Lisa Pearl, UC Irvine

Friday, January 17th at 4pm in Humanities One - Room 210

More learnable than thou? Testing knowledge representations with realistic acquisition data

One (often implicit) motivation for a linguistic knowledge representation (e.g., a set of linguistic parameters or constraints) comes from an argument from acquisition, where language acquisition is assumed to be straightforward if children’s hypothesis space is defined by the correct knowledge representation. Acquisition then becomes the process of selecting the correct language-specific grammar from that hypothesis space, based on the language input encountered. I discuss quantitative metrics based on an argument from acquisition for comparing knowledge representations and the grammars they define. These metrics involve assessing grammar learnability from realistic input data, and I use them to evaluate three prominent knowledge representations in the domain of metrical phonology that each define a grammar for English. Somewhat surprisingly, I discover that learnability issues arise for the English grammars in all three representations. I discuss aspects of the proposed English grammars that may be hurting learnability as well as ways a child may still be able to learn the proposed English grammars from English input.

Go back to Colloquia