Advanced Sarcasm Detection Using Transformer‐Based Models: Analyzing Variations in Formal and Informal Languages

Detecting sarcasm in text is notoriously difficult because it depends on subtle contextual cues that vary widely between communication styles. Informal language, such as satirical headlines, often signals sarcasm through exaggeration or playful wording, whereas formal news headlines embed irony more discreetly, making it harder to detect. Mixed datasets further complicate the task by combining these contrasting styles. To tackle this challenge, we evaluate three transformer‐based models, such as RoBERTa, DistilBERT, and ALBERT, on carefully balanced formal, informal, and mixed headline datasets. Results show that RoBERTa achieves the highest overall performance, reaching 90% accuracy across styles, while DistilBERT and ALBERT remain competitive but less effective in formal contexts. A soft voting ensemble of all three models further boosts performance, achieving 94% accuracy and an F 1‐score of 0.94 on mixed data, demonstrating improved robustness in heterogeneous settings. Beyond predictive strength, we apply LIME to interpret token‐level decisions, revealing that RoBERTa captures subtle contradictions in formal text while lighter models often miss these cues. This integration of cross‐style evaluation, ensemble modeling, and interpretability offers a transparent and adaptable framework for sarcasm detection. The findings not only advance methodological rigor but also carry practical relevance for sentiment analysis, automated moderation, and intelligent media analytics.