yzhangcs / crfsrl

[COLING'22] Code for "Semantic Role Labeling as Dependency Parsing: Exploring Latent Tree Structures Inside Arguments".

Semantic Role Labeling as Dependency Parsing: Exploring Latent Tree Structures Inside Arguments

Yu Zhang¹  Qingrong Xia^1,2  Shilin Zhou¹  Yong Jiang³  Guohong Fu¹  Min Zhang¹ 

¹Soochow University, Suzhou, China

²Huawei Cloud, China

³DAMO Academy, Alibaba Group, China

Citation

If you are interested in our work, please cite

@inproceedings{zhang-etal-2022-semantic,
  title     = {Semantic Role Labeling as Dependency Parsing: Exploring Latent Tree Structures inside Arguments},
  author    = {Zhang, Yu  and
               Xia, Qingrong  and
               Zhou, Shilin  and
               Jiang, Yong  and
               Fu, Guohong  and
               Zhang, Min},
  booktitle = {Proceedings of COLING},
  year      = {2022},
  url       = {https://aclanthology.org/2022.coling-1.370},
  address   = {Gyeongju, Republic of Korea},
  publisher = {International Committee on Computational Linguistics},
  pages     = {4212--4227}
}

Setup

The following packages should be installed:

• PyTorch: >= 1.12.1
• Transformers: >= 4.2

Clone this repo recursively:

git clone https://github.com/yzhangcs/crfsrl.git --recursive

Run the following scripts to obtain the training data. Please make sure PTB and OntoNotes are available:

bash scripts/conll05.sh PTB=<path-to-ptb>             SRL=data
bash scripts/conll12.sh ONTONOTES=<path-to-ontonotes> SRL=data

Run

Try the following commands to train first-order CRF and second-order CRF2o models:

# LSTM
# CRF
python -u crf.py   train -b -c configs/conll05.crf.srl.lstm.char-lemma.ini   -d 0 -f char lemma -p exp/conll05.crf.srl.lstm.char-lemma/model   --cache --binarize
# CRF2o
python -u crf2o.py train -b -c configs/conll05.crf2o.srl.lstm.char-lemma.ini -d 0 -f char lemma -p exp/conll05.crf2o.srl.lstm.char-lemma/model --cache --binarize
# BERT finetuning
# CRF
python -u crf.py   train -b -c configs/conll05.crf.srl.bert.ini   -d 0 -p exp/conll05.crf.srl.bert/model   --batch-size=2000 --encoder bert --bert bert-large-cased --cache --binarize
# CRF2o
python -u crf2o.py train -b -c configs/conll05.crf2o.srl.bert.ini -d 0 -p exp/conll05.crf2o.srl.bert/model --batch-size=2000 --encoder bert --bert bert-large-cased --cache --binarize

To do evaluation:

# end-to-end
python -u crf.py   evaluate -c configs/conll05.crf.srl.bert.ini   -d 0 -p exp/conll05.crf.srl.bert/model
# w/ gold predicates
python -u crf.py   evaluate -c configs/conll05.crf.srl.bert.ini   -d 0 -p exp/conll05.crf.srl.bert/model --prd

To make predictions:

python -u crf.py   predict  -c configs/conll05.crf.srl.bert.ini   -d 0 -p exp/conll05.crf.srl.bert/model
bash scripts/eval.sh pred=pred.conllu gold=data/conll05/test.conllu

Contact

If you have any questions, feel free to contact me via emails.