Assign color hues to a collection of text fragments based on embeddings.
Embeddings are high-dimensional, but can be compared to one another using cosine similarity, which measures the angle between them. Interestingly, points on a colorwheel are similarly separated by angle.
This implementation of embedding coloring works by brute-force: cosine similarity of embeddings is compared against candidate colorspace arrangements to optimize for the correlation between the two. This way, proximity in colorspace corresponds to proximity in embedding space as closely as possible.
The limitation of this methodology is that brute-force puts a cap on the number of embeddings that can be optimally arranged. However, clustering ensures that a larger number of text embeddings can still be semantically arranged. And for the sake of interpretability, clusters will often be what the human eye picks up on anyway.
To run coloring on an example set of words, run:
poetry install
poetry run python -m colorspace