A Review of the Applications of Deep Learning-Based Emergent Communication

TL;DR

This paper surveys how deep learning-based emergent communication can be applied across machine learning, NLP, linguistics, and cognitive science. It categorizes applications into internal goals, task-driven, and knowledge-driven, and provides descriptions, unique roles of emergent communication, current literature, and near-term research directions for each. A core emphasis is on developing evaluation metrics, theoretical models, and tooling to advance the field, as well as identifying gaps in rederiving human language and in knowledge-driven insights. The work aims to guide practitioners and researchers toward standardized metrics, benchmarks, and realistic, scalable avenues for applying emergent communication to real-world tasks and scientific questions, with implications for synthetic data generation, multi-agent coordination, and cognitive science insights.

Abstract

Emergent communication, or emergent language, is the field of research which studies how human language-like communication systems emerge de novo in deep multi-agent reinforcement learning environments. The possibilities of replicating the emergence of a complex behavior like language have strong intuitive appeal, yet it is necessary to complement this with clear notions of how such research can be applicable to other fields of science, technology, and engineering. This paper comprehensively reviews the applications of emergent communication research across machine learning, natural language processing, linguistics, and cognitive science. Each application is illustrated with a description of its scope, an explication of emergent communication's unique role in addressing it, a summary of the extant literature working towards the application, and brief recommendations for near-term research directions.

Figures (9)

• Figure 1: Structure of the applications discussed in this review.
• Figure 2: An illustration of the discrimination variant of the signaling game, one of the simplest and most common environments in emergent communication research.
• Figure 3: Illustration of two spaces with different topographic similarities (toposims). $\mathcal{O}$ and $\mathcal{M}$ represent embedding spaces for the observations and messages, respectively. A high toposim means that distances between any two points in the observation space correlates well with distances between corresponding points in the message space.
• Figure 4: Plots of lexicon entropy ($y$-axis) versus time steps and lexicon size ($x$-axes) comparing a theoretical model against empirical measurements on navigation game with emergent communication (from boldt2022modeling). Theoretical models can help predict the outcomes of emergent communication games much more efficiently than simply running the environment while also providing a conceptual understanding the environment's behavior.
• Figure 5: Self-driving vehicles in complex traffic situations are an important application of multi-agent communication. Diversity in both scenarios as well as the observations themselves indicate that open-ended communication systems could be more appropriate than handcrafted protocols where all scenarios are anticipated ahead of time. Screenshots from documentation of MetaDrive li2023metadrive (Apache-2.0 license).