newbie question: when training networks, what mechanism makes the language's concepts be (almost)orthogonal to each other?