Utilizing Large Language Models for the Generation of Aspect-Based Sentiment Analysis Datasets

With the rapid development of the field of Aspect-Based Sentiment Analysis (ABSA), the importance of Data Set(DS)labeling for training high-quality ABSA models has become increasingly prominent. In the past, the ABSA DS labeling task was usually handed over to crowd workers on the MTurk platform or expert classifiers, but with the development of Large Language Models(LLMs) in natural Language generation tasks, it shows great potential in data labeling. In this paper, we propose a sentiment analysis framework that automatically acquires labeled MOOC's DS to evaluate the accuracy, reliability, and bias of LLMs in ABSA text analysis tasks compared to manual ones. The results show that ChatGPT is superior to humans in ABSA labeling tasks, the zero-sample accuracy and code-to-code consistency both exceed crowdsourced workers. These findings show that LLMs do have the potential to be used for dataset labeling tasks, and can provide a new feasible idea for future DS labeling research in the ABSA field.