RoKGDS: A Robust Knowledge Grounded Dialog System

Jun Zhang, Yuxiang Sun, Yushi Zhang, Weijie Xu, Jiahao Ying, Yan Yang, Man Lan, Meirong Ma, Hao Yuan, Jianguo Zhu

Abstract: In this paper, we propose a pre-training based Robust Know-ledge Grounded Dialog System (RoKGDS) to enhance the performance of the model in unknown scenarios, which is easily generalized to various knowledge grounded dialog tasks, such as persona dialog, knowledge dialog, recommendation dialog. We use a bucket encoder to efficiently extract all kinds of knowledge information (e.g. profile, knowledge graph, and dialog goal). To improve the robustness of the model, we develop a hybrid decoder with a hybrid attention and a copy mechanism. The hybrid attention is an adaptation scheme to apply the pre-trained language model to our model and the copy mechanism is a gate mechanism to control generating a word from generic vocabulary or the input knowledge. Experiments show that our model is more robust than the other baseline models. Furthermore, we use visualization to explain the effectiveness of the hybrid attention compared to other two adaptation schemes. In the 2021 Language and Intelligence Challenge: Multi-Skill Dialog task, our best model ranked 3rd in the automatic evaluation stage and 5th in the human evaluation stage.