OCR Confidences
Hi @Pclanglais and team,
I am heavily working with both PD datasets, and for this Books dataset here I just found some strange dataset entries with weird OCR confidences:
Weird OCR Confidence for: {'file_id': 'bpt6k857075', 'ocr_confidence': '421', 'date': '1771-1777', 'character_count': 640}
Weird OCR Confidence for: {'file_id': 'bpt6k85706t', 'ocr_confidence': '937', 'date': '1771-1777', 'character_count': 932}
Weird OCR Confidence for: {'file_id': 'bpt6k1653770', 'ocr_confidence': '57610', 'date': '1765', 'character_count': 556701}
Weird OCR Confidence for: {'file_id': 'bpt6k1434608', 'ocr_confidence': '50426', 'date': '1897', 'character_count': 454678}
I tested revision 3c8735bd3193aabea12c66e3e4d74e86117de324 of the dataset. Maybe in future versions, the OCR confidences for these books can be fixed :)
Thanks and best,
Stefan
Hi. Thanks a lot for the feedback!
I really think it's a scraping issue: the original sources are in HTML and the first page contains the metadata, including the estimated OCR error rate. Based on the id I'll try to find a fix and regenerate.
On a longer term basis, I plan to implement a standardized estimate: currently we have the one provided by different OCR software at different time period (since the early 2000s) and they are not completely reliable.
Best,
Pierre-Carl