Few-Shot and Zero-Shot Classification with Large Language Models: A Comparative Study

This study systematically compares the performance of large language models in zero-shot and few-shot text classification tasks. Experiments were conducted on a balanced Turkish news dataset, evaluating five different prompting strategies using the ChatGpt model: label list only (zero-shot), one example per label (one-shot), five examples (five-shot), ten examples (ten-shot), and label summaries (summary). The results indicate that increasing the number of examples per label positively affects overall performance, while summary-based prompts emerge as an effective alternative, offering both high accuracy and lower cost. According to macro and micro F1 scores, the highest performance was achieved with the five-shot approach, whereas the summary method demonstrated remarkable effectiveness in zero-shot settings. These findings highlight the critical role of prompt length, content diversity, and contextual clarity in the classification performance of LLM-based system.