Sarcasm Detection and Classification in Kannada Language Using Machine Learning Techniques on a Manually Annotated Dataset

Abstract

Sarcasm, a form of expression where the intended meaning contradicts the literal wording, presents significant challenges for computational interpretation, especially in regional languages like Kannada. Accurate sarcasm detection is vital for applications such as sentiment analysis, opinion mining, and social media monitoring. This study tackles the challenges of detecting sarcasm in Kannada by utilizing advanced techniques. A comparatively large and diverse dataset comprising 7000 annotated sentences is larger compared to previous studies. Standard text preprocessing steps such as data clean ing, tokenization, stop word removal, stemming, and part-of-speech tagging were applied to the dataset. Several machine learning and deep learning models were then trained and evaluated, including SVC, SGD Classifier, Multinomial Naive Bayes, Random Forest, Linear SVC, Logistic Regression, XGBoost, and BERT. Compared to all these techniques, the Logistic Regression model achieved the highest accuracy of 67.4%, highlighting the challenges in capturing the nuances of sarcasm in Kannada. Although the performance is modest compared to smaller-scale studies, the results underscore the importance of dataset size and diversity in model robustness. Future research should focus on exploring more advanced architectures, such as transformer-based models and recurrent neural networks, along with further expansion of annotated datasets to improve accuracy and generalizability in regional sarcasm detection.