CHANGES

v 0.2.0
-------

 - Reworked grammar informed initialization from Baldridge COLING 2008
   paper that never made it in its final form into TexNLP. Results are
   now very much like, though not exactly the same as, the 2008 paper
   results.

 - Added little utility to list all possible category combinations and
   give them a value (possible/impossible, low/heigh weight, prior
   prob values).

 - Minor cleanup and documentation here and there.

v 0.1.6
-------
 - Switched to TexNLP (texnlp) namespace

 - Scripts now begin with 't' rather than 'a'

 - Fixed various minor issues, underspecificed usage of generic classes, 
   variables not read, unused imports

v 0.1.3
-------
 - Computes transition probs given both the possible i and j
   states. Helps both HMM and MEMM in different ways. For HMM, it
   means only looking up the emmission prob once per j word. For MEMM,
   it means we can compute the weights for all features except the
   prev tag feature. This is now handled in TadmClassifier.

 - Tagging of development data also prints out perplexity.

 - Sped up forward-backward by swapping the state->possibleOutputs map
   for an output->possibleStates map, so now values are only computed
   for states that actually matter at each time step.


v 0.1.2
-------
 - Added MEMM and extensive features based on the words for it to
   exploit.

 - Created a hierarchy for MarkovModels so that MEMMs and HMMs could
   share code.

 - Forward-backward now reads all data into memory rather than reading
   from disk on each iteration.

v 0.1.1
-------
 - Created AustinNLP packages and classes from a bunch of code I
   developed for working with CCG grammars and training and using
   HMMs. (Jason)