Dialogue-based Continuous Update of User Portraits

Abstract

With the development of Internet-based services, user portraits, which are used to describe user interests and preferences, have been widely used as precision marketing references for service providers. Meanwhile, Virtual Personal Assistants (VPA) are extensively used in satisfying daily requirements of users by natural language based interactions. Dialogues generated during the interaction between a user and a VPA are conversations that contain rich information on implicit and explicit user demands and decisions on whether recommended services are accepted or not, in other words, these dialogues contain rich user preferences. In this paper, we propose a novel Multi-domain Dialogue-based User Portrait (MDUP) model and the corresponding method for constructing and updating MDUP of a user by historical dialogue mining. Based on the conventional Key-Value structure, we add four key features into MDUP, including: (1) The mentioned frequency of each preference attribute is recorded; (2) Update history of user preferences keeps complete in MDUP and could be used to explore user preference evolution; (3) Quantitative constraints on each preference attribute are imported to accurately express fine-grained user preferences; and (4) External knowledge graphs are utilized to enrich the semantic relationships between user preference attributes belonging to different domains. A text mining based algorithm for mining user preferences from historical dialogues and generating/updating MDUP is elaborately introduced. Experiments are conducted to validate the effectiveness of the proposed MDUP model and the corresponding algorithm. Downstream application scenarios of the MDUP model are briefly discussed to illustrate the usability of the model, including MDUP-based service recommendation, incorporating MDUP into the dialogue strategy of task-based VPA, and perceiving underlying evolution patterns of user preferences.