Abstract
Sarcasm detection is a crucial task in natural language processing (NLP), particularly
in sentiment analysis and opinion mining, where sarcasm can distort sentiment
interpretation. Accurately identifying sarcasm remains challenging due to its contextdependent
nature and linguistic complexity across informal text sources like social media
and conversational dialogues. This study utilizes three benchmark datasets, namely, News
Headlines, Mustard, and Reddit (SARC), which contain diverse sarcastic expressions
from headlines, scripted dialogues, and online conversations. The proposed methodology
leverages transformer-based models (RoBERTa and DistilBERT), integrating context
summarization, metadata extraction, and conversational structure preservation to enhance
sarcasm detection. The novelty of this research lies in combining contextual summarization
with metadata-enhanced embeddings to improve model interpretability and efficiency.
Performance evaluation is based on accuracy, F1 score, and the Jaccard coefficient, ensuring
a comprehensive assessment. Experimental results demonstrate that RoBERTa achieves
98.5% accuracy with metadata, while DistilBERT offers a 1.74x speedup, highlighting the
trade-off between accuracy and computational efficiency for real-world sarcasm detection
applications.