Guest Speaker Seminar Series: Elizabeth Wonnacott

When: Wednesday, April 30, 2025, 4:30 PM - 6:00 PM
Where: Arts Two 2.17, Mile End

Guest Lecture Series: Elizabeth Wonnacott (University of Oxford)

Exploring the mechanisms of language learning and generalization through learning experiments with children and adults

Abstract

Language learning involves abstracting generalizations which operate across linguistic items. According to statistical learning approaches, these emerge on the basis of exposure to the input, but what are the mechanisms that underpin this process? My work explores the extent to which language learning can be explained under a discriminative, error-driven account, based on well-understood principles developed in the study of animal learning. Under this approach, learners continuously seek to reduce uncertainty about language as they hear it, generating prediction error when incoming language violates the listener’s expectations. Prediction error then serves to weaken associations with unhelpful cues and thus, via cue competition, reinforce cues which reduce uncertainty.

My talk presents data from two learning language learning experiments: An experiment in which 7-8 year olds learn novel language structures in a new language (Japanese) via a computerized game, and an eye-tracked artificial language experiment with adults. Overall, the data provides evidence about for a key prediction of the computational model: generalization is boosted by witnessing varied exemplars in the input, which allows linguistic structures to be dissociated from trained instances. In addition, the adult eye-tracking data provide direct evidence for a key assumption under error-driven learning: that learners make predictions which are both correct and incorrect (i.e. based on over-generalization), thus providing opportunity to learn via prediction error. However, there were also large individual differences, including interesting interactions with a cognitive (digit span) measure in the child learners. This highlights the limitations in extrapolating idealized model performance to real human learners in “noisy” environments.