Synthetic Data Made to Order: The Case of Parsing
Implementation of the paper "Synthetic Data Made to Order: The Case of Parsing" by Dingquan Wang and Jason Eisner. EMNLP 2018
Requirements
- Python3
- PyTorch 0.3
Run
-
To permute a source treebank
src.conllu
in UD format towards a target language with only POS-tagstgt.txt
in POS-spaced format (seedata/fr.txt
) and output tosrc~tgt.conllu
:python src/main.py --src src.conllu --tgt tgt.txt --output src~tgt.conllu
For convenience,
--tgt
could also take UD format input (endswith.conllu
) which will ignore everything but the POS-tags. -
To pre-train a self-permutation model and use it as initialization:
python src/main.py --task self_model --src src.conllu --model $(pwd)/pre_model.pkl python src/main.py --src src.conllu --tgt tgt.txt --output src~tgt.conllu --pretrain $(pwd)/pre_model.pkl
-
For more options, please use:
python src/main.py --help
the default hyperparameters are the ones used in the original paper
-
We also release a sample data in
data
and a scriptpermute.sh
Reference
@inproceedings{wang-eisner-2018-emnlp,
author = {Dingquan Wang and Jason Eisner},
title = {Synthetic Data Made to Order: The Case of Parsing},
booktitle = {Proceedings of the Conference on Empirical Methods in
Natural Language Processing (EMNLP)},
year = {2018},
month = nov,
address = {Brussels},
url = {http://cs.jhu.edu/~jason/papers/#wang-eisner-2018-emnlp}
}