Enhancing Personalized Dialogue Generation with Contrastive Latent Variables: Combining Sparse and Dense Persona

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 5456–5468 July 9-14, 2023 ©2023 Association for Computational Linguistics Enhancing Personalized Dialogue Generation with Contrastive Latent Variables: Combining Sparse and Dense Persona Yihong Tang1, Bo Wang2,∗, Miao Fang4, Dongming Zhao3, Kun Huang3, Ruifang He2, Yuexian Hou2 1School of New Media and Communication, Tianjin University, Tianjin, China 2College of Intelligence and Computing, Tianjin University, Tianjin, China 3AI Lab, China Mobile Communication Group Tianjin Co., Ltd. 4School of Computer and Communication Engineering, Northeastern University at Qinhuangdao, Qinghuangdao, China {toyhom, bo_wang}@tju.edu.cn Abstract The personalized dialogue explores the consis- tent relationship between dialogue generation and personality. Existing personalized dialogue agents model persona profiles from three re- sources: sparse or dense persona descriptions and dialogue histories. However, sparse struc- tured persona attributes are explicit but unin- formative, dense persona texts contain rich per- sona descriptions with much noise, and dia- logue history query is both noisy and uninfor- mative for persona modeling. In this work, we combine the advantages of the three resources to obtain a richer and more accurate persona. We design a Contrastive Latent Variable-based model (CLV) that clusters the dense persona descriptions into sparse categories, which are combined with the history query to generate personalized responses. Experimental results on Chinese and English datasets demonstrate our model’s superiority in personalization. 1 Introduction In order to develop personalized dialogue agents, current approaches enhance the personality of gen- erated responses mainly utilizing three kinds of resources: (1) Defined sparse persona attributes (Zhang et al., 2018a; Song et al., 2019; Wolf et al., 2019; Liu et al., 2020; Song et al., 2021); (2) Dense persona description texts (Qian et al., 2018; Zheng et al., 2020; Song et al., 2021); (3) Historical queries of current dialogue (Li et al., 2016b; Ma et al., 2021). Each of the three resources has its advantages and disadvantages. Sparse persona attributes (e.g., gender, age) are highly interpretable and have high information uti- lization, but the information is limited and cannot express complex persona features. Dense persona description text contains rich and flexible persona information but suffers from noisy expressions. ∗Corresponding author. Modeling personality directly from dialogue histo- ries is free of additional persona information, but the persona information in history queries is both noisy and uninformative. To address these issues, in this paper, we im- prove personalized dialogue generation by com- bining the advantages of the three resources. We design a contrastive latent variable (CLV)-based model that clusters the dense persona descriptions into sparse categories, which are combined with the history query to generate personalized responses. Specifically, first, the dialog’s latest query and re- sponse together with dense persona description texts are encoded. Then the recognition distribu- tion of query and response is jointly modeled with a pre-designed dual conditional variational autoen- coder (CVAE (Sohn et al., 2015)). Simultaneously, the persona information is automatically separated into multiple parts to participate in the above pro- cess in parallel. These partitioned persona pieces of information are …

Content truncated to 500 words