An Experimental Survey of ASA on DL Classifiers Using Multi-dialect Arabic Texts

Abstract

Social media services have become a place for people around the world to express their opinions, concerns, and thoughts about universal topics in their different spoken languages. In the Arabic language in general, there are fewer research contributions within the multi-dialect approach. Furthermore, in ASA (Arabic Sentiment Analysis), initiatives addressing the multi-dialects in the Arabic language compared to English are insufficient where the use of classification techniques and natural language processing is at a lower rate. As a result, this research focuses on analyzing sentiments in Arabic text and especially in different dialects within the Arabic language due to its high variation among social media users. This paper focuses on performing sentiment analysis on a multi-dialectical dataset by using several DL (Deep Learning) models: CNN (convolution Neural Network), LSTM (Long Short-Term Memory), RNN (Recurrent Neural Network), and CNN-LSTM. We conducted several experiments using the sequential, attention-based, segregation-aggregation approach on the multi-dialect Arabic text and on the translated version of the dataset. We then compared the results to two benchmark datasets across two word and character sentiment levels. The results obtained throughout the experiments showcased an accuracy of 79% across LSTM on word level on the attention-based approach ASA as well as the accuracy of 86% across both CNN and CNN_LSTM character level ASA using the attention-based approach.