The Challenge of Analyzing a Dynamic Text

[kckluge]

To question 2) Correlation between text and illustrations:
The shift from Currier A to B suggests that the order of sections was: Herbal in Currier A, Pharma in Currier A, Astro, Cosmo, Herbal in Currier B, Stars in Currier B, Biological in Currier B (see section "Shift from <chol>/<chor> to <chedy>/<shedy>" in You are not allowed to view links. Register or Login to view.). This means with the exception of the pages sharing Herbal illustrations, the scribe wrote pages sharing the same illustration together.

IIRC, you interpret the defining characteristics of Lisa Fagin Davis' scribal hands as reflecting changes in the writing of a supposed single creator over time. Herbal B is a thorn in the side of that theory given that the hands she identifies with Scribe 2 (who writes the Balneology section) and Scribe 3 (who writes the Starred paragraphs section) both occur in the Herbal B folios.

Also, any posited gradual transition from the A language dialect pages to the B language dialect pages requires interpreting the label text of the Zodiac folios -- and only the label text -- as the stepping stones between the two (if you didn't follow the thread, see the discussion in You are not allowed to view links. Register or Login to view.). The non-label text on the Zodiac pages (with one exception) falls firmly on the A language side of what I've come to think of as "Currier gulch" in the bigram distribution space.

In short, the two Herbal sections are devastating to the semantic correlation hypothesis. If vocabulary tracked illustration meaning, both Herbal sections should use similar words. Instead, they differ because they were written at different evolutionary stages. Topic modeling detects vocabulary differences between sections and misinterprets them as "topics" when they're actually just temporal snapshots of an evolving pseudo-text generation process.

I agree fully that the Herbal B dialect is a massive problem for efforts to attribute substantial differences in dialect vocabularies to nothing more than topical differences in the sections. To the extent that I remain agnostic at best about spaces being word separators I share your skepticism about what applying topic modeling is actually measuring in the text. 

It's unfortunate there is no preprint of the Yale work Lisa described in a recent talk to refer to for more detail, but if I understand correctly what they are seeing in their Latent Semantic Analysis results is the same type of transition page-to-page in the Starred paragraph section that is found in texts where there is a narrative/argumentative arc (or that you expect it to show from the self-citation method), but they are not seeing that behavior in the herbal section pages (or in the similar genre texts they used as comparisons). Which is not surprising because there is no sustained narrative flow in something like an herbal or bestiary, it's just discrete descriptions of X, Y, & Z one after another. If that's the case, then that requires explanation in the context of the self-citation method -- why should the LSA results differ between the sections?

I remain unconvinced that the various dialects relect a smooth transition of a single process over time as opposed to discrete distributions with some degree of overlap in their tails, but I'll take a closer look at the material in the You are not allowed to view links. Register or Login to view. link you provided.