A Hybrid Semantic–Linguistic Feature Fusion Framework for Cross-Dataset Sarcasm Detection and Contextual Robustness

Abstract

Sarcasm detection is a challenging natural language processing task because sarcastic expressions often convey meanings that differ from their literal form. The challenge becomes more complex in cross-dataset settings, where sarcasm varies across tweets, news headlines, conversational text, and informal social media content. Models trained on one dataset often show reduced performance on another due to semantic drift, vocabulary variation, contextual ambiguity, and domain-specific sarcasm cues. To address this issue, this paper proposes a hybrid semantic–linguistic feature fusion framework for cross-dataset sarcasm detection. The proposed approach combines CNN-based semantic representations with handcrafted contextual linguistic features such as sentiment contrast, hyperbole, punctuation density, capitalization, hashtag usage, and contrast indicators. The framework is evaluated on multiple sarcasm datasets, including news headlines, SPIRS conversational sarcasm data, and tweets with sarcasm and irony. Experimental results show that the proposed hybrid model provides improved cross-domain robustness compared with standalone lexical, machine learning, and deep learning models. The study demonstrates that combining semantic learning with contextual linguistic features improves sarcasm detection under domain shift and heterogeneous textual conditions.