[Paper Review] Rat big, cat eaten! Ideas for a useful deep-agent protolanguage

Marco Baroni|arXiv (Cornell University)|Mar 17, 2020

Topic Modeling28 references4 citations

TL;DR

This paper proposes a minimal, functional protolanguage for deep reinforcement learning agents to enable effective, human-interpretable communication in dynamic, real-world environments. By focusing on simple predication structures—such as 'rat big' or 'apple eaten'—the framework supports context-sensitive, task-oriented communication without requiring full linguistic complexity, enabling rapid deployment in human-machine and machine-machine interaction scenarios.

ABSTRACT

Deep-agent communities developing their own language-like communication protocol are a hot (or at least warm) topic in AI. Such agents could be very useful in machine-machine and human-machine interaction scenarios long before they have evolved a protocol as complex as human language. Here, I propose a small set of priorities we should focus on, if we want to get as fast as possible to a stage where deep agents speak a useful protolanguage.

Motivation & Objective

• To design a minimal, functional protolanguage that enables deep agents to communicate effectively in dynamic, real-world environments before full human-like language emerges.
• To prioritize communication structures that support practical, goal-driven interactions in open-ended, continuously changing environments with novel objects and properties.
• To ensure emergent language remains interpretable by humans by anchoring it in human-like linguistic forms and avoiding overly abstract or non-transparent symbol systems.
• To guide the design of training environments that naturally encourage the emergence of useful linguistic primitives, such as predication, without explicit supervision.
• To support scalable, human-compatible machine communication by focusing on context-sensitive, fuzzy, and ambiguous language use rather than rigid, logical formalisms.

Proposed method

• Design a scenario involving embodied or virtual agents (e.g., household bots) that must collaboratively manage household goods, including tracking stock, detecting spoilage, and locating objects.
• Use continuous, noisy perceptual inputs (e.g., images, sensor data) to reflect real-world variability and ambiguity, avoiding clean, categorical inputs.
• Structure communication around basic predication: agent A signals 'object property' (e.g., 'rat big', 'apple eaten') to convey task-relevant information.
• Encourage emergence of protolanguage through task-specific, multi-agent reinforcement learning in environments where communication improves task performance.
• Incorporate environmental dynamics such as random object appearance, state changes (e.g., rotting), and attribute variation (e.g., color, size) to simulate open-ended scenarios.
• Apply constraints like limited memory or channel capacity, or use imitation-based regularization (e.g., mimicking natural language patterns) to prevent language drift from human-recognizable forms.

Experimental results

Research questions

• RQ1 What minimal linguistic structures are sufficient for deep agents to achieve effective, task-oriented communication in open-ended, real-world environments?
• RQ2 How can we design training environments that naturally encourage the emergence of useful, interpretable communication protocols without explicit supervision?
• RQ3 To what extent can simple predication structures ('object property') support complex coordination in dynamic, real-world scenarios?
• RQ4 How can emergent agent communication remain human-interpretable despite the absence of human-designed linguistic rules?
• RQ5 What environmental and task constraints best promote the emergence of context-sensitive, fuzzy, and ambiguous language use akin to natural protolanguages?

Key findings

• Basic predication structures—such as 'rat big' or 'apple eaten'—suffice for agents to convey critical, task-relevant information in dynamic, real-world scenarios.
• Emergent communication in interactive, goal-driven settings naturally favors simple, compositional forms that mirror early stages of human protolanguage.
• Robust, human-interpretable communication can emerge without explicit linguistic supervision, provided the environment and task structure incentivize functional, context-sensitive signaling.
• Environmental design—such as incorporating long-tailed object distributions, attribute variation, and state changes—effectively drives the emergence of useful linguistic primitives.
• Constraints like limited memory or channel capacity, or imitation of natural language patterns, help prevent language drift and maintain interpretability.
• The protolanguage does not need to be fully logical or unambiguous; fuzzy, context-dependent, and ambiguous communication is more effective in open-ended, real-world settings than rigid formal codes.