Proceedings of the Annual Meeting of the Cognitive Science Society

One of the biggest puzzles in language acquisition is concerned with how children retreat from overgeneralization errorsin valence alternations, for example the ditransitive alternation. Pinker (1989) proposes that children are susceptible toovergeneralization when they acquire broad verb semantic classes initially and they recover when they acquire narrow verbclasses later. To empirically test this hypothesis, we devised a computational framework that automatically induces verbclasses from text data, by combining state-of-art word embeddings (Pennington, Socher & Manning, 2014) with graphalgorithms (Steyvers & Tenenbaum, 2005; Von Luxburg, 2007). We selected three representative valence alternationsfrom Levin (1993) and tested Pinkers hypothesis on five naturalistic language production corpora. Our results demonstratethat contrary to Pinkers predictions, broad verb classes are harder to induce than narrow classes and that semantic classesmay not be the primary mechanism that accounts for childrens retreat from overgeneralization errors.