Less is More: Learning to Refine Dialogue History for Personalized Dialogue Generation

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5808 - 5820 July 10-15, 2022 ©2022 Association for Computational Linguistics Less is More: Learning to Refine Dialogue History for Personalized Dialogue Generation Hanxun Zhong1, Zhicheng Dou2,∗, Yutao Zhu3, Hongjin Qian2, Ji-Rong Wen2 1School of Information, Renmin University of China, Beijing, China 2Gaoling School of Artificial Intelligence, Renmin University of China, Beijing, China 3University of Montreal, Quebec, Canada {hanxun_zhong,dou,ian}@ruc.edu.cn {yutaozhu94,jirong.wen}@gmail.com Abstract Personalized dialogue systems explore the problem of generating responses that are con- sistent with the user’s personality, which has raised much attention in recent years. Exist- ing personalized dialogue systems have tried to extract user profiles from dialogue history to guide personalized response generation. Since the dialogue history is usually long and noisy, most existing methods truncate the dialogue history to model the user’s personality. Such methods can generate some personalized re- sponses, but a large part of dialogue history is wasted, leading to sub-optimal performance of personalized response generation. In this work, we propose to refine the user dialogue history on a large scale, based on which we can handle more dialogue history and obtain more abun- dant and accurate persona information. Specifi- cally, we design an MSP model which consists of three personal information refiners and a personalized response generator. With these multi-level refiners, we can sparsely extract the most valuable information (tokens) from the di- alogue history and leverage other similar users’ data to enhance personalization. Experimental results on two real-world datasets demonstrate the superiority of our model in generating more informative and personalized responses. 1 Introduction Recent years have witnessed great progress in build- ing personalized dialogue systems. In general, pre- vious work explores building a personalized dia- logue system via two pathways: (1) directly mod- eling user personality from predefined persona de- scriptions or user attribute (Qian et al., 2018; Zhang et al., 2018; Song et al., 2019); and (2) implicitly modeling the user personality from the user’s di- alogue history (Li et al., 2016c; Ma et al., 2021). The latter is considered superior as the dialogue history is easy to obtain and comprises rich per- sonalized information. In this paper, we follow the ∗Corresponding author. second pathway that automatically learns implicit user profiles from the user’s dialogue history to assist in personalized response generation. It is challenging to model user personality di- rectly from the dialogue history. The main reason is that a user’s dialogue history may contain mas- sive historical dialogues, which are too heavy to load in the model and likely to be noisy. A straight- forward solution is to truncate the dialogue history, as done by existing work (Ma et al., 2021; Qian et al., 2021a). However, as tremendous information has been wasted, the model’s performance is also influenced. On the other hand, we observe that the dialogue history from other users may also be help- ful in generating a personalized response for the current user. For example, users with …

Content truncated to 500 words