---
title: 'NeLLCom-Lex: Neural Lexical Adaptation'
url: https://www.emergentmind.com/topics/nellcom-lex
type: topic
---

# NeLLCom-Lex: Neural Lexical Adaptation

NeLLCom-Lex is a neural-agent framework for studying the interplay between lexical systems and language use under controlled communicative pressure. It is designed to simulate lexical semantic change not by inducing a symbol system from scratch, but by first grounding agents in an existing human lexicon and then allowing them to adapt through communication. In the reported implementation, the grounded domain is English color naming, the meaning space is represented in CIELAB color space, and the central question is whether agents can both reproduce human-like naming behavior and reorganize their lexical behavior when communicative need changes. The framework models lexical evolution within a single generation rather than multigenerational cultural transmission, and it evaluates both use-level pragmatics and system-level lexical structure [2509.22479].

## 1. Origins, scope, and research problem

NeLLCom-Lex extends the earlier NeLLCom line of work into the lexical domain. The broader NeLLCom family combines supervised learning and reinforcement learning so that agents first learn a language and then continue adapting through communicative interaction; the later "NeLLCom-X" adds role-alternating agents and group communication while preserving that two-phase structure [2407.13999]. NeLLCom-Lex retains the core idea of supervised grounding followed by communication-based adaptation, but shifts the focus from artificial grammatical systems to lexical meaning in a real lexical system [2509.22479].

The framework is motivated by a methodological gap in research on semantic change. Observational approaches such as corpus analysis and distributional semantic modeling can reveal patterns of change, but they do not isolate causal mechanisms. Human experiments can study pragmatic adaptation in real time, but semantic change typically unfolds over extended diachronic timescales. NeLLCom-Lex uses agent-based simulation as an intermediate methodology: communicative pressures can be manipulated directly, the environment can be controlled, and changes in lexical behavior can be observed as agents continue to communicate.

A defining feature of the framework is that it does not treat lexical change as blank-slate emergence. Agents are first trained to “speak” an existing language—here, English color terms from a human dataset—and only afterward exposed to communicative environments that may favor more specific or broader naming. This makes the framework a model of lexical adaptation rather than de novo language invention. The paper is explicit that the timescale is lexical evolution across a single generation, or more precisely adaptation across a lifetime of interaction, rather than historical transmission across generations.

## 2. Grounded task, lexical data, and communicative need

NeLLCom-Lex uses a grounded color-naming referential game based on the Colors in Context paradigm. A speaking agent observes a target color \(c_t\) together with two distractors \(c_{d_1}, c_{d_2}\), and must emit a single-word message \(w\) from a fixed vocabulary \(\mathcal{V}\), with \(w \in \mathcal{V}\). A listening agent receives the same word \(w\) and the same three colors in randomized order, and must identify the intended target. Communication succeeds if the listener selects the correct position [2509.22479].

The communicative environment is organized by context difficulty. The task uses three standard context types: **far**, where both distractors are clearly distinct from the target; **split**, where one distractor is similar and one is distinct; and **close**, where both distractors are similar to the target. Communicative need is operationalized as the granularity of distinctions the speaker must make. A target among perceptually distant distractors requires less specificity than the same target among near neighbors. This operationalization is central to the framework: lexical behavior is expected to shift when hard contexts become more frequent.

The meaning space is represented as 3D vectors in CIELAB, and generated contexts are created by rejection sampling from RGB space and evaluated with CIEDE2000 distance. The appendix states two concrete constraints: all pairwise distances must exceed a lower bound of 5, described as a just noticeable difference, and a threshold \( \theta = 20 \) defines context types. Under that threshold, **close** means all three colors are within \( \theta \), **split** means one distractor is within \( \theta \) and one is farther, and **far** means all colors are farther than \( \theta \).

For supervised grounding, the framework uses the English version of the Colors dataset, processed by Gualdoni and Boleda (2024), based on Monroe et al. (2017). The dataset contains 9,309 far instances, 3,886 split instances, and 2,239 close instances, for 15,434 instances total. Of these, 3K instances are held out as a human-labeled test set \( \text{test}_{hum} \), and the remaining 12.4K are used as supervised training data \( \text{train}_{hum} \). For reinforcement-learning experiments, large synthetic datasets are generated using Monroe et al.’s sampling procedure.

## 3. Agent architecture and learning pipeline

The agents are feedforward neural networks rather than recurrent sequence models, because the messages are single symbols and the task is referential discrimination. The speaker receives a referential context in which the target is always in the first position. Each color is processed by its own feedforward block; the resulting embeddings are concatenated and passed through another feedforward block to produce a joint representation; and a final linear classifier predicts the single-word message. The listener receives the word and the three colors in shuffled order. Each candidate color is projected into a hidden representation, the word is mapped by an embedding layer, and the listener computes a dot-product similarity between the word embedding and each color embedding, followed by a softmax over the three positions [2509.22479].

A central manipulation is whether the speaker is **context-aware** or **context-unaware**. In the context-aware condition, all three colors \(\{c_t, c_{d_1}, c_{d_2}\}\) are encoded normally. In the context-unaware condition, distractor embeddings are zeroed out, so the speaker effectively sees only the target. This factor is used to test whether pragmatic context sensitivity depends on exposure to distractor structure during learning and interaction.

The training procedure has two phases. In supervised learning, speaker and listener are trained independently on human-produced data. The speaker predicts the human color label from the input color context, and the listener predicts the target’s position from the message and shuffled triplet. In reinforcement learning, the speaker and listener interact in the referential game, and both are updated jointly to maximize communicative success. The paper uses REINFORCE, with the speaker updated by policy gradient and the reward defined as the listener’s log-probability assigned to the correct target index. Conceptually, supervised learning performs lexical grounding, and reinforcement learning performs lexical adaptation under communicative pressure.

The appendix specifies the main architectural and optimization settings. Both speaker and listener use feedforward modules with hidden dimension 512; each feedforward module consists of a linear layer, batch normalization, ReLU, and dropout; the optimizer is Adam with learning rate \(1 \times 10^{-5}\); batch size is 32; both supervised learning and reinforcement learning run for 30 epochs; experiments are repeated over 10 random seeds; and reinforcement learning uses entropy regularization parameter 0.15 for both speaker and listener. The implementation is partly based on the EGG toolkit.

## 4. Formalization of lexical behavior and lexical systems

A distinctive aspect of NeLLCom-Lex is that it separates behavior during interaction from the structure of the lexical system. This distinction is operationalized with measures drawn largely from Gualdoni and Boleda (2024), allowing the framework to evaluate both pragmatic adaptation and lexical-system organization [2509.22479].

For a word \(w\), the spread of the objects it denotes in visual space is defined as
\[
S_w=\frac{1}{N}\sum_i\sum_{j\neq i}d(o_i,o_j)
\]
where \(N\) is the number of unique object pairs and \(d(o_i,o_j)\) is Euclidean distance in CIELAB between two objects denoted by the same word. Word informativeness is then the inverse spread:
\[
I_w = 1/S_w.
\]
A word is therefore more informative when it applies to a tighter region of color space.

Context ease \(E_{ctx}\) is defined as the distance between the target and the hardest distractor, that is, the distractor closest to the target in CIELAB. Hard contexts have low ease and easy contexts have high ease. To test pragmatic adaptation, the paper fits a linear mixed-effects model predicting the informativeness of the produced word from context ease, with seed and target chip as random effects. A negative slope of \(I_w\) with respect to \(E_{ctx}\) means that harder contexts elicit more informative words.

At the system level, lexical informativeness is defined as
\[
I_L=\frac{1}{N}\sum_{i=1}^NI_w^i
\]
namely the average informativeness of the words used across \(N\) interactions. Lexical diversity is simply the number of different word types used on an evaluation set, \(|W|\). Semantic drift \(D_L\) measures divergence from the human lexicon by comparing, for each color word, the average CIELAB location of all chips labeled with that word in the agent system and in the human benchmark. The appendix also introduces a supplementary entropy-based measure, with normalized informativeness
\[
P(w_i) = \frac{I(w_i)}{\sum_{j=1}^{N} I(w_j)}
\]
and Shannon entropy
\[
H = - \sum_{i=1}^{N} P(w_i) \log P(w_i).
\]

These definitions make NeLLCom-Lex unusual among neural-agent simulations. The framework is not limited to communicative accuracy; it also asks whether the resulting lexical system has human-like informativeness, lexical diversity, and stability relative to human prototypes.

## 5. Experimental organization and empirical results

The paper reports two main experiments. **Experiment 1** varies whether the speaker has access to distractor context during supervised learning, reinforcement learning, or both. The conditions are **SL\(-\)**, **SL\(+\)**, **SL\(-\)RL\(-\)**, **SL\(-\)RL\(+\)**, and **SL\(+\)RL\(+\)**. **Experiment 2** fixes the best context-aware pipeline, **SL\(+\)RL\(+\)**, and varies the distribution of context difficulty during reinforcement learning: **AllFar**, **HalfHalf**, and **AllClose** [2509.22479].

On the held-out human-labeled test set \( \text{test}_{hum} \), speaker accuracy is reported as about 79% with context and about 75% without it, while listener accuracy reaches about 90% by epoch 30 with little difference across conditions. On generated evaluation data, communication accuracy improves rapidly after reinforcement learning begins. Agents trained without context in supervised learning but given context in reinforcement learning initially struggle and then catch up; agents without context in either phase perform worst.

The clearest behavioral signature is the slope \(\beta(E_{ctx})\) relating context ease to word informativeness. Table 1 reports the following values: **SL\(-\)**, \(0.000\); **SL\(+\)**, \(-0.003\); **SL\(-\)RL\(-\)**, \(-0.000\); **SL\(-\)RL\(+\)**, \(-0.002\); **SL\(+\)RL\(+\)**, \(-0.002\); **Human**, \(-0.008\). Significant negative slopes occur for all context-aware conditions and not for context-unaware ones. The effect is weaker than in humans, but the direction matches human pragmatic modulation: harder contexts elicit more informative words.

At the system level, reinforcement learning markedly changes the lexicon. The paper reports that context-unaware systems produce over-informative yet poor lexicons, with high \(I_L\), low lexical diversity, and weak communication. By contrast, **SL\(+\)RL\(+\)** yields the most human-like profile among the tested pipelines: \(I_L = 2.60\), close to the human value \(2.78\), and \(|W| = 34.5\), much richer than the supervised-only systems though still below the human value 49.0. The corresponding semantic drift is \(D_L = 47.3\). This creates a substantive trade-off: reinforcement learning improves lexical richness and pragmatic adaptability, but also moves word prototypes away from human ones. The paper notes that semantic drift increases sharply after reinforcement learning, from roughly 15–20 before reinforcement learning to roughly 47–48 afterward.

Experiment 2 shows that manipulating communicative need changes naming behavior in the expected direction. On the balanced test set \( \text{test}_{gen,dist50} \), the slopes are **Before RL**, \(-0.007\); **AllClose**, \(-0.005\); **HalfHalf**, \(-0.004\); **AllFar**, \(-0.003\). All are significant at \(p < .001\). Pairwise comparisons show that AllClose is steeper than AllFar by \(-0.002\), steeper than HalfHalf by \(-0.001\), and that HalfHalf is steeper than AllFar by \(-0.001\). Thus, a training environment dominated by hard contexts induces greater sensitivity to contextual difficulty.

Communication accuracy is nearly perfect on far contexts in all conditions, but close contexts remain much harder, at around 70–80% accuracy. AllClose performs best on close contexts, with HalfHalf and especially AllFar trailing. The paper also reports qualitative denotation shifts: for example, **pink** covers a broader region of CIELAB space under AllFar and a narrower one under AllClose, and analogous patterns are observed for *mustard*, *grapes*, *fuchsia*, and *blue*. At the same time, the authors report that overall lexical diversity and system-level informativeness are surprisingly similar across the need-manipulation conditions, and that they do not find strong evidence that AllFar yields a dramatically coarse-grained lexicon. The effect of communicative need is therefore clearer in pragmatic behavior than in the global lexical metrics used here.

## 6. Interpretation, limits, and place in the literature

NeLLCom-Lex is primarily a methodological contribution. It combines grounding in a human lexical system, controlled manipulation of communicative environment, and measurement of both use-level pragmatics and system-level lexical structure. This suggests a way to study lexical adaptation under experimentally controlled pressures without reducing the problem either to corpus observation or to ungrounded emergent communication [2509.22479].

Several misconceptions are explicitly ruled out by the paper’s scope statements. NeLLCom-Lex is not a model of centuries-long historical semantic change; it is a simulation of lexical evolution within a single generation. It is not a model of blank-slate language emergence; the agents begin with supervised grounding in English color naming. It is also not yet a model of population structure, group convergence, or intergenerational transmission. The experiments are dyadic, the domain is narrow, and the task is restricted to color triplets in a referential game.

The limitations are correspondingly specific. The supervised dataset is imbalanced across far, split, and close contexts, which may bias agents toward easier situations. Entropy regularization is fixed at 0.15, and the appendix notes that prior work found message entropy can remain near a lower bound even with strong regularization. The framework does not currently model role alternation within a single agent, group communication, or transmission to new generations, although adjacent NeLLCom work on group communication indicates how such extensions could be formulated [2407.13999]. The paper also emphasizes a substantive caveat: simulated lexical adaptation under communicative pressure is not equivalent to real-world semantic change, because social, historical, cultural, and multimodal factors are abstracted away.

Within those limits, the main result is that neural agents trained to “speak” an existing language can reproduce human-like patterns in color naming “to a remarkable extent,” while also showing systematic adaptation to altered communicative needs [2509.22479]. The authors suggest, rather than directly demonstrate, that this architecture could later be used to investigate mechanisms underlying meaning narrowing, broadening, and related forms of semantic change. A plausible implication is that NeLLCom-Lex is best understood as an experimental platform for mechanism isolation: it does not replace historical explanation, but it provides a controlled setting in which hypotheses about the coupling between lexical use and lexical structure can be tested.

Source: https://www.emergentmind.com/topics/nellcom-lex