Continual Learning for Task-Oriented Dialogue Systems

Task-oriented Dialogue Systems (ToDS) are widely popular now-a-days due to their pervasive usage in real-world applications like fight booking, customer service, virtual assistant services, etc. The main goal of these systems is to understand and complete tasks requested by users through multi-turn dialogues. Traditionally, ToDS are built with a number of subsystems, viz., Automatic Speech Recognition (ASR), Natural Language Understanding (NLU), Dialogue State Tracking (DST), Dialogue Policy (DP) Learning and Natural Language Generation (NLG), each playing its distinctive role in realizing the whole dialogue system (see Sect.  1.1 for more details). This chapter discusses several methods to enable continual learning in task-oriented dialogue systems, which include continual learning in individual subsystems and joint continual learning over all subsystems.