This paper presents a series of reference game experiments(Frank and Goodman, 2012) and fits the results to a numberof Bayesian computational models in order to explore the roleof linguistic and perceptual bias in iterative pragmatic reason-ing. We first discuss the modeling choices made by Franke andJ ̈ager (2016) and others who have used similar frameworks tomodel reference game tasks. We introduce a space of differentplausible Bayesian models based on this work, and comparemodels’ fit to new experimental data to replicate the basic find-ings of Franke and J ̈ager (2016) regarding the strong role forperceptual salience (e.g., the primacy of color over shape asa differentiating property for possible referents) and linguis-tic category (e.g., a preference for nouns over adjectives) inpragmatic reference resolution. We then uncover an additionalpossible effect of what we call labeling, whereby a hearer maysimply ignore non-salient, non-differentiating semantic prop-erties, in a manner similar to how an incremental algorithm(Reiter and Dale, 1992) might ignore certain semantic proper-ties when generating referring expressions.

We present an experimental study investigating the role ofinformation-theoretic factors in determining patterns of redun-dancy and focus in language and other communication sys-tems. Pairs of participants played a simple communicationgame using a non-linguistic visual medium to send messagesto each other. We manipulated noise, effort, and time pres-sures and measured message length, redundancy, and accuracy.Participants behaved as predicted based on an information-theoretic model, with message length and redundancy varyingaccording to circumstance, but accuracy remaining constant.

Speakers can be taken to be committed to utterance con-tent even when that content is contributed in the scope ofan entailment-canceling operator, like negation (e.g., Chier-chia & McConnell-Ginet, 1990). We develop a probabilisticmodel of this phenomenon, called ‘projection’, that relies onthe prosodic realization of utterances. We synthesize exist-ing theoretical claims about prosody, information structure andprojection into a model that assumes a rational speaker (Frank& Goodman, 2012) who produces utterances with prosodicmelodies that can signal which utterance content she is com-mitted to. Predictions of the probabilistic model are comparedto the responses of an experiment designed to test the effect ofprosody on projection in manner adverb utterances. Key be-haviors of the model are borne out empirically, and the quan-titative fit is surprisingly good given that the model has onlyone free parameter. Our findings lend support to analyses ofprojection that are sensitive to the information structure of ut-terances (e.g., Simons, Beaver, Roberts, & Tonhauser, 2017).