Building domain specific NERs by using information from domain-general annotations

From Cohen Courses
Revision as of 19:12, 27 September 2011 by Ysim (talk | contribs) (→‎Team Members)
Jump to navigationJump to search

Building domain-specific Named Entity Recognizers using information from domain-general annotations

Team Members

Project Idea

When adapting Named Entity Recognizers for a new domain, we often need to design and incorporate domain specific new features and retrain existing classifiers. Often, information generated from non-domain specific NERs and annotators are not re-used when building domain specific NERs. In our project, we propose to design a CRF-based NER which uses graphical models to incorporate information obtained from non-domain specific NERs and annotators to perform domain specific named entity recognition.

Comments from William

Seems like a nice idea. The Borthwick et al paper is one prototype for how to use such domain-general annotations. Do you have an idea as to what datasets you will use? You would need multiple datasets from different domains.

This is closely related to the problem of transfer learning, by the way. Andrew Arnold's recent thesis (in CMU MLD Dept) has some survey materiel on that subarea.

--Wcohen 20:48, 22 September 2011 (UTC)