A Document Driven Dialogue Generation Model

Abstract

Most of the current man-machine dialogues are at the two end-points of a spectrum of dialogues, i.e. goal-driven dialogues and non goal-driven chit-chats. Document-driven dialogues provide a bridge between them with the change of documents from structured data to unstructured free texts. This paper proposes a Document Driven Dialogue Generation model (D3G) which generates dialogues centering a given document, as well as answering user’s questions. A Doc-Reader mechanism is designed to locate the content related to user’s questions in documents. A Multi-Copy mechanism is employed to generate document-related responses. And the dialogue history is used in both mechanisms. Experimental results on the CMU_DOG dataset show that our D3G model can not only generate informative responses that are more relevant to the document, but also answer user’s questions better than the baseline models.