Sarcasm Detection in Arabic Tweets: A Comparison Between Deep Learning and Pre-trained Transformer-based Models

Abstract

Sarcasm detection is difficult even in well-resourced languages, and Arabic adds further challenges through dialectal variation and limited annotated data for informal, social media text. This work compares classical deep learning architectures against pre-trained transformer models (such as AraBERT) for detecting sarcasm in Arabic tweets.

Method

Both approaches are trained and evaluated on a dataset of dialectal Arabic tweets labeled for sarcasm, allowing direct comparison of transformer-based transfer learning against models trained from scratch.

Contribution

Quantifies the performance gap between pre-trained transformer models and traditional deep learning approaches for Arabic sarcasm detection, informing model choice for downstream sentiment and opinion mining systems on Arabic social media.