Open Access

ARTICLE

Why Transformers Outperform LSTMs: A Comparative Study on Sarcasm Detection

Palak Bari, Gurnur Bedi, Khushi Joshi, Anupama Jawale^*

Department of Information Technology, Narsee Monjee College of Commerce and Economics, Mumbai, 400056, Maharashtra, India

* Corresponding Author: Anupama Jawale. Email:

Journal on Artificial Intelligence 2025, 7, 499-508. https://doi.org/10.32604/jai.2025.072531

Received 29 August 2025; Accepted 27 October 2025; Issue published 17 November 2025

Abstract

This study investigates sarcasm detection in text using a dataset of 8095 sentences compiled from MUStARD and HuggingFace repositories, balanced across sarcastic and non-sarcastic classes. A sequential baseline model (LSTM) is compared with transformer-based models (RoBERTa and XLNet), integrated with attention mechanisms. Transformers were chosen for their proven ability to capture long-range contextual dependencies, whereas LSTM serves as a traditional benchmark for sequential modeling. Experimental results show that RoBERTa achieves 0.87 accuracy, XLNet 0.83, and LSTM 0.52. These findings confirm that transformer architectures significantly outperform recurrent models in sarcasm detection. Future work will incorporate multimodal features and error analysis to further improve robustness.

---

Keywords

---