A Computationally Efficient Neural Network for Sarcasm Detection in News Headlines

Sarcasm detection is crucial for understanding nuanced human language in applications like sentiment analysis and content moderation, yet computationally challenging due to reliance on context over explicit markers. This study develops an efficient deep learning model for sarcasm detection using a dataset of news headlines. We employ text preprocessing techniques including lowercasing, stopword removal, tokenization, and sequence padding. Pre-trained word embeddings represent semantic features. The model architecture utilizes an embedding layer followed by Global Max Pooling and three dense layers with ReLU activations and dropout for regularization. A final sigmoid layer performs binary classification. The model is trained using binary cross-entropy loss and the Adam optimizer. Performance is evaluated using accuracy, precision, recall, and $\mathbf{F 1}$-score. This research aims to enhance AI's ability to grasp subtle communication, contributing to more robust NLP systems.