Correlated Bigram LSA for Unsupervised LM Adaptation Yik-Cheung (Wilson) Tam & Tanja Schultz (T17)
Graphical Model Topic p ri o r Dirichlet-Tree() Top bigrams sorted by P(u,v|k) k =6 1 k =6 2 topic k =6 3 Large-scale evaluation on Arabic (GALE Phase 3)
dev07



Zi wi

wi-1

25

N M

21 21 20. 7

BG 4-gram
15.415. 314. 9

20 WER 15 (%) 10 5 0 BN BC Overall
12.412. 211.8

+LSA

P(v|u,k)

k-th bigram LM with fractional Kneser-Ney smoothing

+bigram LSA (Kneser-Ney)