Publication

Abstract : This paper will involve a comparative evaluation of the traditional machine-learning models and transformer-based models to classify Kannada news headlines. In the article, Kannada headlines were processed and normalized, and then, they were coded with TF-IDF and trained on five classic classifiers: Logistic Regression, random forest, Decision Tree, Multinomial Naive Bayes, and Linear Support Vector Classifier (Linear SVC). Simultaneously, a BERT model that had been trained on a corpus of relevance was fine-tuned to the target corpus and tested. Empirical findings show that the Linear SVC model will have the best accuracy among the classical models, and the fine-tuned BERT model will far outperform any other traditional methods. These results indicate how transformer architectures can be used with low-resource Indian languages like Kannada.