Learning Personalized End-to-End Task-Oriented Dialogue for Fast and Reliable Adaptation

Personalized task-oriented dialog agents can select responses according to the personalities of users. In this way, they facilitate understanding, and improve the efficiency of the conversation. Existing personalized agents generate matching functions according to human designed persona descriptions. This agent works well in the traditional scenario where sufficient samples are used, but can hardly fast adapt to new personas with few samples. In this paper, we propose to extend meta-learning algorithms to personalized end-to-end task-oriented dialogue learning, and train a switch model to allow for human agent use. Our model learns to quickly adapt to new personas leveraging a few dialogue samples and requesting human agents, while maximizing the task success of users. Empirical results on the Personalized bAbI dataset indicate that our proposal is effective in achieving the desired goals.