Wonnacott, E. . , Miller, C. & Vujovic, M.
University College London
Language learners must identify generalizations which can be used with unattested vocabulary. Discriminative learning suggests this is driven by the structure of the linguistic input: generalization is more likely when learners are exposed to more varied instances, since this allows structures to be disassociated from trained instances (Ramscar et al 2010). We test this hypothesis in a language learning experiment with children.
Methods: 40 7-8-year-olds heard instructions in an unfamiliar language (Japanese) whilst playing a game where they moved pictures within a grid (e.g. move-banana-above -chocolate). We measured learning of word order and two postpositions (equivalent to the English prepositions above/below). There were two between-participant training conditions: (1) high-variability exposure: 28-above and 28-below sentences, each encountered once; (2) low variability exposure: 2-above and 2-below sentences, each repeated 14 times; (number of nouns (8) and total exposure matched across conditions). Training was followed by a generalization test involving untrained nouns.
Results: During training, the low-variability group showed faster learning, presumably due to repeated practice with identical items. However, critically, the high-variability group were stronger in the generalization test. Human performance matches the learning of naïve discriminative learning model. This supports the hypothesis that exemplar variability is key in driving linguistic generalization.