What is it:
============
This is the code for the Easy-first non-directional dependency parser 
described in the papers:

   An Efficient Algorithm for Easy-First Non-Directional Dependency Parsing
   Yoav Goldberg and Michael Elhadad
   NAACL 2010

   Easy First Dependency Parsing of Modern Hebrew
   Yoav Goldberg and Michael Elhadad
   SPMRL 2010 (NAACL Workshop on Statistical Parsing of Morphologically-rich Languages)


research quaility code (== crappy). provided without any guarantees.

if you use this for academic research, we'd appreciate if you cite the NAACL-2010 paper.

Requirements:
=============
- python 2.5 or above (not compatible with python3)
- cython and a c compiler for compiling the ml module

Installation:
=============
- copy to a folder
- install cython:
   easy_install cython
- compile the ml module:
   cd ml
   python setup.py build_ext --inplace

   
Usage:
======
short:

   python train.py -o naaclmodel -f features/engfeatures2.py data/WSJ
   python parse.py -m naaclmodel.model data/TEST > parsed

longer:

Training a model:
``````````````````

      python train.py -o model -f features [options] train_file [dev_file]

   for example, assuming data/WSJ contains the WSJ training data in CoNLL 
   format, training a model with the NAACL-2010 features:

      python train.py -o naaclmodel -f features/engfeatures2.py data/WSJ

   this will produce the files:
      naaclmodel.model
      naaclmodel.weights.N  (for N between 1 and 20)
      naaclmodel.weights.FINAL  
   
      the .N files are the weights after the Nth round of training, and 
      can be deleted if you don't plan on using them.

   if dev_file is supplied, will eval on it whenever weights are saved.
   
   options:
      --help

      --iters N
         run of N iterations [default 20]

      --every N  
         save weights every N iterations [default 1]

Parsing with / Testing a trained model:
```````````````````````````````````````
      python parse.py -m model [options] input_file

   for example, to parse the data/TEST file with the naaclmodel.model 
   created above:
      
      python parse.py -m naaclmodel.model data/TEST > parsed

   this will parse with the FINAL weights.  
   Specifying specific iteration weights is also possible:

      python parse.py -m naaclmodel.model --iter 2 data/TEST > parsed

   options:
      --help

      --eval
          print accuracy information (assuming input file includes 
          parents information).  

      --nopunct 
          ignore punctuation marks (for English) when calculating 
          the accuracy.

      
Support:
========
email yoavg / cs.bgu.ac.il