Discriminative Log-Linear Grammars with Latent Variables
Slav Petrov and Dan Klein, UC Berkeley
Refine the observed grammar categories with latent variables:
S-h NP-h VP-h NP-h DT-h NN-h th e noise .- h .

Poster ID M64

Set parameters to maximize likelihood of correct parse tree given a sentence:

max


PRP-h VBD-h She heard

h

P(y , h|x,  )

tree clusters words parameters Discriminative training is superior to generative training on a full scale parsing task:
F1-score Exact Match
40% 90%

Approximate feature expectations using hierarchical pruning procedure:
... VP NP QP ... ... VP1 VP2 NP1 NP2 ... G0 G1

generative
89% 37%

discriminative

... VP3 VP4 ... NP3 NP4 G2 ... VP6 VP7 ... NP3 NP4 ... G3

88%

34%

87%

31%