What properties are shared by the processes used for learning linguistic and non-linguistic patterns? What properties are different? Research on non-linguistic (mainly visual) pattern learning has found distinct implicit and explicit processes which have different computational architectures, are facilitated by different experimental conditions, and differ in sensitivity to different pattern types. Is the same true of phonological pattern learning (“artificial-language” experiments)?
Experiment 1 asked whether implicit and explicit process are both available to human learners of a phonological pattern, and whether the same experimental conditions favor one over the other as in non-linguistic pattern learning. The pattern involved binary gender assignment conditioned by a single phonological or semantic feature. Training conditions were manipulated in ways which, in analogous non-linguistic experiments, have been found to elicit more implicit or explicit learning. Participants’ responses were duly shifted towards one or the other, as measured by self report of strategy, report of the correct rule, abruptness of learning curves, acceleration of response times after the last error, and bimodality of generalization performance. However, the shift was not categorical; some implicit and some explicit learners were found in both experimental conditions.
Experiment 2 asked whether explicit and implicit processes differed in sensitivity to structurally different patterns. In visual pattern learning, explicit learning has typically been found to be more sensitive to two-feature exclusive-or patterns (red XOR triangle) than to three-feature family-resemblance patterns (at least two of red, triangle, or small), and a large modelling literature has been dedicated to accounting for the exclusive-or advantage. Conditions which favor implicit learning reduce or reverse the exclusive-or advantage (reviewed in Kurtz et al. 2013). Our experiment reversed both of these results: The family-resemblance pattern was the easier one, and explicit learners showed a significantly stronger exclusive-or advantage. The reason for this surprising reversal seems to be that rule-seekers have a much harder time distinguishing relevant from irrelevant features in the exclusive-or condition compared to the family-resemblance condition.
These findings are discussed in the context of the larger question of how much linguistic and non-linguistic learning have in common (Moreton, Pater, & Pertsova, in press), and in connection with the practical question of how to design and interpret phonological-learning experiments.
Elliott Moreton is Professor of Linguistics and Director of Graduate Studies and Admissions at the University of North Carolina at Chapel Hill.
The Linguistic Department hosts colloquium talks by distinguished faculty from around the world.
October 9th: Keith Johnson, UC Berkeley
October 16th: Heidi Harley, University of Arizona
October 30th: Ivano Caponigro, UC San Diego
November 20th: Elliott Moreton, University of North Carolina
January 15th: Sharon Inkelas, UC Berkeley
February 5th: Colin Phillips, University of Maryland
February 6th: N. Goodman, Stanford University and A. Kehler, UC San Diego
March 5th: Linguistics Conference at Santa Cruz Conference
April 15th: Sabine Iatridou, MIT
April 29th: Paul Kiparsky, Stanford University
May 6, 7, 8: Semantics of Under-Represented Languages in the Americas 9
May 20th: Kyle Johnson, University of Massachusetts
May 27th/June 3rd (TBA): Linguistics Undergraduate Research Conference