Authors - Tanvi Pawar, Sachin S. Pande, Emmanuel M Abstract - Sarcasm detection in social media text is a NLP challenge, as sarcastic statements inverse meaning of the statement as sarcastic statements hide the real meaning. This problem intensified on platforms like Reddit by informal phrasing, community-specific references, and implicit cultural knowledge. This paper introduces a RoBERTa-based classification framework which addresses three core issues: contextual impoverishment of isolated comments, unstable training caused by random initialization, and catastrophic forgetting during fine-tuning. These are handled via inline textual metadata fusion (encoding subreddit identity and upvote score into the input sequence), a structured multi-layer classification head, and a biphasic two-stage training method with differential learning rates. Trained on a balanced 500,000-sample subset of the SARC dataset, the model achieves 68.36% accuracy with stable, monotonic convergence across all training epochs. Near-symmetric false positive and false negative rates shows that the model does not favor a single class. Future directions include knowledge graph integration, model distillation, multi-class sarcasm taxonomy, and multilingual extension.
Open Zoom