Source codes used in EMNLP 2020 paper, Adversarial Attack and Defense of Structured Prediction Models.
- python 2 & python 3
- anaconda
- pytorch 0.4.1 & pytorch >= 1.0
- transformers
- gensim
- numpy
- bert-score
- pytorch-pretrained-bert
- nltk
-
Clone this repository and two anaconda environments: one is python 2 and the other is python 3
-
For python 2 env, install pytorch 0.4.1, gensim, numpy, nltk.
For python 3 env, install pytorch >= 1.0, transformers, bert-score, pytorch-pretrained-bert and numpy
You also can download the python2 enviroments from here. -
Downloading sskip embedding and conllu format PTB dataset.
# pretrain victim model
$ sh examples/run_graphParser.sh
# pretrain reference parser stackPtr
$ sh examples/run_stackPtrParser.sh
# pretrain reference parser bist
$ cd bist_parser
$ sh test.sh- Move biaffine parser and stackPtr parser to
./models/parsing/{biaffine, stack_ptr}. Move pretrained bist parser tobist_parser/pretrained/model1. - Pretrain seq2seq sentence generator. Or you also can get our trained seq2sq model here.
$ /path/to/python2/envs/python examples/pretrain_seq2seq.py --cuda --mode LSTM \
--num_epochs 30 --batch_size 64 --hidden_size 512 --num_layers 3 --pos_dim 100 --char_dim 100 --num_filters 100 \
--learning_rate 1e-3 --decay_rate 0.05 --schedule 5 --gamma 0.0 \
--p_in 0.33 --p_rnn 0.33 0.5 --p_out 0.5 \
--word_embedding sskip --word_path data/sskip/sskip.eng.100.gz \
--train data/ptb/train.conllu \
--dev data/ptb/dev.conllu \
--test data/ptb/test.conllu \
--char_embedding random \
--model_path models/parsing/biaffine- RL training:
$ sh examples/run_rl_graph_parser.py- You also can download our trained model here. And run eval to get results reported in paper.
$ sh examples/eval_rl_graph_parser.py - Pretrain the victim model. Or you also can get our pretrained version here.
$ sh run_posCRFTagger.sh- Download and unzip reference parser: stanford-postagger, senna.
- Pretrain the seq2seq model. Or you also can download our pretrained version here
/path/to/python2/envs/python examples/pretrain_seq2seq.py --cuda --mode LSTM \
--num_epochs 30 --batch_size 64 --hidden_size 256 --num_layers 1 --char_dim 30 --num_filters 30 \
--learning_rate 1e-3 --decay_rate 0.05 --schedule 5 --gamma 0.0 \
--p_in 0.33 --p_rnn 0.33 0.5 --p_out 0.5 \
--word_embedding sskip --word_path data/sskip/sskip.eng.100.gz \
--train data/ptb/train.conllu \
--dev data/ptb/dev.conllu \
--test data/ptb/test.conllu \
--char_embedding random \
--model_path path/to/victim/model/- RL training.
$ sh examples/run_adv_tagger.sh- You also can download our trained model here. And run eval to get results reported in paper.
$ sh examples/eval_adv_tagger.py