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%)