USASD: A SENTIMENT AND SARCASM ANALYSIS TECHNIQUE FOR URDU LANGUAGE SCRIPT

Abstract

Human beings are accustomed to expressing positive or negative comments in their language. The beauty of natural language is in its crispy sarcastic way of expression. Traditionally, sarcasm detection and sentiment analysis are treated as entirely different tasks in Natural Language Processing (NLP), though their relationship is crucial to understanding the contextual meaning. Urdu's richness in morphology and complicated syntax would indeed create something unique in that area. This paper presents a hybrid deep learning model named USASD (Urdu Sentiment Analysis and Sarcasm Detection) incorporating BERT, CNN, LSTM, and fine-tuned T5 for Urdu sentiment analysis and sarcasm detection. It provides contextual embedding, CNN captures the local text pattern, LSTM processes sequential dependence, and T5 puts it in a unified text-to-text paradigm. Pretrained on Urdu datasets fine-tuned using a custom-annotated Urdu tweets dataset, USASD correctly identifies sentiment polarity and nuances of sarcasm with 89% accuracy in sentiment analysis and 83% in sarcasm detection against the XGBoost and XLM-RoBERTa baseline. This work advances NLP for low-resource languages, thus allowing for better text analysis in Urdu for social media monitoring, opinion mining, and multilingual NLP research.