PROJECT

THEME

FY2025: Multimodal AI for Kuzushiji Recognition in Classical Books

代表者
Shinichiro Omachi (Graduate School of Engineering)
分担者
  • Yoshiyasu Sakura (Graduate School of Arts and Letters)

INTRODUCTION

AI-Generated Comment on the Study



Many classical Japanese books are written in kuzushiji, a cursive script whose transcription demands specialized expertise and considerable effort. Converting the roughly 300,000 digitized classical books into searchable text has therefore remained a long-standing challenge for humanities research.

This project addresses the challenge by fine-tuning a multimodal large language model—an AI that understands images and text together—specifically for kuzushiji. The approach substantially improves accuracy in both single-character and full-line recognition, and the study shows that preserving the aspect ratio of line images is particularly important. In parallel, an analysis of war tales (gunki monogatari) examines how differences in orthography relate to differences in expressive content, reflecting the project's interdisciplinary bridge between information science and the humanities.

The results lay a foundation for accurate search, deeper understanding of classical texts, and automatic modern-language translation, and may open new perspectives on Japanese cultural and literacy history.




Prompt Used for AI-Generated Comment
Please read the following research summary and write a short comment on the study.
The comment should be suitable for inclusion in an academic research report. It should evaluate the significance, originality, interdisciplinary value, and future potential of the study. Please keep the tone formal, balanced, and positive, but avoid exaggerated claims. The comment should be understandable to a general academic audience and should not be too technical.
Preferred length: 100–150 words.
(Claude)



Research Details



Link: https://www.aisogochi.tohoku.ac.jp/en/archives/1201