Representing the word2vec Models

As I did few month ago, I was wondering how could I make another representation of word2vec Spaces. Instead of looking near a defined term, as I did, I wanted to show all the word2vec system.

I wanted to make following representation of the model trained using french rap. Thus I made visuals following a PCA reduction (from 100 of dimensions two 4 major one), so two are used to pose the word in the canvas, two are used to show variation in the font settings (via variable font — into its weights and its diacritics size).

It shows a kind of black hole in a a center corner of the page (in around one third of the image).

Showing all the term in the models was two big (around 7k) so kind of reduced the representation to a single artist Heuss l’enfoiré, reducing the amount of words.

It still has a kind of “black hole” in it.

It continued to zoom in less and less data, and wanted to show vector representation of Khapta a famous song of the artist. Somehow I still got issues about lots of overlapping so I decided to randomize the position of each element and “encode” the dimensionnality into more graphic way : scaling and rotation of element with the variation of font.