Key points are not available for this paper at this time.
This study investigated the potential of enhancing the performance of text classification by augmenting the training dataset with external knowledge samples generated by a generative AI, specifically ChatGPT. The study conducted experiments on three models - CNN, HiSAN, and BERT - using the Reuters dataset. First, the study evaluated the effectiveness of incorporating ChatGPT-generated samples and then analyzed the impact of various factors such as sample size, sample variability, and sample length on the models’ performance by varying the number, variety, and length of the generated samples. The models were assessed using macro and micro-averaged scores, and the results revealed that the macro-averaged scores improved significantly across all three models, with the BERT model showing the greatest improvement (from 49.87% to 65.73% in macro F1 score). The study further found that adding 30 distinct samples produced better results than adding 6 duplicates of 5 samples, and samples with 150 and 256 words had similar performance, while those with 50 words performed slightly worse. These findings suggest that incorporating external knowledge samples generated by a generative AI is an effective approach to enhance text classification models’ performance. The study also highlights that the variability of articles generated by ChatGPT positively impacted the models’ accuracy, and longer synthesized texts convey more comprehensive information on the subjects, leading to higher classification accuracy scores. Additionally, we conducted a comparison between our results and those obtained from EDA, a widely used data augmentation generator. The findings clearly demonstrate that our method surpasses EDA and offers additional advantages by reducing computational costs and solving zero-shot problem. Our code is available on GitHub. 1
Zhao et al. (Sat,) studied this question.