When do we think strategically?

Abstract

Michael Franke and Gerhard Jäger (this volume), henceforth F&J, offer a programmatic illustration of the benefits of a Bayesian rational approach to pragmatic analysis. F&J outline three test cases: (i) computational modeling of experimental results involving so-called reference games (Frank and Goodman 2012), (ii) prediction of gradient acceptability judgments of quantifiers, and (iii) a sketch of a theoretical model of indirect speech act production/interpretation in a negotiation scenario. In this comment, I will focus mostly on the first of these test cases, which I believe raises some interesting deeper questions about how to model gradient linguistic behavior, and about the explanatory power of different approaches to cognitive modeling in general. The overall arc of the argument can be summarized as follows. 1. Subjects' responses in F&J's reference game experiment could be interpreted in two ways: First, they could be interpreted as they are by F&J -as the product of some inherently probabilistic approximation of optimal reasoning about speaker intention. Alternatively, they could be interpreted as the interaction of two distinct "winner-takes-all" forms of reasoning. For example, hearer responses in F&J's reference game could be analyzed as the result of an interaction between one mechanism which dumbly zeroes in on a visually salient object and checks whether that object is consistent with the speaker's message, and another (which only kicks in when the first fails) for figuring out which object is strategically optimal under a simple iterated best response (IBR) model (Franke 2009). 2. A simple parameter-free computational model along these lines not only predicts results in the right neighborhood, but also predicts the priors elicited in F&J's prior elicitation experiment with minimal assumptions. Moreover, adding a single free parameter, a probability of defecting from an IBRbased strategy back to a salience-based strategy, results in a near-perfect fit to the observed data. 3. Though both models are only meant to be illustrative, the differences between them highlight the need to consider some important questions:

Reevaluating pragmatic reasoning in language games

“…Much recent work has focused on constructing formal models of pragmatic reasoning (e.g., [11][12][13][14][15][16][17][18][19][20][21][22][23][24]; see [25], for a review) Critically, the predictions of many of these models have been extensively tested against empirical data on human reasoning behavior. A highly influential paper in this regard is the Science article by Frank & Goodman [20] (henceforth, F&G), which introduces the Rational Speech Act (RSA) model of human pragmatic reasoning and evaluates its performance against behavioral data in a referential communication game [26].…”

Reconsidering Miranda rights: Modeling strategic action during the invocation stage of a police interrogation

“…In the analysis of reference, Stevens (2016) discusses, through a Bayesian and cognitive lens, whether strategic and non-strategic thinking play independent roles in explaining pragmatic behavior. Stevens notes that a model that explains interaction must account for why people think strategically and, alternatively, why they may not.…”