A Comparative Study of Pre-trained Word Embeddings for Arabic Sentiment Analysis

In this paper, we conduct a series of experiments to systematically study both context-independent and context-dependent word embeddings for the purpose of Arabic sentiment analysis. We use pretrained word embeddings as fixed features extractors to provide input features for a CNN model. Experimental results with two different Arabic sentiment analysis datasets indicate that the pre-trained contextualized AraBERT model is the most suitable for such tasks. AraBERT reaches an accuracy score of 91.4% and 95.49% on the large Arabic book reviews dataset (LABR) and the hotel Arabic-reviews dataset (HARD), respectively.