21-05-2026, 04:20 AM
(20-05-2026, 04:06 PM)oshfdk wrote: You are not allowed to view links. Register or Login to view.While existing templates don't capture all words and the existing words only, hypothetically, if someone finds a relatively simple set of rules that would cover >95% of existing word types and would only produce < 5% of unregistered word types, is there a possibility to resolve all of this?
There may be, but until one is found, that remains a problem for all "gibberish" theories. Is there a simple enough model that generates the Voynichese words with the observed frequencies? Even if it is not the kind of model that a Medieval author could have devised and used?
For any lexicon and any probability distribution of the words, can build a tree-like automaton with probabilities on the transitions that generates precisely those words with precisely those probabilities. But that model is as complex as a table with the frequency of each word.
To reduce the complexity one could merge sub-trees, creating an automaton which is not a tree but a directed acyclic graph (a DAG). In that model, any state S that can be reached from the root by M different paths and leads to N final states represents a subset of the lexicon whose words consist of M prefixes combined with N suffixes, chosen with various probabilities but independently of each other. How many such subsets can we find in the VMS lexicon? Even allowing for, say, spelling and transcription errors on 15% of the words? I don't expect that such a model can approximate the whole lexicon with the observed frequencies.
Maybe a model that is not a finite state automaton can do better. In my word model, the constraints on the total counts of dealers and benches can be implemented by keeping counters, reducing the size of the corresponding automaton by a large factor. But my model does not predict the word probabilities, only whether words are "valid" or not; and it still allows many "valid" words that don't occur at all.
All the best, --stolfi