Introduction

For installation and docs please refer to release 0.6 of pytorch_pretrained_bert.

The current fork adds the jupyter notebook on the attention analysis of the 12 layer BERT model. For details please refer to the paper.

Note that for the extraction of attention weights the source code of was modified (this functionality was added in later realeases of the forked repo).

Requirements and usage

1. Install the requirements as

pip install -r requirements.txt

2. The current implementation assumes you have GLUE datasets downloaded and fine-tuned BERT model weights saved to a directory of your choice. You can download GLUE data as described here. To fine-tune BERT, run .

3. The code for analysis is contained in the jupyter notebook.

4. To repeat the results of experiments, make sure to change the path_to_model and path_to_data in the notebook.

References

@article{kovaleva2019revealing,
  title={Revealing the Dark Secrets of BERT},
  author={Kovaleva, Olga and Romanov, Alexey and Rogers, Anna and Rumshisky, Anna},
  journal={arXiv preprint arXiv:1908.08593},
  year={2019}
}