It Takes Two to Lie: One to Lie, and One to Listen

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3811–3854 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 3811 It Takes Two to Lie: One to Lie, and One to Listen Denis Peskov, Benny Cheng Ahmed Elgohary, Joe Barrow Computer Science, University of Maryland {dpeskov, bcheng96, elgohary, jdbarrow}@umd.edu Cristian Danescu-Niculescu-Mizil Information Science Cornell University cristian@cs.cornell.edu Jordan Boyd-Graber iSchool, Language Science, UMIACS, LSC University of Maryland jbg@umiacs.umd.edu Abstract Trust is implicit in many online text conversations—striking up new friendships, or asking for tech support. be betrayed through deception.But trust can We study the language and dynamics of deception in the negotiation-based game Diplomacy, where seven players compete for world domination by forging and breaking alliances with each other. Our study with players from the Diplomacy community gathers 17,289 messages annotated by the sender for their intended truthfulness and by the receiver for their perceived truthfulness. Unlike existing datasets, this captures deception in long-lasting relationships, where the interlocu- tors strategically combine truth with lies to advance objectives. A model that uses power dynamics and conversational contexts can predict when a lie occurs nearly as well as human players. 1 Introduction A functioning society is impossible without trust. In online text interactions, users are typically trusting (Shneiderman, 2000), but this trust can be betrayed through false identities on dating sites (Toma and Hancock, 2012), spearphishing attacks (Dhamija et al., 2006), sockpuppetry (Ku- mar et al., 2017) and, more broadly, disinforma- tion campaigns (Kumar and Shah, 2018). Beyond such one-off antisocial acts directed at strangers, deception can also occur in sustained relationships, where it can be strategically combined with truth- fulness to advance a long-term objective (Cornwell and Lundgren, 2001; Kaplar and Gordon, 2004). We introduce a dataset to study the strategic use of deception in long-lasting relationships. To col- lect reliable ground truth in this complex scenario, we design an interface for players to naturally gen- erate and annotate conversational data while play- ing a negotiation-based game called Diplomacy. Message Sender’s intention Receiver’s percep. If I were lying to you, I’d smile and say “that sounds great.” I’m honest with you because I sin- cerely thought of us as partners. Lie Truth You agreed to warn me of un- expected moves, then didn’t ...You’ve revealed things to England without my permission, and then made up a story about it after the fact! Truth Truth ...I have a reputation in this hobby for being sincere. Not be- ing duplicitous. It has always served me well. ...If you don’t want to work with me, then I can understand that . . . Lie Truth (Germany attacks Italy) Well this game just got less fun Truth Truth For you, maybe Truth Truth Table 1: An annotated conversation between Italy (white) and Germany (gray) at a moment when their re- lationship breaks down. Each message is annotated by the sender (and receiver) with its intended or perceived truthfulness; Italy is lying about ...lying. A full tran- script of this dialog is …

Content truncated to 500 words