[PS-3.41] Inferring phonetic category structure from skewed input distributions

Olejarczuk, P. & Kapatsinski, V.

University of Oregon

We present a phonetic category learning experiment on how the shape of the input distribution is reflected in the internal structure of the learned category. Particularly, we ask: what are the most typical instances of a category?

Two groups learned novel categories comprising instances of a monosyllable with a rising-falling tone. Pitch excursion magnitude varied along a continuum, with values sampled from a different unimodal distribution for each group. These distributions had identical means and variances, differing only in the direction of skew. Following training, subjects rated typicality of tokens with pitch excursions within and outside the experienced range.

We expected maximum typicality to align with some measure of central tendency: the mean, median or mode of the distribution. Asymmetric distributions allowed us to tease the predictions of these measures apart.

Unlike mean pitch excursion, peak typicality differed between groups. However, it shifted towards the long tail and away from the median and mode. The distribution of typicality was also more symmetrical than the input. We suggest that the category is represented by a detector with a floating symmetrical receptive field, and that unusual, unexpected examples in the long tail draw attention, demanding that the receptive field shift towards them.