Minimal generalisations in short-term morphological convergence

The Minimal Generalisation Learner is an algorithmic learning model that has seen extensive use in modelling morphological variation. We take a new approach to apply the learner to morphological convergence, lexical priming, and word integration. We build on data from an online experiment in which participants have to play a word-matching game using nonwords with an artificial co-player. These data show that participants pick up lexical, as opposed to word-level, patterns from training with the co-player and apply these patterns to previously unseen nonwords in subsequent testing. We show that the Minimal Generalisation Learner can capture this shift in participant behaviour if, instead of building new generalisations on the nonwords in the training data, we use these words to update existing generalisations, built on real language data. This result breaks new ground in the use of the Minimal Generalisation Learner in modelling elicitation and learning tasks that go beyond Wug tasks and corpus studies.