Workshop "Experimental Pragmatics meets Game Theory"

Recent work on experimental pragmatics by Noah Goodman and co-workers (Frank & Goodman 2012, Science; Goodman & Stuhlmueller 2012, CogSci; Bergen, Godman & Levy, 2012; CogSci) has shown a remarkable convergence with the Iterated Best Response model of game theoretic pragmatics (Degen & Franke 2012, SemDial; Degen, Franke & Jäger 2012, XPrag; Franke 2009, PhD thesis Amsterdam; Franke 2011, Semantics & Pragmatics; Franke & Jäger 2012, JoLLI; Jäger & Ebert 2009, Sinn+Bedeutung, Jäger 2012 Handbook Semantics, Jäger to appear, Erkenntnis). The workshop serves to identify points where these two research traditions agree and where they differ, and to explore whether a joint agenda for future research can be developed.

• Judith Degen (University of Rochester, Department of Brain and Cognitive Sciences)
• Michael Franke (University of Amsterdam, Institute of Logic, Language and Computation)
• Noah Goodman (Stanford University, Department of Psychology)

14:00 - 15:00: Gerhard Jäger, The Iterated Best Response Model of game theoretic pragmatics and its closest relatives
15:15 - 16:15: Judith Degen, Reasoning about Referential Expressions
16:45 - 17:45: Michael Franke, The Use of Quantifiers: Set Size & Typicality
18:15 - 19:15: Noah Goodman, Uncertainty in Pragmatics and Semantics

The Iterated Best Response (IBR) model is a game-theoretic approach to formal pragmatics that spells out pragmatic reasoning as back- and-forth reasoning about interlocutors’ rational choices and beliefs (Franke, 2011; Jäger, 2011). Two studies investigated the production and comprehension of referential expression within this framework. Manipulating the complexity of inferences involved in production and comprehension of referential expressions yielded an intriguing asymmetry: comprehension performance is better than production in corresponding complex inference tasks, but worse on simpler ones. This is not predicted by standard formulations of IBR, which makes categorical predictions about rational choices. However, a variant of IBR that uses a stochastic instead of a categorical choice function and assumes that players assume a distribution over lower-level strategic types (Iterated Quantal Response, IQR) is found to provide a good fit to the data. The relation of these results to Frank & Goodman (2011)'s model of pragmatic inference is discussed.

Two recent empirical studies have investigated intuitive typicality/naturalness judgements of the scalar quantifier /some/ in connection with referent sets of different cardinality (Degen & Tanenhaus 2011/submitted, van Tiel 2012). In distinction to previous interpretations of these typicality data, I conjecture that the mean typicality rating of a quantifier given a set reflects the degree of likelihood that a speaker would use that quantifier to describe that set. Support for this conjecture comes from a quantitative model that seems capable of predicting the data almost perfectly. Assuming that numerals within the subitizing range, as well as quantifiers /most/, /many/ and /all/ are available but differently costly alternatives, we obtain a tight numerical fit from a noise-affected model of a truth-conforming speaker.

Viewing language as a form of rational action suggests applying tools developed for Bayesian models of social cognition. I follow this idea to build models of language understanding in simple reference games and scalar implicature. I show that these models capture the quantitative details of human judgements in several behavioral experiments. I will then explore the different ways to treat semantic free variables in a Bayesian setting. By fixing these variables at the level of pragmatic listener, we can derive the principle of division of pragmatic labor and explain some puzzling effects in the semantics of gradeable adjectives.

The talk will give an overview of the Iterated Best Response model of game theoretic pragmatics. After presenting the model and going through basic examples, I will discuss scenarios where the interests of the interlocutors are not aligned, or if they are boundedly rational. Special attention will be devoted to the "Iterated Quantal Response" (IQR) variant of the model. This variant is inspired by recent work in behavioral game theory and makes probabilistic predictions. Therefore it lends itself readily for modeling experimental data.