论文记录：A Survey on Dialogue Systems: Recent Advances and New Frontiers

综述：对话系统（未完成，不想写了，原文介绍太粗略--感兴趣自己看看也还是可以了解各种思路）

对话系统分类：

1. 面向特定任务的系统
2. 面向非特定任务的系统

面向特定任务（Task-oriented dialogue system）的对话系统的总体架构（pipeline）
NLU-> Dialogue state tracking -> Policy learning -> NLG
NLU: map natural language into semantic slots.
Two sorts of slots in NLU:

1. utterance level representation: user’s intent or utterance category. This type slot filling is done by classification of intents or content categories.
2. word level representation: NER etc. This information is attained by sequence labeling models such as CRF, etc.

Dialogue state tracking: slot filling for intent-recognition, or maintain a distribution over different goals.

Policy learning: Reinforcement learning or supervised learning. KB can be used to guide the conversation.

NLG: …

End2End: …