A relationship between the structure of emergent languages and compositional generalization ability in representation learning.

Abstract

The world is realized by combining various factors such as entities and concepts, and vast amount of states can be realized by multiplying the number and types of factors. Therefore, even if an agents can experience each factor almost exhaustively, it is virtually impossible for them to experience all of the nearly infinite combinations, thus generalization to unknown combinations is important. Such generalization to unknown combinations is called compositional generalization. There are many approaches to compositional generalization, and among them, there is a study that considers the relationship with language emergence from the viewpoint of compositionality. Language emergence is often realized through communication between multiple agents. In deep learning, the encoder and decoder of an autoencoder are considered to be agents respectively, and a language-like structure is assumed for the embeddings exchanged between them. In this study, we examine language emergence using deep generative models, and investigate the relationship between the classes of emergent languages based on the Chomsky Hierarchy and their compositional generalization abilities.