2026-09-23 Full-Text Search Data Updated (Change of OCR Engine)
We have revised the OCR (optical character recognition) engine used to generate the text data for this database's full-text search, and have replaced part of the published text data accordingly.
Previously, both line detection and character recognition relied on NDL Kotenseki OCR-Lite. We have now replaced the character-recognition component with PaddleOCR-VL, while keeping the same line detection, and confirmed that this substantially reduces the character error rate. Based on these results, we have replaced the full-text search data with the output of re-running OCR with PaddleOCR-VL, for all ranges except those where proofreading has already been completed, as listed below.
Excluded (proofreading already completed; not included in this update):
- 001-01 to 015-08 (including some as-yet unpublished items)
- 177-03 to 177-05
For all other ranges, the full-text search results are now based on the new text data. As this text is the unproofread output of automatic character recognition, it may still contain errors.
For details of the comparison and evaluation of the OCR engines, please see the following article (in Japanese):
「TEI の校正履歴を正解データにして、PaddleOCR-VL と NDL古典籍OCR を比べる」 ["Comparing PaddleOCR-VL and NDL Kotenseki OCR Using TEI Proofreading History as Ground Truth"]