Learning Personalized End-to-End Task-Oriented Dialog Agent in Low Resource Scenarios

Abstract:
Learning personalized dialog systems is an interesting topic, as the personalized agents can better perceive the dialog contents and thus improve the communication efficiency. As for the retrieval-based dialog systems, existing personalized variant selects correct responses according to the persona labels collected manually. However, when few samples are obtained, the personalized agents can maintain the original performance. To fill this gap, we propose to apply MAML and its variants to train task-oriented dialogue model, in order to transfer to arbitrary persona by learning few samples. We implement the proposed models on the personalized bAbI dataset, and the experimental results validate the effectiveness of our solution.