Understanding the Rational Speech Act model
- Arianna Yuan, Department of Psychology, Stanford University, Stanford, California, United States
- Will Monroe, Department of Computer Science, Stanford University, Stanford, California, United States
- Yu Bai, Department of Statistics, Stanford University, Stanford, California, United States
- Nate Kushman, Microsoft Research, Microsoft Research, Cambridge, United Kingdom
AbstractThe Rational Speech Act (RSA) model, which proposes that probabilistic speakers and listeners recursively reason about each other’s mental states to communicate, has been successful in explaining many pragmatic reasoning phenomena. However, several theoretical questions remain unanswered. First, will such a pragmatic speaker–listener pair always outperform their literal counterparts who do not reason about each others mental states? Second, how does communication effectiveness change with the number of recursions? Third, when exact inference cannot be performed, how does limiting the computational resources of the speaker and listener affect these results? We systematically analyzed the RSA model and found that in Monte Carlo simulations pragmatic listeners and speakers always outperform their literal counterparts and the expected accuracy increases as the number of recursions increases. Furthermore, limiting the computation resources of the speaker and listener so they sample only the top k most likely options leads to higher expected accuracy. We verified these results on a previously collected natural language dataset in color reference games. The current work supplements the existing RSA literature and could guide future modeling work.
Return to previous page