Hi Michelle,
I hoped to find the time to make a few experiments about this, but I didn't. I am not an expert and I may have misunderstood something. Anyway, here are some thoughts:
1. As pointed out by Anton, the best fit found by Jaskiewicz is not a Caucasian language but a Romance one (Moldavian / Romanian).
2. I also agree with Anton about the fact that the huge differences between the five best matches show that the method is unstable. This is understandable, because the method is extremely simple (see next point).
3. Jaskiewicz has addressed one of the problems that were ignored by Darrin: we don't know how to match Voynichese characters with characters in other alphabets. His solution is quite simple: sort characters by decreasing frequency, so that you compare the most frequent character in language A with the most frequent character in language B, the second most frequent with the second most frequent and so on. For instance, you would compare the frequency of EVA:o with Greek:ν, English:e, Latin:i (these can vary depending on the specific reference text). He than computes the overall difference between character histograms.
4. It is not clear how he treats diacritics (for Darrin 'à' and 'a' are the same). Here, no choice is optimal. One should experiment with various choices for all languages. In principle, one should also use different texts for each language.
5. As David pointed out, Jaskiewicz uses a single Voynich transliteration (v101). For a study like this, one should include several different transliterations.
6. Jaskiewicz wrote that in computational linguistics "a distribution of pairs and triples of letters may be used to automatically recognize a language of an unknown document". Indeed Darrin's trigram-based analysis could be more robust than Jaskiewicz's approach. If Darrin incorporated the kind of character ordering proposed by Jaskiewicz, his method would probably be improved (something along these lines was suggested by Jonas in the other thread). Anyway, points 4 and 5 should also be addressed, at least by experimenting with different options. It is also true that, as others said, there are strong factors that suggest that the VMS is not a direct phonetic encoding / simple substitution, so these exercises are not very likely to find the underlying language (if there is any and if it isn't an otherwise undocumented artificial language, as Friedman concluded).
I hoped to find the time to make a few experiments about this, but I didn't. I am not an expert and I may have misunderstood something. Anyway, here are some thoughts:
1. As pointed out by Anton, the best fit found by Jaskiewicz is not a Caucasian language but a Romance one (Moldavian / Romanian).
2. I also agree with Anton about the fact that the huge differences between the five best matches show that the method is unstable. This is understandable, because the method is extremely simple (see next point).
3. Jaskiewicz has addressed one of the problems that were ignored by Darrin: we don't know how to match Voynichese characters with characters in other alphabets. His solution is quite simple: sort characters by decreasing frequency, so that you compare the most frequent character in language A with the most frequent character in language B, the second most frequent with the second most frequent and so on. For instance, you would compare the frequency of EVA:o with Greek:ν, English:e, Latin:i (these can vary depending on the specific reference text). He than computes the overall difference between character histograms.
4. It is not clear how he treats diacritics (for Darrin 'à' and 'a' are the same). Here, no choice is optimal. One should experiment with various choices for all languages. In principle, one should also use different texts for each language.
5. As David pointed out, Jaskiewicz uses a single Voynich transliteration (v101). For a study like this, one should include several different transliterations.
6. Jaskiewicz wrote that in computational linguistics "a distribution of pairs and triples of letters may be used to automatically recognize a language of an unknown document". Indeed Darrin's trigram-based analysis could be more robust than Jaskiewicz's approach. If Darrin incorporated the kind of character ordering proposed by Jaskiewicz, his method would probably be improved (something along these lines was suggested by Jonas in the other thread). Anyway, points 4 and 5 should also be addressed, at least by experimenting with different options. It is also true that, as others said, there are strong factors that suggest that the VMS is not a direct phonetic encoding / simple substitution, so these exercises are not very likely to find the underlying language (if there is any and if it isn't an otherwise undocumented artificial language, as Friedman concluded).
![[+]](https://www.voynich.ninja/images/collapse_collapsed.png)