Unnormalized Models

This is the recipe for this.
Random fields,
exponential models,
motivated from (turn

your head
and say natural language
processing

). Segmenting and
labeling sequences. A
framework

                  based on
conditional random fields
offering several

advantages over
hidden Markov models and
stochastic grammar.

(she was thin
I thought
not normal I
liked her segments
enough to fill
the universe with a 2-d
string)

Second, we derive an equivalence
between the well-known
technique of boosting and maximum
likelihood for exponential
models. The idea of
unnormalized models plays
a key role.