Predicting Pronunciation of “Kirakira-Names” Using Language Models

In recent years, unusual names such as “Kirakira-names”, which have attracted attention in Japan, often make headlines due to their unique kanji combinations and pronunciations. These names are born from the parents' special feelings or the aim of differentiating their child from other children, but in daily life, there are many disadvantages such as bullying due to the name, and it is seen as a problem. In order to deal with these problems, the government is planning to revise the Family Registration Act, set naming standards, and require the notification of furigana. However, there is a risk that the local government will register incorrect furigana due to this revision. In this study, we propose a method for accurately predicting how to read names and preventing mispronunciation by constructing a dataset of “Kirakira-names” and fine-tuning multiple language models. This is intended to reduce the risk of misregistration and reduce the burden on the person registering the furigana.