I’ll say I’m currently neutral on the Q of picking the extra specs by hand vs. using some sort of AI algorithm that almost never makes a mistake in its conservative mode relative to assessing KR context. Developing such an algorithm is a tougher research project than doing the hand selection well. So payoff would be more long term.

Heuristically, one expects that homonym (not synonym) pairs of words imply different categories/contexts of interpretation – some cut planes in the middle somewhere. Should be a clue to what is ez why for a human intelligence.