Loading ...
Sorry, an error occurred while loading the content.

Re: Ancient languages reconstructed by computer program

Expand Messages
  • Matthew Turnbull
    Thanks for pointing out this study, I read an article about it in Nature, but the link helped me actually find it. I m not much for math, but it looks really
    Message 1 of 10 , Feb 13 5:25 PM
    • 0 Attachment
      Thanks for pointing out this study, I read an article about it in Nature,
      but the link helped me actually find it. I'm not much for math, but it
      looks really cool. There is a part in the paper where they describe how to
      build a tree no? page 2 paragraph 7, or did I misunderstand that? Also they
      do a test where they do a random tree and the reconstruction fails
      considerably more, with only about 68% of words agreeing instead of 85%. It
      appears to be more affected by the size of the dataset than the tree, as
      shown in figure 1 A, no? (I understand the concepts here but not the actual
      implementation, so any help is appreciated if I've got this wrong). I
      couldn't find a link to the program, my guess is they will publish it with
      the paper if they do so, this is published ahead of print.

      -Matt


      On Wed, Feb 13, 2013 at 8:36 AM, Петр Кларк <pyotr.klark@...> wrote:

      > The BBC has an article (http://www.bbc.co.uk/news/science-
      > environment-21427896) about computer reconstructions of proto-languages
      > (specifically Austronesian), with an 85% match with what linguists had
      > reconstructed "by hand". The full report is published in the Proceedings of
      > the National Academy of Science
      > (http://www.pnas.org/content/early/2013/02/05/1204678110) -- anyone have
      > access to it?
      > I'm a little curious if they did any sort of checking for
      > accuracy; for
      > instance, inputing modern Romance langauges and seeing if the program spat
      > out
      > Latin.
      > :Peter
      >
    • Armin Buch
      That s great. And less work for us to do, because in the next years I ll be working on just that. (My professor of course sees this as a missed chance for a
      Message 2 of 10 , Feb 14 2:37 AM
      • 0 Attachment
        That's great. And less work for us to do, because in the next years I'll
        be working on just that. (My professor of course sees this as a missed
        chance for a publication of his own. He considers it well done.)




        On 14.02.2013 02:25, Matthew Turnbull wrote:
        > Thanks for pointing out this study, I read an article about it in Nature,
        > but the link helped me actually find it. I'm not much for math, but it
        > looks really cool. There is a part in the paper where they describe how to
        > build a tree no? page 2 paragraph 7, or did I misunderstand that? Also they
        > do a test where they do a random tree and the reconstruction fails
        > considerably more, with only about 68% of words agreeing instead of 85%. It
        > appears to be more affected by the size of the dataset than the tree, as
        > shown in figure 1 A, no? (I understand the concepts here but not the actual
        > implementation, so any help is appreciated if I've got this wrong). I
        > couldn't find a link to the program, my guess is they will publish it with
        > the paper if they do so, this is published ahead of print.
        >
        > -Matt
        >
        >
        > On Wed, Feb 13, 2013 at 8:36 AM, Петр Кларк <pyotr.klark@...> wrote:
        >
        >> The BBC has an article (http://www.bbc.co.uk/news/science-
        >> environment-21427896) about computer reconstructions of proto-languages
        >> (specifically Austronesian), with an 85% match with what linguists had
        >> reconstructed "by hand". The full report is published in the Proceedings of
        >> the National Academy of Science
        >> (http://www.pnas.org/content/early/2013/02/05/1204678110) -- anyone have
        >> access to it?
        >> I'm a little curious if they did any sort of checking for
        >> accuracy; for
        >> instance, inputing modern Romance langauges and seeing if the program spat
        >> out
        >> Latin.
        >> :Peter
        >>


        --
        ----------------------------------------------------------------------
        Dr. Armin Buch armin.buch@...
        Zimmer 1.17, Blochbau (Wilhelmstr. 19) Tel. 07071/29-73960
        Seminar für Sprachwissenschaft
        Universität Tübingen
        ----------------------------------------------------------------------
      • Roger Mills
        (rejected the other day because I was over my limit....) ... Thanks for pointing out this study, I read an article about it in Nature, but the link helped me
        Message 3 of 10 , Feb 15 4:57 AM
        • 0 Attachment
          (rejected the other day because I was over my limit....)

          --- On Wed, 2/13/13, Matthew Turnbull <ave.jor@...> wrote:
          Thanks for pointing out this study, I read an article about it in Nature,
          but the link helped me actually find it. I'm not much for math, but it
          looks really cool.
          ===========================
          In
          principle it sounds great. But I'm not much for higher math/statistics
          either, nor computerese. Consequently _in the article itself_ I really
          couldn't understand what they were getting at.

          In the
          Supplementary material, where they go into the actual protoforms and the
          tree diagrams, I'm not impressed. I simply don't get, for ex., their
          finding on the word for "wind"-- Fiji "cagi" = /DaNi] has cognate
          relatives ALL OVER THE PLACE < AN *haNin
          (except for the initial /D/ and the loss of the original final (both
          explainable), but they apparently managed to find a couple Oceanic
          languages that didn't, and so came up with a weird protoform. And
          judging from the way they format their various protoforms, I wonder if
          they have any knowledge at all of the established forms, and what all
          the various symbols mean.

          Another chart in the Supplement
          discusses differences between "parent" and "child" langauges, but in
          many cases they're the wrong way round, and many of them would be
          "obvious upon inspection" to a trained Austronesianist. Nothing new
          here.

          As for the tree diagrams (in the big circular format) I
          strongly suspect they've used trees already devised by traditional
          comparative work. One cavil I have is that the nodes are simply
          numbered, not labeled, and the numbering doesn't seem to be explained
          anywhere.

          Unless someone really understands all the jargon etc., I personally
          can't imagine how it would be useful in my work..
          ====================================

          There is a part in the paper where they describe how to
          build a tree no? page 2 paragraph 7, or did I misunderstand that?
          ==========================
          I'll double check that in the morning.
          (TODAY: sorry, I still haven't gotten round to that....)
        • Alex Fink
          ... No, they re not building the tree there. It s this: the reconstruction process would be easiest if you had _cognate_ sets in all the languages in
          Message 4 of 10 , Feb 15 5:15 AM
          • 0 Attachment
            On Wed, 13 Feb 2013 19:25:30 -0600, Matthew Turnbull <ave.jor@...> wrote:

            >There is a part in the paper where they describe how to
            >build a tree no? page 2 paragraph 7, or did I misunderstand that?

            No, they're not building the tree there. It's this: the reconstruction process would be easiest if you had _cognate_ sets in all the languages in question. If all you have instead is a big multilingual dictionary, then semantic change and other replacement events mean that some of the corresponding words won't be cognates. This is what they're talking about in that paragraph: after you know what the tree is, you'll find that sometimes the word at one node of the tree is not the regular sound change descendant of its parent but is just some completely unrelated replacement.

            Anyway, I only just noticed the supplementary information part of the paper
            http://www.pnas.org/content/suppl/2013/02/06/1204678110.DCSupplemental/sapp.pdf
            so I've got some more reading to do...

            Alex
          • Roger Mills
            ... No, they re not building the tree there.  It s this:  the reconstruction process would be easiest if you had _cognate_ sets in all the languages in
            Message 5 of 10 , Feb 15 6:46 AM
            • 0 Attachment
              --- On Fri, 2/15/13, Alex Fink <000024@...> wrote:

              No, they're not building the tree there.  It's this:  the reconstruction process would be easiest if you had _cognate_ sets in all the languages in question.  If all you have instead is a big multilingual dictionary, then semantic change and other replacement events mean that some of the corresponding words won't be cognates.  This is what they're talking about in that paragraph: after you know what the tree is, you'll find that sometimes the word at one node of the tree is not the regular sound change descendant of its parent but is just some completely unrelated replacement. 

              RM: and that's sort of obvious to anyone, "elementary", isn't it?
            Your message has been successfully submitted and would be delivered to recipients shortly.