Incremental RSA Model Explains Adjective Ordering Preferences by Communicative Efficiency Across Contexts

Hening Wang, Fabian Schlotterbeck, Michael Franke

Keywords: Adjective Ordering Preferences, Rational Speech Act Model, Bayesian Quantitative Analysis, Simulation of Communicative Contexts

TL;DR: We show that an incremental Rational Speech Act model accounts for how communicative efficiency shapes adjective ordering preferences, both within experimental contexts and in broader simulated communicative scenarios.

Abstract: Adjective ordering preferences reflect constraints on rational communication. Previous accounts link AOP to cognitive factors such as subjectivity or discriminatory strength by proposing an incremental RSA model that implements these hypotheses and qualitatively accounts for results from a slider rating experiment. We extend this model with a novel implementation of Bayesian Data Analysis using RSA models within the state-of-the-art machine learning framework NumPyro. This substantially improves the efficiency of MCMC inference, particularly when handling intermediate marginal distributions over large discrete utterance and state spaces involved in recursive computations. Furthermore, we run simulations with the model in a large set of randomly generated contexts. We show that an incremental RSA model accounts for how communicative efficiency shapes adjective ordering preferences, both within experimental contexts and in broader simulated communicative scenarios.

Submission Number: 57