Learning to Express in Knowledge-Grounded Conversation

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2258 - 2273 July 10-15, 2022 ©2022 Association for Computational Linguistics Learning to Express in Knowledge-Grounded Conversation Xueliang Zhao1,2, Tingchen Fu3, Chongyang Tao4, Wei Wu5, Dongyan Zhao1,2∗, Rui Yan3∗ 1Wangxuan Institute of Computer Technology, Peking University 2Center for Data Science, AAIS, Peking University 3Gaoling School of Artificial Intelligence, Renmin University of China 4Microsoft 5Meituan, Beijing, China {xl.zhao,zhaody}@pku.edu.cn ruiyan@ruc.edu.cn {lucas.futingchen,chongyangtao,wuwei19850318}@gmail.com Abstract Grounding dialogue generation by extra knowl- edge has shown great potentials towards build- ing a system capable of replying with knowl- edgeable and engaging responses. Existing studies focus on how to synthesize a response with proper knowledge, yet neglect that the same knowledge could be expressed differently by speakers even under the same context. In this work, we mainly consider two aspects of knowledge expression, namely the structure of the response and style of the content in each part. We therefore introduce two sequential latent variables to represent the structure and the content style respectively. We propose a segmentation-based generation model and op- timize the model by a variational approach to discover the underlying pattern of knowledge expression in a response. Evaluation results on two benchmarks indicate that our model can learn the structure style defined by a few exam- ples and generate responses in desired content style. 1 Introduction Building an open domain dialogue system has at- tracted increasing attention from the community of AI and NLP. Despite the impressive progress, existing models are notorious for replying with generic and bland responses. To bridge the gap, researchers resort to ground dialogue generation by extra knowledge such as unstructured documents (Zhou et al., 2018c; Dinan et al., 2019). By this means, the documents serve as content sources and make a dialogue system knowledgeable regarding various concepts in a discussion. However, existing studies focus on how to syn- thesize a response with proper knowledge (Dinan et al., 2019; Kim et al., 2020; Zhao et al., 2020b), but pay little attention to the fact that the same knowledge could be expressed differently even un- der the same context. These models usually em- *Corresponding authors: Dongyan Zhao and Rui Yan. Knowledge • MovieName: Frozen • Year: 2013 • Rating: Rotten Tomatoes: 89% , Metacritics: 74/100, CinemaScore: A+ • Cast: Kristen Bell as Anna, the 18-year-old Princess of Arendelle and Elsa's younger sister, Livvy Stubenrauch as 5-year-old Anna, Katie Lopez as 5-year-old Anna (singing), Agatha Lee Monn as 9-year-old Anna … • … Conversations User1: I was really surprised that disney chose Kristen Bell to be the voice of Anna in Frozen I was really surprised that disney chose Kristen Bell to be the voice of Anna in Frozen User2: Yes, I didn't imagine it'd be her! Yes, I didn't imagine it'd be her! User2: What do you think about the rating? What do you think about the rating? User1: 74 in Metacritics. I believe it deserves, indeed. The rating is 74 in Metacritics. …

Content truncated to 500 words