In recent years, with the explosive growth of user-generated content in online spaces such as social media and review platforms, automated text analysis has become crucial for corporate marketing and understanding societal trends. Sarcasm poses a significant challenge for sentiment analysis and opinion mining, as its surface meaning often contradicts its true intent, substantially impairing accuracy. To address this issue, we propose the Multi-Scale Convolutional Fusion with Contrastive Feature Alignment (MSCF-CFA), a model designed to capture semantic discrepancies while effectively handling imbalanced data classification. The model integrates Multi-Scale Convolutional Fusion (MSCF) to capture dependencies at various scales, Contrastive Feature Alignment (CFA) to enhance class-specific feature representation using contrastive learning, and adjustments to the loss function, aiming to effectively detect the complex patterns unique to sarcasm while addressing data imbalance. Experiments using a sarcasm detection dataset demonstrated that MSCF-CFA outperformed existing methods and models designed to handle imbalanced data across all evaluation metrics. Furthermore, feature visualizations and ablation studies confirmed that the MSCF, CFA, and loss function adjustments significantly enhance class separation, making them critical components of the model’s effectiveness. This study highlights the efficacy of multi-layered feature extraction and contrastive learning for advanced language understanding tasks, including sarcasm detection, and opens new possibilities for further applications.