Sarcasm Detection in Arabic Text Using Contextualized Models

One of the most difficult issues in text classification is detecting sarcasm, especially in informal Arabic with a lot of syntactic and semantic ambiguity. Our goal is to identify whether the provided text is sarcastic or not. We have tested 19 models that are pre-trained on Arabic Language for two datasets which are ArSarcasm-v2 and Saudi-Dialect-Irony-Dataset. The highest accuracy we achieved is 78.73% accuracy using AraBERT v2 base [1] on ArSarcasm-v2 and 71.18% using mBERT [2] on Saudi-Dialect-Irony-Dataset.