Evaluating Multilingual Approaches to Sarcasm Detection: Insights from English and Arabic Texts

This study focuses on sarcasm detection across languages by leveraging the Sarcasm Headlines and ArSarcasm datasets to assess three leading transformer-based models: BERT, DistilBERT, and RobERTa. Evaluations are conducted using metrics such as accuracy, precision, recall, and F1-score, with findings indicating strong sarcasm detection capabilities. Notably, the models achieve up to 0.99 accuracy on English headlines and 0.84 on Arabic content, underscoring the effectiveness of these models for detecting sarcasm in diverse linguistic contexts. This research provides valuable insights into the role of dataset-specific features in cross-lingual sarcasm detection and opens avenues for further exploration.