Hierarchical history based information selection for document grounded dialogue generation

Meiqi Wang, Shiyu Tian, Ziwei Bai, Caixia Yuan, Xiaojie Wang

Abstract: Selecting appropriate information from the dialogue history and the document is a prerequisite for a high-quality response in document grounded dialogue generation task. Most of the existing works take dialogue history information as a sequence to interact with documents. In fact, dialogue history has an internal hierarchical structure, which can provide constraints on the selection of information and the interaction with document information. Therefore, this paper proposes a model that uses the hierarchical structure of dialogue history for key information selection. The main idea is to locate important information in history and document by merging both word-level and utterance-level attention of history, and then to generate a better response. The experimental results on two public data sets show that our method significantly outperforms the baseline models.