Potentials of ChatGPT for Annotating Vaccine Related Tweets

Abstract

This study evaluates ChatGPT's performance in annotating vaccine-related Arabic tweets by comparing its annotations with human annotations. A dataset of 2,100 tweets representing various factors contributing to vaccine hesitancy was examined. Two domain experts annotated the data, with a third resolving conflicts. ChatGPT was then employed to annotate the same dataset using specific prompts for each factor. The ChatGPT annotations were evaluated through zero-shot, one-shot, and few-shot learning tests, with an average accuracy of 82.14%, 83.85%, and 85.57%, respectively. Precision averaged around 86%, minimizing false positives. The average recall and F1-score ranged from 0.74 to 0.80 and 0.65 to 0.93, respectively. AUC for zero-shot, one-shot, and few-shot learning was 0.79, 0.80, and 0.83. In cases of ambiguity, both human annotators and ChatGPT faced challenges. These findings suggest that ChatGPT holds promise as a tool for annotating vaccine-related tweets.