I was curious on how the chatbot actually works, so I found an arXiv paper about it: https://arxiv.org/abs/1812.08989 (The Design and Implementation of XiaoIce, an Empathetic Social Chatbot)
Thank you for this, I was wondering that as well. I thought it was a bit interesting that one of the metrics was "expected Conversation-turns Per Session". I would think most of the success in the chat bot comes from a proper metric for the goodness of a chat. It seems like this was the first paper where that measure was used, does anyone know more about this?