To use my implementation of the Yarowsky's algorithm, please
follow the steps below.

0. Download and untar yarowsky.tar to your working directory.

1. Prepare Data
   1.0 Pick a polysemous word and some of its senses to
   disambiguate. Number the senses as 1, 2, ..., where 1 stands
   for the default sense.
   1.1 Training data -- cluster of discourses containing one or
   more instances of the polysemous word, one discourse per line.
   1.2 Seed examples -- chosen from the training data that are
   representative of each sense to disambiguate. The number of
   seed examples for each sense should be roughly 1% of the size
   of training set. Tag the instances with corresponding senses
   by putting "(X)" next to each instance where X is the sense
   number for that instance.
   1.3 Test data -- cluster of discourses containing one or more
   instances of the polysemous word, one discourse per line. Test
   set is untagged, and so is the training set.

2. Run Yarowsky's Algorithm on Training Set
   % perl yarowsky.pl [word] [num_senses] [window_size] [seed_set] [train_set]
   For example: at /clair6/projects/msa/siwei/NLPCourse/
   % perl yarowsky.pl plant 2 5 plant-seed-2 plant-train
   would disambiguate word "plant" for two of its meanings with
   window size 5 based on seed examples in file "plant-seed-2"
   and training set "plant-train". The user does not have to tell
   yarowsky.pl about the senses being used as long as they tag
   the seed examples accordingly (refer to Step 1 above). 

   The outputs of this step would be a list of decision list files
   and an output-train file under the same directory where you run the
   command. 

   decision_list_X are the decision lists produced at each
   iteration of the algorithm. The one with the maximum value for
   X is the final version generated. 

   output-train is a copy of the training set with all instances of the
   polysemous word tagged with one of the candidate senses. 

3. Run Yarowsky's Algorithm on Test Data
   % perl yarowsky-test.pl [word] [num_senses] [window_size] [test_set] [decision_list]
   For example: at /clair6/projects/msa/siwei/NLPCourse/
   % perl yarowsky-test.pl plant 2 5 plant-test decision-list

   Note that the decision-list in this step could be any of the
   ones generated in Step 2. Usually (and most practically), the
   final decision list of Step 2 should be used in this step.  

   The output of this step is an output-test file under the same
   directory where you run the command. It is a copy of the test
   set with all instances of the polysemous word tagged with one
   of the candidate senses. 

For questions and concerns, contact Siwei Shen at shens@umich.edu