ABSTRACT
Most learning theories agree that the productivity of a rule or a pattern relies on regular exemplars being dominant over exceptions; the threshold for productivity is, however, unclear; moreover, gradient productivity levels are assumed for different rules/patterns, regular or irregular. One theory by Yang, the Tolerance Principle (TP), specified a productivity threshold applicable to all rules, calculated by the numbers of total exemplars and exceptions of a rule; furthermore, rules are viewed as quantal, either productive or unproductive, with no gradient levels. We evaluated the threshold and gradience-quantalness questions by investigating infants' generalization. In an implicit learning task, 14-month-olds heard exemplars of an artificial word-order rule and exceptions; their distributions were set closed to the TP-threshold (5.77) on both sides: 11 regular exemplars vs. 5 exceptions in Condition 1 (productiveness predicted), and 10 regular exemplars vs. 6 exceptions in Condition 2 (unproductiveness predicted). These predictions were pitted against those of the statistical majority threshold (50%), a common assumption which would predict generalization in both conditions (68.75, 62.5%). Infants were tested on the trained rule with new exemplars. Results revealed generalization in Condition 1, but not in Condition 2, supporting the TP-threshold, not the statistical majority threshold. Gradience-quantalness was assessed by combined analyses of Conditions 1-2 and previous experiments by Koulaguina and Shi. The training across the conditions contained gradually decreasing regular exemplars (100, 80, 68.75, 62.5, 50%) relative to exceptions. Results of test trials showed evidence for quantalness in infants (productive: 100, 80, 68.75%; unproductive: 62.5, 50%), with no gradient levels of productivity.
ABSTRACT
Infants use statistics-based word segmentation strategies from the preverbal stage. Statistical segmentation is, however, constrained by the Onset Bias, a language-universal principle that disfavors segmentation that harms syllable integrity. Children eventually learn language-specific exceptions to this principle. For instance, sub-syllabic parsing occurs for vowel-initial words in French liaison contexts, that is, when a word's final consonant surfaces as the following word's syllabic onset (e.g., /n/ in un /n/éléphant). In past research, French-learning 24-month-olds succeeded in parsing a vowel-initial pseudo-word surfacing with variable liaison consonants. This study further investigated infants' liaison representation, its potential impacts on parsing, and its interaction with the Onset Bias. In Experiments 1 and 2, French-learning 24-month-olds were familiarized with pseudo-words with variable liaison-like versus nonliaison-like onset consonants, preceded by words that cannot trigger those onsets (e.g., un zonche; un gonche). We found no mis-segmentation as vowel-initial and successful segmentation as consonant-initial. In Experiment 3, when the preceding words could trigger a liaison consonant that matched the onset of the following word (e.g., un nonche), infants showed a vowel-initial mis-interpretation, against the Onset Bias, revealing an effect of liaison knowledge. These results demonstrate that toddlers balance their use of language-general principles/strategies and language-specific knowledge during early acquisition.
