Visualizing embeddings is difficult because of their high dimensionality. They are usually projected with PCA or t-SNE into 2 or 3 dimensions, but this approach has shortcomings: it does not support fine-grained analysis and does not reveal what makes two embeddings similar or different.
Parallax is a tool that uses explicit algebraic formulae as the basis for projection, enabling fine-grained analysis across multiple dimensions. It also supports comparisons across different datasets. It is built with Bokeh and provides a simple interface for such analyses.
We released the code on GitHub and we presented a paper describing the system at ACL 2019.