[Paper Review] Emergence of Compositional Language with Deep Generational Transmission

Michael Cogswell, Jiasen Lu+3 more|arXiv (Cornell University)|Apr 19, 2019

Language and cultural evolution47 references36 citations

TL;DR

The paper shows that periodic replacement (generational transmission) of neural dialog agents induces cultural transmission and yields more compositional generalization in emergent language.

ABSTRACT

Recent work has studied the emergence of language among deep reinforcement learning agents that must collaborate to solve a task. Of particular interest are the factors that cause language to be compositional -- i.e., express meaning by combining words which themselves have meaning. Evolutionary linguists have found that in addition to structural priors like those already studied in deep learning, the dynamics of transmitting language from generation to generation contribute significantly to the emergence of compositionality. In this paper, we introduce these cultural evolutionary dynamics into language emergence by periodically replacing agents in a population to create a knowledge gap, implicitly inducing cultural transmission of language. We show that this implicit cultural transmission encourages the resulting languages to exhibit better compositional generalization.

Motivation & Objective

• Motivate and model cultural transmission as periodic agent replacement to induce language evolution.
• Show that generational dynamics encourage expressive and compressible language for compositionality.
• Evaluate whether replacement leads to better generalization to compositionally novel attribute combinations.

Proposed method

• Extend a cooperative Task & Talk dialog game with a population of Q-bots and A-bots.
• Introduce periodic re-initialization (agent replacement) to simulate cultural transmission.
• Train with policy gradients in a multi-agent setting and compare replacement strategies (uniform random, epsilon-greedy, oldest).
• Use overcomplete and memory settings to test compositional generalization on novel pair and novel instance datasets.
• Measure language similarity across agents to evidence cultural transmission via KL-divergence between token distributions.
• Visualize emergent languages to illustrate compositional structure in agent communication.

Experimental results

Research questions

• RQ1 Does periodic agent replacement induce cultural transmission in neural dialog agents?
• RQ2 Does implicit cultural transmission improve compositional generalization compared to baselines without transmission?
• RQ3 How do different replacement strategies influence language compositionality and learning efficiency?
• RQ4 Is the emergent language compositional and interpretable from a qualitative visualization?
• RQ5 How does memory and vocabulary size interact with generational transmission to affect outcomes?

Key findings

• Cultural transmission via periodic replacement improves compositional generalization, with test accuracy rising to up to 46% in the best configuration on a harder dataset.
• Multi-agent populations with replacement show more similar languages across agents than non-replacement baselines, indicating transmission.
• Replacement strategies consistently outperform no-replacement baselines under various vocabulary and memory settings.
• Memoryless or small-vocabulary configurations display strongest compositional gains, consistent with prior work on language emergence.
• Language visualizations demonstrate that emergent tokens map to disentangled shape and color attributes, evidencing compositional structure.