The Rational Speech Acts (RSA) model treats language use as a recursive process in which probabilistic speaker and listener agents reason about each other's intentions to enrich, and negotiate, the semantics of their language along broadly Gricean lines. RSA builds on early work by the philosopher David Lewis and others on signaling systems as well as more recent developments in Bayesian cognitive modeling. Over the last five years, RSA has been shown to provide a unified account of numerous core phenomena in pragmatics, including metaphor, hyperbole, sarcasm, politeness, and a wide range of conversational implicatures. Its precise, quantitative nature has also facilitated an outpouring of new experimental work on these phenomena. However, applications of RSA to large-scale problems in NLP and AI have so far been limited, because the exact version of the model is intractable along several dimensions. In this talk, I'll report on recent progress in approximating RSA in ways that retains its core properties while enabling application to large datasets and complex environments in which language and action are brought together.
展开▼