mt_idioms

Code & data for "Can Transformer be Too Compositional? Analysing Idiom Processing in Neural Machine Translation", published at ACL 2022.

@inproceedings{dankers2022can,
  title={Can Transformer be Too Compositional? Analysing Idiom Processing in Neural Machine Translation},
  author={Dankers, Verna and Lucas, Christopher and Titov, Ivan},
  booktitle={Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  pages={3608--3626},
  year={2022}
}

Install

1 - Install libraries using requirements.txt.

2 - Download a Spacy model for English using python -m spacy download en_core_web_sm.

3 - Custom install of the transformers library: transformers==4.3.0, with modeling_marian.py replaced with the adapted version from this repository. To clone transformers and replace MarianMT with our adapted version, carry out the following instructions:

git clone https://github.com/huggingface/transformers.git
cd transformers
git checkout 800f385d7808262946987ce91c158186649ec954
cd ..
mv modeling_marian.py transformers/src/transformers/models/marian/modeling_marian.py
cd transformers
pip install -e .

Other non-standard libraries are provided in the requirements.txt file.

Experiments

For results from the paper's sections, visit the following folders, that have their own README.

• Section 3: See the data folder.
• Section 4: See the attention folder.
• Section 5: See the hidden_representations folder.
• Section 6: See the amnesic_probing folder.
• Appendix B: See the data folder.
• Appendix C: See the attention folder.
• Appendix D: See the attention and hidden_representations folder.
• Appendix E: See the hidden_representations folder.
• Appendix F: See the amnesic_probing folder.
• Appendix G: See the data folder.