aded with grey) were shared Po e S n act c Topic Models ledsttr M4 role ofya tgenieric part he Jordan Boyd-Graber and David Blei ic specification. OPrheetotoUniversity t inc r n pics, t'osumeemantic constraints. s nts are collections of parse trees. Dc

z1 w1:START z4 w4:START

 k

D d

DP(D , ) de: k  M ent z0  d start  ,0  mult(1, z0 ) The latent class depends on the parent node Learned topics are consistent with both and adocument's topic pn , n . . . and theme. d,nthe nd parentdistribution.  {1, syntaxdn } ent zd,n  d zp(d,n)
z3 w3:they

}:

z2

w2:slept

z5

w2:ran

T