Contextual Understanding and Reprocessing in Sarcasm Detection: a Study of BERT Vs Logistic Regression

Detecting sarcasm in natural language presents distinct challenges, primarily due to its reliance on contextual clues and the use of subtle, contradictory expressions. This research examines how the BERT model and Logistic Regression (LR) perform in detecting sarcasm, emphasizing the importance of contextual understanding and various preprocessing methods. Using a social media dataset, we evaluate both models on sentiment analysis tasks after applying different preprocessing methods including tokenization, elimination of stop words, and noise reduction. We assess the models' performance using evaluation metrics like F1 scores and confusion matrices to determine their accuracy in recognizing sarcastic expressions. Our findings suggest that although BERT's bidirectional architecture provides a deeper understanding of linguistic context, it does not always outperform the simpler LR model when noise reduction preprocessing is applied. This indicates that preprocessing techniques can affect model performance in varying ways, and selecting NLP models should align with the unique needs of the sentiment analysis objective. This study provides meaningful perspectives on the role of contextual models and preprocessing methods in improving the precision of sarcasm detection.