Domain Adaptation with Structural Correspondence Learning

Google Tech TalksSeptember, 5 2007ABSTRACTStatistical language processing tools are being applied to anever-wider and more varied range of linguistic data. Researchers andengineers are using statistical models to organize and understandfinancial news, legal documents, biomedical abstracts, and weblogentries, among many other domains. Because language varies so widely,collecting and curating training sets for each different domain isprohibitively expensive. At the same time, differences in vocabularyand writing style across domains can cause state-of-the-art supervisedmodels to dramatically increase in error.This talk describes structural correspondence learning (SCL), a methodfor adapting models from resource-rich source domains to resource-poortarget domains. SCL uses unlabeled data from both domains to induce acommon feature representation for domain adaptation. We demonstrateSCL for two NLP tasks: sentiment classification and part of speechtagging. For each of these tasks, SCL significantly reduces the errorof a state-of-the-art discriminative model.Speaker: John Blitzer

Channel: People & Blogs
Uploaded: November 30, 1999 at 12:00 am
Author: Google

Length: 59:50
Rating: 5.00
Views: 3269

Tags: engedu  google  tachtalk  techtalks  

Video Comments