Friday, February 27, 2015

Modeling the changes of natural vowel systems

Back in 2013, I posted about my plans to write a computer simulation that could model the way vowel systems are learned, transmitted, and changed.  I am happy to report that the first phase of this project was completed, and we presented a poster at the AAAS meeting this month.  I was able to find an excellent student of computer science, Hannah Scott, who is now my coauthor on the project.  She programmed the entire simulation in Python, leveraging methods from the "imitation game" described by Luc Steels.

In our simulation, we begin with a vowel system, which is just a set of F1 and F2 values on a vowel chart. Then some number of agents we call "parents" says the vowels a number of times to each of the "children."  The children then try to repeat the vowels, and we provide feedback based on how accurate the repetition is.  All of this has random noise introduced, of course, and the main parameters are the children's precision and the tolerance of the feedback system.

We had originally conceived of the feedback as coming from the parents, but then colleagues pointed out that in reality, parents' feedback is not used by children learning the sounds of language.  The real feedback comes from the child's own brain, which has already learned and stored an excellent perceptual model of the correct vowels.  The child is perfectly capable of monitoring and critiquing his/her own production efforts.

The simulation is, as might be expected, very sensitive to the values of the precision and tolerance parameters.  When either of these is too large or too small, the vowel system goes haywire or is completely extinct in one or two generations.  The sweet spot for these parameters, so far, is around 1 Equivalent Rectangular Bandwidth (an auditory unit of frequency).

We obtained some interesting little results so far.  E.g. the simulation preserves the Spanish 5-vowel system perfectly for eternity, but with the same parameters applied to the English vowels there are changes, with vowels merging etc.  These changes are then mitigated by introducing vowel duration as an additional attribute.  In my opinion this behavior of the simulation is incredibly realistic for a first try.

And this indeed seems to be *the first* attempt to do this.  I could find nothing in literature that presented a simulation of sound change, whether for vowels or anything at all. There is, of course, the literature by Steels and Bart de Boer which models how vowel systems could have emerged in the first place, at the dawn of language, but that is really a different type of question.

5 comments:

  1. This is very interesting work! Neat to find that you're geographically fairly nearby (I'm in Sacramento).

    At the risk of horn-tooting, I'll point out that my PhD work looked at computational modeling of the evolution and acquisition of vowel harmony.
    http://dl.acm.org/citation.cfm?id=2048953

    Also, Mark Dras and Marc Ettlinger have done work on vocalic sound change with computational models.

    ReplyDelete
  2. Can you please provide specific citations for the sound change modeling you mention?

    ReplyDelete
  3. OK, I found a paper by Ettlinger that appears never to have been published, which did some computational modeling, but not in any way similar to our methods. Not to say it is not of interest, but it is just addressing a different question in a different way.

    ReplyDelete
  4. Andrew Wedel at Arizona has done some work that might be relevant.

    ReplyDelete
  5. I've read Wedel's work and found it to be in a very different vein. It was so dissimilar to ours that I could find no reason to cite it, in fact.

    ReplyDelete