IsoVec: Controlling the Relative Isomorphism of Word Embedding Spaces

This is an implementation of the experiments and combination system presented in:

Kelly Marchisio, Neha Verma, Kevin Duh, and Philipp Koehn. 2022. IsoVec: Controlling the Relative Isomorphism of Word Embedding Spaces. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 6019–6033, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.

If you use this software for academic research, please cite the paper above.

Requirements

Download third party packages: cd third_party && sh get_third_party.sh && cd ..
- Note: If you're on Mac with an M1 chip, word2vec might not build. You can fix this by changing -march=native to -mcpu=apple-m1 in word2vec's makefile, and subbing in getc_unlocked and putc_unlocked for fgetc_unlocked/fputc_unlocked. You'll also need to use gshuf instead of shuf within src/train.py.
Download and make data: cd data && sh make_data.sh
Download and make train/dev/test dictionaries: cd data/dicts && sh create_dicts.sh

To reproduce Table 1 in the paper (Baselines), run:

sh baseline.sh $system $lang $seed
- For instance, run sh baseline.sh w2v uk for offical word2vec trained on Ukrainian.
- system choices: {isovec, w2v}
- lang choices: {uk, bn, ta, en}
After you train English and Ukrainian baseline w2v spaces, for instance, you can map them and evaluate the dictionary precision with: sh map-and-eval.sh baseline w2v uk en dev
- Results will be in exps/baseline/w2v/uk-en/*out

To run IsoVec in reference to a fixed embedding space (main experiments):

Example Goal: Train a Ukrainian embedding space with RSIM-U, in reference to a fixed English space.
Step 1: Train the fixed English space with sh baseline.sh isovec en
Step 2: Train the Ukrainian space with: sh run-isovec.sh rsim-u uk en
- Choices of Isovec training algorithm are l2, proc-l2, proc-l2-init, rsim, rsim-init, rsim-u, evs-u for L2, Proc-L2, Proc-L2+Init, RSIM, RSIM-U, and EVS-U as detailed in Section 4.3 and 4.4 of the paper.
Step 3: Map & Evaluate the spaces with: sh map-and-eval.sh isovec rsim-u uk en dev