Establishing Ground Truth
Yesterday and today, I spent some hours manually transcribing pages from the 1674 print of the Ethica Complementoria. The edition, printed in Copenhagen by Christian Wering and published by Wolf Lambrecht, survived in only one copy, held by the State and University Library Hamburg. The copy has been digitised a while ago and can be accessed and downloaded freely from their website. I have argued in my book about the print and transmission history of the Ethica Complementoria that this edition is the source for the Danish translation published in 1678 in Copenhagen. I’m preparing a bilingual digital scholarly edition of the German and the Danish Ethica.
However, with the print being digitised and available, you might wonder why I would manually transcribe it! Actually, I don’t plan to transcribe the entire print manually. I am using Transkribus to do the work for me. I have already established a transcription for the Danish print using the Transkribus app and the NorFraktur HTR model provided by the National Library of Norway. The performance was OK; however, there’s no way around a manual quality check when preparing a transcription for a scholarly edition. The NorFraktur model was trained on multilingual materials: 77 Danish-Norwegian and 23 German small blackletter prints from the mid-16th to the mid-17th century. With a CER of 2.0%, it’s quite good. However, with our print from the late 17th century in Danish, we found it underperformed a bit. We decided to improve the model by training it on manually transcribed pages from the digitised copy held by the Norwegian University of Science and Technology library. This improved the quality slightly, but a few issues remained. Since we were going for a quality check of the entire transcription anyway, we corrected all ‘misreadings’. A couple of pages are left; the first round of manual quality checks will be done by the end of September!
For the German print, we initially thought of using NorFraktur, too. However, there are now models for German print which perform significantly better than NorFraktur. Results will be significantly better. For what it’s worth, I will compare two approaches: Firstly, testing the quality of a German model, like the Transkribus Print Multi-Language model, with a CER of 1.6%. Secondly, I will train my model with manually transcribed pages to see if it can outperform the Transkribus Print model: for only this particular print.
Once I have the results, I will report back!