Abstract
<p>Although grammatical gender assignment often correlates probabilistically with semantics, training these distinctions in artificial languages has proven challenging. Notably, unsuccessful training paradigms typically employ passive presentation, which may not encourage the predictive, error-driven processing that typifies naturalistic learning. Using a novel ‘error-inducing’ paradigm (N=77), we examined learning of an artificial language comprising nouns (vegetables) split into two gender-classes marked by affixes associated with a mass/count distinction. In training trials, participants viewed two vegetables and heard a noun-phrase of the form prefix–noun–suffix, and were asked to click on the vegetable matching the audio (without feedback). Despite the ambiguous way items were presented in training, in post-tests participants classified the intended referents correctly, and were above chance in assigning gender to novel referents, demonstrating they had learned the appropriate semantic cues. Critically, analysis of eye-movements in training provided evidence that participants (i) fixated referents faster when foil and target came from different gender-classes (cf Lew-Williams, &amp; Fernald, (2010)) (ii) showed this effect more strongly than in a control condition where the relationship between semantic and gender was eliminated. These results suggest that learners can use the prediction error inherent in ambiguity to discriminate the cues to referents and support generalization.</p>