A Bayesian LDA-Based model for semisupervised part-of-speech tagging
Kristina Toutanova and Mark Johnson
Unlabeled data

Spotlight Id M65

Bayesian generative model
Ambiguity Class Model + Context Model
P(t|walks,{noun,verb}) ~ Dir(,) walks
(Sparse prior)

Partial supervision

{noun verb}

v She bursts into tears and walks away . v

government : {noun} benefit : {noun,verb} last: {noun, verb, adverb, adjective} net: {adjective, noun, verb} at: {noun, prep., part.}

if something walks like a duck no no s

n
really fun walks all around the area

40 to 50% error reduction over the state of the art