Visualizing embeddings is difficult because of their high dimensionality. usually they are projected using either PCA or t-SNE into 2 or 3 dimensions and then visualized, but this has some shortcomings: it does not allow for fine-grained analysis and does not to figure out what makes two embeddings similar and what makes them different.
Parallax is a tool that exploits eplxicit algebraic formulae as the basis for projection and allows for such fine grained analysis over multiple dimensions. It moreover allows for comparisons across different datasets. It is buildt using Bokeh and provides a simple user interface to perform such analses.
We released the code on GitHub and we presented a paper describing the system at ACL 2019.