Source code for COLING 2024 paper titled "Attack Named Entity Recognition by Entity Boundary Interference".
Demonstration of Entity Boundary Interference.
- python >= 3.7.9
- torch >= 1.7.1
- transformers >= 4.25.1
- scikit-learn >= 1.0
- tqdm >= 4.49.0
- numpy >= 1.20.2
- The code for finetuning BERT-style model for NER is in
firun_tok_clas.py. run_tok_clas_at.pyintroduces the FreeLB AT training strategy.run_tok_clas_bound.pyandrun_tok_clas_cls.pycan train the victim models in Table4 and Table5 respectively.
We provide the conll-03, msra, ontonotes-5-ch, ontonotes-5-en, wnut-17 datasets in CoNLL format such as:
-- O
basically O
, O
it O
was O
unanimously O
agreed O
upon O
by O
the O
various O
relevant O
parties O
. O
You can finetuning a BERT-style model by the script like:
python run_tok_clas.py \
--cache_dir ../cache/ \
--data_dir YOUR_DATA_PATH \
--do_eval \
--do_train \
--eval_batch_size 64 \
--fp16 \
--learning_rate 5e-5 \
--model_type bert-base-uncased \
--num_train_epochs 6 \
--output_dir YOUR_OUTPUT_DIR \
--seed 42 \
--task_name ontonotes-5-en \
--train_batch_size 32 \
--weight_decay 0.8You can hack a well trained victim model by running scripy like:
python hack.py \
--hack_file YOUR_DATA_FILE \
--task_name ontonotes-5 \
--load_model_path bert-base-chinese \
--do_lower_case \
--model_file YOUR_VICTIM_MODEL_FILE \
--cache_dir /home/yangyf/NER-Attack/cache \
--output_dir YOUR_OUTPUT_DIRIf you find our work helpful, please consider citing:
@article{yang2023attack,
title={Attack named entity recognition by entity boundary interference},
author={Yang, Yifei and Wu, Hongqiu and Zhao, Hai},
journal={arXiv preprint arXiv:2305.05253},
year={2023}
}