Testing a computational model of causative overgeneralizations: Child judgment and production data from English, Hebrew, Hindi, Japanese and K’iche’

How do language learners avoid the production of verb argument structure overgeneralization errors (*The clown laughed man c.f. The made laugh), while retaining ability to apply such generalizations productively when appropriate? This question has long been seen as one that is both particularly central acquisition research and challenging. Focussing on causative this type, a previous study reported computational model learns, basis corpus data human-derived verb-semantic-feature ratings, predict adults’ by-verb preferences for less- versus more-transparent forms (e.g., * vs laugh) across English, Hebrew, Hindi, Japanese K’iche Mayan. Here, we tested (and an expanded version with multiple hidden layers) explain binary grammaticality judgment from children aged 4;0-5;0, elicited-production 4;0-5;0 5;6-6;6 (N=48 per language). In general, successfully simulated children’s data, correlations r=0.5-0.6 r=0.75-0.85, respectively, also generalized unseen verbs. Importantly, all five languages showed some evidence making types – in judgments previously observed naturalistic studies English *I’m dancing it). Together findings, present demonstrates simple learning can (a) continuous (b) (c) (with no training these datasets), therefore constitutes plausible mechanistic account verbs’ restrictions.