Datasets:
Under-filtering.
UN and EU not banned.
Many modern brands not in the list.
Suggested is to snowball such words: Use current list to collect other keywords in the nearest context, say 100 words around the banned term.
Banned bigrams will improve filtering. Half of 20 century specific bigrans contain of 2 valid 19-century words.
extremely rare in the 19-century words. Google was a valid word before 1900, Latin virus was known, but they only occur in 3-4 books in phrases, which we can add manually as exceptions and filter out all huge modern virusology and "Digitized by Google" garbage.
The banned list started from around 30 words and it's now ~800. Sure I could add way more, I have ~700k more modern words that are almost sure not to exist in vintage books, but I won't eyeball all of them.
The logic for pre-1900 is not perfect, but it's quite decent. Now that we have the updated word lists v2 and bigrams, I could think about this a bit more.
I think these guys : https://huggingface.co/datasets/jbduran/think-dataset-clean-1930s/tree/main/_banned are doing a much better job than me, with their 3-tier filtering.
I think there are also benefits of just replacing modern words with vintage alternatives (if they exist), instead of dropping the whole entry. For now if my filter finds 1 single modern word like Einstein, the whole entry is dropped, and in case of Einstein, I think it should be, but there are definitely modern words that have vintage replacements and that could salvage a bunch of texts.
I'll give it some thought.