Predictive structure emerges during generalisation of kin terms to new referents

Maisy Hallam, Fiona M. Jordan, Simon Kirby, Kenny SmithPsyArXiv 2025 · DOI

Abstract

Despite crosslinguistic diversity in how kin relations map to terminology, there are constraints on which kin may be categorised together. But what are the constraints on kin term variation, and where do they come from? One proposed constraint is internal co-selection -- an evolutionary process where terminological changes in one generation of kin co-occur with parallel changes in other generations. This results in kin terms which are predictable on the basis of other kin terms, a property we call predictive structure. To determine the strength of this constraint, we measured the predictive structure of systems of kinship terminology from 1135 languages. We found that kinship systems exhibit a significant degree of predictive structure, and we argue that its prevalence reflects a pressure for simplicity. We expected that this pressure would be imposed during the generalisation of known kin terms to new referents, and we tested this claim using an artificial kin term generalisation task. Our results suggest that generalisation of kin terms does indeed favour predictive structure, but that people's preference for predictive structure faces interference from other pressures to distinguish kin by features like gender.

Introduction

Human languages exhibit a remarkable variety in kinship terminology. While English uses a single term for "uncle," other languages like Hindi distinguish between paternal and maternal uncles of different ages. Conversely, Hawaiian merges "father" with all of a parent's male siblings. Despite this diversity, the variation is not infinite. Certain patterns, such as grouping "grandfather" and "brother" under one term, are virtually non-existent.

Researchers hypothesize that kinship systems are constrained by a trade-off between expressivity (precision) and simplicity (ease of learning). This study investigates internal co-selection as a mechanism for simplicity. Internal co-selection suggests that changes in one generation of kin (e.g., merging "father" and "uncle") trigger parallel changes in another (e.g., merging "brother" and "cousin"), maintaining a "predictive structure" that makes the system easier to navigate and learn.

Methods

The research employed a multi-methodological approach:

  1. Typological Analysis: The authors quantified "predictive structure" using Mutual Information (MI). This information-theoretic metric measures how much knowing a term in one generation (G+1, e.g., parents) reduces uncertainty about a term in the next (G0, e.g., siblings/cousins). Data was sourced from Kinbank, covering 1,135 languages across 21 major language families.
  2. Null Hypothesis Testing: To validate the significance of their findings, the authors used Monte Carlo methods. They compared real-world languages against 1,000 "shuffled" versions of each language, where terms were randomly reassigned to relatives.
  3. Behavioral Experiment: An artificial language task was conducted with 40 native English-speaking participants. Participants were shown family trees with some members labeled with novel kin terms (e.g., popingi) and were asked to generalize these terms to unlabelled relatives. This tested whether humans naturally favor predictive structure when learning new systems.

Key Insights: The Logic of Predictive Structure

The core concept of predictive structure is illustrated by the "funcle" logic. If a language merges "father" and "uncle" into a single term (a "funcle"), it is highly likely to also merge "brother" and "cousin" into a single term (a "brousin"). This makes the system highly compressible. Instead of memorizing disparate rules, a speaker only needs to know that "children of a funcle are brousins."

Illustration of internal co-selection: merging terms for father and uncle into 'funcle' predicts the merging of brother and cousin into 'brousin'.

Importantly, this predictability is independent of the sounds of the words themselves. The study found that predictive structure exists even in languages where the words are phonologically unrelated, suggesting that the constraint operates on the mapping of meanings rather than linguistic form.

Results

The analysis revealed that 87.5% of natural languages exhibit significantly higher predictive structure than their randomized baselines. Many languages exceeded their baseline by more than five standard deviations, supporting the theory that kinship systems are optimized for simplicity and efficient communication.

In the behavioral experiments, participants were significantly more likely to generalize terms in a way that maximized predictive structure. However, this bias faced interference from feature bias, specifically gender. When gender was a salient feature in the input, participants often prioritized gender-based distinctions over structural predictability, even if the input did not explicitly distinguish between male and female relatives.

Discussion and Conclusion

The prevalence of predictive structure across diverse languages suggests a universal cognitive pressure for simplicity. This pressure appears to manifest during the generalization of known terms to new referents. However, the experimental results indicate that this structural pressure competes with other cognitive biases, such as the tendency to categorize individuals by gender.

Future research should address the limitations of current datasets, as some languages lacking predictive structure may simply have incomplete records in Kinbank. Additionally, expanding behavioral studies to speakers of non-bilateral kinship systems (like the unilateral systems found in Hindi) will be crucial to determine if cultural background influences the generalization of predictive structures. Finally, integrating social factors like marriage and inheritance customs with information-theoretic measures remains a vital next step in understanding the evolution of kinship terminology.

Summary updated

Abstract

Despite crosslinguistic diversity in how kin relations map to terminology, there are constraints on which kin may be categorised together. But what are the constraints on kin term variation, and where do they come from? One proposed constraint is internal co-selection -- an evolutionary process where terminological changes in one generation of kin co-occur with parallel changes in other generations. This results in kin terms which are predictable on the basis of other kin terms, a property we call predictive structure. To determine the strength of this constraint, we measured the predictive structure of systems of kinship terminology from 1135 languages. We found that kinship systems exhibit a significant degree of predictive structure, and we argue that its prevalence reflects a pressure for simplicity. We expected that this pressure would be imposed during the generalisation of known kin terms to new referents, and we tested this claim using an artificial kin term generalisation task. Our results suggest that generalisation of kin terms does indeed favour predictive structure, but that people's preference for predictive structure faces interference from other pressures to distinguish kin by features like gender.
BibTeX
@misc{hallam2025predictive,
  author = {Maisy Hallam and Fiona M. Jordan and Simon Kirby and Kenny Smith},
  title = {{Predictive structure emerges during generalisation of kin terms to new referents}},
  year = {2025},
  doi = {10.31234/osf.io/t6prw},
  howpublished = {PsyArXiv},
}

Loading related content...

Figure