Automated Topic Analysis of the Voynich Manuscript

[quimqu]

Intersting!

But it seems that, with all methods, the "topics" are based on the words that are most common overall, which, not coincidentally, are those with least information.  The really interesting topics would be about the conditions that the plants are supposed to treat.

And another unfortunate thing is that the plants in the herbal section are arranged alphabetically, rather than by similarity of purpose.  Thus any topic will be uniformly scattered through the whole herbal section, rather than lumped in interesting ways...

Yes, exactly! That’s the challenge. 

With Culpeper, we know what’s important and what’s filler, so the topics end up reflecting structure or common language — but not always the really interesting stuff like conditions treated. And since the plants are ordered alphabetically, the useful groupings are scattered.

With the Voynich, we don’t even have that baseline. We don’t know which words are meaningful, which ones are structural, or even if there’s a true "meaning" at all in the usual sense. That makes it hard to distinguish signal from noise.

Still, the Culpeper exercise is helpful, I think, as it shows that topic models tend to pick up structure and repetition, not necessarily themes. I can say (tell me if I am wrong), that we can see the "same" thing in the Voynich through its topics clustering by section type. I think it tell us there’s a consistent underlying structure, even if we can’t read it yet.

It’s not a translation, but it’s a start...