Sarcasm detection in text remains one of the trickier problems in NLP, mainly because sarcastic statements often depend on subtle contradictions and contextual clues that aren’t immediately obvious. We propose a dual pathway approach that leverages RoBERTa-based contextual embeddings alongside emoji-derived affective features, which helps the system better understand sarcastic content whether it appears in tweets or longer texts. Low data availability and the possibility of model overfitting was mitigated by the implementation of data augmentation. The developed model was trained and tested on five different benchmark datasets which includes Twitter contents, headlines and various other online forums. The proposed model demonstrated high performance by scoring up to 6% and 8% gains in F1 scores and accuracy. Additionally, on further testing, the model showed good generalization capabilities across different datasets.