Detecting Sarcasm in Tweets: A Comparative Study of Deep Learning and Traditional Approaches

Detecting sarcasm in social media text remains a challenging task due to the informal, brief, and contextdependent nature of user-generated content. This study proposes a deep learning framework based on Convolutional Neural Networks (CNNs) to effectively identify sarcasm in tweets. The model was trained and evaluated on a balanced dataset of $\mathbf{2 0, 0 0 0}$ manually labeled tweets, including informal language elements such as emojis, hashtags, and slang. The architecture leverages $\mathbf{2 0 0}$-dimensional word embeddings, a convolutional layer with 128 filters, and dropout regularization to extract both local and semantic features. Comparative analysis was performed against traditional machine learning classifiers-Naive Bayes (NB) and Support Vector Machines (SVM)-using the same dataset. The CNN-based model significantly outperformed the baselines, achieving an F1-score of 0.99, compared to 0.78 for NB and 0.76 for SVM. Evaluation was conducted through a train-validation-test split and further validated using a confusion matrix, which demonstrated high accuracy and low false positive rates. Despite its strong performance, the model exhibited signs of overfitting, as evidenced by the divergence between training and validation accuracy, indicating the need for improved regularization or more diverse data. The findings underscore the limitations of traditional methods in capturing contextual cues and highlight the robustness of CNNs in modeling sarcasm. This work contributes to sentiment analysis, brand monitoring, political discourse mining, and content moderation by providing a scalable and accurate sarcasm detection approach tailored to the dynamics of social media platforms.