BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Recent work on POMDP-based dialog systems at AT&T - Jason Williams
  (At&T)
DTSTART:20080512T120000Z
DTEND:20080512T130000Z
UID:TALK12168@talks.cam.ac.uk
CONTACT:Dr Marcus Tomalin
DESCRIPTION:Building spoken dialog systems is difficult because speech rec
 ognition\nerrors are common and user's behavior is unpredictable\, which i
 ntroduces\nuncertainty in the current state of the conversation.  At AT&T\
 , we have\nbeen applying partially observable Markov decision processes (P
 OMDPs) to\nbuilding these systems.  We model the uncertainty in the dialog
  state\nexplicitly as a Bayesian network and apply machine learning techni
 ques\nto determine what the system should say or do.\n\nIn this talk\, I'l
 l review the overall approach of applying statistical\ntechniques and then
  describe two recent advances: first\, because the\nsystem must operate in
  real-time\, efficient Bayesian inference is\ncrucial\, yet the set of pos
 sible dialog states is enormous.  To solve\nthis\, I'll present a techniqu
 e which uses a particle filter to perform\napproximate inference in real-t
 ime.  Second\, to choose actions\, ideally\nwe would like to combine the r
 obustness of machine optimization with the\nexpertise of human designers. 
  To tackle this\, I'll present a method\nwhich unifies human expertise wit
 h automatic optimization.\n\nTo illustrate these techniques\, I'll provide
  examples of two dialog\nsystems: a voice dialer\, and a troubleshooting s
 ystem that helps users\nrestore connectivity on a failed DSL connection.  
 Graphical displays\nillustrate the operation of the techniques\, and quant
 itative results\nshow that applying statistical techniques outperforms the
  traditional\nmethod of building systems by hand.
LOCATION:LR5\, Engineering Department\, Baker Building
END:VEVENT
END:VCALENDAR