Learning high quality sentence embeddings from dialogues has drawn increasing
attentions as it is essential to solve a variety of dialogue-oriented tasks
with low annotation cost. However, directly annotating and gathering utterance
relationships in conversations are difficult, while t