Pragmatic Reasoning.pdf

The Rational Speech Act (RSA) framework is a prominent approach to pragmatics, formulating pragmatic reasoning as probabilistic speakers and listeners recursively reasoning about each other. RSA recursion implements an alternating maximization algorithm, optimizing a tradeoff between expected utility and communicative effort, rather than maximizing expected utility as previously conjectured. This tradeoff can be generalized to a Rate-Distortion (RD) tradeoff, yielding a modified model of pragmatic reasoning, RD-RSA. RD-RSA is similar to RSA in accounting for human behavior, while avoiding a bias toward non-informative random utterance production. The RSA framework is based on a reference game involving a speaker and a listener, with a lexicon determining literal meanings of utterances. The speaker and listener are related recursively, with the speaker rationalizing about the listener's inferences and the listener being Bayesian with respect to the speaker's distribution. The dynamics of RSA recursion are illustrated, and the RSA tradeoff is shown to improve with each speaker and listener update. The RD-RSA model suggests that human pragmatic reasoning may be understood in terms of RD theory, providing a mathematical understanding of RSA models and their ability to account for human behavior. Key findings include: (1) RSA recursion implements an alternating maximization algorithm, (2) the RSA tradeoff can be generalized to an RD tradeoff, and (3) RD-RSA is similar to RSA in accounting for human behavior while avoiding biases. Overall, the study provides a novel analysis of the RSA framework, grounding it in RD theory and furthering the mathematical understanding of RSA models.