Emergent Communication in Agents Generating Messages Using Different Pretrained Deep Generative Models

Abstract

Compositionality is one of the most important properties of human language. In the study of emergent communication, methods to induce compositionality in emergent language learned by neural network agents have been actively studied. Humans have individually articulated prior knowledge about the environment, and it is said to be important for compositional human language by communicating based on this prior knowledge. We implement this idea as a Reassembly game, a reconstruction task-based Lewis signaling game in which communicating agents have different pre-trained VAE. Experimental results show that the emergent language learned in the Reassembly game showed high compositionality on several metrics than the emergent language learned in the other reconstruction task-based Lewis signaling game settings. In addition, compared to other factors inducing compositional emergent language such as the length of message and the number of vocabulary, the emergent language learned in the Reassembly game achieve a high level of both communication success rate and compositionality.