This is really cool. The next logical thing is to extend it to allow multiple meanings for the same word with the same type. For example "lead is a heavy element" vs "I own a dog lead". Not sure how you'd do that without explicitly giving each unique word an ID and manually annotating the training data.