Three Exemplars Allow at Least Some Linguistic Generalizations: Implications for Generalization Mechanisms and Constraints

Abstract The mechanism that allows learners to generalize over linguistic input and its relation to constraints on possible generalizations was explored in three experiments. Infants were familiarized briefly with words exhibiting stress patterns generated by a set of ordered principles, and then tested on new words that were either consistent or inconsistent with those principles. In Experiment 1, 9-month-olds generalized from three different heavy syllables (those ending in a consonant) heard during familiarization, to a new heavy syllable at test, but did not generalize from multiple tokens of one heavy syllable. Experiment 2 demonstrated that, although 9-month-olds were able to use three syllables to generalize the linguistically natural principle assigning stress to heavy syllables, they failed to use them to generalize a principle that does not occur in natural languages, suggesting that the mechanism requiring three input examples does not apply to all input equally. Experiment 3 demonstrated that although 9-month-olds failed to generalize the unnatural principle, 7.5-month-olds succeeded, suggesting that constraints on generalization can be acquired over development. This research was supported by NIH HD042170 to LouAnn Gerken. The authors warmly thank Rebecca Gómez, Tom Shultz, Andy Wedel, Josh Tenenbaum, and two anonymous reviewers for helpful comments and suggestions on earlier drafts. Notes 1Presenting a single heavy syllable in multiple word contexts is similar to presenting a single label to several dalmations, as was done by CitationXu and Tenenbaum (2007b). 2To ensure that the words containing heavy syllables were roughly equivalent in the two familiarization conditions, acoustic duration measurements were made of all familiarization words containing heavy syllables. The mean duration in msec. (and SE) of words containing heavy syllables in the 1 Heavy Syllable condition was 924 (24) and 911 (24) in L1 and L2, respectively. The comparable values in the 3 Heavy Syllables condition were 930 (32) and 936 (31) in L1 and L2, respectively. Based on the duration measure, the words in the two conditions appear to be acoustically very similar, with a mean duration difference of only 16 msec., less than one standard error. 3Interestingly, adults exposed to the same familiarization stimuli as infants in the 3 Heavy Syllables condition and then asked at test whether each test word was consistent or inconsistent with what they had heard during familiarization showed a strong effect of learning for the stress pattern test items, but no effect of learning at all for the Abstraction test items (also see CitationGuest et al., 2000). One interpretation of these findings is that, as native English-speakers, adults have realized that syllable position (e.g., first syllable of a word) is the most reliable associate of lexical stress and therefore have come under most circumstances to ignore more subtle correlations (e.g., stress and ending in a consonant). 4Two anonymous reviewers also did counts on two other databases, CHILDES and CELEX. The data from the CHILDES database supported our finding of an asymmetry in patterning of stressed syllables ending in consonants vs. starting in /t/. The data from the CELEX database did not reveal the same asymmetry. Perhaps one reason for the difference is that our analysis and the CHILDES-based analysis used words that occur in young children's utterances. A related fact is that our analysis was restricted to mono- and disyllabic words. The reviewer using the CHILDES database also noted the large number of monosyllabic words. In contrast, the analysis using CELEX was unrestricted by number of syllables. The specific input that young learners receive and what they notice in that input clearly warrants further study.