Measuring Idiomaticity in Text Embedding Models with 𝜀-compositionality

The data is included as a git submodule, from there the python module epsilon_compositionality runs all the necessary code and outputs a results directory containing the LaTeX code we included in the article. To reproduce everything, run the following:

git clone --recurse-submodules THIS_REPOSITORY
python -m epsilon_compositionality

For a more decomposed (pun intended) approach, this is equivalent to:

python -m epsilon_compositionality.build_dataset
python -m epsilon_compositionality.extract_similarities
python -m epsilon_compositionality.compute_statistics