A general purpose tagger and a greedy transition-based dependency parser, both with CNN as character composition model.
The tagger is described in the SCLeM2017 Paper: https://arxiv.org/abs/1706.01723
The parser is described in the ACL2017 paper (with slight improvement): https://arxiv.org/abs/1705.10814
Requirements:
- Python 2.7
- Numpy
- Theano 0.9.0
- Lasagne 0.2-dev1
Basic usages:
# Train the tagger
python main.py -tool tagger -train [train_file] -dev [dev_file] -model_to [model_dir] -source_feat word,char -target_feat utag
# Test the tagger
python main.py -test [test_file] -model_from [model_dir] -out [output_file]
# Train the parser
python main.py -tool parser -train [train_file] -dev [dev_file] -model_to [model_dir] -source_feat word,char
# Test the parser
python main.py -test [test_file] -model_from [model_dir] -out [output_file]