Dual-Encoder Incongruity Transformer for Context-Aware Sarcasm Detection with Data Augmentation

Sarcasm is hard for NLP systems because people rarely announce it directly—its meaning usually comes from how a sentence fits (or clashes) with what was said before. Many existing sarcasm detectors still rely on surface cues such as sentiment words, punctuation, or handcrafted lexical statistics, which often miss this “context vs. utterance” mismatch that drives sarcasm. In this paper, we introduce a Dual-Encoder Incongruity Transformer (DEIT) that treats sarcasm as an incongruity problem: the context and the target utterance are encoded separately using a shared Transformer, and their relationship is modeled using simple but effective interaction signals such as the absolute difference and element-wise product of their embeddings. Because sarcasm datasets are typically small, we also enlarge the training set to about 2000 samples through a hybrid augmentation strategy that mixes lightweight lexical edits with masked language model substitutions while keeping the class distribution balanced. On the evaluated dataset, DEIT achieves 98.33% accuracy, 98.33% F1-score, and an MCC of 0.9667, showing that explicitly modeling context–utterance incongruity leads to more reliable sarcasm detection.