We have recently seen many successful applications of sentence embedding methods. It has not been well understood, however, what kind of properties are captured in the resulting sentence embeddings, depending on the supervision signals. In this paper, we focus on two types of sentence