Leibniz MMS Days 2024 - Abstract

Sauerland, Uli

Linguistic Meaning and Bayesian Modeling

(joint work with Anton Benz (ZAS Berlin), Michael Franke (University of Tübingen), Ari Joshi (ZAS Berlin / University of Potsdam), and Hening Wang (University of Tübingen))

Recent linguistic research has developed models integrating game theoretic models of communication with Bayesian statistical modelling that provide successful accounts of human language production in several simple domains. But while extant models have been successfully applied to data from controlled laboratory experiments, the computational complexity of such models has so far impeded wide-spread application to larger, possibly open-ended domains such as naturally occurring conversation. We present ongoing research within the LM-Bayes project on three core obstacles for scaling Bayesian models of language use: multiple sources of uncertainty, inference over large structured sets of discrete categories, and individual-level differences.