Arabic sentiment analysis using GCL-based architectures and a customized regularization function

•We propose a new architecture called Gated Convolution Long (GCL) for Arabic sentiment analysis (ASA), to address the issue of long training samples, extract the best features for binary and n-ary classification, and boost the effectiveness of ASA.•We develop a custom regularization function (CRF), which helps to improve the performance of the proposed model. Furthermore, we perform an ablation study which demonstrates that the improved results are due to CRF.•We conduct a comparison study with a standard loss function, and show that our custom regularization aids in optimizing the loss function’s performance.•We show that the proposed method offers the best classification performance relative to current baselines. We also use the proposed method to investigate the link between sentiments in Modern Standard Arabic and those in five different Arabic dialects.•Finally, we compare the proposed technique with standard regularization functions on massive Arabic datasets; our model plus CRF is more effective, performs better, and takes less time.