paperweekly 李纪为 对话机器人直播。

chatbot资料全集

1.Generating interesting and informative responses

• seq2seq loss = -logp(response|message)
• Always generating dull responses -> i don’t know problem
  
  • solution:加规则 无法考虑语义相似
  • mutual information 互信息从相应推测问句。
    
(1-a)logp(response|message)+alogp(message|response)


2.Preserve speaker consistency

• how old are you?
• what’s your age? 理想情况下回复相同
• 加入user embedding 在input中，personal seq2seq

3.long term conversational

• 强化学习 reward: no dull easy to answer
  information flow
  meaningfulness

4.task oriented conversation