The (~ n GB) word embeddings should support fast queries to find similar words. Will try using R-Tree or similar structures.
