Why Language Learning Requires Miscalibration
Abstract: As children learn language, they try to mimic adult speech, while
leaving room for change. Furthermore, manuscript data shows that
language is fuzzy, in that speakers use alternative forms to express
the same meaning and that the frequency with which they use those
alternatives varies. Thus, child language acquisition requires
children to learn usage frequencies for various constructions, as well
as which constructions are possible. I will present an ODE model, a
more complicated PDE model, and a detailed discrete simulation that
all show that children cannot be accurate in replicating the usage
frequencies of the adult population. If they are, then the population
tends to an incoherent mixture of all possible constructions rather
than a single language dominated by one alternative. Instead,
children must use a learning algorithm that is miscalibrated so as to
prefer usage frequencies close to 0 or 1. With proper miscalibration, the population can settle on a dominant language.