Rough grained model
Slightly different
model of splitting than
earlier work
Smaller amount of
training examples
No significant gain
from co-training but
comparable to other
work (19.5% error vs.
14-18 error%)