Holistic Exploration on UDS Parsing [paper]
mkdir datasets
cd datasets
wget 'https://data.statmt.org/news-crawl/en/news.2021.en.shuffled.deduped.gz'
wget 'https://nlp.stanford.edu/data/glove.840B.300d.zip'
unzip glove.840B.300d.zip
gzip -d news.2021.en.shuffled.deduped.gzFirst install PredPatt and decomp following the instructions in the link, then run:
pip install -r requirements.txtFor naive model training, run:
python heuds/main.py train --task UDSTask --arch Bert_UDS --save-dir 'Bert_naive' --encoder-output-dim 1024 --layer-in-use 0,0,1,1,1,1,1For model training with additional syntactic information, run:
python heuds/main.py train --task UDSTask --arch Bert_UDS --save-dir 'Bert_incorpsyn' --encoder-output-dim 1024 --contact-ud --syntax-edge-gcnFor our best model training with additional syntactic information and data augmentation method, run:
python heuds/main.py train --task UDSTask --arch Bert_Syntactic --save-dir 'Bert_syntactic' --encoder-output-dim 1024
python heuds/main.py generate --task ConlluTask --arch Bert_Syntactic --save-dir 'Bert_syntactic' --encoder-output-dim 1024 --mono-file datasets/news.2021.en.shuffled.deduped --conllu-file datasets/news.conllu
python heuds/main.py train --task PredPattTask --arch Bert_UDS --save-dir 'Bert_best_pretrained' --max-epoch 30 --encoder-output-dim 1024 --layer-in-use 1,1,1,1,1,0,0 --conllu datasets/news.conllu --name news --validate-interval -1 --contact-ud --syntax-edge-gcn
python heuds/main.py train --task UDSTask --arch Bert_UDS --save-dir 'Bert_best' --pretrained-model-dir 'Bert_best_pretrained' --encoder-output-dim 1024 --lr 2e-5 --pretrained-lr 1e-6 --contact-ud --syntax-edge-gcnReplace "train" to "test" for model evaluation.
If you find this work helpful, please consider citing as follows:
@article{deng2023hexp4uds,
title = {Holistic Exploration on Universal Decompositional Semantic Parsing: Architecture, Data Augmentation, and LLM Paradigm},
author = {Hexuan Deng, Xin Zhang, Meishan Zhang, Xuebo Liu, Min Zhang},
journal = {arXiv preprint arXiv:2307.13424},
year = {2023}
}
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent