Abstract
Adults, infants, and other species are able to learn and generalize abstract patterns from sequentially presented stimuli. Rule learning of this type may be involved in children's acquisition of linguistic structure, but the nature of the mechanisms underlying these abilities is unknown. While inferences regarding the capabilities of these mechanisms are commonly made based on the pattern of successes and failures in simple artificial-language rule-learning tasks, failures may be driven by memory limitations rather than intrinsic limitations on the kinds of computations that learners can perform. Here we show that alleviating memory constraints on adult learners through concurrent visual presentation of stimuli allowed them to succeed in learning regularities in three difficult artificial rule-learning experiments where participants had previously failed to learn via sequential auditory presentation. These results suggest that memory constraints, rather than intrinsic limitations on learning, may be a parsimonious explanation for many previously reported failures. We argue that future work should attempt to characterize the role of memory constraints in natural and artificial language learning.
ACKNOWLEDGMENTS
We gratefully acknowledge Denise Ichinco and Kelly Drinkwater for help with data collection, and Charles Kemp, Talia Konkle, LouAnn Gerken, Sharon Goldwater, Noah Goodman, Gary Marcus, Rebecca Saxe, Josh Tenenbaum, Ed Vul, and three anonymous reviewers for valuable discussion. This research was supported by a Jacob Javits Graduate Fellowship and NSF DDRIG #0746251.
Notes
1Because the test items in rule-learning tasks are usually novel stimuli that participants have never seen before, this kind of task contrasts with artificial segmentation (CitationSaffran et al., 1996b, Citation1996a) and word learning (CitationVouloumanos, 2008; CitationYu & Smith, 2007; L. CitationSmith & Yu, 2008) tasks in which learners are generally tested on items or pairings that are present in the familiarization stimuli.
2Imagine a language of the form aXbY with a Y element that had a variability equivalent to the highest variability of the X element. This manipulation would equate the number of string types across conditions with different variability of the X element, dissociating memory and variability. (Thanks to Charles Kemp for this observation.) Note that covariation of memory demands and variability does not compromise Gómez's result; if anything the result is strengthened, since even under conditions with high memory demands, variability still leads to the extraction of meaningful regularities.
3Note, however, that this result would not support the contention that retention demands are the only barrier to success; as in the case of the CitationGómez (2002) experiment, we expect that other factors will also play a role.
4Previously available at http://www.research.att.com
5We use z as a test statistic for the Wilcoxon rank-sum test; this approximation is appropriate when both samples have N > 10.
6A previous version of this model was reported in CitationFrank, Ichinco, and Tenenbaum (2008).
7Thanks to an anonymous reviewer for careful discussion of this issue.