Enhanced Arabic Sentiment Analysis Using a Novel Stacking Ensemble of Hybrid and Deep Learning Models

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