/span-based-dependency-parsing

Source code of "Headed Span-Based Projective Dependency Parsing" and "Combining (second-order) graph-based and headed span-based projective dependency parsing

Primary LanguageHTML

span-based-dependency-parsing

Source code of "Headed Span-Based Projective Dependency Parsing" and "Combining (second-order) graph-based and headed span-based projective dependency parsing"

Setup

prepare environment

conda create -n parsing python=3.7
conda activate parsing
while read requirement; do pip install $requirement; done < requirements.txt

prepare dataset:

you can download the datasets I used from link.

Run

python train.py +exp=base  datamodule=a model=b seed=0
a={ptb, ctb, ud2.2}
b={biaffine, biaffine2o, span, span1o, span1oheadsplit, span2oheadsplit}

multirun example:

python train.py +exp=base datamodule.ud2.2 model=b datamodule.ud_lan=de,it,en,ca,cs,es,fr,no,ru,es,nl,bg seed=0,1,2 --mutlirun

For UD, you also need to prepare the JAVA environment for the use of MaltParser.

TODO

  • Clean code (e.g. add comments)
  • Add eval.py, now we only support training from scratch.
  • Release pre-trained model.

Contact

Please let me know if there are any bugs. Also, feel free to contact bestsonta@gmail.com if you have any questions.

Citation

@misc{yang2021headed,
      title={Headed Span-Based Projective Dependency Parsing}, 
      author={Songlin Yang and Kewei Tu},
      year={2021},
      eprint={2108.04750},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

@misc{yang2021combining,
      title={Combining (second-order) graph-based and headed span-based projective dependency parsing}, 
      author={Songlin Yang and Kewei Tu},
      year={2021},
      eprint={2108.05838},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Acknowledge

The code is based on lightning+hydra template. I use FastNLP as dataloader. I use lots of built-in modules (LSTMs, Biaffines, Triaffines, Dropout Layers, etc) from Supar.