Proceedings of the Annual Meeting of the Cognitive Science Society

Children learning their first language face multiple problemsof induction: how to learn the meanings of words, and howto build meaningful phrases from those words according tosyntactic rules. We consider how children might solve theseproblems efficiently by solving them jointly, via a computa-tional model that learns the syntax and semantics of multi-word utterances in a grounded reference game. We select awell-studied empirical case in which children are aware of pat-terns linking the syntactic and semantic properties of words –that the properties picked out by base nouns tend to be relatedto shape, while prenominal adjectives tend to refer to otherproperties such as color. We show that children applying suchinductive biases are accurately reflecting the statistics of child-directed speech, and that inducing similar biases in our compu-tational model captures children’s behavior in a classic adjec-tive learning experiment. Our model incorporating such biasesalso demonstrates a clear data efficiency in learning, relative toa baseline model that learns without forming syntax-sensitiveoverhypotheses of word meaning. Thus solving a more com-plex joint inference problem may make the full problem of lan-guage acquisition easier, not harder.