Revealing the Blind Spot of Sentence Encoder Evaluation by HEROS