Proceedings of the Annual Meeting of the Cognitive Science Society

One way people deal with uncertainty is by asking questions.A showcase of this ability is the classic 20 questions gamewhere a player asks questions in search of a secret object. Pre-vious studies using variants of this task have found that peopleare effective question-askers according to normative Bayesianmetrics such as expected information gain. However, so far,the studies amenable to mathematical modeling have used onlysmall sets of possible hypotheses that were provided explic-itly to participants, far from the unbounded hypothesis spacespeople often grapple with. Here, we study how people eval-uate the quality of questions in an unrestricted 20 Questionstask. We present a Bayesian model that utilizes a large data setof object-question pairs and expected information gain to se-lect questions. This model provides good predictions regardingpeople’s preferences and outperforms simpler alternatives.