UVa-NLP / HEDGE

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

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}
}