Role estimation of the speaker and the target from the player’s estimated statements in a werewolf game

Abstract

In a Werewolf game, which is a game with imperfect information and communication, it is reportedly important to estimate the hidden roles of other players to win. Several proposed estimation methods are based on machine learning, but the timing of estimation used for these methods remains limited. They do not estimate player roles by individual statements that express knowledge or intent. This paper proposes a system that estimates a speaker’s role and a target’s role from each ESTIMATE statement using selective desensitization neural networks. We actually built a role estimation system and examined its performance. Results show that roles such as “MEDIUM”, “POSSESSED”, and “WEREWOLF” can be estimated with high accuracy from ESTIMATE sentences.