Similarity and confidence in artificial grammar learning

Three experiments examined the relationship between similarity ratings and confidence ratings in artificial grammar learning. In Experiment 1 participants rated the similarity of test items to study exemplars. Regression analyses revealed these to be related to some of the objective measures of similarity that have previously been implicated in categorization decisions. In Experiment 2 participants made grammaticality decisions and rated either their confidence in the accuracy of their decisions or the similarity of the test items to the study items. Regression analyses showed that the grammaticality decisions were predicted by the similarity ratings obtained in Experiment 1. Points on the receiver operating characteristics (ROC) curves for the similarity and confidence ratings were closely matched. These data suggest that meta-cognitive judgments of confidence are predicated on structural knowledge of similarity. Experiment 3 confirmed this by showing that confidence ratings to median similarity probe items changed according to the similarity of preceding items.