TUPA is a transition-based parser used as a baseline system in the CoNLL 2019 Shared Task on Cross-Framework Meaning Representation Parsing. It was originally built for Universal Conceptual Cognitive Annotation (UCCA), and extended to support DM, PSD, EDS and AMR.
- Python 3.6+
Install the latest code from GitHub:
pip install git+https://github.com/danielhers/tupa.git@mrp
Having a directory with MRP graph files (for example, the MRP 2019 UCCA data), run:
python -m tupa -t <train_dir> -d <dev_dir> -m <model_filename>
Alternatively, download any of the pre-trained models for MRP 2019 and MRP 2020.
Preprocess a text file (here named example.txt) using UDPipe:
udpipe --tag --parse --input horizontal --tokenizer "ranges;presegmented;normalized_spaces" --output conllu english-ewt-ud-2.4-190531.udpipe < example.txt > example.conllu
Convert the output to mrp using mtool:
tool/main.py --read conllu --write mrp < example.conllu > example.mrp
Run the parser using a trained model:
python -m tupa example.mrp -m <model_filename>
An mrp file will be created per instance (separate by newlines in the text file).
If you already have a preprocessed mrp file (for example, from the shared task data), then there is no need to run UDPipe and mtool.
- Daniel Hershcovich: [email protected]
- Ofir Arviv: [email protected]
If you make use of this software, please cite the following paper:
@InProceedings{hershcovich-arviv-2019-tupa,
author = {Hershcovich, Daniel and Arviv, Ofir},
title = {{TUPA} at {MRP} 2019: A Multi-Task Baseline Syste},
booktitle = {Proc. of CoNLL MRP Shared Task},
year = {2019},
pages = {28--39},
url = {https://www.aclweb.org/anthology/K19-2002},
doi = {10.18653/v1/K19-2002}
}
This package is licensed under the GPLv3 or later license (see LICENSE.txt).
