Skip to content

Latest commit

 

History

History
88 lines (49 loc) · 2.99 KB

README.md

File metadata and controls

88 lines (49 loc) · 2.99 KB
Raw

It’s not Greek to mBERT:
Inducing Word-Level Translations from Multilingual BERT

This project includes the experiments described in the paper:

"It’s not Greek to mBERT: Inducing Word-Level Translations from Multilingual BERT"

Hila Gonen, Shauli Ravfogel, Yanai Elazar, Yoav Goldberg, BlackboxNLP, 2020.

Prerequisites

  • Python 3
  • Required INLP code is already in this repository

Data

Please download relevant data files from this folder

Rowspace (language specific) and Nullspace (language neutral) matrices

You can either create these matrices using the script inlp-oop/create_inlp_matrix_lang.py.

Example:

python create_inlp_matrix_lang.py --extract_data --train_P

Alternatively, you can download them from this folder:

  • rowspace_embeddings: Rowspace (language specific) for embeddings
  • nullspace_embeddings: Nullspace (language neutral) for embeddings
  • rowspace_context: Rowspace (language specific) for representations in context
  • nullspace_context: Nullspace (language neutral) for representations in context

Details:

  • Languages used (15 most common in TED): zh-tw, zh-cn, pt-br, en, ar, he, ru, ko, it, ja, es, fr, nl, ro, tr

  • 20 iterations of INLP

  • 5000 examples per language

  • Embeddings matrices: embedding representations of a random token from each sentence, excluding token that start with "##"

  • In-context matrices: representation in context of a random token from each sentence, may include "CLS" or "SEP" tokens

Analogies method

Use the script source/analogies.py.

Template-based method

Use the script source/translator.py. This will create a file with the matching representations for the Euralex data: ../data/representations_embed_translator_template.npy that will be used in source/compare_translator_analogies.py for evaluation.

Misc.

For comparison between the methods, use source/compare_translator_analogies.py.

For language prediction, use the notebook source/translator_langs.ipynb.

Cite

If you find this project useful, please cite the paper:

@inproceedings{gonen_mbert20,
    title = "It{'}s not {G}reek to m{BERT}: Inducing Word-Level Translations from Multilingual {BERT}",
    author = "Gonen, Hila and Ravfogel, Shauli and Elazar, Yanai and Goldberg, Yoav",
    booktitle = "Proceedings of the Third BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP",
    year = "2020",
}

Contact

If you have any questions or suggestions, please contact Hila Gonen.

License

This project is licensed under Apache License - see the LICENSE file for details.