A probabilistic account of first language acquisition from string-meaning pairs
Speaker: Tom Kwiatkowski
Abstract:
When learning their first language children infer the rules of syntax on the basis of strings observed in conjunction with contextually afforded meaning representations. This problem is closely related to the well explored task of inducing a grammar from a corpus of sentences paired with database queries and I will show how methods developed for this related task can be generalised to provide a language independent, pscycholinguistically plausible, model of first language acquisition.
I will present some preliminary results gained from training this model on a small set of child directed utterances paired with unambiguous and correct meaning representations before then going on to explore possible extensions of both the model and experimental setting to account for the added problem of referential uncertainty.