Deep Convolutional Neural Network-Based Sarcasm Detection and Sentiment Analysis on Social Media

The rapid growth of social media platforms has led to an abundance of user-generated content, posing significant challenges for sentiment analysis, especially in detecting sarcasm. Traditional sentiment analysis methods often struggle to detect sarcasm, as it involves communication where the intended meaning differs from the literal interpretation. This paper proposes a dual-task Convolutional Neural Network (CNN)-based sarcasm detection approach that integrates sarcasm recognition with sentiment classification. The proposed model employs context-aware feature extraction, pre-trained word embeddings (GloVe), and an attention mechanism to identify sarcasm-indicative words and phrases. The model captures both local and global contextual information by combining multi-layer convolutional filters and hierarchical feature fusion techniques. Experiments conducted on Twitter and Reddit datasets demonstrate that the proposed model outperforms traditional CNN, LSTM, and BERT-based sentiment classifiers and achieved notable improvements in F1-score for sarcasm detection and overall sentiment classification accuracy. This approach provides a scalable, robust solution for applications like opinion mining, brand monitoring, and user sentiment tracking on social media.