Symposium on Machine Learning in Speech and Language Processing (MLSLP)
September 14, 2012
Portland, Oregon, USA
Speaker: Hal Daumé III (U. Maryland)
Title: Transfer Learning in Language
Abstract:
Human language is messy, and machine learning has done a lot to tame
this messiness. There are many facets to language processing, and
while the common approach is to run a bunch of component systems in a
pipeline, there is mounting evidence that this is a bad idea. Enter
transfer learning and multitask learning. Unfortunately, there are many
open-ended problems in transfer learning for language due to the
sorts of data and annotations that we have easy access to in the
language domain. This talk will highlight some successful attempts to
use transfer learning (generative and otherwise) in language, but will
also talk a good deal about what is unsolved, and point to some
interesting current avenues of research.