An incongruity-aware gated CNN–BiLSTM model for sarcasm detection in news headlines

The problem of sarcasm detection in the short and textual information, which is the news headline is still very challenging due to the tacit and contradictory nature of the sarcastic terms. Out of the literal linguistic forms, sarcasm often follows as a result of semantic sentimental incongruity, whereby the purportedly positive lexical expressions carry a negative encoding, or the reverse. Conventional machine learning models mainly rely on surface-level lexical features. These features are often insufficient to detect hidden contradictions. Deep learning models attempt to improve the performance. To overcome these limitations, this paper presents a novel Incongruity-Aware Gated CNN-BiLSTM (IAG-CB) that is specific to news headlines sarcasm detection. The proposed framework employs a dual-branch architecture in which a CNN-BiLSTM module captures contextual semantic representations, while a TextBlob-based sentiment encoder extracts polarity-aware sentiment features. A clear mechanism of incongruity-modeling then measures the deviation between semantic and sentiment representations. Semantic and incongruity features are then adaptively combined by a learnable gated fusion module and then finally classification takes place. Extensive experiments conducted on benchmark news headline datasets demonstrate that the proposed model consistently outperforms traditional machine learning classifiers, state-of-the-art deep learning and Transformer baselines. These results support the claim that the explicit modeling of semantic -sentiment discrepancy significantly improves the performance of sarcasm detection and increases the interpretability at the same time. The suggested framework is therefore a structurally principled approach to robust sarcasm detection of short textual data that is linguistically based.