Trust is implicit in many online text conversations---striking up new friendships, or asking for tech support---but that trust can be betrayed through deception (e.g., catfishing, social engineering attacks, phishing).  To study the language of deception, we focus on the game Diplomacy where seven players who negotiate through chats forge and break alliances. We conduct a human study with players from the Diplomacy community to gather over 10,000 messages annotated by the sender for their ground truth values and by the receiver for their perceived truth values. Unlike most previous deception datasets, this contains deception in the context of long-lasting relationships, where the interlocutors need to strategically combine truth with lies to advance towards an objective. To demonstrate the relevance of this context, we develop detection-prediction models that can exploit information about the power differentials between the interlocutors and their past history to detect deception at par with human players.

@InProceedings{Peskov+al:2020,

  author={Denis Peskov, Benny Cheng, Ahmed Elgohary, Joe Barrow,

          Cristian Danescu-Niculescu-Mizil and Jordan Boyd-Graber},

  title={It Takes Two to Lie: One to Lie, and One to Listen},

  booktitle={Proceedings of ACL},

  year={2020}

}