Reinforcement of Semantic Representations in Pragmatic Agents Leads to the Emergence of a Mutual Exclusivity Bias

Abstract

We present a novel framework for building pragmatic artificial agents with explicit and trainable semantic representations, using the Rational Speech Act model. We train our agents on supervised and unsupervised communication games and compare their behavior to literal agents lacking pragmatic abilities. For both types of games pragmatic but not literal agents evolve a mutual exclusivity bias. This provides a computational pragmatic account of mutual exclusivity and points out a possible direction for solving the mutual exclusivity bias challenge posed by Gandhi and Lake (2019). We find that pragmatic reasoning can cause the bias either by promoting lexical constraints during learning, or by affecting online inference. In addition we show that pragmatic abilities lead to faster learning and that this advantage is even stronger when meanings to be communicated follow a more natural distribution as described by Zipf's law. © 2020 The Author(s)