Interesting point. Do you think this is more about training data limitations or evaluation methods?