Continual Learning in Chit-Chat Systems

Chit-chat dialogue systems, also known as open-domain dialogue systems, focus on carrying out chit-chat type of conversations with users on any topic without specific goals to complete (see Sect.  1.1 for more details). The need to support such free-flow conversations often makes it challenging to build dialogue systems that can perform well in practice. Existing methods for building chit-chat systems mainly rely on collecting a dialogue corpus with context and response pairs and training a response generation model. However, no matter how big a corpus is collected for the training, it will always be limited in terms of the knowledge that the chatbot needs to successfully model relevant responses in post-deployment real-world conversations with users. The ability to continually learn from post-deployment conversation experiences and self-improve the response generation performance thus become essential for the success of commercial chit-chat systems. This chapter focuses on the topic.