Some cunning NLP research which has found that really simple statistical models (e.g. does the sentence contain "not"?) is sufficient to answer a good number of text comprehension datasets: https://thegradient.pub/nlps-clever-hans-moment-has-arrived/