Compositional Languages Emerge in a Neural Iterated Learning Model

Abstract

The principle of compositionality, which enables natural language to represent complex concepts via a structured combination of simpler ones, allows us to convey an open-ended set of messages using a limited vocabulary. If compositionality is indeed a natural property of language, we may expect it to appear in communication protocols that are created by neural agents in language games. In this paper, we propose an effective neural iterated learning (NIL) algorithm that, when applied to interacting neural agents, facilitates the emergence of a more structured type of language. Indeed, these languages provide learning speed advantages to neural agents during training, which can be incrementally amplified via NIL. We provide a probabilistic model of NIL and an explanation of why the advantage of compositional language exist. Our experiments confirm our analysis, and also demonstrate that the emerged languages largely improve the generalizing power of the neural agent communication.

Learning to Draw: Emergent Communication through Sketching

“…Studies on language origins [28,32] consider cooperation to be a key prerequisite to language evolution as it implies multiple agents having to self-organise and adapt to the same convention. Studies on the emergence of communication in cooperative multi-agent environments from recent years have focused on (natural) language learning [21,22] and its inherent properties such as compositionality and expressivity [12,13,29].…”

“…Spurred by innovations in artificial neural networks, deep and reinforcement learning techniques, recent work in multi-agent emergent communication [3,12,14,22,29] pursues interactions in the form of gameplay between agents to induce human-like communication. Artificial communicating agents can collaborate to solve various tasks: image referential games with realistic visual input [14,21,22], negotiation [1], navigation of virtual environments [6,15] or reconstruction of missing input [3,12].…”

Compositionality Through Language Transmission, using Artificial Neural Networks

“…(Griffiths and Kalish, 2007) proved that for Bayesian agents, that the iterated learning method converges to a distribution over languages that is determined entirely by the prior, which is somewhat aligned with the result in (Locatello et al, 2019) for disentangled representations. (Li and Bowling, 2019), (Cogswell et al, 2020), and (Ren et al, 2020) extend ILM to artificial neural networks, using symbolic inputs. Symbolic input vectors are by na-ture themselves compositional, typically, the concatenation of one-hot vectors of attribute values, or of per-attribute embeddings (e.g.…”