Go above and beyond: does input variability affect children's ability to learn spatial adpositions in a novel language?

Abstract Human language is characterized by productivity, that is, the ability to use words and structures in novel contexts. How do learners acquire these productive systems? Under a discriminative learning approach, language learning involves using cues to predict and discriminate linguistic outcomes and ‘generalization’ involves dissociating idiosyncratic irrelevant cues in favour of informative, invariant cues. The current work tests the predictions of this account using the learning of spatial adpositions as a test case. Spatial adpositions describe the location of one object in relation to another (e.g. English prepositions ‘above’ and ‘below’) and may occur in reversible sentences, such as the picture is above the window; generalization involves using these terms in novel contexts, such as with unattested nouns. Computational simulations implementing an error-driven, discriminative learning process, demonstrate that broadening the irrelevant cues associated with the stimuli may boost the discovery of invariant cues, i.e. the association between the adposition and the spatial relation. We explored the predictions of these models in human learners by adapting a training paradigm introduced by Hsu & Bishop (Hsu, Bishop 2014 PeerJ2, e656 (doi:10.7717/peerj.656)) to teach typically developing 7–8-year olds spatial adpositions in an unfamiliar language (Japanese) using a computerized learning game. We manipulated the cue variability by comparing groups of children trained with more variable sentences (HV) with a condition with repetition of the same sentences (LV). A third condition (skew) tested whether learning and generalization are boosted when learning from a heavy tailed distribution that more closely resembles that of natural language. We examined the following predictions: (i) for sentences with novel nouns, participants trained with variable sentences will show better performance (i.e. stronger generalization) than those trained with repeated sentences; (ii) in contrast, those trained with repeated sentences will show stronger performance in training itself (i.e. stronger item learning); and (iii) training with a heavy tailed distribution—more closely resembling the natural one—will lead to the strongest item learning and generalization. In our main analyses, for (i) we found clear evidence that the HV condition outperformed the LV condition in generalization, in line with predictions of the computational model when trained on the same datasets. However, for (ii) the frequency advantage was not clearly observed and for (iii) skewed input did not provide an additional benefit over variability (with Bayesian evidence for the null that it was beneficial for generalization). Interestingly, the fact that the skew condition did not outperform the high variability benefit was in fact consistent with the computational modelling, although skew was found to be supportive in other domains. Finally, exploratory analyses indicate interesting individual differences in how learners respond to variability and frequency in their input, which may be owing to their current environment as well as learner characteristics.