Leveraging Zero-Shot Learning in Large Language Models for Sentiment Analysis: A Comparative Study on the Indonesian Language

This research assesses the performance of large language models (LLMs), specifically GPT-3.5 and GPT-4o mini, in sentiment analysis of Indonesian compared to traditional models like IndoBERT and DistilBERT. Using the IndoLAM dataset, we benchmarked these models on precision, recall, and F1 score metrics, focusing on their zero-shot learning capabilities. Our results show that GPT-3.5 achieved the highest F1 score of 0.83, reflecting a strong balance between precision (0.81) and recall (0.86). GPT-4o Mini excelled in precision (0.88) but had a lower recall (0.72), resulting in an F1 score of 0.76. In contrast, traditional models, particularly DistilBERT, exhibited weaker performance with an F1 score of 0.59. Additionally, we applied these models to analyze sentiment during the 2024 Indonesian presidential election using unlabeled Twitter data, demonstrating their effectiveness in real-world zero-shot scenarios. The study highlights the superiority of LLMs in contextual understanding and their capability to handle complex linguistic tasks with minimal labeled data, contributing to the discourse on the practical applications of LLMs in sentiment analysis, especially in linguistically diverse contexts.