Hierarchical Explanations for Text Classification Models: Fast and Effective

The main implementation for the paper Hierarchical Explanations for Text Classification Models: Fast and Effective

Requirement:

• pytoch == 1.9.1
• python == 3.7.11
• numpy == 1.21.2
• matplotlib == 3.4.3
• nltk == 3.6.5

Model and data

The accuracy of the pre-trained model for each dataset is shown in Table2 of our paper.

Download the pre-trained models on each dataset Model and Data

After Download, you can put these files to the folder Cla_datasets and TrainedModel for the dataset and pre-trained models.

Generate explanation

1. Run the following command to generate explanations.

   python HETSG_main.py --task_name sst-2 --start_pos 0 --end_pos -1

   The results of the explanation for each text will be saved in the folder Explain_results.

2. Run the following command to visualize the explanation on the given examples.

   python HETSG_main.py --task_name sst-2 --visualize 1

   The figures of the hierarchical explanation will be saved in the main folder.

Acknowledgement

Thanks for the following two repositories for their help in the implementation of our method.

https://github.com/zdgithub/Interpretable_Interaction_Trees#requirements

https://github.com/UVa-NLP/HEDGE#model-and-data