Joint vs. Conditional Models
Joint (generative) models p lac e p rob ab ilities over
b oth ob served data and th e h idden stu f f (gene-
rate th e ob served data f rom h idden stu f f ):
A ll th e b est k now n S tatN L P models:
n- gram models, Naive Bayes classifiers, hidden
M ark ov models, p rob ab ilist ic cont ex t -free grammars
Discriminative (conditional) models tak e th e data
as g iven, and p u t a p rob ab ility over h idden
stru ctu re g iven th e data:
L ogist ic regression, condit ional loglinear models,
max imu m ent rop y mark ov models, ( S V M s,
p ercep t rons)
P (c, d)
P (c| d)
