Skip to content

Latest commit

 

History

History
26 lines (25 loc) · 997 Bytes

README.md

File metadata and controls

26 lines (25 loc) · 997 Bytes
Raw

Lexical-AT

This is the codes of our paper: LexicalAT: Lexical-Based Adversarial Reinforcement Trainingfor Robust Sentiment Classification. The architecture of our Lexical-AT

Requirements

  • Ubuntu 16.0.4
  • Python 3.6
  • Tensorflow 1.12
  • NLTK 3.4.3

Training Examples

For CNN or LSTM

cd cnn_lstm
CUDA_VISIBLE_DEVIVES=0 bash sst_adv_cnn.sh

For Bert

Before executing the command below, you need to download the pretrained model and vocabulary of bert. The download url and other detailed information can be found in bert.

cd bert 
CUDA_VISIBLE_DEVICES=0 bash rt_adv_bert.sh

Note

  • Before running the code, you need to download wordnet via nltk.
  • The code is currently non-deterministic due to various GPU ops, so you are likely to end up with a slightly better or worse evaluation.

Citation

If you use the above codes for your research, please cite our paper.