Pereira, Tishby and Lee

Talks specifically about clustering to relieve data sparseness.

I had a bit trouble seeing whether the measures they used showed an
advantage for clustering.

Baker and McCallum

Provides a method for using known class labels to identify sets of
words that appear in similar distributions of documents.  The
resulting word clusters/categories are used for text classification in
Naive Bayes.

Can we get the benefits of clustering (shared statistics) without
hard clusters?  Yes, I think so---that's what we're trying to do for
Markov models.  We estimate the statistics for a set of items that are
sufficiently similar.  This is done without making a hard partition,
but instead using kernels.  It remains to be seen whether this is more
accurate and/or whether it can be implemented accurately.

Slomin and Tishby

Aren't words already "clusters"?  How can we bootstrap a process like
this?

Have people looked at MDL clustering?  Could that be a way to choose
the number of clusters?

All the papers from today used real evaluations, which is nice.