Abstract
Sarcasm detection in natural language processing (NLP) remains a complex challenge, especially in social media, where contextual clues are often subtle. This study addresses this challenge by leveraging transformer-based models, including BERT, GPT-3, Claude-2, and Llama-2, for sarcasm detection on a large dataset from the Self-Annotated Reddit Corpus (SARC). The proposed method utilizes multi-head attention mechanisms to enhance model performance by capturing nuanced contextual relationships in the text. Fine-tuning of BERT, GPT-3, and Llama-2 was conducted to ensure a fair comparison and to provide a more detailed understanding of sarcasm in context. Our BERT-based model achieved state-of-the-art performance, with precision, recall, F1 score, and accuracy of 0.918, 0.917, 0.917, and 0.917, respectively, outperforming the other models. The effectiveness of our approach is demonstrated through rigorous statistical validation, ablation studies, and error analysis, providing robust evidence of its superiority. This study also highlights the significance of fine-tuning, machine translation, and multi-head attention in improving sarcasm detection.