Homework 3: Machine Learning
		Due:    Dec 4 2003
		Points: 133

For this homework assignment, you will be asked to perform some
machine learning experiments on the TDT data you worked with in
homework 2.  You are expected to use 'ripper', a program available on
the cluster machines for inducing classification rules from a set of
preclassified examples.  The ripper executable is located here:

  /proj/nlp/tools/ML/ripper2/ripper

Here are some resorces to help get you aquainted with ripper:

 1. '/proj/nlp/users/cs4705/ripper_howto.txt' is a step-by-step guide
    to running ripper, which may be a bit easier to follow at first
    than the man page.

 2. To view the man page, type: 'man /proj/nlp/tools/ML/ripper.man'.

 3. '/proj/nlp/users/cs4705/ml95-paper.ps' shows use of ripper
    by it's creator in some classification experiments of his own.

 4. For some sample data, you may want to check out these directories:
        /proj/nlp/users/cs4705/pdata/
        /proj/nlp/users/cs4705/tdata/

========================================================================
OBJECTIVE

Run a classification experiment to automatically learn rules to
predict the topic of a document in the TDT corpus (in ripper
terminology, a 'topic' is synonomous with a 'class'.  The two terms
will be used interchangeably).  You may extract whatever features you
chose.

You will be working with these two files:
    /proj/nlp/users/cs4705/train.sgml
    /proj/nlp/users/cs4705/test.sgml
  
You will be training on train.sgml and testing on test.sgml.  Once you
have obtained the best performance you can, supply a write up
describing the feature set you used, the error rate on test.sgml, and
the ripper .hyp file.  Describe some surprises any surprises you
encountered in the process.

You must also submit one program that takes a training file and a
testing file in .sgml format and produces the three necessary ripper
files: .names, .data, and .test.  This is necessary because we will be
testing how well your feature set performs on a validation test set,
which you will not have access to.

========================================================================
FURTHER GUIDELINES

Each document in the TDT corpus has one of the following classes
(topics):

 Conspiracy
 Crime and criminals
 Finance, Public
 Outer space--Exploration
 Strikes and lockouts

For ease of processing and compatibility between homeworks, when
reporting your results and rule sets, apply the following mapping:

 Conspiracy                => CONSPIRACY 
 Crime and criminals       => CRIME
 Finance, Public           => FINANCE
 Outer space--Exploration  => SPACE
 Strikes and lockouts      => STRIKES