I have been working further on the two topics detected by the model. As you might remember, at the folio level we observe a close match with the Currier A/B language distinction. At the paragraph level, the fit is not as strong, but still reasonably good. This is to be expected, since Currier assigned a single topic or language label to an entire folio, whereas at the paragraph level the model operates with finer granularity and can identify paragraphs with different dominant topics within the same folio.
These are the results vs. Currier's languages:
The mixture found suggests that the detected topics might actually represent linguistic styles or sublanguages. If they were truly different languages, they would have to be very close to each other (like Spanish and Portuguese) since they share quite a lot of words, or perhaps they are dialects. But let’s assume they are topics instead. In that case, topic 0 "talks" about A and topic 1 "talks" about B. Temember that for each paragraph, we get a mixture of topics, and from this we can infer or imagine how it "sounds" or what it is "about".
So... how does the manuscript "sound" using topic modelling? This is a graphical experiment. Each paragraph is weighted and the black trend line (a bit smoothed) shows if it tends to topic A or topic B. The result is curious to see:
Following the black line you can see through the Manuscript how the topics detected are shown by paragraph. Note that as the plot is by paragraph, herbal section is much shorter than stars (herbal has 2/3 paragraphs by folio and stars much more).
These are the results vs. Currier's languages:
The mixture found suggests that the detected topics might actually represent linguistic styles or sublanguages. If they were truly different languages, they would have to be very close to each other (like Spanish and Portuguese) since they share quite a lot of words, or perhaps they are dialects. But let’s assume they are topics instead. In that case, topic 0 "talks" about A and topic 1 "talks" about B. Temember that for each paragraph, we get a mixture of topics, and from this we can infer or imagine how it "sounds" or what it is "about".
So... how does the manuscript "sound" using topic modelling? This is a graphical experiment. Each paragraph is weighted and the black trend line (a bit smoothed) shows if it tends to topic A or topic B. The result is curious to see:
Following the black line you can see through the Manuscript how the topics detected are shown by paragraph. Note that as the plot is by paragraph, herbal section is much shorter than stars (herbal has 2/3 paragraphs by folio and stars much more).