CaSiNo: A Corpus of Campsite Negotiation Dialogues for Automatic Negotiation Systems

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3167–3185 June 6–11, 2021. ©2021 Association for Computational Linguistics 3167 CaSiNo: A Corpus of Campsite Negotiation Dialogues for Automatic Negotiation Systems Kushal Chawla1 Jaysa Ramirez2∗ Rene Clever3∗ Gale Lucas1 Jonathan May4 Jonathan Gratch1 1&4University of Southern California, Los Angeles, USA 2Rollins College, Winter Park, USA 3CUNY Lehman College, Bronx, USA 1{chawla,lucas,gratch}@ict.usc.edu 2jramirez@rollins.edu 3rene.clever@lc.cuny.edu 4jonmay@isi.edu Abstract Automated systems that negotiate with hu- mans have broad applications in pedagogy and conversational AI. To advance the devel- opment of practical negotiation systems, we present CaSiNo: a novel corpus of over a thou- sand negotiation dialogues in English. Partici- pants take the role of campsite neighbors and negotiate for food, water, and firewood pack- ages for their upcoming trip. Our design re- sults in diverse and linguistically rich negoti- ations while maintaining a tractable, closed- domain environment. Inspired by the liter- ature in human-human negotiations, we an- notate persuasion strategies and perform cor- relation analysis to understand how the dia- logue behaviors are associated with the ne- gotiation performance. We further propose and evaluate a multi-task framework to rec- ognize these strategies in a given utterance. We find that multi-task learning substantially improves the performance for all strategy la- bels, especially for the ones that are the most skewed. We release the dataset, anno- tations, and the code to propel future work in human-machine negotiations: https:// github.com/kushalchawla/CaSiNo. 1 Introduction Negotiations are highly prevalent in our interac- tions, from deciding who performs the household chores to high-stake business deals to maintaining international peace. Automatic negotiation systems are helpful in providing cost-effective social skills training (Johnson et al., 2019) and for advanced capabilities of AI assistants such as Google Du- plex (Leviathan and Matias, 2018). A negotiation requires understanding the part- ner’s motives along with effective reasoning and communication, which is challenging for an auto- mated system. Prior work in human-machine nego- tiations primarily uses strict communication proto- cols such as a pre-defined menu of options (Mell ∗Work done when authors were interns at USC ICT and Gratch, 2016). Systems involving free-form di- alogue are limited due to a lack of interdisciplinary efforts in NLP and Computational Social Science in this direction. Initial efforts in building dia- logue systems for negotiations looked at game en- vironments (Asher et al., 2016; Lewis et al., 2017). DealOrNoDeal (Lewis et al., 2017) involves two negotiators who split given quantities of three arbi- trary items: books, balls, and hats. This provides a concrete structure to the negotiation, keeps the design tractable, and ensures a reliable evaluation based on final points scored. Many practical solu- tions in negotiations follow similar closed-domain designs (Mell and Gratch, 2016). However, most of the dialogues in these game settings reduce to merely an exchange of offers from both sides. For instance, ‘i need the book and the balls you can have the hat’ or ‘i want the ball and 2 books’ in DealOrNoDeal. One reason …

Content truncated to 500 words