The cultural evolution of language in the lab

1. The cultural evolution of language in the lab Kenny Smith Department of Psychology Northumbria University

2. The general approach Language is socially-learned Languages evolve to facilitate their transmission Consequences: • Adaptation to biases of language learners • Adaptation to other pressures (e.g. poverty of the stimulus) Structural properties of language not necessarily hard-wired into language learners (Implications for brain stuff later)

3. Language learning in context A question: how do the kinds of languages we see in populations relate to the individual-level processes of learning and production? utterances Learning grammar Production utterances grammar utterances grammar

4. One way to address these questions Mathematical/computational modelling • Build models of the processes of learning and production • With known biases • Build a population model • A chain, a series of non-overlapping generations, … • Run it • How do the biases of individuals influence the structure of the languages that develop in these populations? • Can other things (e.g. amount of data you see, frequency, structure of population you are in, …) have an impact?

5. Another way Do the same thing in the lab • Artificial language learning • Participants attempt to learn a simple target language • Diffusion chain methodology • Participant n’s produced language becomes training data for n+1 Two areas of current work • Elimination of unpredictable variation • Smith & Wonnacott, forthcoming, Cognition • Creation of structure

6. An observation: languages are structured Languages contain generalizable structure • This allows learners to learn a large system after exposure to only part of it Why is language like this? Can structure emerge from transmission? • Kirby, Cornish & Smith, 2008, PNAS

7. An artificial language learning experiment • 27 pictures • 3 colours, 3 shapes, 3 movements • Adult participant repeatedly trained on labels for subsetof 13 • Tested on all 27 pictures • Learners required to generalize • Poverty of stimulus problem

8. Iterated Learning picture-label pairs Learner 1 Evolution? picture-label pairs Learner 2 picture-label pairs Learner 3

9. A random initial language

10. A 9th generation language

11. Why is it becoming more structured? • Language finds a structural generalization which allows it to be learned and transmitted • Compositionality • Given different transmission pressures, you get different structures • E.g. systematic underspecification (ambiguity) • Structure potentially emerges through cultural (as opposed to biological) evolutionary processes • Our participants already know stuff about how languages should work • But models illustrate same effects in populations of learners lacking these biases

12. Implications Because language is socially learned • It evolves • It probably adapts to us faster than we adapt to it • Its structure probably reflects (weak) biases of domain-general learning mechanisms • I would expect high overlap between brain regions / processes for language and related non-linguistic tasks