Impact of oxymoron features and deep learning techniques in the detection of sarcastic contents.

This article focuses on developing an automatic sarcasm classification system using a multi-domain English dataset of 67,850 sarcastic and non-sarcastic sentences collected from various online sources. The study introduces oxymoron features—figurative expressions combining contradictory terms—alongside fastText embeddings, syntactic, semantic, lexical n-gram, sentiment, and punctuation features to improve sarcasm detection. Multiple machine learning (SVM, Multinomial Naïve Bayes, Random Forest), deep learning (CNN, LSTM, MLP), and an ensemble CNN+LSTM model were evaluated, with the CNN+LSTM model incorporating all features achieving the highest accuracy of 92.01% and a kappa of 0.84. Experimental results demonstrate that including oxymoron features enhances sarcasm classification performance across datasets, suggesting their significance in detecting figurative language in text. The study also outlines future plans to expand the dataset, incorporate additional deep learning architectures, and address multilingual sarcasm detection.