I've been looking at the Voynich word graph (in CUVA, but I'm showing the examples in EVA here) and comparing it to various natural texts, and I think the difference is not so much in whether there is a giant component, but in how it is organized internally.
In the Voynich, almost all of the vocabulary is contained within a single giant component. This means that almost any word can be connected to another by very small changes (maximum 2). In natural texts this is also true for the most frequent tokens in the corpus, but it is not usually the case for almost all of the vocabulary (set of tokens).
When you look at the internal structure of the Voynich, three large, very dominant groups emerge. They are not small groups, but blocks that concentrate the majority of words and are very densely connected. A first group revolves around shapes like chedy, chey, qokeedy, qokeey, okey, with many small variations especially at the end of the word. A second group is that of daiin and variant as aiin, dar, dal, okaiin, otaiin, where there are many prefix changes and small substitutions. The third group is much more compact and short, with forms like ol, chol, or, cho, chor, which are combined almost systematically.
| Group | EVA Examples | Comment |
|---|
| 1 | chedy, chey, qokeedy, qokeey, okey | Variation mainly in endings, many very close forms |
| 2 | daiin, aiin, dar, dal, okaiin | Prefix changes and small substitutions |
| 3 | ol, chol, or, cho, chor | Very compact system with short words |
Furthermore, these groups are not completely separate. There are very short words like y, dy, or, al, ar that act as a bridge between different areas of the graph, as if they were highly reusable pieces.
When you compare this with natural texts, the picture is different. There are also large giant components, especially in long texts, but the vocabulary is not that much absorbed by the giant component and, above all, the internal part is much more fragmented. Instead of a few large blocks, many smaller communities appear.
To see it better, this is a comparison with the corpora I tested, all normalized to max 50,000 tokens:
| Corpus | Tokens | Vocab | % vocab in giant | % tokens in giant | Comunities |
|---|
| Voynich | ~38k | ~7.8k | ~94% | ~98% | 10 |
| Tirant lo Blanch | 50k | 5361 | 79.0% | 94.6% | 87 |
| Culpepper | 50k | 4486 | 67.4% | 91.2% | 50 |
| De docta ignorantia | 37121 | 5974 | 58.5% | 84.1% | 68 |
| Plató (Apologia) | 8437 | 2754 | 59.3% | 82.2% | 32 |
| Portuguès antic | 4053 | 1387 | 60.7% | 81.7% | 18 |
The Tirant is the closest case to the Voynich in percentage, especially in tokens, but there are still clear differences. Although 79% of the vocabulary is connected, the network is divided into many more communities, more than eighty, and not into three or four dominant blocks. In addition, there are many more words that fall outside the giant component.
The general feeling is that in natural language you have many relatively independent word families, whereas in the Voynich it seems that almost all the vocabulary is built from a few very large nuclei, with small and very repetitive variations. It seems not so much a question of how connected it is, but of how compact and homogeneous this system is.
quimqu