Transformer-Based Sarcasm Detection in News Headlines Using BERT

Sarcasm detection is a challenging problem in sentiment analysis because sarcastic expressions often convey meanings that differ from their literal wording. Detecting sarcasm is important for improving the reliability of opinion mining, social media analysis, and customer feedback interpretation. This study proposes a transformer-based sarcasm detection system using the News Headlines Sarcasm Dataset. A baseline model using TF-IDF features with Logistic Regression is first implemented to establish a reference performance, achieving an accuracy of 82.84%. Subsequently, a fine-tuned BERT (Bidirectional Encoder Representations from Transformers) model is applied to capture contextual relationships and semantic dependencies within headlines. The dataset is preprocessed and divided into training, validation, and testing sets to ensure reliable evaluation. Experimental results show that the proposed BERT-based model achieves an accuracy of 93.85%, significantly outperforming the baseline model. The results highlight the effectiveness of transformer-based contextual language models for identifying subtle linguistic patterns in sarcastic text.