PeaCoK: Persona Commonsense Knowledge for Consistent and Engaging Narratives

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 6569–6591 July 9-14, 2023 ©2023 Association for Computational Linguistics PEACOK: Persona Commonsense Knowledge for Consistent and Engaging Narratives Silin Gao1, Beatriz Borges1∗, Soyoung Oh1∗, Deniz Bayazit1∗, Saya Kanno2, Hiromi Wakaki2, Yuki Mitsufuji2, Antoine Bosselut1† 1NLP Lab, IC, EPFL, Switzerland, 2Sony Group Corporation, Tokyo, Japan 1{silin.gao,beatriz.borges,soyoung.oh,deniz.bayazit}@epfl.ch 2{saya.kanno,hiromi.wakaki,yuhki.mitsufuji}@sony.com 1antoine.bosselut@epfl.ch Abstract Sustaining coherent and engaging narratives re- quires dialogue or storytelling agents to under- stand how the personas of speakers or listeners ground the narrative. Specifically, these agents must infer personas of their listeners to produce statements that cater to their interests. They must also learn to maintain consistent speaker personas for themselves throughout the narra- tive, so that their counterparts feel involved in a realistic conversation or story. However, personas are diverse and complex: they entail large quantities of rich intercon- nected world knowledge that is challenging to robustly represent in general narrative sys- tems (e.g., a singer is good at singing, and may have attended conservatoire). In this work, we construct a new large-scale persona common- sense knowledge graph, PEACOK, containing ∼100K human-validated persona facts. Our knowledge graph schematizes five dimensions of persona knowledge identified in previous studies of human interactive behaviours, and distils facts in this schema from both exist- ing commonsense knowledge graphs and large- scale pretrained language models. Our analysis indicates that PEACOK contains rich and pre- cise world persona inferences that help down- stream systems generate more consistent and engaging narratives.1 1 Introduction Interlocutors or storytellers in narrative scenarios often exhibit varying behaviours, which are af- fected by their own diverse personas, but also the personas of the counterparts they are interacting with. For example, an adventurous architect may be interested in talking about outdoor explorations with his friends who have similar hobbies, but may prefer to discuss architectural design ideas with his *Equal contribution. †Corresponding author. 1We release our data and code to the community at https: //github.com/Silin159/PeaCoK I am a singer who want to release an album study music at college good at singing write songs win a Grammy award have more fans Characteristic Experience Routine or Habit Goal or Plan I am a student who major in music composition I am a famous pianist Relationship skilled in playing the piano talk with music mentor Routine or Habit Routine or Habit Relationship Experience Characteristic Figure 1: Illustration of world persona knowledge grounded on commonsense reasoning. colleagues at work. Narrative systems must know when such behaviours should be exhibited, requir- ing them to learn and represent the rich personas of characters based on self-introductions, biographies and other background profiles. This goal of modeling diverse persona attributes is at the heart of research in the areas of persona- grounded dialogue (Zhang et al., 2018; Zhong et al., 2020; Xu et al., 2022), story generation (Chandu et al., 2019; Zhang et al., 2022) and narrative under- standing (Brahman et al., 2021). However, the com- plex nature of real-world personas, which involve …

Content truncated to 500 words