Abstract
Sentiment analysis (SA) is a machine learning application that drives people's opinions from text using natural language processing (NLP) techniques. Implementing Arabic SA is challenging for many reasons, including equivocation, numerous dialects, lack of resources, morphological diversity, lack of contextual information, and hiding of sentiment terms in the implicit text. Deep learning models such as convolutional neural networks (CNN) and long short-term memory (LSTM) have significantly improved in the Arabic SA domain. Hybrid models based on CNN combined with long short-term memory (LSTM) or gated recurrent unit (GRU) have further improved the performance of single DL models. In addition, the ensemble of deep learning models, especially stacking ensembles, is expected to increase the robustness and accuracy of the previous DL models. In this paper, we proposed a stacking ensemble model that combined the prediction power of CNN and hybrid deep learning models to predict Arabic sentiment accurately. The stacking ensemble algorithm has two main phases. Three DL models were optimized in the first phase, including deep CNN, hybrid CNN-LSTM, and hybrid CNN-GRU. In the second phase, these three separate pre-trained models' outputs were integrated with a support vector machine (SVM) meta-learner. To extract features for DL models, the continuous bag of words (CBOW) and the skip-gram models with 300 dimensions of the word embedding were used. Arabic health services datasets (Main-AHS and Sub-AHS) and the Arabic sentiment tweets dataset were used to train and test the models (ASTD). A number of well-known deep learning models, including DeepCNN, hybrid CNN-LSTM, hybrid CNN-GRU, and conventional ML algorithms, have been used to compare the performance of the proposed ensemble model. We discovered that the proposed deep stacking model achieved the best performance compared to the previous models. Based on the CBOW word embedding, the proposed model achieved the highest accuracy of 92.12%, 95.81%, and 81.4% for Main-AHS, Sub-AHS, and ASTD datasets, respectively.