Event

Contextual Bootstrapping for Grammar Learning

Eva Mok

Friday, May 02, 2008
12:30

In this talk, I will describe a computational model of child grammar learning within the construction grammar paradigm. As has long been argued in the field of first language acquisition, the input to grammar learning is impoverished: the grammatical structures of utterances are not present in the input, explicit corrections are rare, and in the case of pro-drop languages such as Mandarin Chinese, verb arguments are often omitted in the utterances. While this observation has in the past led to postulations that various aspects of language are innate, only in the more recent history have children been studied with an eye towards the ability they bring to the task of language learning. In my model, I combine rich, embodied meaning representation with data-driven structural learning in a Bayesian learning framework to demonstrate that early argument structure constructions can be successfully learned from situated language input. Context bootstraps learning by helping the learner to partially infer the the intended meaning of an input utterance, even in cases where arguments are omitted. Iterative learning of new constructions can then be performed to maximize the completeness of understanding achievable using the grammar. A few learning operations have been implemented within this framework to propose new phrasal constructions and generalize across these constructions. I will describe these learning operations in my talk along with preliminary results I have obtained.