💬 NLP

Modeling Human-Like Color Naming Behavior in Context

This paper introduces upsampling of rare color terms and multi-listener reinforcement learning interactions to the NeLLCom-Lex framework, demonstrating that combining these factors with supervised learning produces neural agent color lexicons that are both more informative and geometrically convex, closely resembling human-like color categories.

Original authors: Yuqing Zhang, Ecesu Ürker, Tessa Verhoef, Gemma Boleda, Arianna Bisazza

Published 2026-04-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yuqing Zhang, Ecesu Ürker, Tessa Verhoef, Gemma Boleda, Arianna Bisazza

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot how to talk about colors. You want it to sound like a human, using words like "green," "red," or "blue" in a way that feels natural and logical.

This paper is about an experiment where researchers tried to build a robot that learns color names not just by reading a dictionary, but by playing a game with other robots.

The Game: "Guess the Chip"

Think of the setup like a party game.

The Speaker Robot sees a specific colored chip (the target) and two other similar chips (the distractors).
The Speaker has to say a single word (like "green") to tell the Listener Robot which one is the target.
The Listener has to pick the right chip based on that word.

If they get it right, they get a point. If they get it wrong, they learn from their mistake. This is called a "Referential Game."

The Problem: Robots Make "Weird" Maps

The researchers found that while these robots became very good at the game (they could guess the right chip most of the time), their internal "maps" of colors were weird.

Human Maps: When humans think of "green," we imagine a neat, solid blob of green colors. If you draw a line between any two green things, everything in between is also green. In math terms, our color categories are convex (like a smooth, round ball).
Robot Maps: The robots, however, created "swiss cheese" maps. Their idea of "green" was scattered. They might call a dark green chip "green," a bright green chip "green," but skip the middle shades, calling them something else. Their categories were non-convex (full of holes and jagged edges).

The robots had developed secret, weird codes that worked for them and their specific partner, but didn't look like how humans actually organize colors.

The Solution: Two Tweaks to the Recipe

The researchers tried two specific changes to fix the robots' "weird maps" and make them more human-like.

1. The "Rare Candy" Tweak (Upsampling)

The Problem: In the training data, some colors (like "teal" or "magenta") appear very rarely, while others (like "red" or "blue") appear constantly.
The Analogy: Imagine a teacher who only gives you a few examples of "rare" candies but hundreds of examples of "common" candies. You will become an expert at common candies but will barely understand the rare ones. The robots were ignoring the rare colors.
The Fix: The researchers took the rare color examples and copied them (upsampled) so the robots saw them more often during their initial learning phase.
The Result: This forced the robots to learn a wider variety of words. They stopped ignoring the rare colors and built a richer vocabulary.

2. The "Crowded Room" Tweak (Many Listeners)

The Problem: In the original setup, one Speaker robot played with just one Listener robot.
The Analogy: Imagine you are playing a secret code game with only one friend. You two might develop a weird, private shorthand that only the two of you understand (like calling a specific shade of green "banana"). It works for you, but it's not a real language.
The Fix: The researchers made the Speaker robot play with many different Listener robots (5 or even 30) instead of just one.
The Result: The Speaker couldn't rely on a private code with just one friend. To communicate with everyone in the crowd, the Speaker had to use clear, standard, "safe" words that everyone could understand. This forced the robots to create neat, solid color categories (convex shapes) instead of scattered, private ones.

The "Goldilocks" Discovery

The researchers found that you need the right mix of these two tweaks to get the best results.

Too little help: If you don't copy the rare colors, the robots miss them.
Too much help: If you copy the rare colors too much, the robots get confused and create too many tiny, specific words, breaking their neat categories again.
Just right: The "Goldilocks" setting was moderate copying of rare colors combined with playing in a crowded room (many listeners).

The Final Outcome

When the robots played with many listeners and got a moderate boost of rare color examples, their color maps looked almost exactly like human maps:

Neat Shapes: Their "green" was a solid, round blob, not a scattered mess.
Rich Vocabulary: They knew more words than before.
Stable Meaning: They didn't drift away from human meanings as much.

Summary

The paper shows that to make AI sound and think like humans, you can't just teach them facts. You have to design their social environment carefully. By giving them a diverse crowd to talk to and ensuring they practice the rare things enough, you can guide them to develop language that is not just efficient, but also geometrically and logically similar to how humans see the world.

1. Problem Statement

The paper addresses a critical limitation in computational models of language emergence: while neural agents can develop effective communication systems, the resulting lexicons often diverge systematically from human linguistic properties. Specifically, using the NeLLCom-Lex framework (Zhang et al., 2025), agents trained via Supervised Learning (SL) on human data and Reinforcement Learning (RL) in referential games develop non-convex color categories.

Human color categories are known to be convex in perceptual space (CIELAB), meaning the set of colors denoted by a single term forms a contiguous, convex region. In contrast, the agents in the baseline model develop "conceptual pacts" (partner-specific agreements) that result in fragmented, non-convex regions. The authors aim to identify the specific training conditions necessary to induce convexity and other human-like properties (e.g., appropriate vocabulary size and informativeness) in emergent lexicons.

2. Methodology

The study builds upon the NeLLCom-Lex framework, where feed-forward neural networks act as speakers and listeners. The agents are first pretrained via SL on human color naming data and then fine-tuned via RL in referential games (selecting a target color among distractors based on a spoken label).

The authors introduce two primary experimental manipulations to test their effects on lexicon geometry:

Listener Diversity (Many-Listener Setup):
- Baseline: One speaker interacts with a single listener.
- Manipulation: One speaker interacts with 5 or 30 independently initialized listeners.
- Hypothesis: Interacting with diverse listeners prevents the formation of idiosyncratic "conceptual pacts," forcing the speaker to adopt more systematic, listener-agnostic, and convex categories.
Data Upsampling (Frequency Imbalance Correction):
- Problem: Human color data is highly imbalanced (frequent terms like "red" vs. rare terms). This leads to poor learning of rare terms, which are then forgotten or absorbed during RL.
- Manipulation: Rare color samples are upsampled during the SL phase to reach target counts ( $N=100$ and $N=200$ ).
- Hypothesis: Increasing exposure to rare terms prevents semantic drift and allows for a richer, more balanced vocabulary without sacrificing geometric coherence.

Evaluation Metrics:

Communication Accuracy ( $Acc_{comm}$ ): Success rate in referential games.
Pragmatic Adaptation: Correlation between context difficulty and word informativeness.
Lexical Diversity ( $|W|$ ): Number of unique color words used.
System-Level Informativeness ( $I_L$ ): Average informativeness of the lexicon.
Convexity: A geometric measure defined as the ratio of the area of a color term's denotation to the area of its convex hull in CIELAB space.
Semantic Drift ( $D_L$ ): Euclidean distance between agent prototypes and human prototypes.

3. Key Contributions

Identification of Convexity Drivers: The paper isolates listener diversity and data sparsity as the two critical factors governing the geometric coherence (convexity) of emergent color lexicons.
The "Goldilocks" Principle: The authors demonstrate that neither extreme upsampling nor extreme listener diversity is optimal alone. A moderate upsampling level combined with a many-listener setup yields the most human-like results.
Quantification of Geometric Coherence: The study rigorously applies a convexity metric to neural agent lexicons, providing a quantitative link between communication efficiency and the geometric structure of meaning.
Trade-off Analysis: The work elucidates the trade-off between informativeness/diversity (favored by single listeners and high upsampling) and convexity/stability (favored by multiple listeners and moderate upsampling).

4. Results

Communication Accuracy: All conditions achieved high accuracy (>90%), confirming that non-convexity does not hinder basic communication but is a structural deviation from human norms.
Lexical Diversity:
- Upsampling: Significantly increased vocabulary size ( $|W|$ ) and system-level informativeness ( $I_L$ ).
- Listeners: Increasing the number of listeners decreased vocabulary size, forcing agents to use a more consistent, focused set of words.
Convexity:
- SL Phase: Agents showed high convexity initially due to low diversity.
- RL Phase: Convexity dropped significantly in the baseline (1 listener, no upsampling).
- Listener Effect: Increasing listeners consistently increased convexity (e.g., from 0.24 with 1 listener to 0.31 with 30 listeners).
- Upsampling Effect: The effect was non-monotonic. Moderate upsampling (N=100) improved convexity in many-listener setups, while excessive upsampling (N=200) caused convexity to drop again, likely due to the introduction of too many fine-grained distinctions that fragmented semantic regions.
Optimal Configuration: The combination of 30 listeners and moderate upsampling (N=100) produced lexicons with the highest convexity (0.37) and the best overall balance across all metrics, closely resembling human systems.
Semantic Drift: Upsampling significantly reduced semantic drift, stabilizing the lexicon. However, interacting with many listeners increased drift relative to the human baseline, suggesting that group communication prioritizes internal consistency over strict alignment with human conventions.

5. Significance

This research provides a mechanistic explanation for why neural agents fail to replicate human-like semantic structures and offers a solution to fix it.

Theoretical Impact: It supports the hypothesis that population structure (interacting with many diverse interlocutors) is a crucial evolutionary pressure for the emergence of convex, systematic categories, preventing the "drift" into partner-specific idiosyncrasies.
Practical Impact: It demonstrates that simply training on human data is insufficient; the distribution of training data (handling frequency imbalance) and the social structure of training (diversity of interlocutors) must be carefully engineered to model human language evolution accurately.
Future Directions: The findings suggest that future emergent communication models should move beyond dyadic (one-on-one) interactions and incorporate data balancing strategies to better simulate the dynamics of real-world language acquisition and change.