BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Unsupervised Word Alignment and Part of Speech Induction with Undi
 rected Models - Chris Dyer\, Carnegie Mellon University
DTSTART:20111028T110000Z
DTEND:20111028T120000Z
UID:TALK34214@talks.cam.ac.uk
CONTACT:Thomas Lippincott
DESCRIPTION:This talk explores unsupervised learning in undirected\ngraphi
 cal models for two problems in natural language processing.\nUndirected mo
 dels can incorporate arbitrary\, non-independent features\ncomputed over r
 andom variables\, thereby\novercoming the inherent limitation of directed 
 models\, which require\nthat features factor according to the conditional 
 independencies of an\nacyclic generative process. Using word alignment (fi
 nding lexical\ncorrespondences in parallel texts) and bilingual part-of-sp
 eech\ninduction (jointly learning syntactic categories for two languages\n
 from parallel data) as case studies\, we show that relaxing the\nacyclicit
 y requirement lets us formulate more succinct models that\nmake fewer coun
 terintuitive independence assumptions. Experiments\nconfirm that our undir
 ected alignment model yields consistently better\nperformance than directe
 d model baselines\, according to both intrinsic\nand extrinsic measures. W
 ith POS tagging\, we find more tentative\nresults. Analysis reveals that o
 ur parameter learner tends to get\ncaught in shallow local optima correspo
 nding to poor tagging\nsolutions. Switching to an alternative learning obj
 ective (contrastive\nestimation\; Smith and Eisner\, 2005) improves the st
 ability and\nperformance\, but it suggests that non-convex objectives may 
 be a\nlarger problem in undirected models than with directed models.
LOCATION:FW26\, Computer Laboratory
END:VEVENT
END:VCALENDAR