Etudiant : ES-SADEQY AYOUB

Filière : Master Systèmes Intelligents et Décisionnels (MSID)

Encadrant : Pr. EL FAHSSI KHALID

Annèe : 2023

Résumé : Sarcasm has become a prevalent form of communication on social media platforms like Twitter and Reddit. Detecting sarcasm is crucial for understanding the true sentiments of individuals, as it often involves expressing emotions contrary to the literal meaning of words. This report presents two distinct approaches to sarcasm detection. In the first part, several transformer models, such as BERTweet, RoBERTa, and others, were implemented for English binary sarcasm detection using three distinct datasets. Additionally, the MARBERT model was employed for Arabic sarcasm detection. The models were thoroughly analyzed and compared, assessing their performance and effectiveness in detecting sarcasm. The second part of the project focused on integrating commonsense knowledge into the sarcasm detection process. Specifically, the pre-trained COMET model was utilized to generate relevant commonsense knowledge. Two different knowledge selection strategies were employed to investigate the influence of commonsense knowledge on performance. Furthermore, a knowledge-text integration module was designed to effectively model both the text and knowledge. Experimental results demonstrated the effectiveness of the proposed model across the three datasets. Through the implementation and evaluation of these two approaches, this report aims to shed light on their respective strengths and limitations in detecting sarcasm. The findings provide valuable insights into the effectiveness of transformer-based models and the implementation of commonsense knowledge for the accurate identification of sarcasm in social media contexts.