RSA model of referential expression production and comprehension under noise

Ivan Rygaev, Martin V. Butz, Asya Achimova

Keywords: RSA, Bayesian model, prior expectations, referential expressions

TL;DR: We present an RSA model of the production and comprehension of referential expressions which is able to cope with an apparent contradiction to RSA in the experimental data

Abstract: This study investigates how speakers and listeners manage referential ambiguity in event descriptions, focusing on situations where prior expectations about event plausibility influence language production and comprehension. We examine how speakers choose between zero anaphors, pronouns, and explicit noun phrases when describing outcomes that either align with or contradict listeners’ expectations. Prior work suggests that speakers avoid ambiguous expressions when outcomes are surprising, anticipating listeners' reliance on plausibility-based priors. Through a series of perception, production, and priors-assessment experiments, we confirm this tendency. To account for these patterns, we present a Rational Speech Act (RSA) model incorporating listener priors and potential misinterpretation under noise. The model successfully predicts the experimentally observed rates of referential expression production and comprehension, including a surprising fact in our data that listeners frequently interpret pronouns as referring to the patient of the previous event, even though speakers very rarely use pronouns to refer to the patient. This is explained by listeners’ prior expectations and the observation that speakers also rather rarely use pronouns to refer to the agent. This work advances our understanding of pragmatic reasoning under uncertainty.

Submission Number: 23