Two of the main problems concerning language include how we use
language, and how we learn to so do. Most work in linguistics
explicitly concerns itself with the first question, at various
levels of abstraction. Still, even here, considerations of
learnability are often implicitly appealed to in theory
construction. Outside of linguistics, in the domain of natural
language processing, it has become recognized as practically
impossible to hand craft wide coverage grammars, and much energy has
been devoted to their automatic induction from raw data.
This course is devoted to understanding the literature on grammatical inference, especially as it pertains to learning linguistic structures from data. We will be primarily interested in matters of principle: under which conditions can we guarantee that a learner will acquire a grammar with which kind of relation to the input; and only secondarily in matters of practice: how provably correct algorithms can serve as the foundation for practical learning systems.
We study two of the most influential learning paradigms, the categorical Gold paradigm and the stochastic PAC, with attention to how they can inform our ideas about how humans learn language, and thus about grammar.
|Lectures||MW 1:30-2:50, RO 208|
|Office Hours||by appointment|
|Notes||Notes and papers will be handed out.|
|Exams||A final paper will be due after finals week|