shape and intensity on the internet. As a result, numerous research efforts have explored the possibility of recognizing hate speech in several languages using social media data to train predictive machine learning models. This study explores the use of three embedding methods, ArabGlossBERT, AraBERT, and AraVec, to investigate the role of the word embedding in enhancing the efficiency of deep learning models in detecting hate speech in the Arabic language. We used a labeled Twitter dataset of 4 classes. Our experiments consisted of two stages. The first stage examined the impact of coupling word embeddings from the three aforementioned techniques with LSTM and CNN models for classification. ArabGlossBERT demonstrated the best performance with both deep learning models, achieving accuracies of 60% and 52%, respectively. We evaluated the performance of the different embedding techniques with other ML techniques to compare with results reported on the same dataset, showing the superiority of ArabGlossBERT over alternative embedding methods more so when coupled with deep learning models.