Multi-sarcasm and consarcasm: a contextual dataset and hybrid model for fine-grained sarcasm detection

Sarcasm often relies on prior discourse context and subtle lexical cues, which makes it challenging for conventional models to capture. This paper presents ConSarcasm, a fully end-to-end sarcasm recognition framework that incorporates hierarchical conversational context modeling, sarcasm-aware preprocessing, and a hybrid Transformer-based classification strategy. To preserve pragmatic indicators such as emojis and elongated characters, the framework adopts a custom Byte Pair Encoding (BPE) tokenizer designed to handle out-of-vocabulary tokens. We also introduce Multi-Sarcasm, a context-rich dataset annotated with three levels of sarcasm intensity using a hybrid zero-shot LLM-assisted annotation scheme followed by systematic human validation. The originality of ConSarcasm lies in the principled integration of conversational context modeling, sarcasm-sensitive tokenization, and hybrid Transformer–CNN learning within a unified end-to-end framework for graded sarcasm detection. Experimental results on Multi-Sarcasm show that ConSarcasm achieves an accuracy of 89.99% and an F1-score of 90.02%, outperforming strong baselines such as BERT and RoBERTa. The model also generalizes well to binary sarcasm detection, achieving accuracies of 80.29% on IAC V2 and 93.70% on the Binary-Sarcasm dataset, demonstrating its robustness beyond binary classification settings.