Skip to Content

Unsupervised embedding quality evaluation

paper · pdf


We can approximately know the quality of a given set of embeddings.

In this paper:

  • We identify three different perspectives on evaluation of embedding quality in unsupervised manner and introduce four metrics based on these perspectives.
  • We experimentally study two novel settings for embedding quality evaluation, showing that standard metrics often fail when shallow models are being studied.
  • We conduct a study on computational stability of all metrics and identify the minimum viable sample sizes.