Emergence of Recursive Language through Bootstrapping and Iterated Learning

Recursive structures are fundamental aspects of many human languages, allowing the embedding of concepts within other concepts. These structures are thought to be key factors in the expressiveness and flexibility of human communication. Such structures evolve through continuous and iterated learning, transmitted across generations via the bottleneck of language transmission. In this paper, we study language acquisition and the emergence of groundedness and recursive linguistic structures through neural iterated learning, where expressing a goal requires multiple levels of communication. We model this process as a language game within the framework of a decentralized, multi-agent deep reinforcement learning setting, where agents with local learning and neural cognitive faculties interact through a series of dialogues. Our examinations reveal the emergence of a shared depth-1 recursive language, where agents are able to acquire and generalize their bootstrapped language for expressing complex concepts.