Poster W9

Agreement-Based Learning
Percy Liang, Dan Klein, Michael I. Jordan

Problem: learning complex hidden-variable models Traditional solution: approximate EM E
one intractable model
µ

M



E ...

Our solution: product EM (train submodels to agree) E 
two tractable submodels
µ1 µ2 µ ¯

M

1 2

¯ 

E ...

+µ ¯

+ ¯

E M E Applications: unsupervised NLP, phylogenetic HMMs