Multilingual sentiment analysis of Arabic, Bahraini dialects and English

Abstract

Sentiment analysis is a crucial natural language processing (NLP) task to analyze the user’s emotions and opinions towards entities such as events, services, or products. Arabic NLP faces numerous challenges, some of which include: (1) the scarcity of resources, especially in modern standard Arabic and Arabic dialects, particularly the Bahraini one; (2) the lack of multilingual deep learning models; and (3) insufficient transfer learning studies on Arabic dialects in general and Bahraini dialects specifically. This research aims to create a balanced dataset of Bahraini dialects that covers product reviews by translating English Amazon product reviews to modern standard Arabic, which were then converted to Bahraini dialects. Another aim of this research is to provide a multilingual deep learning long short-term memory (LSTM) model to analyze the parallel dataset of English, modern standard Arabic, and Bahraini dialects, which differ in linguistic properties. Many experiments were conducted using train-validate-test split and k-fold cross-validation to evaluate the model performance using accuracy, F1 score, and AUC metrics. The average accuracy of the model on all datasets ranged from 96.72% to 97.04% and 97.91% to 97.93% in the F1 score, while in AUC was 98.46% to 98.7% when utilizing an augmentation technique. The LSTM model was incorporated in a stacking ensemble learning process that includes other LSTM architectures as base learners and a decision tree (DT) as a meta-learner. Interestingly, promising results were obtained, such as 99.52%, 99.25%, and 98.52% of mean accuracy for English, MSA, and BDs datasets. Moreover, the LSTM model was utilized as a pre-trained model in the transfer learning process to exploit the knowledge gained from analyzing the product reviews in Bahraini dialects to perform another sentiment analysis task on a small dataset of movie comments in the same dialects. The pre-trained model performance was 96.97% accuracy, 96.65% F1 score, and 97.94% AUC.