Proceedings of the Annual Meeting of the Cognitive Science Society

The Rational Speech Act (RSA) model, which proposesthat probabilistic speakers and listeners recursively reasonabout each other’s mental states to communicate, hasbeen successful in explaining many pragmatic reasoningphenomena. However, several theoretical questions remainunanswered. First, will such a pragmatic speaker–listenerpair always outperform their literal counterparts who donot reason about each others mental states? Second, howdoes communication effectiveness change with the number ofrecursions? Third, when exact inference cannot be performed,how does limiting the computational resources of the speakerand listener affect these results? We systematically analyzedthe RSA model and found that in Monte Carlo simulationspragmatic listeners and speakers always outperform theirliteral counterparts and the expected accuracy increases asthe number of recursions increases. Furthermore, limitingthe computation resources of the speaker and listener so theysample only the top k most likely options leads to higherexpected accuracy. We verified these results on a previouslycollected natural language dataset in color reference games.The current work supplements the existing RSA literature andcould guide future modeling work.