You Impress Me: Dialogue Generation via Mutual Persona Perception

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1417–1427 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 1417 You Impress Me: Dialogue Generation via Mutual Persona Perception Qian Liu†∗, Yihong Chen♦∗, Bei Chen§, Jian-Guang Lou§, Zixuan Chen♠∗, Bin Zhou†, Dongmei Zhang§ †School of Computer Science and Engineering, Beihang University, China ♦UCL Centre for Artificial Intelligence, University College London, United Kindom ♠School of Computer Science, Fudan University, China §Microsoft Research, Beijing, China †{qian.liu, zhoubin}@buaa.edu.cn; §{beichen, jlou, dongmeiz}@microsoft.com; ♦yihong.chen@cs.ucl.ac.uk; ♠remch183@outlook.com Abstract Despite the continuing efforts to improve the engagingness and consistency of chit-chat di- alogue systems, the majority of current work simply focus on mimicking human-like re- sponses, leaving understudied the aspects of modeling understanding between interlocutors. The research in cognitive science, instead, suggests that understanding is an essential signal for a high-quality chit-chat conversa- tion. Motivated by this, we propose P2 BOT, a transmitter-receiver based framework with the aim of explicitly modeling understanding. Specifically, P2 BOT incorporates mutual per- sona perception to enhance the quality of per- sonalized dialogue generation. Experiments on a large public dataset, PERSONA-CHAT, demonstrate the effectiveness of our approach, with a considerable boost over the state-of-the- art baselines across both automatic metrics and human evaluations. 1 Introduction Thanks to the advance in neural models and the accessibility of massive datasets, open-domain di- alogue (i.e. chit-chat) systems have made great progress towards mimicking human-like responses. Nevertheless, there still exist some serious chal- lenges in building personalized chatbots that can deliver engaging conversations and gain user trust (Song et al., 2019). For example, current chit-chat systems tend to generate uninformative responses (Li et al., 2016b). Moreover, they are usually lack of coherent personality traits due to the fact that training dialogues actually come from a diverse set of speakers (Zhang et al., 2018b). ∗Work done during an internship at Microsoft Research. Hello how are you, I am new to the Springfield area. I bought my first home. I love to barbecue. I live in Springfield. I’m a writer. I weight 300 pounds. I am not healthy. I am a man. I like The Godfather. Persona Persona Hi! Seen any good movies lately? I have been to the movies. I love The Godfather, one of my favorites! Was that filmed? I don’t believe so. I don’t watch movies more of a writer. What do you write? Any diet books ? I am not very healthy. Figure 1: A clippled dialogue from PERSONA-CHAT. Several attempts have been made to alleviate the above issues. Methods like special reward shap- ing to reduce generic responses (Li et al., 2016b) and representing the speakers with latent variables (Li et al., 2016a) were introduced to improve the engagingness of chit-chat systems. A more straight- forward approach, which equips chit-chat systems with predefined personas, was proposed accompa- nied by a novel dataset, PERSONA-CHAT (Zhang et al., 2018b). Figure 1 shows a clipped dialogue from PERSONA-CHAT. Two interlocutors meet for the first time and are having a conversation in order …

Content truncated to 500 words