Are word embeddings universal?
The latent variable space is normally non-identifiable up to an orthonormal transformation, so no. But, can we estimate that transformation given a small set of paired points?
Does it work for less explicitly paired data eg different language word embeddings? What about sentence embeddings?
- Use better corpus eg europarl
- Enforce orthonormality
- Transform between languages