Evolving constraints and rules in Harmonic Grammar
Elliott Moreton · ScholarWorks@UMassAmherst (University of Massachusetts Amherst) · 2020
An evolutionary model of pattern learning in the MaxEnt OT/HG framework is described in which constraint induction and constraint weighting are consequences of reproduction with variation and differential fitness. The model is shown to fit human data from published experiments on both unsupervised phonotactic and supervised visual pattern learning, and to account for the observed reversal in difficulty order of exclusive-or vs. gang-effect patterns between the two experiments. Different parameter settings are shown to yield gradual, parallel, connectionist- and abrupt, serial, symbolic-like performance.