Can anyone make it clear why the  ``label  bias  problem''  is  a
problem?    The   authors'  explanation  is  lost  on  me:  ``the
transitions leaving a given state compete only against each other
[sic],  rather than against all other transitions in the model.''
Why is this a problem? In a given state, only transitions leaving
that state can be taken, right?

I hoped the example in figure 1 would help, but it just  seems  a
useless  model for distinguishing between ``rib'' and ``rob'': it
forces a commitment to one or the other on the first letter, when
the  two  words  simply  can't  be  disambiguated. Why use such a
model? What is the point here? What is this ``score  mass''  they
invoke? (Anyone read Bottou 1991?)