Introduction

This repository compares different methods of obtaining interpretable dimension in word embedding spaces.

More specifically it compares:

• Densifier
• DensRay: A method closely related to Densifier, but computable in closed form.
• Support Vector Machines / Regression
• Linear / Logistic Regression.

The evaluation tasks are lexicon induction and set-based word analogy.

For more details see the Paper.

Note that this repo does not include an implementation of the Densifier, but relies on the original Matlab implementation by the authors of Densifier.

Usage

For an example how to use the code see example.sh.

References

If you use the code, please cite

@article{dufter2019analytical,
  title={Analytical Methods for Interpretable Ultradense Word Embeddings},
  author={Dufter, Philipp and Sch{\"u}tze, Hinrich},
  journal={arXiv preprint arXiv:1904.08654},
  year={2019}
}