A Comparative Study of BERT-Based Models for Sarcasm Detection in Social Media Texts

Abstract

Social media has transformed communication, facilitating the rapid exchange of emotions and ideas between users. This shift has created the necessity for the development of sentiment analysis tools, enabling businesses to gain insights into audience reactions. However, detecting sentiment remains a challenging task due to the presence of informal language, abbreviations, and, notably, sarcasm, which can modify the intended message. Sarcasm, often conveyed through ironic or contrary statements, is particularly difficult to identify in text, as it lacks the non-verbal cues typically present in face-to-face communication. Recent advancements in deep learning, particularly the emergence of BERT (Bidirectional Encoder Representations from Transformers) based models, have significantly enhanced sarcasm detection by capturing nuanced contextual meanings of words. This paper compares several BERT-based models—BERT, RoBERTa, ALBERT, DistilBERT, and DeBERTa—to assess their effectiveness in sarcasm detection on iSarcasmEval and Sarcasm Corpus V2 datasets. Key performance metrics, including accuracy, computational efficiency, and the ability to capture complex contextual relationships, are analyzed to identify the most suitable model for sarcasm detection tasks. DeBERTa achieved the best performance on both datasets in this challenging task. Smart citations: https://scite.ai/reports/10.61467/2007.1558.2026.v17i2.853Dimensions.Open Alex.