Attention is not not Explanation

Code for the EMNLP 2019 paper Attention is not not Explanation by Wiegreffe & Pinter.

When using this codebase, please cite:

@inproceedings{wiegreffe-pinter-2019-attention,
    title = "Attention is not not Explanation",
    author = "Wiegreffe, Sarah  and
      Pinter, Yuval",
    booktitle = "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)",
    month = nov,
    year = "2019",
    address = "Hong Kong, China",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/D19-1002",
    doi = "10.18653/v1/D19-1002",
    pages = "11--20"
}

We've based our repository on the code provided by Sarthak Jain & Byron Wallace for their paper Attention is not Explanation.

Dependencies

Please refer to the installation instructions for the repository provided by Jain & Wallace. We use the same dependencies. Also, make sure to export the meta-directory into which you clone attention to your PYTHONPATH in order for the imports to work correctly. For example, if the path to the cloned directory is /home/users/attention/, then run export PYTHONPATH='/home/users'.

Data Preprocessing

Please perform the preprocessing instructions provided by Jain & Wallace here. We replicated these instructions for the Diabetes, Anemia, SST, IMDb, AgNews, and 20News datasets.

Running Baselines

We replicate the reported baselines in Jain & Wallace's paper (as reported in our paper in Table 2) by running the following commands:

• ./run_baselines.sh [Diabetes, Anemia, AgNews, 20News_sports, imdb, sst]

Freezing the Attention Distribution (Section 3.1)

• ./run_frozen_attention.sh [Diabetes, Anemia, AgNews, 20News_sports, imdb, sst]

Running Random Seeds Experiments (Section 3.2)

• parallel ./run_seeds.sh :::: seeds.txt ::: sst AgNews imdb 20News_sports Diabetes Anemia
• Code for constructing the violin plots in Figure 3 can be found in seed_graphs.py and Seed_graphs.ipynb.

Running BOWs Experiments (Section 3.3)

• To run the Bag of Words model with trained (MLP) attention weights: ./run_bows_baselines.sh [Diabetes, Anemia, AgNews, 20News_sports, imdb, sst]
• To run the Bag of Words model with uniform attention weights: ./run_bows_frozen_attn.sh [Diabetes, Anemia, AgNews, 20News_sports, imdb, sst]
• To run the Bag of Words model with frozen attention weights from another model: ./run_bows_set_to_pretrained_distribution.sh [Diabetes, Anemia, AgNews, 20News_sports, imdb, sst] [path/to/saved/model/with/attention/weights]

Running Adversarial Model Experiments (Section 4)

• ./run_adversarial.sh [Diabetes, Anemia, AgNews, 20News_sports, imdb, sst] [lambda_value] [path/to/saved/model/with/gold/attentions/and/predictions]