Abstract
In this study, we performed the task of simplifying standard news broadcasts by NHK into plain Japanese using ChatGPT. Specifically, we quantitatively analyzed four types of news articles: 1) standard news, 2) news simplified by the GPT-4 model, 3) news simplified by the GPT-4o model, and 4) 'Easy Japanese News' crafted by humans, based on the readability analysis results. The analysis revealed that the articles generated by GPT-4 were the most readable, while the standard news was the least readable. Furthermore, the text generated by GPT-4o was closer to the 'Easy Japanese' created by humans. Additional analysis of textual features indicated that GPT-4o simplified the text by adjusting factors such as average sentence length, while faithfully reflecting the lexical attributes of the standard news. These results suggest that generative AI, through large language models, is not only improving in surface task performance but also evolving towards human-like reasoning and language generation capabilities.
https://ror.org/00ntfnx83 
