Ensemble Method for OCR Error Handling in Early-Modern Japanese Texts
Abstract Recently, the digitization of books has been actively promoted to improve accessibility. However, early modern Japanese books published from the Meiji to early Showa era (1868--1935) exhibit numerous characteristics that differ from contemporary publications. Current research continues to advance specialized character--recognition systems for these books. Nevertheless, recognition errors remain prevalent. In this study, we introduce a method for textual error detection and correction in Optical Character Recognition (OCR)--processed early--modern Japanese texts. The proposed approach employs an ensemble framework that integrates a Bidirectional Gated Recurrent Unit (Bi--GRU) model and a dictionary for error detection, alongside BERT for error correction. Evaluation experiments demonstrate that our method achieves an F--measure of \((74.9%)\), outperforms baseline methods, and effectively identifies erroneous characters with high accuracy while minimizing false detections.
Authors
- Masami Takata (ORCID: https://orcid.org/0000-0003-3475-7565)
- Miho Chiyonobu (ORCID: https://orcid.org/0009-0000-8043-5079)
- Yu Ishikawa
- Hiyori Kanetaka
- Yuki Takemoto
Institutions
- Shiga University (JP)
- Kyoto Women's University (JP)
- Nara Women's University (JP)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-12
- DOI
- https://doi.org/10.21203/rs.3.rs-11275171/v1
- Primary Topic
- Handwritten Text Recognition Techniques
- Type
- preprint
Funders
- Japan Society for the Promotion of Science