13 Mar 2015 - Michael Franke: Seminar

Abstract

Recent advances in probabilistic pragmatics have achieved considerable success in modeling speakers' and listeners' pragmatic reasoning as probabilistic inferences. Usually models are applied to aggregate data, implicitly suggesting a homogeneous population without individual differences. Here, I would like to call this implicit homogeneity presupposition into question, by comparing the popular Rational Speech Act model (e.g. Frank and Goodman, Science, 2012; Goodman and Stuhlmüller, TopiCS, 2013; Kao et al., PNAS, 2014) to a heterogeneous, mixed-type variant inspired by cognitive hierarchy models from behavioral economics (e.g., Camerer, "Behavioral Game Theory", 2003). The latter assumes that populations are mixtures of player types of varying sophistication and contains the Rational Speech Act model as a special case. I show by Bayesian comparison of nested models that the homogeneous model is a very poor predictor of subjects' behavior in repeated measures reference games that require drawing ad hoc Quantity implicatures of varying complexity. This raises the interesting question (impossible to answer here) whether other Bayesian models are likewise good predictors of average population-level data, but bad predictors of individual-level behavior.

Bio

Michael Franke is a Junior Research Group Leader at the Department of Linguistics of the University of Tübingen, Germany. He holds a PhD in Philosophy from the University of Amsterdam on applications of game theory within linguistics. His research interests intersect formal models of human interaction, especially game- and decision theory, theoretical linguistics and cognitive science. He has worked on topics concerning logic, philosophy of language, game theory, language evolution, as well as formal and experimental pragmatics.

Mar 13 2015 -

13 Mar 2015 - Michael Franke: Seminar

Reasoning in Referential Games: Individual- vs. Population-Level Probabilistic Modeling

Informatics Forum 4.31/4.33