We present a novel framework for building pragmatic artificialagents with explicit and trainable semantic representations, us-ing the Rational Speech Act model. We train our agents onsupervised and unsupervised communication games and com-pare their behavior to literal agents lacking pragmatic abilities.For both types of games pragmatic but not literal agents evolvea mutual exclusivity bias. This provides a computational prag-matic account of mutual exclusivity and points out a possi-ble direction for solving the mutual exclusivity bias challengeposed by Gandhi and Lake (2019). We find that pragmaticreasoning can cause the bias either by promoting lexical con-straints during learning, or by affecting online inference. In ad-dition we show that pragmatic abilities lead to faster learningand that this advantage is even stronger when meanings to becommunicated follow a more natural distribution as describedby Zipf’s law.

The study of emergent languages in deep multi-agent simulations has become an important research field. While targeting different objectives, most studies focus on analyzing properties of the emergent language—often in relation to the agents’ inputs—ignoring the influence of the agents’ perceptual processes. In this work, we use communication games to investigate how differences in perception affect emergent language. Using a conventional setup, we train two deep reinforcement learning agents, a sender and a receiver, on a reference game. However, we systematically manipulate the agents’ perception by enforcing similar representations for objects with specific shared features. We find that perceptual biases of both sender and receiver influence which object features the agents’ messages are grounded in. When uniformly enforcing the similarity of all features that are relevant for the reference game, agents perform better and the emergent protocol better captures conceptual input properties.