Domain-adaptive training and LLM-based post-processing for improving OCR quality in Korean investigative documents
Journal of Data Forensics Research · 30.06.2026 · 3 (1) · 93-112
DOI: 10.12972/jdfr.2026.3.1.5 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Data Forensics Research, online erschienen am 30.06.2026. Verfasst von Joonseong Kang und 3 weiteren Autor:innen.
Abstract (Originalsprache)
This study combines domain-adaptive training with conservative LLM-based post-processing to improve OCR quality for Korean investigative documents containing both printed and handwritten text. The OCR model is fine-tuned on investigative-document domain data and public Korean OCR datasets, and remaining errors are corrected with a lightweight LLM. To preserve evidentiary strings such as account numbers, phone numbers, URLs, and case numbers, the post-processing corrects only clear errors and suppresses over-correction through rule-based checks for ending changes and number insertions. Evaluated against manually transcribed and verified ground truth, the proposed method reduces character and word error rates by more than half compared with the base model, achieving CER 7.14% and WER 11.52%. These results show that the proposed approach effectively improves OCR for Korean investigative documents.
Zitieren
APA
Kang, J., Park, M., Choi, H., & Song, K. (2026). Domain-adaptive training and LLM-based post-processing for improving OCR quality in Korean investigative documents. Journal of Data Forensics Research, 3(1), 93-112. https://doi.org/10.12972/jdfr.2026.3.1.5
Vancouver
Kang J, Park M, Choi H, Song K. Domain-adaptive training and LLM-based post-processing for improving OCR quality in Korean investigative documents. Journal of Data Forensics Research. 2026;3(1):93-112. doi:10.12972/jdfr.2026.3.1.5
BibTeX
@article{Kang2026,
author = {Kang, Joonseong and Park, Minhoi and Choi, Hosik and Song, Kyungwoo},
title = {Domain-adaptive training and LLM-based post-processing for improving OCR quality in Korean investigative documents},
journal = {Journal of Data Forensics Research},
year = {2026},
volume = {3},
number = {1},
pages = {93--112},
doi = {10.12972/jdfr.2026.3.1.5}
}
EUVIMED