Iterated learning and the evolution of compositional structure
This reading plus practical covers a paper plus experiment featuring an iterated learning paradigm, where one participant’s test output from a training-plus-test paradigm becomes the training input to another participant (via written text, playing a simple director-matcher game) - if you think this kind of paradigm might be useful for you, or you just want to see how it’s done, this is the one to look at.
The reading Beckner et al. (2017), which provides a reanalysis and online replication of Experiment 2 of Kirby, Cornish & Smith (2008). Beckner et al. start with a reanalysis of data from Kirby et al., but the more relevant part here is their experiment, a large online iterated artificial language learning experiment. Participants are asked to learn an artificial language which provides labels/descriptions for groups of novel objects, where those objects come in 3 shapes, 3 colours, and appear in groups of 1, 2 or 3. After training on a language, participants are tested, producing a new set of labels/descriptions. Those descriptions are then used as training for another participant: the language is passed from participant to participant down a chain of transmission, potentially changing as it goes as a result of accumulated errors and innovations made by participants. Beckner et al. impose a bottleneck on transmission (participants are trained on labels/descriptions for a subset of the objects, but required to generalise to the entire object set at test time), and filter out ambiguous labels (to discourage the language from collapsing down to a single label, which was a result we got in Experiment 1 of our 2008 paper). They find, as in the original experiment, that compositional structure develops through this iterated learning process: the first participant in each chain is trained on a language where each label/description is random and completely idiosyncratic, but the languages gradually evolve regularities, e.g. where shape and number are consistently encoded separate morphemes which are combined to form a complex (well, 2-part) signal.
As usual, in the associated practical you’ll get a chance to look at a similar experiment in jsPsych, which will involve code from our word learning experiment again but also the infrastructure to run an iterated learning design, which involves manipulating CSV files on the server.
Read:
A couple of things to note:
All aspects of this work are licensed under a Creative Commons Attribution 4.0 International License.