Abstract
Sentiment analysis (SA) is a natural language processing (NLP) application that aims to analyse and identify sentiment within a piece of text. Arabic SA started to receive more attention in the last decade with many approaches showing some effectiveness for detecting sentiment on multiple datasets. While there have been some surveys summarising some of the approaches for Arabic SA in literature, most of these approaches are reported on different datasets, which makes it difficult to identify the most effective approaches among those. In addition, those approaches do not cover the recent advances in NLP that use transformers. This paper presents a comprehensive comparative study on the most effective approaches used for Arabic sentiment analysis. We re-implement most of the existing approaches for Arabic SA and test their effectiveness on three of the most popular benchmark datasets for Arabic SA. Further, we examine the use of transformer-based language models for Arabic SA and show their superior performance compared to the existing approaches, where the best model achieves F-score scores of 0.69, 0.76, and 0.92 on the SemEval, ASTD, and ArSAS benchmark datasets. We also apply an extensive analysis of the possible reasons for failures, which show the limitations of the existing annotated Arabic SA datasets, and the challenge of sarcasm that is prominent in Arabic dialects. Finally, we highlight the main gaps in Arabic sentiment analysis research and suggest the most in-need future research directions in this area.