Skip to content

UVa-NLP/HEDGE

master
Switch branches/tags
Code

Latest commit

 

Git stats

Files

Permalink
Failed to load latest commit information.
Type
Name
Latest commit message
Commit time
 
 
cnn
 
 
 
 
 
 

HEDGE

Code for the paper "Generating Hierarchical Explanations on Text Classification via Feature Interaction Detection"

Requirement:

  • torchtext == 0.4.0
  • gensim == 3.4.0
  • pytorch == 1.2.0
  • numpy == 1.16.4

Model and data:

Download well-trained models and data.

Generate explanations:

We provide the example code of HEDGE interpreting the LSTM, CNN and BERT model on the IMDB dataset. We adopt the BERT-base model built by huggingface: https://github.com/huggingface/transformers.

In each folder, run the following command to generate explanations on the test data for a well-trained model.

python hedge_main_model_imdb.py --save /path/to/your/model

We save the start-end word indexes of text spans in a hierarchy (in the order of importance) into the "hedge_interpretation_index.txt" file.

To visualize the hierarchical explanation of a sentence, run

python hedge_main_model_imdb.py --save /path/to/your/model --visualize 1(the index of the sentence)

Reference:

If you find this repository helpful, please cite our paper:

@inproceedings{chen2020generating,
  title={Generating hierarchical explanations on text classification via feature interaction detection},
  author={Chen, Hanjie and Zheng, Guangtao and Ji, Yangfeng},
  booktitle={ACL},
  url={https://arxiv.org/abs/2004.02015},
  year={2020}
}

About

Code for the paper "Generating Hierarchical Explanations on Text Classification via Feature Interaction Detection"

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages