BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Multiple Instance Learning for Natural Language Tasks - Mark Crave
 n\, University of Cambridge (visiting)
DTSTART:20061013T110000Z
DTEND:20061013T120000Z
UID:TALK5571@talks.cam.ac.uk
CONTACT:NLIP Seminars
DESCRIPTION:Many state-of-the-art methods for text and natural-language pr
 ocessing\nemploy supervised learning algorithms.  A key obstacle to the\na
 pplication of supervised learning methods\, however\, is that labeled\ntra
 ining instances are usually expensive to acquire.  One way around this\nob
 stacle\, I argue\, is to exploit data that can be readily and\ninexpensive
 ly labeled at a coarse level of granularity.  Such\nsituations are well su
 ited to multiple-instance learning.  In\nmultiple-instance learning\, indi
 vidual instances are not given labels\, but\ninstead bags of instances are
  labeled.  Whereas a negative bag is assumed to\ncontain only negative ins
 tances\, a positive bag need contain only one\npositive instance.\n\nI wil
 l describe the multiple-instance setting\, discuss its\napplicability to n
 atural language tasks\, and present several recent\nresults on (i) an empi
 rical comparison of multiple-instance learning\nto ordinary supervised lea
 rning\, (ii) a method for learning to combine\npredictions in a multiple-i
 nstance setting\, and (iii) active\nmultiple-instance learning.
LOCATION:SW01 Computer Laboratory
END:VEVENT
END:VCALENDAR