Individual Differences in the Role of Prior Knowledge in Word Learning: An Inferred Bayesian Model
Hannah Marlatte · TSpace (University of Toronto) · 2019
Individuals may gain information about the meaning of novel words by using contextual clues. A Bayesian framework of such inferential word learning suggests that learners integrate prior knowledge about possible word meanings with the statistical structure of the observed instances to make rational inductive inferences. We examined this process by creating a mechanistic Bayesian Observer model. Forty-five adults used contextual information to infer the names of twenty novel animals and tested on next-day memory. The informativeness of the contextual information was calculated from a preliminary norming study (INF rating, likelihood). INF ratings, with participants’ accuracy (posterior distribution), were used to retroactively infer participants’ prior distribution, an estimate of evidence supporting their hypotheses throughout the task. Individual’s initial prior distribution and how incrementally it was updated across the task were predictive for next-day memory. Exploratory cluster analyses suggest individuals’ strategies in the task were associated with different prior and posterior distributions.