Performance documentation
  - Misuse of "comprise" -- c.f. "The Elements of Style" or a good
    dictionary.  (Common error.)
  - "through", not "thru" (informal for a document like this)
  - Learning curves -- excellent.
  - Usually, we don't report accuracy rates averaged over all
    training sizes, but only at asymptote (the other number doesn't
    necessarily really reflect the accuracy we would see "in
    practice" after fully training the system).
  - Token counts vs # emails -- nice.
  - Nice analysis of Krusty vs. Whitespace.  Hard to tell if Krusty
    is really doing "better" than Whitespace, though.  At asymptote
    of Fig 1, there's still some variability in accs.  The two might
    be the same, or might be diff.  Remember, as well, that you're
    only testing on 40 emails -> you only have a resolution of 2.5%
    in measuring accuracy -- the results for Krusty and WhiteSpace
    look pretty close according to that.
  - NGram -- did you consider using n!=5 (e.g., 6, 7, 8, ...)?  Any
    thoughts on how that might effect performance?
  - What about different types of errors?  (E.g., spam->normal
    vs. normal->spam)
  - "good enough" -> "well enough"
  - Good observations on "future work", esp. on tokenizers and
    training.  More details on "better algorithm"?
  - Good observation on serialization -- might be possible to
    serialize these models faster, or might be better off just
    replacing it with a direct I/O system.
User/API docs
  - README.TXT
    - Nice description of compilation.  Good to specify dependency on
      getopt.
    - You can probably get away with something like "getopt must be on
      your CLASSPATH.  Either add it to the CLASSPATH, or use the
      -classpath option to javac and java.  For further details,
      refer to..." and then ignore it thereafter.
  - USERDOC.TXT
    - Tool overviews.  Good.
    - "BSFTrain accepts its training files either from standard input
      or from direct access." -- "direct access" ==?  From files
      specified on the command line?
    - Should say that in training mode, it generates and saves the
      models to disk for use by BSFTest.  (And in dump mode, that it
      doesn't save any.)
    - Should specify which tokenizers are available, what each does,
      how to identify them on the command line, and possibly which
      one is recommended (might refer to performance docs for
      additional info on these)
    - If you provide dynamic instantiation of third party tokenizers
      (cool, BTW), you should describe that in the user docs, telling
      the user how to _add_ a 3rd party tokenizer.  You should also
      provide a ref to API docs that describes how an independent
      developer might make a new tokenizer.
    - Nice examples of usage.  Interpreted output for reader --
      good.  Explained dump -- excellent.
    - Shouldn't output of BSFTest include the whole email (plus
      X-Spam-Status header), not just the header?