On the robustness of arabic aspect-based sentiment analysis: A comprehensive exploration of transformer-based models

Abstract

In the era of rapid technological advancement, users generate an overwhelming volume of data on social media networks and e-commerce platforms daily. This data, rich in opinions, sentiments, values, and habits, holds immense value for both consumers and businesses. Leveraging this unstructured data manually is error-prone and time-consuming. The field of Sentiment Analysis automates the process of analyzing human opinions from this data. Sentiment Analysis classifies text into positive, negative, or neutral sentiments. However, it confines text classification to a single sentiment polarity, providing a broad overview without accounting for specific aspects. With the growing demand for data analysis, this standard sentiment polarity classification is no longer sufficient. Aspect-Based Sentiment Analysis has emerged to dig deeper into the text, uncovering perspectives and points of view. It can identify multiple aspects in text with corresponding sentiment polarity. Therefore, interest in this field has increased and many research efforts have been devoted recently to tackle this problem for the English language. Unfortunately, there is a scarcity of Arabic research in this field. This study will address the aforementioned deficiency by investigating the potential of four transformer models namely, AraBERT v2.0, ArBERT, MARBERT, and Multilingual BERT in enhancing the accuracy of Aspect-Based Sentiment Analysis for Arabic texts using two dedicated corpora (AraMA and AraMAMS). The extensive experiments revealed that the proposed approach achieved its expected effect surpassing the results of previous studies in the field. The best results of Aspect Category Detection and Aspect Sentiment Classification tasks in AraMA corpus were obtained by using AraBERT v2.0 with F1-Measure result equals to 95.75% and 92.83% respectively. In addition, the best result of Aspect Category Detection and Aspect Sentiment Classification tasks in AraMAMS corpus were achieved by using AraBERT v2.0 with F1-Measure result equals to 95.54% and 89.52% respectively.