If I understand your approach correctly, you could represent relations between words as graphs and use graph/network similarity measures (of which there are ton

If I understand your approach correctly, you could represent relations between words as graphs and use graph/network similarity measures (of which there are tons) to possibly get over the 92%. (Or not, I have never tried it.)