Zero-Shot and Few-Shot Learning for Telugu News Classification: A Large Language Model Approach

Abstract

Telugu news classification involves categorizing news articles written in the Telugu language into five categories: business, editorial, entertainment, nation, and sport. Zero-shot classification enables categorization into unseen classes without any prior training on those specific classes. Few-shot classification generalizes with minimal training data by leveraging a limited number of labeled examples per class. This work leverages Large Language Models (LLMs) such as mBERT, Indic-BERT, XLM-Roberta, and Flan-T5 Large for zero-shot classification. For few-shot classification, models like Flan-T5 and BART are used in 1 -shot, 3 -shot, and 5 -shot scenarios. Traditional machine learning models including Support Vector Machines (SVM), Naive Bayes, AdaBoost, Random Forest, and Logistic Regression are also explored for comparison. Prompt engineering is employed with various prompts to enhance zero-shot classification performance. Among the LLMs, mBERT achieves the highest F1-score of 0.58 in zero-shot classification. For few-shot classification, Flan-T5 achieved an F1-score of $\mathbf{0. 2 3}$.