In this paper, we propose several specific embeddings in Arabic sentiment analysis (SA) framework. Indeed, Arabic is characterized by its agglutination and morphological richness contributing to great sparsity that could affect embedding quality. This work presents a rigorous study that compares different types of Arabic-specific embeddings. We evaluate them with 2 neural architectures: one based on convolutional neural network (CNN) and the other one based on Bidirectional Long Short-Term Memory Bi-LSTM. Experiments are done on the Large Arabic-Book Reviews corpus LABR. Our best results boost previous published accuracy by 1.9%. Moreover, we experiment combination of our individual systems defining very confident decision, reaching an accuracy of 92.2% on 98.25% of LABR test dataset.