On the Optimality of Kinship Naming: an Information-theoretic Approach

AI summary

Overview

• Research area: Computational linguistics and information theory, specifically the cross-linguistic structure of kinship naming systems and the efficiency trade-off between informativeness and complexity. Also touches on emergent communication and graph neural networks.
• Technical level: Advanced. The paper relies on mutual information, entropy, Kullback-Leibler divergence, the Information Bottleneck principle, and graph neural network training.
• Scope: The paper extends an information-theoretic account of kinship naming systems to non-universal communicative needs and non-Bayesian listeners, then tests whether the predicted optimality emerges in neural network communication agents.

What This Paper Is About

Languages partition the meaning space of family relations differently — for example, English uses one term "grandmother" where Vietnamese distinguishes maternal from paternal and older from younger relatives. Prior work argued that such systems sit near an optimal trade-off between informativeness and complexity, but assumed a single universal communicative need across languages and idealized Bayesian listeners. This paper removes both assumptions by collecting communicative-need data from four languages and by analyzing how listener errors shift the trade-off away from the theoretical optimum, and it shows the same optimum arises in learned communication systems.

Key Contributions

1. An adapted information-theoretic framework for discrete domains. Inspired by multiclass classification, the authors redefine information loss as the expected cross-entropy between the intended referent and the listener's prediction. This lets them apply the framework to discrete domains and derive the optimal trade-off curve analytically, avoiding extra parameters that would otherwise have to be estimated empirically.
2. A novel dataset of communicative need in four languages. The authors estimate a separate communicative need distribution for English, Dutch, Spanish, and Vietnamese from corpus frequency counts, rather than assuming one universal need distribution.
3. An explicit treatment of non-Bayesian listeners. The framework quantifies how far a communication system lies from the optimal curve as a function of the KL divergence between the speaker's Bayesian decoder and the listener's actual decoder, and the paper simulates large populations of noisy listeners to test this.
4. A referential game demonstrating emergent optimality. Neural network speaker and listener agents playing a kinship naming game develop terminologies whose informativeness–complexity trade-off approaches the theoretical optimum.

Main Findings

• Bayesian listeners sit exactly on the optimal curve. Across all four languages, communication with the Bayesian HP-Listener lies exactly on the optimal curve (zero distance), while communication with non-Bayesian listeners deviates progressively further from it.
• Listener error rates predict the size of the deviation. Non-Bayesian HP-Listeners with flip rates of 0.1%, 0.5%, 1%, and 2% produce trade-offs that are approximately parallel to the optimal curve, and lower error rates consistently yield more favorable trade-offs than higher error rates.
• Indo-European languages cluster together, Vietnamese differs. English, Dutch, and Spanish exhibit similar levels of adjusted complexity and information loss (and similar accuracy), whereas Vietnamese shows higher information loss but lower adjusted complexity, indicating more remaining capacity to increase complexity in order to improve informativeness.
• Optimal human communication occupies a narrow band. The results suggest that optimal human communication tends to cluster within an information loss range of approximately [0.2, 0.4] and an adjusted complexity range of [-0.4, -0.2].
• Neural agents approach the theoretical optimum. The NN communication system achieves trade-offs closer to the optimal curve than human communication with the non-Bayesian HP-Listener at the 1% flipping rate.
• Emergent naming evolves toward optimality over training. Recorded every 10 epochs, agents initially show moderate complexity but high information loss, relying on only a few vocabulary terms; information loss then decreases gradually at the cost of increasing complexity, with the trajectory approaching the optimal curve.
• Early stopping acts as a complexity constraint. Because the NN models are trained only to minimize information loss, they are not inherently constrained by complexity; selecting the checkpoint with accuracy closest to that of human communication serves as an effective mechanism for limiting complexity.
• A distribution-free way to compare languages. By defining adjusted complexity as C̃ = C − H(U), the optimal curve simplifies to L = −C̃ and no longer depends on the communicative need distribution, enabling cross-linguistic comparison while preserving Euclidean distance.

Methodology in Plain English

The authors model kinship naming as a communication game. A speaker wants to talk about a specific family member drawn from a communicative need distribution, picks a kinship term, and a listener tries to guess which family member was meant. Complexity is measured as how much information the term carries about the intended person (the mutual information between referent and word). Information loss is measured as how badly the listener does at recovering the intended person, using the standard cross-entropy loss from multiclass classification.

With these two quantities, the authors prove an identity: information loss equals the entropy of the need distribution minus complexity, plus a term measuring how much the listener's decoder differs from the speaker's Bayesian decoder. Because that last term can never be negative, the curve L = H(U) − C defines a lower bound, and a system sits exactly on it only when the listener decodes in a Bayesian-optimal way. Subtracting the entropy term gives an adjusted complexity that makes the optimal curve the same across languages with different need distributions.

For the human side, the authors take a family structure of 33 relatives across five generations with either Alice or Bob as the reference person (ego). They estimate, for English, Dutch, Spanish, and Vietnamese, how often each kinship term is used for each relative by searching corpora for first-person possessive constructions such as "my mother." Polysemous terms are split evenly among plausible referents. Listener behavior cannot be read off corpora, so the authors simulate large populations of listeners that occasionally flip their decision, with flip rates of 0.1%, 0.5%, 1%, and 2%.

For the neural side, they train a graph neural network speaker and a separate graph neural network listener on a pruned kinship graph containing only the primitive relations parent-of and child-of, so that relations like sibling must be inferred compositionally. The speaker produces a one-token message from a vocabulary of 128 items; the listener scores all 33 nodes. Training uses 10,000 game turns with an 80/20 train/validation split, the Adam optimizer at a learning rate of 1×10⁻³, 500 epochs, and mini-batches of 50. During training the listener picks among the target plus five distractors; at evaluation it must identify the target among all 32 non-ego family members, giving 2×32 = 64 evaluation points. Training and evaluation are repeated 50 times.

Why This Matters

Impact on research. The paper removes two simplifying assumptions that have shaped prior work on semantic universals: that communicative need is identical across languages, and that listeners are perfectly Bayesian. It supplies an analytically tractable optimality curve and a distance measure for discrete domains, which makes it possible to compare communication systems across languages with different cultural profiles on a common scale. It also connects information-theoretic accounts of language to emergent communication, showing that a single pressure — minimizing information loss — combined with early stopping is enough to reproduce human-like trade-offs.

Real-world applications.

• Designing and evaluating terminology systems in domains with small, discrete referent spaces, such as kinship terms in legal, genealogical, or anthropological databases.
• Analyzing communication robustness when the channel is noisy or the receiver is impaired, since the framework explicitly models imperfect decoders.
• Cross-cultural natural language processing and translation, where kinship terms map many-to-one or one-to-many between languages.
• Diagnosing and improving learned communication protocols in multi-agent systems, where vocabulary size trades off against task success.

Industry relevance. The results matter for anyone building or evaluating reference systems where a small set of labels must serve many distinct referents: dialogue systems, database schemas, ontologies, and multi-agent reinforcement learning pipelines. The paper's finding that optimizing for accuracy alone, combined with early stopping, yields near-optimal complexity suggests a practical recipe for training communication agents without an explicit complexity penalty.

Future Directions

• Expand the comparison across more languages and linguistic families beyond English, Dutch, Spanish, and Vietnamese, and analyze the observed clustering patterns in more depth.
• Allow compositional messages: the current neural agents are limited to one-token messages and cannot develop expressions like the English prefix "grand-" or the Spanish gender suffixes "-a" and "-o." Allowing synonymy is also flagged as an open issue, since the deterministic neural speaker produces none.
• Reconsider the complexity measure itself, which is upper-bounded by the entropy of the communicative need distribution and therefore does not capture a language's ability to grow its vocabulary without bound.
• Broaden the corpora used to estimate communicative need, since the current datasets are domain-specific and may not represent all domains of language use.

Target Audience

This paper suits researchers in computational linguistics, information-theoretic models of language, semantic typology, and emergent communication, as well as graduate students comfortable with information theory and neural network training. Readers interested in cross-cultural variation in kinship terminology, or in multi-agent communication systems, will find the framework and the simulation setup directly relevant, though the derivations and model architecture require an advanced background.

Authors’ abstract

The structure of naming systems in natural languages hinges on a trade-off between high informativeness and low complexity. Focusing on the domain of kinship naming, we analyze such trade-off while addressing simplifying assumptions of prior work, namely: (i) universal communicative need across languages, and (ii) optimal listeners. To that aim, we collect data from four different languages, and analyze how different communicative needs and variations in the listener model influence the informativeness--complexity trade-off. Adopting a referential game setup from emergent communication, we further show that trade-off optimality is not only theoretically achievable but also emerges empirically in learned communication systems.