Writing a fake Voynich by rules: copies, habits and the link between words

[npcompl33t]

I just trained a GRU to write voynich, like the one in the article from the other day Torsten shared. I did not give it any information about section, hand , gradients, or anything else.

Yet when i feed it words from specific sections, i found I can train a simple linear model on the GRU's internal memory, and predict where on the gradient it is within about 12%. So it is learning information about the gradient, beyond section (Herbal A is about 40% of the entire curve), even though I'm not explicitly giving it that information.

I ran npcompl33t’s new code (unlabelled, one point per bifolio) on my text, which has no order at all, five times. The overall order it finds for the whole manuscript is almost the same in my text as in the real one (rank correlation 0.86–0.92), so that order needs no chronology. In Q13 it finds an order in my text that is more “secure” than in the real one (0.65–0.76 against 0.54). Only in Q20 is the real text more stable than all five of my runs (0.95 against 0.59–0.89), though two of them come close. So Q20 may hold something my skribe doesn’t make, and that is worth a closer look, but I don’t see it as evidence of time.

A GRU is an AI model that generates text and learns patterns. The point is, his model learns different patterns from the different scribes, which correspond to different points on the curve. So it replicates the horeseshoe. The quote of mine that you used, explains exactly why this occurs.

A horeshoe appearing at all requires a continuous gradient.  Anytime you see a horesshoe, it means there is some strict underlying continuous gradient. However you also need to check the MDS1vs 3 dimensions, because there should be a second order shape there, either an S or a W. if there is not, it is not a Guttman horseshoe. 

It's not at all surprising your model would learn the gradient. As is said, i just trained a model and it learned the gradient without being told. Sections are not continuous, so they wouldn't be forced into a horeshoe. My guess would be your model actually performs poorly at distinguishing within Herbal A, i can test this if you think that would be valuable! Just share your code, i don't see it in your repository. Ultimately without seeing the model i can't say for sure what it is doing! 

@Muaro -- my method got different results than Lisa's, but it based on a much broader uncertainty analysis. Some pages it simply is not possible to determine the exact ordering of. That was why i gave the error brackets. They say in their paper that there are several different possible orderings.