Authors: G.Sudheer Kumar, K. Sindhu
Abstract: Sarcasm is a sophisticated form of linguistic expression in which the intended meaning differs from the literal interpretation of the text, making automatic sentiment analysis a challenging task for conventional Natural Language Processing (NLP) systems. News headlines frequently employ sarcastic language to attract readers, convey criticism, or present humorous perspectives, thereby increasing the complexity of accurate text classification. This paper presents an intelligent sarcasm detection framework that integrates Machine Learning (ML) and Deep Learning (DL) techniques to improve the automatic identification of sarcastic news headlines. The proposed methodology utilizes comprehensive text preprocessing followed by feature representation using CountVectorizer and Word2Vec embeddings. Multiple learning models, including AdaBoost, Extreme Gradient Boosting (XGBoost), Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), Bidirectional Long Short-Term Memory (BiLSTM), and a Hybrid CNN–BiLSTM architecture, are trained and comparatively evaluated using the Kaggle Sarcasm Detection News Headlines Dataset. To enhance model transparency, Local Interpretable Model-Agnostic Explanations (LIME) is incorporated for interpreting prediction outcomes, while t-distributed Stochastic Neighbor Embedding (t-SNE) is employed for feature visualization. Experimental analysis demonstrates that the RNN model delivers the highest overall performance, achieving an accuracy of 79.88% together with an F1-score of 0.76, indicating its superior capability for identifying contextual patterns associated with sarcastic language. The proposed framework offers an effective and explainable solution for sarcasm detection and can support numerous natural language understanding applications, including sentiment analysis, opinion mining, intelligent recommendation systems, and social media analytics.
