Detection of sarcasm is a subtle task considering its reliance on irony and figurative language. In this study, we presented Contextual Collocation Augmentation (C2A) methodology and we compare its effect on English and Hindi sarcasm detection using two different datasets: an English news headline dataset and a Hindi Twitter-based corpus. Our findings indicate that C2A significantly improves baseline models and beats the state-of-the-art results in sarcasm detection, with Ridge, Logistic Regression, and LaBSE models achieving F1-scores of up to 97%, which perform better than certain deep learning–based methods. Even though transformer models inherently capture deep contextual signals, the relative impact of C2A was higher in baseline models, particularly for Hindi, where sarcasm tends to be phrase-based and structure-dependent. Comparative analysis reveals that C2A outperforms prior approaches, including deep learning and ensemble architectures such as CNN-BiLSTM hybrids, XGBoost, and earlier BERT-based methods, especially in recall and balanced F1-scores. ROC-AUC analysis further confirms the enhanced discriminatory power of feature-augmented models. These results suggest that C2A remains a crucial component in sarcasm detection, offering competitive performance for traditional machine learning models against computationally expensive transformers while providing better interpretability and efficiency.
T.O. et al. (Thu,) studied this question.