Leveraging the insights and tools of cognitive science, this project aims to bridge the gap between useful but unrealistic models of CG and current models of human psychology, to arrive at a cognitively plausible model of how humans represent, and reason about, CG. Theories of CG which make implausible assumptions—e.g., mental representations of infinitely-nested beliefs, reasoning over infinite possibilities, or ungrounded coordination among multiple agents —are powerful, but it is unclear how they could be implemented by humans. We instead explore minimal, resource-rational models which take constraints on human cognition (e.g., memory, unawareness) as a starting point and then implement the insights on conversational dynamics provided by CG therein. Our aim is to develop a cognitively plausible model of a single agent who uses resource-rational heuristics and Bayesian pragmatics to perform boundedly-rational joint reasoning about other agents and the state of the discourse, i.e., what is CG. We also plan to develop models of non-default reasoning, as agents confronted with conflicting information must revise their internal models, in ways consistent with research on metacognitive control.
In support of these aims, our project includes empirical investigations into how humans track and use information which is CG. Using an attribution task, we explore which mental or dispositional states an observer attributes to a speaker; this allows us to gauge which (levels of) beliefs or commitments actually correlate with CG status (and thus informs us about the core components of CG), and how those might differ across information encodings (e.g., whether content was provided explicitly vs. implicitly, or as at-issue vs. not-at-issue). Using a recruitment task that investigates the interpretation of indirect answers to polar questions, we explore which information in CG hearers make use of when resolving utterance uncertainty; this allows insight into whether all information in CG is treated equally (along the lines of a true ‘bag of worlds’ model) or whether hearers distinguish among (and so, store representations of) different sources or statuses of information. Both tasks help develop a reliable diagnostic for CG status.
