Kind of like a distance/similarity metric for words, but not an embedding. Joint probability. I once used it to automatically group words into multiword tokens. The wiki page is informative. But wouldn't know what they were using.