- definitions:
-
features -- any statistic in the data can be expressed as an
indicator function
-
constraints -- force the model to agree with the training data
on how often the model exhibits the feature
- knowledge sources
-
each knowledge source gives rise to a constraint
-
suppose we want to estimate P(Know/You).
-
simple estimate is K = C(YOU,KNOW)/C(YOU)
-
in ME framework require that above estimate is true
"on an average" -- this becomes a constraint
-
feature is an indicator function which is 1 when KNOW
follows YOU and 0 otherwise