End-to-End Task-oriented Dialogue System Using Knowledge Filter and Attention Memory Pointer

Abstract

The end-to-end neural model provides a more robust solution to generate responses than the traditional pipe-line method in the task-oriented dialogue system. However, it is challenging to incorporate the proper knowledge into the gen-erated response, especially when there are substantially related knowledge tuples. This paper proposes a knowledge filter and an attention memory pointer to improve the task-oriented dia-logue model. Specifically, the model uses the knowledge filter to obtain the knowledge tuples most relevant to the keywords of dialog context and builds the knowledge vector. Besides, the task-oriented dialogue model usually needs to copy objects from the correct knowledge tuples to form the question's an-swer. We define an attention memory pointer to help the model choose the correct knowledge tuples. Finally, we conduct ex-periments on the In-Car Assistant dataset. The experimental results show that our model can generate more accurate re-sponses than baseline models in automatic and human evaluations.