hevalazizoglu / parser-v2

An updated version of the Parser-v1 repo, used for Stanford's submission in the CoNLL17 shared task.

Parser v2

This is the parsing system that currently (as of 06/18/2017) holds state-of-the-art performance at the CoNLL2017 shared task on Universal Dependency parsing. It's under active development, so new features are likely to break old models, and it contains a bit of experimental stuff that hasn't been published or isn't working yet.

Training the parser

If you don't care about hyperparameter settings at all, the basic commands to train and run the parser are the following:

python main.py --save_dir saves/default train
python main.py --save_dir saves/default parse /path/to/treebank/*.conllu

Parsed files are by default saved in the save directory with the same name as the original file. You can also specify the output directory with the --output_dir flag, and when parsing a single file you can give it a new name with the --output_file flag.

Of course, you need to tell the parser where the train and validation data is located, and there are a lot of hyperparameters to play with. You can check them all out in config/defaults.cfg, but the most important ones (including the location of the training datasets) have been condensed down into a separate configuration file called config/template.cfg. What you probably want to do is make your own copy of config/template.cfg (we'll say config/my_config.cfg), which you can then freely modify. Any parameters not specified here are loaded in from config/defaults.cfg. Once you've tweaked the settings to your liking, you can train a model that uses it with the following command:

python main.py --save_dir saves/my_model train --config_file config/my_config.cfg

The model saves all the configuration settings in the save directory, so you don't need to re-specify this file when running the model:

python main.py --save_dir saves/my_model parse /path/to/treebank/*.conllu

You might want to keep most settings the same but change one or two on the command line without re-editing the configuration file. To do this, you specify --config_heading setting1=value1 setting2=value2 .... For example, to only use the first 500,000 entries of a pretrained embedding matrix (which can speed up loading time for test runs), you would run the following:

python main.py --save_dir saves/my_model train --config_file config/my_config.cfg \
                                               --pretrained_vocab max_rank=500000

Again, the model saves this information, so you don't need to specify it again when parsing:

python main.py --save_dir saves/my_model parse /path/to/treebank/*.conllu

More documentation to come!