Train a shallow neural model, and project each document onto this vector embedding space.