Transition Based Dependency Parsers

These are implementations of the (unlabeled) arc-eager and arc-standard dependency parsing algorithms.
These parsers are very fast and are reasonably accurate.
In particular, the arc-standard parser with the features described in [1] (the default feature set) can achieve very competitive accuracies.

The input file for both training and parsing should be in CoNLL format (see conll.example).
Columns 8,9,10 are always ignored (but must be present).
When parsing new text, you can put whatever you want in column 7, the parser will overwrite it (it uses this column to report accuracy scores)

Compiling:
==========
Speed is achieved using a c/cython extension module. 
This needs to be compiled using either cython or a c compiler.
See instructions in ml/README

Training the parsers:
=====================

   ./eager.py -o model_file [options] conll_input_file

   or

   ./standard.py -o model_file [options] conll_input_file

   (use -f instead of -o to create feature vector files for training with an external classifier.  If you don't know what it means,
    just ignore this option.  The model file format is the same as Megam's.)

Parsing new text with the trained model:
========================================

   ./eager.py -m model_file [options] conll_file_to_parse > output

   or 

   ./standard.py -m model_file [options] conll_file_to_parse > output


References:
~~~~~~~~~~~
[1] Liang Huang, Wenbin Jiang and Qun Liu. 2009.
    Bilingually-Constrained (Monolingual) Shift-Reduce Parsing.