
-> We submitted the Hindi projected data that we extracted from TIDES parallel corpora.
	Bulgarian and Spanish data had license issues and were not distributable.
	The data is in 10 column CoNLL format with the unconnected words taking a "-1" parent.

-> The data\ directory has three files
		1. "hindi-projected-full-TIDES.conll"
			- The entire projected treebank extracted from TIDES parallel corpus in CoNLL format.
		2. "hindi-projected-10k.conll"
			- Contains 10,000 sentences randomly extracted from the full projected treebank. 
			This dataset was used for all our experiments.
		3. "hindi-test-gold.conll" 
			- Test data used for evaluating our system.


-> Two java packages GNPPA and EGNPPA corresponding to the two parsers described 
	in the paper are available.

-> Run the two parsers using the following commands :
	$ java -cp .:trove.jar gnppa.Parser <specification-file>
	$ java -cp .:trove.jar egnppa.Parser <specification-file>


->To emulate the results obtained in the paper, please use the specification files
	readily provided in the specs/ directory. They train as well as evaluate on the test data.

		specs/hindi-egnppa.train  
		specs/hindi-gnppa.train  

	For example,
		$ java -cp .:trove.jar gnppa.Parser specs/hindi-gnppa.train
		$ java -cp .:trove.jar gnppa.Parser specs/hindi-egnppa.train

-> 
	All the parameter models are saved in models/ directory. The outputs of the
	parsers on the test data after evaluation are stored in outputs/ directory


-> "eval07.pl" in the parent directory is the CoNLL Shared Task 2007 evaluation script.
	Please use it to get the accuracies without punctuation.
	After training the parsers using the commands given above. Execute 
	$ perl eval07.pl -q -p -g data/hindi-test-gold.conll -s outputs/hindi-gnppa-i1-k1-pos1-local0.evaloutput
	$ perl eval07.pl -q -p -g data/hindi-test-gold.conll -s outputs/hindi-egnppa-i2-k1-pos2-local0.evaloutput

	Remove the "-p" option from the above commands to get the accuracies on the outputs including punctuations.


Please report problems to prashanth@research.iiit.ac.in.
