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

Abstract

Recent advances in probabilistic pragmatics have achieved considerable success in modeling speakers’ and listeners’ pragmatic reasoning as probabilistic inference. However, these models are usually applied to population-level data, and so implicitly suggest a homogeneous population without individual differences. Here we investigate potential individual differences in Theory-of-Mind related depth of pragmatic reasoning in so-called reference games that require drawing ad hoc Quantity implicatures of varying complexity. We show by Bayesian model comparison that a model that assumes a heterogenous population is a better predictor of our data, especially for comprehension. We discuss the implications for the treatment of individual differences in probabilistic models of language use.

Pragmatic reasoning through semantic inference

“…In general, estimating the pragmatics model that best fits an individual's or population's inferences requires jointly estimating these parameters. Some existing empirical work suggests that the recursion depth is low (about L 1 ) and l is moderate but greater that one (Goodman & Stuhlmüller 2013, Franke & Degen 2015. More work is needed to pin this down completely, especially for the complex implicatures dealt with here.…”

Probabilistic pragmatics, or why Bayes’ rule is probably important for pragmatics

“…For example, Qing and Franke (2015) show by statistical model comparison that variants of the approach sketched here do worse overall, if the comprehension rule does not consider empirically measured salience. Franke and Degen (2015) show by the same method that, if we take individual-level data into account, other production and comprehension rules appear credible as well.…”

“…Reference games are heavily simplified laboratory tasks designed to investigate production and comprehension of referential expressions in a confined, controlled environment. Reference games with different kinds of stimuli, different kinds of experimental measures, and slightly different empirical goals have been studied in the recent literature (Stiller, Goodman, and Frank 2011;Degen and Franke 2012;Frank and Goodman 2012;Degen, Franke, and Jäger 2013;Baumann, Clark, and Kaufmann 2014;Carstensen, Kon, and Regier 2014;Franke and Degen 2015;Qing and Franke 2015).…”

When do we think strategically?

“…This is an empirical question, and the issue of between-subject variation in pragmatics experiments is raised by Franke and Degen (2015), who conclude that some such variation does indeed come into play. But I'd like to abstract away from that, and consider a second possibility: that even within the individual, there are distinct "winner-takes-all" decision-making heuristics that can in principle be "turned on" or "turned off", which need not reflect any inherently probabilistic reasoning.…”