[Article] Topic Modeling in the Voynich Manuscript

[Torsten]

I especially like the figures 6, 8, 10, 12, and 14. The figures clearly visualize the gradual evolution of a single system from "state A" to "state B" as suggested in Timm & Schinner (see Timm & Schinner 2020, p. 6).

The results of the LDA topic clustering method "created one dominant topic, with the majority of pages belonging to topic 2" (after figure 3 the topic 2 belongs to Currier B). This way "there is a clear distinction between Language A and B" (Sterneck et. al. 2021, p. 7). The reason for this result is that "LDA learns the topic representation of each document, as well as the words associated with each topic" (Sterneck et. al. 2021, p. 2). Words typical for Currier A also exist in Currier B, but not the other way round. Therefore it is possible to distinguish Currier B from Currier A based on frequency counts of tokens containing the sequence "ed". It is therefore reasonable to assume that the LDA algorithm has learned words typical for Currier B like "chedy", "shedy", "qokeedy" etc.

A result of the NMF topic clustering method is "The star paragraphs are close to NMF topic 1 and the visual balneological topic, but it also is close to NMF topic 6." This confirms that "the starred paragraphs and balneological may have overlapping content as distinct topics" (Sterneck et. al. 2021, p. 11). 

Another result is "NMF topics 0 and 4 are almost identical with the botanical section close by. It’s interesting to see NMF topics 0 and 4 nearly collapse into one topic" (Sterneck et el, p. 11). Again it become apparent that on a smaller scale the "topics" seem to overlap, whereas on a larger scale the massive differences between Currier A and B are immediately apparent. Such results demonstrate that it not that easy to find a clear topic structure in the manuscript.

However, Sterneck et al interpret the results differently than I do. For instance they dismiss the results of the LDA analyses since "the LDA topics don’t seem to cluster in any significant way" (Sterneck et al 2021, p. 7). 

The authors themselves warn because of the existence of two different 'languages' or 'dialects' A and B: "From a statistical perspective, this discovery has complicated our understanding of the Voynich Manuscript because topic modeling relies on word frequencies and expects consistency across texts" (Sterneck et al 2021, p. 4). But if this is the case their conclusion is contradicting itself: "We find that computationally derived clusters match closely to a conjunction of scribe and subject matter (as per the illustrations)" (Sterneck et al 2021, p. 1). Either the topics are a result of scribal differences or a result of subject changes. Therefore it is at least problematic to use two contradicting explanations together.