Olejarczuk, P. & Kapatsinski, V.

University of Oregon

Research on distributional learning has shown that categories can be learned from probability distributions (Maye & Gerken, 2000). However, little attention has focused on the knowledge acquired in this process. Parametric models of categorization have claimed that learners summarize experienced distributions with a measure of central tendency and a measure of variance (e.g. Buz et al., 2016; Flanagan et al., 1986). However, there is little agreement on the right measure of central tendency. Computational models of speech perception have assumed that within-category probability distributions are normal, and therefore that the mean, the mode and the median are equal. However, this is not true of real within-category distributions, which are often skewed (e.g. Koenig, 2001). We examined the knowledge acquired by English speakers from exposure to skewed distributions of tone contours. Skewed distributions allow us to decouple the various measures of central tendency. Knowledge of the category was examined using typicality ratings of exemplars (Miller, 1994). Our results rule out the mode and median as appropriate measures of category center (cf. Buz et al., 2016). Peak typicality corresponds to the mean of the frequency distribution, provided that frequency is logarithmically transformed to take into account the salience of relatively surprising exemplars.