Words and documents

There ought to be a difference between things which are frequent in all documents (e.g of the) and those which are frequent in some only (e.g. sherlock holmes).

• The binomial model, and its relative the Poisson distribution don't take account of the ``burstiness'' of words.
• The negative binomial does, used by Church to find ``interesting words'' and by Mosteller and Wallace to discriminate authorship