There Are a Thousand Hamlets in a Thousand People’s Eyes: Enhancing Knowledge-grounded Dialogue with Personal Memory

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 3901 - 3913 May 22-27, 2022 c⃝2022 Association for Computational Linguistics There Are a Thousand Hamlets in a Thousand People’s Eyes: Enhancing Knowledge-grounded Dialogue with Personal Memory Tingchen Fu1†, Xueliang Zhao2†, Chongyang Tao3, Ji-Rong Wen1, Rui Yan1∗ 1Gaoling School of Artificial Intelligence, Renmin University of China 2Wangxuan Institute of Computer Technology, Peking University 3Microsoft Corporation {lucas.futingchen,zhaoxlpku,chongyangtao}@gmail.com {jrwen,ruiyan}@ruc.edu.cn Abstract Knowledge-grounded conversation (KGC) shows great potential in building an engaging and knowledgeable chatbot, and knowledge se- lection is a key ingredient in it. However, pre- vious methods for knowledge selection only concentrate on the relevance between knowl- edge and dialogue context, ignoring the fact that age, hobby, education and life experience of an interlocutor have a major effect on his or her personal preference over external knowl- edge. Without taking the personalization issue into account, it is difficult to select the proper knowledge and generate persona-consistent re- sponses. In this work, we introduce personal memory into knowledge selection in KGC to address the personalization issue. We propose a variational method to model the underlying relationship between one’s personal memory and his or her selection of knowledge, and de- vise a learning scheme in which the forward mapping from personal memory to knowledge and its inverse mapping is included in a closed loop so that they could teach each other. Experi- ment results show that our method outperforms existing KGC methods significantly on both automatic evaluation and human evaluation. 1 Introduction Open-domain dialogue system often suffers from safe response (Li et al., 2015; Zhang et al., 2019) problem as they could only refer to the context when generating a response. To alleviate this, knowledge-grounded conversation (KGC) is pro- posed to introduce external fact and real-world commonsense as prior knowledge (Zhou et al., 2018a; Dinan et al., 2019; Zhao et al., 2020a), such that a dialogue system is able to ground the conver- sation with the provided knowledge and therefore †The first two authors contribute equally. Xueliang Zhao is responsible for the design of the methodology and algo- rithm. Tingchen Fu is responsible for the implementation and experiment. The order is decided by a coin flip. *Corresponding author: Rui Yan (ruiyan@ruc.edu.cn) generate informative and engaging responses. As external knowledge supplements the background to the inputs and decides what to say, knowledge selection is a key ingredient in KGC. Numerous methods have been developed to tackle the knowledge selection problem by sequen- tial latent variables (Kim et al., 2020; Meng et al., 2020), reinforcement learning (Zhao et al., 2020b), or expectation maximization algorithm (Li et al., 2020). In spite of the progress in this task, knowl- edge selection remains an unsolved problem as the precision is still far from satisfactory in Wiz- ard of Wikipedia (Dinan et al., 2019) and other benchmarks in KGC (Gopalakrishnan et al., 2019), which also hinders the optimization of subsequent response generation models. A crucial point is, they often make assumption that the …

Content truncated to 500 words