There are Indo-European cognate pairs that are phonetically exact and regular in the sense that their phonematic make-up is completely explained by systematic application of the relevant sound rules to a reconstructed PIE form that does not involve any consideration of paradigmatic levelling, contamination, irregular assimilations etc.

An example of such a cognate pair would be work and έργο, which are connected somehow like this:

work < weorc < *werk < *werką < *wérǵom > *wérgon > ἔργον > έργο

(The point being, every stage in the above derivation uses sound shifts that might be considered 'exceptionless' in the analysis of this pair, and the derivation is exhausted by such regular sound shifts. I have not checked the OE > ME stage myself for ignorance of Germanic linguistics but I assume that even if this particular pair is not entirely regular then surely there are other cognate pairs across IE which would be.)

Having a list of sound changes for the relevant (proto-)languages to hand, we could theoretically automatize the process of their application to Modern English and Modern Greek, so that a computer can run the whole thing for us and, when we input 'wɜːk', to output 'ˈeɾɣo' (along with other phonetically exact and regular cognates which could have been possible), all this with no reference to a dictionary of Modern English, Modern Greek or a PIE lexicon but relying solely on the list of sound changes we program it with. The same could be done with other pairs of IE languages, obviously.

Has such a computer program ever been written?

  • 1
    I don't know of any, but this sounds like a good place to apply FSTs.
    – Draconis
    Commented Apr 6, 2021 at 23:42
  • 1
    I ran across a DOS program that generated Spanish forms from Latin input. Many many years ago; but even then Romance linguistics was pretty well developed.
    – jlawler
    Commented Apr 7, 2021 at 2:07
  • 1
    Sounds amazing @jlawler! Please lmk if you come across it again, that's similar to what I am looking for. Commented Apr 8, 2021 at 13:28
  • @SimonKorneev getting Spanish from Latin is easy because you're only dealing with sound changes going forward. I know several tools have been developed for conlanging purposes that apply sound changes to an input. zompist.com/sca2.html is probably the best known and when you first open it has several of the changes between Latin & Portuguese programmed in (although it's missing quite a few others)
    – Tristan
    Commented Apr 8, 2021 at 14:35

1 Answer 1


Sound changes are not generally reversible and so it is, in general, impossible to produce a single unique cognate

As an example, in Old English, i-mutation causes /e/ to merge into /i/ when followed by an /i/ that is earlier eCi iCi > iCi. As many final -i were lost, many i in final syllables could derive from an earlier e or an earlier i and so we would have to consider multiple possible proto-forms for a given English term

The more sound changes you want to reverse, the more proto-forms you will get out. Exponentially more in fact. In general, going back as far as PIE is likely to be intractable (especially as Modern English has undergone heavy dialect levelling resulting in a lot of unpredictable vowels, or various other unexpected or missing sound changes)

The only reason reconstruction is possible is through comparison; either between different languages (the comparative method), between different words/forms within a language (internal reconstruction), or a mixture of both and there is no known way to do this programatically (and in all likelihood no such method exists)

  • There is the dissertation from Jouna Pyysalo and the project from Pyysalo et al. at pielexicon.humhelsinki.fi. The sound changes are written as logical rules, that are used to develop words from roots according o compare with the lexical stock of respective branches.
    – vectory
    Commented Apr 8, 2021 at 5:16
  • @vectory logical rules are still in general information destroying, so it is still not generally possible to derive a unique proto-form from a single language. It's also worth noting Pyysalo's reconstruction of PIE is... idiosyncratic
    – Tristan
    Commented Apr 8, 2021 at 9:36
  • Thanks a lot @Tristan; I would like to stress though that I am not looking for a way to produce actual cognates, nor to reconstruct any PIE forms; the question merely concerns an algorithm which may, as a side-effect, produce cognates in cases of phonetically regular and exact development, but is mainly intended as an illustration to the sound-change patterns. Commented Apr 8, 2021 at 11:01
  • @SimonKorneev so from two word-lists pick out cognates? If you ignore semantics that probably could be done, but doing it exactly would be computationally intensive. As a heuristic, you could just look at the most common correspondence sets and ignore the conditioning environments. This would at least be a decent first pass for a human to look over
    – Tristan
    Commented Apr 8, 2021 at 11:04
  • changes are reversible. the complexity of the operation you propose is bounded in the lower regimen. It is approximately C^n. The number C of possible sound changes at any step is low and irrelevant. The number of steps in a row is even lower, approximately 6.Of course this is not how sound change works, in neatly sequenced steps, but that seems to be the premisses of the quizstion to begin with.
    – vectory
    Commented Apr 9, 2021 at 14:30

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.