A Hybrid Computational Linguistics Approach to Figurative Language and Sarcasm Interpretation in Text

Abstract

Identifying sarcasm, irony, and figurative language in text is a challenging problem in Natural Language Processing (NLP) because such expressions obscure true sentiments. While transformer-based models are effective, they are computationally expensive. This paper proposes SarcasmNet, a hybrid deep learning ensemble model that combines Long Short-Term Memory (LSTM), Recurrent Neural Networks (RNN), and BERT embeddings to classify tweets into sarcasm, irony, figurative, and regular categories. Leveraging sequential learning from LSTM and RNN along with contextual richness from BERT, our model achieves 73.79% accuracy on a Kaggle dataset of 81,408 tweets, outperforming traditional baselines. Experimental results demonstrate that SarcasmNet effectively captures contextual nuances in sarcastic and ironic statements, making it suitable for sentiment analysis, opinion mining, and misinformation detection. Future work will focus on integrating graph-based reasoning and multimodal sarcasm detection to enhance performance. The experimental findings suggest that SarcasmNet successfully acquires contextual representations of sarcastic and ironic phrases, and a promising tool for sentiment analysis, opinion mining, and misinformation detection.