Our system is written in Haskell with some human interface features in written in bash. The Haskell component depends on the following modules.

• Data.Char
• Data.List
• Data.List.Split
• Data.Map
• Data.Maybe
• Data.Ratio
• Data.Set
• System.IO
• System.IO.Unsafe
• Data.MemoCombinators

The following sort of command will run the system

sh run_test.sh <training corpus size in words> > <output file>.

For example

sh run_test.sh 100000 > foo.pos.

Because we did not develop a command-line interface for passing parameters, the Probabilities.hs file must be recompiled to test different component arrangements. The relevant lines are below.

    let unk = True
        smooth = False -- add-one smoothing
        ...
        taggrams = safe_ngram_tally 2 tags

unk and smooth control unknown word handling and smoothing respectively. Setting either to True will turn it on, and setting it to False will turn it off.

The number in safe_ngram_tally 2 tags controls the n-gram length. Set it to 2 to use bigrams, 3 to use trigrams, and so on.

After adjusting the parameters, compile the executable like so

$ ghc --make Probabilities.hs

and then run run_test.sh as explained before.

You can submit the resulting predictions like this. (This depends on curl.)

$ sh submit_predictions.sh <netid> results/foo.pos

Save the emails with the submission results in plain text as separate files (.eml) or as one file (.mbox), and run the following command to put them in a convenient format.

$ sh tabulate_predictions.sh emails.mbox > results/results.csv

The email filename(s) may not contain spaces.