OCR of historical printings with an application to building diachronic corpora: A case study using the RIDGES herbal corpus

This article describes the results of a case study that applies Neural\nNetwork-based Optical Character Recognition (OCR) to scanned images of books\nprinted between 1487 and 1870 by training the OCR engine OCRopus\n[@breuel2013high] on the RIDGES herbal text corpus [@OdebrechtEtAlSubmitted].\nTraining specific OCR models was possible because the necessary *ground truth*\nis available as error-corrected diplomatic transcriptions. The OCR results have\nbeen evaluated for accuracy against the ground truth of unseen test sets.\nCharacter and word accuracies (percentage of correctly recognized items) for\nthe resulting machine-readable texts of individual documents range from 94% to\nmore than 99% (character level) and from 76% to 97% (word level). This includes\nthe earliest printed books, which were thought to be inaccessible by OCR\nmethods until recently. Furthermore, OCR models trained on one part of the\ncorpus consisting of books with different printing dates and different typesets\n*(mixed models)* have been tested for their predictive power on the books from\nthe other part containing yet other fonts, mostly yielding character accuracies\nwell above 90%. It therefore seems possible to construct generalized models\ntrained on a range of fonts that can be applied to a wide variety of historical\nprintings still giving good results. A moderate postcorrection effort of some\npages will then enable the training of individual models with even better\naccuracies. Using this method, diachronic corpora including early printings can\nbe constructed much faster and cheaper than by manual transcription. The OCR\nmethods reported here open up the possibility of transforming our printed\ntextual cultural heritage into electronic text by largely automatic means,\nwhich is a prerequisite for the mass conversion of scanned books.\n