Ensemble Stacking Model for Sentiment Analysis of Emirati and Arabic Dialects

Abstract

Sentiment analysis is the process of examining people's opinions and emotions towards goods, services, organizations, individuals, and other things, through the use of textual data. It involves categorizing text as positive, negative, or neutral to quantify people's beliefs. Social media platforms have become an important source of sentiment analysis data due to their widespread use for sharing opinions and information. As the number of social media users continues to grow, the amount of data generated for sentiment analysis also increases. Previous research on sentiment analysis for the Arabic language has mostly focused on Modern Standard Arabic and various dialects such as Egyptian, Saudi, Algerian, Jordanian, Tunisian, and Levantine. However, there has been no research on the utilization of deep-learning approaches for sentiment analysis of the Emirati dialect, which is an informal form of the Arabic language spoken in the United Arab Emirates. It's important to note that each country in the Arab world has its dialect, and some dialects may even have several sub-dialects.The primary aim of this research is to create a highly advanced deep-learning model that can effectively perform sentiment analysis on the Emirati dialect. To achieve this objective, the authors have proposed and utilized seven different deep-learning models for sentiment analysis of the Emirati dialect. Then, an ensemble stacking model was introduced to combine the best-performing deep learning models used in this study. The ensemble stacking deep learning model consisted of deep learning models with a meta-learner layer of classifiers. The first model combined the two best-performing deep learning models, the second combined the four best-performing models, and the final model combined all seven trained deep learning models in this research. The proposed ensemble stacking deep learning model was assessed on four datasets, three versions of the ESAAD Emirati Sentiment Analysis Annotated Dataset, two versions of Twitter-based Benchmark Arabic Sentiment Analysis Dataset (ASAD), an Arabic Company Reviews dataset, and an English dataset known as a Preprocessed Sentiment Analysis Dataset PSAD. The results of the experiments demonstrated that the proposed ensemble stacking model presented an outstanding performance in terms of accuracy and achieved an accuracy of 95.54% for the ESAAD dataset, 96.71% for the ASAD benchmark dataset, 96.65% for the Arabic Company Reviews Dataset, and 98.53% for the Preprocessed Sentiment Analysis Dataset PSAD.