A deep CNN architecture with novel pooling layer applied to two Sudanese Arabic sentiment data sets

Arabic sentiment analysis has become an important research field in recent years. Initially, work focused on Modern Standard Arabic (MSA), which is the most widely used form. Since then, work has been carried out on several different dialects, including Egyptian, Levantine and Moroccan. Moreover, a number of data sets have been created to support such work. However, up until now, no work has been carried out on Sudanese Arabic, a dialect which has 32 million speakers. In this article, two new public data sets are introduced, the two-class Sudanese Sentiment Data set (SudSenti2) and the three-class Sudanese Sentiment Data set (SudSenti3). In the preparation phase, we establish a Sudanese stopword list. Furthermore, a convolutional neural network (CNN) architecture, Sentiment Convolutional MMA (SCM), is proposed, comprising five CNN layers together with a novel Mean Max Average (MMA) pooling layer, to extract the best features. This SCM model is applied to SudSenti2 and SudSenti3 and shown to be superior to the baseline models, with accuracies of 92.25% and 85.23% (Experiments 1 and 2). The performance of MMA is compared with Max, Avg and Min and shown to be better on SudSenti2, the Saudi Sentiment Data set and the MSA Hotel Arabic Review Data set by 1.00%, 0.83% and 0.74%, respectively (Experiment 3). Next, we conduct an ablation study to determine the contribution to performance of text normalisation and the Sudanese stopword list (Experiment 4). For normalisation, this makes a difference of 0.43% on two-class and 0.45% on three-class. For the custom stoplist, the differences are 0.82% and 0.72%, respectively. Finally, the model is compared with other deep learning classifiers, including transformer-based language models for Arabic, and shown to be comparable for SudSenti2 (Experiment 5).