Sarcasm Detection in Arabic Text using Artificial Intelligence-Based Modeling

Abstract

Social media's pervasiveness in our daily lives has made it possible for people to voice their thoughts without facing consequences. As a result, comments made online about anything or anyone frequently include sarcastic overtones. To better understand attitudes and social expressions in digital communication, researchers are growing more and more interested in sarcasm detection. However, because sarcastic remarks are subjective, context-dependent, and culturally nuanced, identifying them can be challenging. This task becomes even more difficult for Arabic, a language that is morphologically rich and diverse, with numerous dialects. Consequently, an effective sarcasm detection model for Arabic must be developed. This paper proposes an AraBERT transformer-based model for sarcasm detection in Arabic text. A dataset of approximately 10,500 human-annotated Arabic tweets was collected from publicly available digital media platforms. The dataset has a skewed distribution of sarcastic and non-sarcastic statements, and various strategies were applied to address this imbalance during training. The model was trained and tested on unseen datasets, achieving an accuracy of 92.46% and specificity of 95.82%. Its performance was also measured using precision, recall, F1-score, specificity, and accuracy. For Arabic sarcasm detection, the suggested model performs better than a number of alternative models already in use. The suggested model performs admirably when it comes to identifying irony in Arabic text.