A Study of Using GPT-3 to Generate a Thai Sentiment Analysis of COVID-19 Tweets Dataset

Abstract

This study evaluated the effectiveness of using synthetic text datasets generated by GPT-3 for sentiment analysis with deep learning models, namely Bi-GRU and Bi-LSTM. The study compares the performance of these models on both synthetic text and Label Tweet datasets using GPT-3 and reveals that deep learning model performance is dependent on the dataset's nature. The results indicate that using synthetic text generated by GPT-3 significantly enhances the accuracy of both models, with Bi-LSTM achieving an accuracy of 0.84 and Bi-GRU achieving an accuracy of 0.85. The study underscores the importance of meticulous dataset selection and preparation for developing precise and effective deep learning models for various sequential data types. The findings demonstrate that synthetic text datasets generated by GPT-3 can serve as a valuable resource for developing deep learning models as they are labeled and save researchers time and effort in manual labeling of large datasets.