Banded quality values by year diagram.
The development looks suspiciously smooth, which it would not be in reality since typefaces do not change gradually. I suspect that the only thing the graph really shows is language slowly conforming to the standards of the dictionary.
In ENO we have a custom dictionary for the period around 1800. When we run it forward in time, it looks the same. Gradual decline (so the opposite direction). What you want to see, if you want to assess actual quality of text recognition, is something more uneven. So you should consider a dictionary that includes more history.
Also, the degree of wrong matters in some contexts – especially for processes dealing with sub-word tokenizers. So an approach could be to measure how many changes are needed on average to make words conform to closest match in the dictionary. This will reveal passages that are totally garbled vs. what is half-right. This might be important also because there might be a point at which it is half-right enough to be fixable with an LLM.
Also, word length and character can serve you as proxies. Two tell-tale signs of garbled OCR (of the type you have here at least): wildly inconsistent word length and a different distribution of characters across the alphabet from clean texts. Both would require cleaner corpora to compare with. You have that in ENO (where you can just filter aggressively on pwa), Grundtvig and Trykkefriheds-korpus.
Two other things, relating to the comparison with ENO.
I like the general idea of comparing based on the two, but there is a caveat. ENO has a lot more proper nouns than typical book genres. These are not in the dictionaries. These include places for the news reporting, and a LOT of personal names for advertisements, adresses etc.. I would suggest that you only compare the book corpus to news articles. It would look more equivalent in terms of both quality and distribution of word types etc. You can filter on this in the version that Alie created in the chcaa repo. And maybe consider actually modifying the dictionary with common proper nouns. Some of these can be found in various geoname-ressources and personal names can be found from censurs data.
I think you should filter on the pwa-variable from the original ENO-release rather than the dynaword indicator. Maybe both. While the latter is generally meaningful, it also has an anachronistic bend. The pwa is based on a custom dictionary (that does not fit after 1850), and if you take only texts with pwa above 97% the texts are generally really clean, in terms of looking like something from the period that birthed them.
And another thing: one of the texts you references as above the ref line is quite broken:
Kongernes Konge, Du Herrernes Herre! Dybt boie vi os for Dig, som holder Almagtens Scepter i DinHaand; men i det vi boie os for Dig med dybeste LErefrygt, hceve vi ogsaa Blikket til Dig med den barnligste Tillid, fordi vi vide, at Du, paa hvis Vink Troner henveires og Herskerspire knuses, er Algodhedens Gud, som Du er Alvceldens. Men derfor tor vi ogsaa for Dig, barmhjertige Fader, udtale vore Klager,
This has 9-10 errors on word level. And it certainly not something that I would expect to be above a ref defined by the best parts of ENO. Based on the ENO test, this would have a pwa in the 80s (doing the math in my head, but something like that). So this sort of suggests the kind of dictionary issue, I outlined previously.
Thanks @JohanHeinsen , we will def. take a look at this. It is not on the very top of the list though (sadly)